INFO 04-13 07:18:48 [__init__.py:216] Automatically detected platform cuda.
2026-04-13 07:18:50,586 | 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, 9442.81it/s]
`torch_dtype` is deprecated! Use `dtype` instead!

Loading checkpoint shards:   0%|          | 0/3 [00:00<?, ?it/s]
Loading checkpoint shards:  33%|███▎      | 1/3 [00:06<00:13,  6.53s/it]
Loading checkpoint shards:  67%|██████▋   | 2/3 [00:12<00:06,  6.24s/it]
Loading checkpoint shards: 100%|██████████| 3/3 [00:15<00:00,  4.87s/it]
Loading checkpoint shards: 100%|██████████| 3/3 [00:15<00:00,  5.27s/it]
The model is already on multiple devices. Skipping the move to device specified in `args`.
INFO 04-13 07:20:37 [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 04-13 07:20:38 [model.py:547] Resolved architecture: LlamaForCausalLM
INFO 04-13 07:20:38 [model.py:1510] Using max model len 1536
INFO 04-13 07:20:38 [parallel.py:380] Using external launcher for distributed inference.
INFO 04-13 07:20:38 [parallel.py:420] Disabling V1 multiprocessing for external launcher.
INFO 04-13 07:20:38 [scheduler.py:205] Chunked prefill is enabled with max_num_batched_tokens=4096.
INFO 04-13 07:20:39 [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 04-13 07:20:40 [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 04-13 07:20:40 [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 04-13 07:20:40 [gpu_model_runner.py:2602] Starting to load model Navid-AI/Yehia-7B-preview...
INFO 04-13 07:20:40 [gpu_model_runner.py:2634] Loading model from scratch...
INFO 04-13 07:20:40 [cuda.py:366] Using Flash Attention backend on V1 engine.
INFO 04-13 07:20:41 [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,  2.19it/s]

Loading safetensors checkpoint shards:  67% Completed | 2/3 [00:01<00:00,  1.68it/s]

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

Loading safetensors checkpoint shards: 100% Completed | 3/3 [00:01<00:00,  1.58it/s]

INFO 04-13 07:20:43 [default_loader.py:267] Loading weights took 2.10 seconds
INFO 04-13 07:20:43 [gpu_model_runner.py:2653] Model loading took 13.0406 GiB and 2.532173 seconds
INFO 04-13 07:20:49 [backends.py:548] Using cache directory: /root/.cache/vllm/torch_compile_cache/8b012be8b8/rank_0_0/backbone for vLLM's torch.compile
INFO 04-13 07:20:49 [backends.py:559] Dynamo bytecode transform time: 5.81 s
INFO 04-13 07:20:53 [backends.py:164] Directly load the compiled graph(s) for dynamic shape from the cache, took 3.482 s
INFO 04-13 07:20:54 [monitor.py:34] torch.compile takes 5.81 s in total
INFO 04-13 07:20:55 [gpu_worker.py:298] Available KV cache memory: 26.11 GiB
INFO 04-13 07:20:55 [kv_cache_utils.py:1087] GPU KV cache size: 53,456 tokens
INFO 04-13 07:20:55 [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):  60%|██████    | 3/5 [00:00<00:00, 21.21it/s]
Capturing CUDA graphs (mixed prefill-decode, PIECEWISE): 100%|██████████| 5/5 [00:00<00:00, 18.90it/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, 24.82it/s]
Capturing CUDA graphs (decode, FULL): 100%|██████████| 4/4 [00:00<00:00, 25.40it/s]
INFO 04-13 07:20:57 [gpu_model_runner.py:3480] Graph capturing finished in 1 secs, took 0.09 GiB
INFO 04-13 07:20:57 [core.py:210] init engine (profile, create kv cache, warmup model) took 13.10 seconds
INFO 04-13 07:20:57 [llm.py:306] Supported_tasks: ('generate',)
2026-04-13 07:20:57,395 | INFO | train_grpo_train | dataset_id=Shaer-AI/ashaar-enhanced-desc-baseform-final-sft-lte20-min500-splits-grpo-meter-count-v1 train_size=9955 eval_size=80
2026-04-13 07:20:57,395 | INFO | train_grpo_train | source_dataset_id=Shaer-AI/ashaar-with-enhanced-descriptions-baseform-final-sft-lte20-min500-splits test_size=80 hard_diagnostic_size=3126
2026-04-13 07:20:57,395 | INFO | train_grpo_train | active_rewards=['meter', 'count_adherence', 'hard_gate', 'judge_quality', 'repeat_soft', 'total_composite']
2026-04-13 07:20:57,395 | INFO | train_grpo_train | reward_weights={'meter': 0.0, 'count_adherence': 0.0, 'hard_gate': 0.0, 'judge_quality': 0.0, 'repeat_soft': 0.0, 'total_composite': 1.0}
2026-04-13 07:20:57,395 | INFO | train_grpo_train | output_repo=Shaer-AI/Shaer-adapters-grpo-short1k-no-trio-v3
2026-04-13 07:20:57,396 | INFO | train_grpo_train | train_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/cap_1000/selected_manifest.csv
2026-04-13 07:20:57,396 | INFO | train_grpo_train | hard_diagnostic_manifest_path=/root/workspace/Shaer/grpo/outputs/curated_meter_count_easyfirst_short_drop_trio/hard_diagnostic_cap_256/selected_manifest.csv
2026-04-13 07:20:57,396 | INFO | train_grpo_train | trainable_params=159907840 total_params=7160467456 trainable_ratio=0.022332
2026-04-13 07:20:57,396 | INFO | train_grpo_train | resume_decision={"timestamp_utc": "2026-04-13T07:20:37Z", "resume_mode": "fresh", "requested_spec": "", "result": "fresh", "reason": "explicit_fresh", "local_resume_path": null, "remote_repo": null, "remote_prefix": null, "compatibility": "unchecked", "config_fingerprint": "002103f46be89ed6311e30cbbf1952cf127698400c47c48f018bf0a5f0eb1c65"}
2026-04-13 07:20:57,396 | 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 04-13 07:20:58 [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 [00:04<4:31:16,  4.93s/it]2026-04-13 07:21:02,713 | INFO | train_grpo_train | metrics_logged mode=train step=1

                                                  
{'loss': 0.1336, 'grad_norm': 23.196989059448242, 'learning_rate': 1e-05, 'num_tokens': 1670.0, 'completions/mean_length': 41.75, 'completions/min_length': 29.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6725290417671204, 'rewards/meter/std': 0.40340036153793335, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957022070884705, 'rewards/repeat_soft/std': 0.00409209867939353, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.09941796213388443, 'rewards/total_composite/mean': 0.5380368828773499, 'rewards/total_composite/std': 0.16582772135734558, 'reward': 0.5380368828773499, 'reward_std': 0.1658277064561844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22603146731853485, 'sampling/sampling_logp_difference/max': 1.5626206398010254, 'sampling/importance_sampling_ratio/min': 0.20958609879016876, 'sampling/importance_sampling_ratio/mean': 1.013262391090393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.01923768222332, 'clip_ratio/low_mean': 0.10917956382036209, 'clip_ratio/low_min': 0.10917956382036209, 'clip_ratio/high_mean': 0.12637650407850742, 'clip_ratio/high_max': 0.12637650407850742, 'clip_ratio/region_mean': 0.23555606789886951, 'reward_total_mean': 0.5380368828773499, 'reward_meter_mean': 0.6725290417671204, 'reward_meter_std': 0.40340036153793335, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957022070884705, 'reward_repeat_soft_std': 0.00409209867939353, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.09941796213388443, 'reward_total_composite_mean': 0.5380368828773499, 'reward_total_composite_std': 0.16582772135734558, 'epoch': 0.0}

  0%|          | 1/3300 [00:04<4:31:16,  4.93s/it]INFO 04-13 07:21:02 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 2/3300 [00:13<6:14:10,  6.81s/it]2026-04-13 07:21:10,827 | INFO | train_grpo_train | metrics_logged mode=train step=2

                                                  
{'loss': 0.0546, 'grad_norm': 8.660832405090332, 'learning_rate': 9.996969696969698e-06, 'num_tokens': 4114.0, 'completions/mean_length': 141.5, 'completions/min_length': 100.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.5, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.6691720485687256, 'rewards/meter/std': 0.3117619752883911, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969235062599182, 'rewards/repeat_soft/std': 0.003796192118898034, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5947020649909973, 'rewards/total_composite/std': 0.15181177854537964, 'reward': 0.5947020649909973, 'reward_std': 0.15181177854537964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19878226518630981, 'sampling/sampling_logp_difference/max': 1.3262004852294922, 'sampling/importance_sampling_ratio/min': 0.2654840648174286, 'sampling/importance_sampling_ratio/mean': 1.028130292892456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9687132090330124, 'clip_ratio/low_mean': 0.08277223259210587, 'clip_ratio/low_min': 0.08277223259210587, 'clip_ratio/high_mean': 0.09919708035886288, 'clip_ratio/high_max': 0.09919708035886288, 'clip_ratio/region_mean': 0.18196931295096874, 'reward_total_mean': 0.5947020649909973, 'reward_meter_mean': 0.6691720485687256, 'reward_meter_std': 0.3117619752883911, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969235062599182, 'reward_repeat_soft_std': 0.003796192118898034, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5947020649909973, 'reward_total_composite_std': 0.15181177854537964, 'epoch': 0.0}

  0%|          | 2/3300 [00:13<6:14:10,  6.81s/it]INFO 04-13 07:21:11 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 3/3300 [00:20<6:26:46,  7.04s/it]2026-04-13 07:21:18,146 | INFO | train_grpo_train | metrics_logged mode=train step=3

                                                  
{'loss': 0.1089, 'grad_norm': 23.679033279418945, 'learning_rate': 9.993939393939395e-06, 'num_tokens': 5877.0, 'completions/mean_length': 43.375, 'completions/min_length': 31.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.44267401099205017, 'rewards/meter/std': 0.36050936579704285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.992233157157898, 'rewards/repeat_soft/std': 0.012951970100402832, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.5438969135284424, 'rewards/total_composite/std': 0.21351033449172974, 'reward': 0.5438969135284424, 'reward_std': 0.21351033449172974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2622106671333313, 'sampling/sampling_logp_difference/max': 2.150146245956421, 'sampling/importance_sampling_ratio/min': 0.11646712571382523, 'sampling/importance_sampling_ratio/mean': 1.0123623609542847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2060932368040085, 'clip_ratio/low_mean': 0.13107881788164377, 'clip_ratio/low_min': 0.13107881788164377, 'clip_ratio/high_mean': 0.08155973814427853, 'clip_ratio/high_max': 0.08155973814427853, 'clip_ratio/region_mean': 0.2126385560259223, 'reward_total_mean': 0.5438969135284424, 'reward_meter_mean': 0.44267401099205017, 'reward_meter_std': 0.36050936579704285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.992233157157898, 'reward_repeat_soft_std': 0.012951970100402832, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.5438969135284424, 'reward_total_composite_std': 0.21351033449172974, 'epoch': 0.0}

  0%|          | 3/3300 [00:20<6:26:46,  7.04s/it]INFO 04-13 07:21:18 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 4/3300 [00:28<6:43:52,  7.35s/it]2026-04-13 07:21:25,975 | INFO | train_grpo_train | metrics_logged mode=train step=4

                                                  
{'loss': 0.1969, 'grad_norm': 17.794963836669922, 'learning_rate': 9.990909090909093e-06, 'num_tokens': 7527.0, 'completions/mean_length': 49.25, 'completions/min_length': 31.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.514047384262085, 'rewards/meter/std': 0.4770835041999817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9997227191925049, 'rewards/repeat_soft/std': 0.0006712718168273568, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6147944927215576, 'rewards/total_composite/std': 0.264877587556839, 'reward': 0.6147944927215576, 'reward_std': 0.264877587556839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2241060584783554, 'sampling/sampling_logp_difference/max': 3.2926888465881348, 'sampling/importance_sampling_ratio/min': 0.03715381398797035, 'sampling/importance_sampling_ratio/mean': 1.0513813495635986, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8340110927820206, 'clip_ratio/low_mean': 0.08712000586092472, 'clip_ratio/low_min': 0.08712000586092472, 'clip_ratio/high_mean': 0.10127950087189674, 'clip_ratio/high_max': 0.10127950087189674, 'clip_ratio/region_mean': 0.18839950673282146, 'reward_total_mean': 0.6147944927215576, 'reward_meter_mean': 0.514047384262085, 'reward_meter_std': 0.4770835041999817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9997227191925049, 'reward_repeat_soft_std': 0.0006712718168273568, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6147944927215576, 'reward_total_composite_std': 0.264877587556839, 'epoch': 0.0}

  0%|          | 4/3300 [00:28<6:43:52,  7.35s/it]INFO 04-13 07:21:26 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 5/3300 [00:35<6:51:29,  7.49s/it]2026-04-13 07:21:33,716 | INFO | train_grpo_train | metrics_logged mode=train step=5

                                                  
{'loss': 0.1922, 'grad_norm': 8.504620552062988, 'learning_rate': 9.987878787878788e-06, 'num_tokens': 9871.0, 'completions/mean_length': 116.0, 'completions/min_length': 77.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.400786817073822, 'rewards/meter/std': 0.38270607590675354, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9964909553527832, 'rewards/repeat_soft/std': 0.004073499236255884, 'rewards/judge_quality/mean': 0.7074999809265137, 'rewards/judge_quality/std': 0.247487410902977, 'rewards/total_composite/mean': 0.526107668876648, 'rewards/total_composite/std': 0.1974838376045227, 'reward': 0.526107668876648, 'reward_std': 0.1974838227033615, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17819346487522125, 'sampling/sampling_logp_difference/max': 1.6136455535888672, 'sampling/importance_sampling_ratio/min': 0.19916023313999176, 'sampling/importance_sampling_ratio/mean': 1.0168025493621826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.27456446737051, 'clip_ratio/low_mean': 0.10275345481932163, 'clip_ratio/low_min': 0.10275345481932163, 'clip_ratio/high_mean': 0.0539776710793376, 'clip_ratio/high_max': 0.0539776710793376, 'clip_ratio/region_mean': 0.15673112589865923, 'reward_total_mean': 0.526107668876648, 'reward_meter_mean': 0.400786817073822, 'reward_meter_std': 0.38270607590675354, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9964909553527832, 'reward_repeat_soft_std': 0.004073499236255884, 'reward_judge_quality_mean': 0.7074999809265137, 'reward_judge_quality_std': 0.247487410902977, 'reward_total_composite_mean': 0.526107668876648, 'reward_total_composite_std': 0.1974838376045227, 'epoch': 0.0}

  0%|          | 5/3300 [00:35<6:51:29,  7.49s/it]INFO 04-13 07:21:33 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 6/3300 [00:43<6:56:27,  7.59s/it]2026-04-13 07:21:41,483 | INFO | train_grpo_train | metrics_logged mode=train step=6

                                                  
{'loss': 0.0954, 'grad_norm': 11.160085678100586, 'learning_rate': 9.984848484848485e-06, 'num_tokens': 12514.0, 'completions/mean_length': 131.375, 'completions/min_length': 108.0, 'completions/max_length': 158.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.8979466557502747, 'rewards/meter/std': 0.27314630150794983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976483583450317, 'rewards/repeat_soft/std': 0.0014516275841742754, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.2650572657585144, 'rewards/total_composite/mean': 0.7199420928955078, 'rewards/total_composite/std': 0.20595994591712952, 'reward': 0.7199420928955078, 'reward_std': 0.20595994591712952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20203472673892975, 'sampling/sampling_logp_difference/max': 1.973508358001709, 'sampling/importance_sampling_ratio/min': 0.13896843791007996, 'sampling/importance_sampling_ratio/mean': 1.0401043891906738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2473559081554413, 'clip_ratio/low_mean': 0.0936892218887806, 'clip_ratio/low_min': 0.0936892218887806, 'clip_ratio/high_mean': 0.10637771897017956, 'clip_ratio/high_max': 0.10637771897017956, 'clip_ratio/region_mean': 0.20006694085896015, 'reward_total_mean': 0.7199420928955078, 'reward_meter_mean': 0.8979466557502747, 'reward_meter_std': 0.27314630150794983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976483583450317, 'reward_repeat_soft_std': 0.0014516275841742754, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.2650572657585144, 'reward_total_composite_mean': 0.7199420928955078, 'reward_total_composite_std': 0.20595994591712952, 'epoch': 0.0}

  0%|          | 6/3300 [00:43<6:56:27,  7.59s/it]INFO 04-13 07:21:41 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 7/3300 [00:51<6:54:15,  7.55s/it]2026-04-13 07:21:48,954 | INFO | train_grpo_train | metrics_logged mode=train step=7

                                                  
{'loss': -0.0688, 'grad_norm': 11.945133209228516, 'learning_rate': 9.981818181818183e-06, 'num_tokens': 14844.0, 'completions/mean_length': 107.25, 'completions/min_length': 81.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.3544602692127228, 'rewards/meter/std': 0.3453842103481293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.999114990234375, 'rewards/repeat_soft/std': 0.0010229938197880983, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.372326135635376, 'rewards/total_composite/std': 0.16378259658813477, 'reward': 0.372326135635376, 'reward_std': 0.16378259658813477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2678569257259369, 'sampling/sampling_logp_difference/max': 1.720468521118164, 'sampling/importance_sampling_ratio/min': 0.17898227274417877, 'sampling/importance_sampling_ratio/mean': 1.018048644065857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.666846424341202, 'clip_ratio/low_mean': 0.05837517976760864, 'clip_ratio/low_min': 0.05837517976760864, 'clip_ratio/high_mean': 0.2030274923890829, 'clip_ratio/high_max': 0.2030274923890829, 'clip_ratio/region_mean': 0.26140267215669155, 'reward_total_mean': 0.372326135635376, 'reward_meter_mean': 0.3544602692127228, 'reward_meter_std': 0.3453842103481293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.999114990234375, 'reward_repeat_soft_std': 0.0010229938197880983, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.372326135635376, 'reward_total_composite_std': 0.16378259658813477, 'epoch': 0.0}

  0%|          | 7/3300 [00:51<6:54:15,  7.55s/it]INFO 04-13 07:21:49 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 8/3300 [00:57<6:37:21,  7.24s/it]2026-04-13 07:21:55,540 | INFO | train_grpo_train | metrics_logged mode=train step=8

                                                  
{'loss': 0.0948, 'grad_norm': 24.52655029296875, 'learning_rate': 9.97878787878788e-06, 'num_tokens': 16305.0, 'completions/mean_length': 26.625, 'completions/min_length': 18.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7332162261009216, 'rewards/meter/std': 0.33431848883628845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9578598141670227, 'rewards/repeat_soft/std': 0.01113725546747446, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.2620762288570404, 'rewards/total_composite/mean': 0.7266648411750793, 'rewards/total_composite/std': 0.22824539244174957, 'reward': 0.7266648411750793, 'reward_std': 0.22824537754058838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19474588334560394, 'sampling/sampling_logp_difference/max': 3.4640512466430664, 'sampling/importance_sampling_ratio/min': 0.031302690505981445, 'sampling/importance_sampling_ratio/mean': 1.0509170293807983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5002107545733452, 'clip_ratio/low_mean': 0.07552917674183846, 'clip_ratio/low_min': 0.07552917674183846, 'clip_ratio/high_mean': 0.0975095797330141, 'clip_ratio/high_max': 0.0975095797330141, 'clip_ratio/region_mean': 0.17303875647485256, 'reward_total_mean': 0.7266648411750793, 'reward_meter_mean': 0.7332162261009216, 'reward_meter_std': 0.33431848883628845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9578598141670227, 'reward_repeat_soft_std': 0.01113725546747446, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.2620762288570404, 'reward_total_composite_mean': 0.7266648411750793, 'reward_total_composite_std': 0.22824539244174957, 'epoch': 0.0}

  0%|          | 8/3300 [00:57<6:37:21,  7.24s/it]INFO 04-13 07:21:55 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 9/3300 [01:05<6:54:12,  7.55s/it]2026-04-13 07:22:03,772 | INFO | train_grpo_train | metrics_logged mode=train step=9

                                                  
{'loss': 0.169, 'grad_norm': 8.872598648071289, 'learning_rate': 9.975757575757577e-06, 'num_tokens': 19099.0, 'completions/mean_length': 143.25, 'completions/min_length': 92.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.25, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.8696469068527222, 'rewards/meter/std': 0.18565556406974792, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950947761535645, 'rewards/repeat_soft/std': 0.006243214942514896, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/total_composite/mean': 0.6252528429031372, 'rewards/total_composite/std': 0.0826655700802803, 'reward': 0.6252528429031372, 'reward_std': 0.0826655700802803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1975843459367752, 'sampling/sampling_logp_difference/max': 3.2445778846740723, 'sampling/importance_sampling_ratio/min': 0.03898501768708229, 'sampling/importance_sampling_ratio/mean': 1.0382981300354004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.174727141857147, 'clip_ratio/low_mean': 0.16696189157664776, 'clip_ratio/low_min': 0.16696189157664776, 'clip_ratio/high_mean': 0.02835051529109478, 'clip_ratio/high_max': 0.02835051529109478, 'clip_ratio/region_mean': 0.19531240686774254, 'reward_total_mean': 0.6252528429031372, 'reward_meter_mean': 0.8696469068527222, 'reward_meter_std': 0.18565556406974792, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950947761535645, 'reward_repeat_soft_std': 0.006243214942514896, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.6252528429031372, 'reward_total_composite_std': 0.0826655700802803, 'epoch': 0.0}

  0%|          | 9/3300 [01:06<6:54:12,  7.55s/it]INFO 04-13 07:22:04 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 10/3300 [01:12<6:37:52,  7.26s/it]2026-04-13 07:22:10,367 | INFO | train_grpo_train | metrics_logged mode=train step=10

                                                   
{'loss': -0.0406, 'grad_norm': 18.636837005615234, 'learning_rate': 9.972727272727274e-06, 'num_tokens': 20677.0, 'completions/mean_length': 35.25, 'completions/min_length': 24.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.878533124923706, 'rewards/meter/std': 0.1565551459789276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9728596210479736, 'rewards/repeat_soft/std': 0.036961715668439865, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/total_composite/mean': 0.6499476432800293, 'rewards/total_composite/std': 0.41783440113067627, 'reward': 0.6499476432800293, 'reward_std': 0.41783440113067627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21180443465709686, 'sampling/sampling_logp_difference/max': 2.064453125, 'sampling/importance_sampling_ratio/min': 0.12688764929771423, 'sampling/importance_sampling_ratio/mean': 0.9985918402671814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1395761594176292, 'clip_ratio/low_mean': 0.08045504428446293, 'clip_ratio/low_min': 0.08045504428446293, 'clip_ratio/high_mean': 0.10815705172717571, 'clip_ratio/high_max': 0.10815705172717571, 'clip_ratio/region_mean': 0.18861209601163864, 'reward_total_mean': 0.6499476432800293, 'reward_meter_mean': 0.878533124923706, 'reward_meter_std': 0.1565551459789276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9728596210479736, 'reward_repeat_soft_std': 0.036961715668439865, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.6499476432800293, 'reward_total_composite_std': 0.41783440113067627, 'epoch': 0.0}

  0%|          | 10/3300 [01:12<6:37:52,  7.26s/it]INFO 04-13 07:22:10 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 11/3300 [01:19<6:25:14,  7.03s/it]2026-04-13 07:22:16,876 | INFO | train_grpo_train | metrics_logged mode=train step=11

                                                   
{'loss': 0.0526, 'grad_norm': 22.84332847595215, 'learning_rate': 9.96969696969697e-06, 'num_tokens': 22159.0, 'completions/mean_length': 35.25, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.42980778217315674, 'rewards/meter/std': 0.4602866768836975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9971010684967041, 'rewards/repeat_soft/std': 0.004350362345576286, 'rewards/judge_quality/mean': 0.6025000214576721, 'rewards/judge_quality/std': 0.1976107507944107, 'rewards/total_composite/mean': 0.517084002494812, 'rewards/total_composite/std': 0.2049219310283661, 'reward': 0.517084002494812, 'reward_std': 0.2049219161272049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25400835275650024, 'sampling/sampling_logp_difference/max': 1.6263856887817383, 'sampling/importance_sampling_ratio/min': 0.19663900136947632, 'sampling/importance_sampling_ratio/mean': 1.044650673866272, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.648899346590042, 'clip_ratio/low_mean': 0.11001225002110004, 'clip_ratio/low_min': 0.11001225002110004, 'clip_ratio/high_mean': 0.0788961062207818, 'clip_ratio/high_max': 0.0788961062207818, 'clip_ratio/region_mean': 0.18890835624188185, 'reward_total_mean': 0.517084002494812, 'reward_meter_mean': 0.42980778217315674, 'reward_meter_std': 0.4602866768836975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9971010684967041, 'reward_repeat_soft_std': 0.004350362345576286, 'reward_judge_quality_mean': 0.6025000214576721, 'reward_judge_quality_std': 0.1976107507944107, 'reward_total_composite_mean': 0.517084002494812, 'reward_total_composite_std': 0.2049219310283661, 'epoch': 0.0}

  0%|          | 11/3300 [01:19<6:25:14,  7.03s/it]INFO 04-13 07:22:17 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 12/3300 [01:25<6:15:19,  6.85s/it]2026-04-13 07:22:23,318 | INFO | train_grpo_train | metrics_logged mode=train step=12

                                                   
{'loss': 0.1019, 'grad_norm': 23.60201072692871, 'learning_rate': 9.966666666666667e-06, 'num_tokens': 23601.0, 'completions/mean_length': 20.25, 'completions/min_length': 15.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5795812606811523, 'rewards/meter/std': 0.4281229078769684, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9527608156204224, 'rewards/repeat_soft/std': 0.02547159232199192, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5066647529602051, 'rewards/total_composite/std': 0.11989112198352814, 'reward': 0.5066647529602051, 'reward_std': 0.11989112198352814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21970289945602417, 'sampling/sampling_logp_difference/max': 0.9635438919067383, 'sampling/importance_sampling_ratio/min': 0.3823465406894684, 'sampling/importance_sampling_ratio/mean': 1.056498408317566, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.821111336350441, 'clip_ratio/low_mean': 0.10659844242036343, 'clip_ratio/low_min': 0.10659844242036343, 'clip_ratio/high_mean': 0.08916170708835125, 'clip_ratio/high_max': 0.08916170708835125, 'clip_ratio/region_mean': 0.19576014950871468, 'reward_total_mean': 0.5066647529602051, 'reward_meter_mean': 0.5795812606811523, 'reward_meter_std': 0.4281229078769684, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9527608156204224, 'reward_repeat_soft_std': 0.02547159232199192, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5066647529602051, 'reward_total_composite_std': 0.11989112198352814, 'epoch': 0.0}

  0%|          | 12/3300 [01:25<6:15:19,  6.85s/it]INFO 04-13 07:22:23 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 13/3300 [01:31<6:06:41,  6.69s/it]2026-04-13 07:22:29,654 | INFO | train_grpo_train | metrics_logged mode=train step=13

                                                   
{'loss': 0.0329, 'grad_norm': 17.67831039428711, 'learning_rate': 9.963636363636364e-06, 'num_tokens': 25165.0, 'completions/mean_length': 38.5, 'completions/min_length': 24.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5992826819419861, 'rewards/meter/std': 0.3339884281158447, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9907096028327942, 'rewards/repeat_soft/std': 0.014991275034844875, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5917410850524902, 'rewards/total_composite/std': 0.22081266343593597, 'reward': 0.5917410850524902, 'reward_std': 0.22081266343593597, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24050824344158173, 'sampling/sampling_logp_difference/max': 1.6661157608032227, 'sampling/importance_sampling_ratio/min': 0.18897968530654907, 'sampling/importance_sampling_ratio/mean': 0.9785161018371582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0621970295906067, 'clip_ratio/low_mean': 0.14360823296010494, 'clip_ratio/low_min': 0.14360823296010494, 'clip_ratio/high_mean': 0.09686947427690029, 'clip_ratio/high_max': 0.09686947427690029, 'clip_ratio/region_mean': 0.24047770723700523, 'reward_total_mean': 0.5917410850524902, 'reward_meter_mean': 0.5992826819419861, 'reward_meter_std': 0.3339884281158447, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9907096028327942, 'reward_repeat_soft_std': 0.014991275034844875, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5917410850524902, 'reward_total_composite_std': 0.22081266343593597, 'epoch': 0.0}

  0%|          | 13/3300 [01:31<6:06:41,  6.69s/it]INFO 04-13 07:22:29 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 14/3300 [01:40<6:38:13,  7.27s/it]2026-04-13 07:22:38,259 | INFO | train_grpo_train | metrics_logged mode=train step=14

                                                   
{'loss': -0.0512, 'grad_norm': 11.103429794311523, 'learning_rate': 9.960606060606062e-06, 'num_tokens': 27784.0, 'completions/mean_length': 133.375, 'completions/min_length': 89.0, 'completions/max_length': 209.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 133.375, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 209.0, 'rewards/meter/mean': 0.7462164163589478, 'rewards/meter/std': 0.40208691358566284, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9947032928466797, 'rewards/repeat_soft/std': 0.006884792819619179, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.1527603417634964, 'rewards/total_composite/mean': 0.6377272009849548, 'rewards/total_composite/std': 0.19227305054664612, 'reward': 0.6377272009849548, 'reward_std': 0.19227305054664612, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19112931191921234, 'sampling/sampling_logp_difference/max': 1.8332371711730957, 'sampling/importance_sampling_ratio/min': 0.18594108521938324, 'sampling/importance_sampling_ratio/mean': 1.0424418449401855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.17301943898201, 'clip_ratio/low_mean': 0.0681159496307373, 'clip_ratio/low_min': 0.0681159496307373, 'clip_ratio/high_mean': 0.11231567524373531, 'clip_ratio/high_max': 0.11231567524373531, 'clip_ratio/region_mean': 0.18043162487447262, 'reward_total_mean': 0.6377272009849548, 'reward_meter_mean': 0.7462164163589478, 'reward_meter_std': 0.40208691358566284, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9947032928466797, 'reward_repeat_soft_std': 0.006884792819619179, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.1527603417634964, 'reward_total_composite_mean': 0.6377272009849548, 'reward_total_composite_std': 0.19227305054664612, 'epoch': 0.0}

  0%|          | 14/3300 [01:40<6:38:13,  7.27s/it]INFO 04-13 07:22:38 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 15/3300 [01:47<6:38:15,  7.27s/it]2026-04-13 07:22:45,540 | INFO | train_grpo_train | metrics_logged mode=train step=15

                                                   
{'loss': 0.141, 'grad_norm': 17.71084976196289, 'learning_rate': 9.957575757575757e-06, 'num_tokens': 29510.0, 'completions/mean_length': 50.75, 'completions/min_length': 33.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.3207266628742218, 'rewards/meter/std': 0.35546308755874634, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9936358332633972, 'rewards/repeat_soft/std': 0.010752102360129356, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.4819847345352173, 'rewards/total_composite/std': 0.1911517083644867, 'reward': 0.4819847345352173, 'reward_std': 0.1911516934633255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2281627207994461, 'sampling/sampling_logp_difference/max': 2.519559860229492, 'sampling/importance_sampling_ratio/min': 0.08049502968788147, 'sampling/importance_sampling_ratio/mean': 1.0143530368804932, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7476065307855606, 'clip_ratio/low_mean': 0.17134212143719196, 'clip_ratio/low_min': 0.17134212143719196, 'clip_ratio/high_mean': 0.0547595527023077, 'clip_ratio/high_max': 0.0547595527023077, 'clip_ratio/region_mean': 0.22610167413949966, 'reward_total_mean': 0.4819847345352173, 'reward_meter_mean': 0.3207266628742218, 'reward_meter_std': 0.35546308755874634, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9936358332633972, 'reward_repeat_soft_std': 0.010752102360129356, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.4819847345352173, 'reward_total_composite_std': 0.1911517083644867, 'epoch': 0.0}

  0%|          | 15/3300 [01:47<6:38:15,  7.27s/it]INFO 04-13 07:22:45 [block_pool.py:378] Successfully reset prefix cache

  0%|          | 16/3300 [01:54<6:31:04,  7.15s/it]2026-04-13 07:22:52,386 | INFO | train_grpo_train | metrics_logged mode=train step=16

                                                   
{'loss': -0.0027, 'grad_norm': 17.285341262817383, 'learning_rate': 9.954545454545456e-06, 'num_tokens': 31067.0, 'completions/mean_length': 43.625, 'completions/min_length': 31.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4554264545440674, 'rewards/meter/std': 0.4676746129989624, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939950108528137, 'rewards/repeat_soft/std': 0.004853702150285244, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.5386902093887329, 'rewards/total_composite/std': 0.21134473383426666, 'reward': 0.5386902093887329, 'reward_std': 0.21134471893310547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23214435577392578, 'sampling/sampling_logp_difference/max': 1.6480565071105957, 'sampling/importance_sampling_ratio/min': 0.1924235224723816, 'sampling/importance_sampling_ratio/mean': 1.0233776569366455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.098197489976883, 'clip_ratio/low_mean': 0.10589986108243465, 'clip_ratio/low_min': 0.10589986108243465, 'clip_ratio/high_mean': 0.10519339889287949, 'clip_ratio/high_max': 0.10519339889287949, 'clip_ratio/region_mean': 0.21109325997531414, 'reward_total_mean': 0.5386902093887329, 'reward_meter_mean': 0.4554264545440674, 'reward_meter_std': 0.4676746129989624, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939950108528137, 'reward_repeat_soft_std': 0.004853702150285244, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.5386902093887329, 'reward_total_composite_std': 0.21134473383426666, 'epoch': 0.0}

  0%|          | 16/3300 [01:54<6:31:04,  7.15s/it]INFO 04-13 07:22:52 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 17/3300 [02:01<6:34:38,  7.21s/it]2026-04-13 07:22:59,755 | INFO | train_grpo_train | metrics_logged mode=train step=17

                                                   
{'loss': -0.0109, 'grad_norm': 17.762605667114258, 'learning_rate': 9.951515151515152e-06, 'num_tokens': 32694.0, 'completions/mean_length': 44.375, 'completions/min_length': 32.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4266776740550995, 'rewards/meter/std': 0.3314104378223419, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950998425483704, 'rewards/repeat_soft/std': 0.007583285681903362, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.47071021795272827, 'rewards/total_composite/std': 0.09301955997943878, 'reward': 0.47071021795272827, 'reward_std': 0.09301956743001938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2339089810848236, 'sampling/sampling_logp_difference/max': 1.5349130630493164, 'sampling/importance_sampling_ratio/min': 0.2154744267463684, 'sampling/importance_sampling_ratio/mean': 1.0189992189407349, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1940976977348328, 'clip_ratio/low_mean': 0.1057262234389782, 'clip_ratio/low_min': 0.1057262234389782, 'clip_ratio/high_mean': 0.07310517132282257, 'clip_ratio/high_max': 0.07310517132282257, 'clip_ratio/region_mean': 0.17883139476180077, 'reward_total_mean': 0.47071021795272827, 'reward_meter_mean': 0.4266776740550995, 'reward_meter_std': 0.3314104378223419, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950998425483704, 'reward_repeat_soft_std': 0.007583285681903362, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.47071021795272827, 'reward_total_composite_std': 0.09301955997943878, 'epoch': 0.0}

  1%|          | 17/3300 [02:01<6:34:38,  7.21s/it]INFO 04-13 07:22:59 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 18/3300 [02:08<6:17:50,  6.91s/it]2026-04-13 07:23:05,954 | INFO | train_grpo_train | metrics_logged mode=train step=18

                                                   
{'loss': -0.1495, 'grad_norm': 15.353915214538574, 'learning_rate': 9.948484848484849e-06, 'num_tokens': 34119.0, 'completions/mean_length': 29.125, 'completions/min_length': 5.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.125, 'completions/min_terminated_length': 5.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7113527059555054, 'rewards/meter/std': 0.41699108481407166, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.32732686400413513, 'rewards/total_composite/mean': 0.5885436534881592, 'rewards/total_composite/std': 0.2998615503311157, 'reward': 0.5885436534881592, 'reward_std': 0.29986152052879333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23092404007911682, 'sampling/sampling_logp_difference/max': 1.4778499603271484, 'sampling/importance_sampling_ratio/min': 0.22812765836715698, 'sampling/importance_sampling_ratio/mean': 1.0105531215667725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8655916303396225, 'clip_ratio/low_mean': 0.05056818202137947, 'clip_ratio/low_min': 0.05056818202137947, 'clip_ratio/high_mean': 0.15987520338967443, 'clip_ratio/high_max': 0.15987520338967443, 'clip_ratio/region_mean': 0.2104433854110539, 'reward_total_mean': 0.5885436534881592, 'reward_meter_mean': 0.7113527059555054, 'reward_meter_std': 0.41699108481407166, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.32732686400413513, 'reward_total_composite_mean': 0.5885436534881592, 'reward_total_composite_std': 0.2998615503311157, 'epoch': 0.0}

  1%|          | 18/3300 [02:08<6:17:50,  6.91s/it]INFO 04-13 07:23:06 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 19/3300 [02:15<6:20:47,  6.96s/it]2026-04-13 07:23:13,047 | INFO | train_grpo_train | metrics_logged mode=train step=19

                                                   
{'loss': -0.0096, 'grad_norm': 14.933284759521484, 'learning_rate': 9.945454545454546e-06, 'num_tokens': 35830.0, 'completions/mean_length': 58.875, 'completions/min_length': 43.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.7577532529830933, 'rewards/meter/std': 0.42325934767723083, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9988213777542114, 'rewards/repeat_soft/std': 0.0029925131238996983, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.24859607219696045, 'rewards/total_composite/mean': 0.7470488548278809, 'rewards/total_composite/std': 0.2544901967048645, 'reward': 0.7470488548278809, 'reward_std': 0.2544901967048645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17060860991477966, 'sampling/sampling_logp_difference/max': 1.4674263000488281, 'sampling/importance_sampling_ratio/min': 0.23051801323890686, 'sampling/importance_sampling_ratio/mean': 1.0229806900024414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2996538653969765, 'clip_ratio/low_mean': 0.08071060664951801, 'clip_ratio/low_min': 0.08071060664951801, 'clip_ratio/high_mean': 0.09031699690967798, 'clip_ratio/high_max': 0.09031699690967798, 'clip_ratio/region_mean': 0.171027603559196, 'reward_total_mean': 0.7470488548278809, 'reward_meter_mean': 0.7577532529830933, 'reward_meter_std': 0.42325934767723083, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9988213777542114, 'reward_repeat_soft_std': 0.0029925131238996983, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.24859607219696045, 'reward_total_composite_mean': 0.7470488548278809, 'reward_total_composite_std': 0.2544901967048645, 'epoch': 0.0}

  1%|          | 19/3300 [02:15<6:20:47,  6.96s/it]INFO 04-13 07:23:13 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 20/3300 [02:22<6:25:34,  7.05s/it]2026-04-13 07:23:20,309 | INFO | train_grpo_train | metrics_logged mode=train step=20

                                                   
{'loss': 0.0622, 'grad_norm': 13.815991401672363, 'learning_rate': 9.942424242424244e-06, 'num_tokens': 37941.0, 'completions/mean_length': 76.875, 'completions/min_length': 56.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.4842698574066162, 'rewards/meter/std': 0.3375895917415619, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9973297119140625, 'rewards/repeat_soft/std': 0.002192606683820486, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.18946824967861176, 'rewards/total_composite/mean': 0.36717891693115234, 'rewards/total_composite/std': 0.2395496368408203, 'reward': 0.36717891693115234, 'reward_std': 0.23954962193965912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2359773963689804, 'sampling/sampling_logp_difference/max': 1.638615608215332, 'sampling/importance_sampling_ratio/min': 0.19424878060817719, 'sampling/importance_sampling_ratio/mean': 1.0300812721252441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.34169939160347, 'clip_ratio/low_mean': 0.05780553072690964, 'clip_ratio/low_min': 0.05780553072690964, 'clip_ratio/high_mean': 0.18343774788081646, 'clip_ratio/high_max': 0.18343774788081646, 'clip_ratio/region_mean': 0.2412432786077261, 'reward_total_mean': 0.36717891693115234, 'reward_meter_mean': 0.4842698574066162, 'reward_meter_std': 0.3375895917415619, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9973297119140625, 'reward_repeat_soft_std': 0.002192606683820486, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.18946824967861176, 'reward_total_composite_mean': 0.36717891693115234, 'reward_total_composite_std': 0.2395496368408203, 'epoch': 0.0}

  1%|          | 20/3300 [02:22<6:25:34,  7.05s/it]INFO 04-13 07:23:20 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 21/3300 [02:29<6:20:01,  6.95s/it]2026-04-13 07:23:27,032 | INFO | train_grpo_train | metrics_logged mode=train step=21

                                                   
{'loss': 0.0865, 'grad_norm': 32.94882583618164, 'learning_rate': 9.939393939393939e-06, 'num_tokens': 39483.0, 'completions/mean_length': 33.75, 'completions/min_length': 26.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.2751612365245819, 'rewards/meter/std': 0.3222569227218628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9974137544631958, 'rewards/repeat_soft/std': 0.007314986549317837, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.39775148034095764, 'rewards/total_composite/std': 0.20206686854362488, 'reward': 0.39775148034095764, 'reward_std': 0.20206685364246368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2578023374080658, 'sampling/sampling_logp_difference/max': 2.409052848815918, 'sampling/importance_sampling_ratio/min': 0.0899004116654396, 'sampling/importance_sampling_ratio/mean': 1.0084861516952515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.245336264371872, 'clip_ratio/low_mean': 0.10033134557306767, 'clip_ratio/low_min': 0.10033134557306767, 'clip_ratio/high_mean': 0.0975898988544941, 'clip_ratio/high_max': 0.0975898988544941, 'clip_ratio/region_mean': 0.19792124442756176, 'reward_total_mean': 0.39775148034095764, 'reward_meter_mean': 0.2751612365245819, 'reward_meter_std': 0.3222569227218628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9974137544631958, 'reward_repeat_soft_std': 0.007314986549317837, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.39775148034095764, 'reward_total_composite_std': 0.20206686854362488, 'epoch': 0.0}

  1%|          | 21/3300 [02:29<6:20:01,  6.95s/it]INFO 04-13 07:23:27 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 22/3300 [02:36<6:22:07,  6.99s/it]2026-04-13 07:23:34,121 | INFO | train_grpo_train | metrics_logged mode=train step=22

                                                   
{'loss': 0.261, 'grad_norm': 14.979035377502441, 'learning_rate': 9.936363636363638e-06, 'num_tokens': 41219.0, 'completions/mean_length': 58.0, 'completions/min_length': 37.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8595972657203674, 'rewards/meter/std': 0.2375323474407196, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9979074001312256, 'rewards/repeat_soft/std': 0.0036863330751657486, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6432663798332214, 'rewards/total_composite/std': 0.14628660678863525, 'reward': 0.6432663798332214, 'reward_std': 0.14628660678863525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19536568224430084, 'sampling/sampling_logp_difference/max': 1.2948904037475586, 'sampling/importance_sampling_ratio/min': 0.2739278972148895, 'sampling/importance_sampling_ratio/mean': 1.0419034957885742, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9579142928123474, 'clip_ratio/low_mean': 0.08774491120129824, 'clip_ratio/low_min': 0.08774491120129824, 'clip_ratio/high_mean': 0.06056361272931099, 'clip_ratio/high_max': 0.06056361272931099, 'clip_ratio/region_mean': 0.14830852393060923, 'reward_total_mean': 0.6432663798332214, 'reward_meter_mean': 0.8595972657203674, 'reward_meter_std': 0.2375323474407196, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9979074001312256, 'reward_repeat_soft_std': 0.0036863330751657486, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6432663798332214, 'reward_total_composite_std': 0.14628660678863525, 'epoch': 0.0}

  1%|          | 22/3300 [02:36<6:22:07,  6.99s/it]INFO 04-13 07:23:34 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 23/3300 [02:44<6:35:53,  7.25s/it]2026-04-13 07:23:41,963 | INFO | train_grpo_train | metrics_logged mode=train step=23

                                                   
{'loss': 0.1544, 'grad_norm': 17.582691192626953, 'learning_rate': 9.933333333333334e-06, 'num_tokens': 42894.0, 'completions/mean_length': 49.375, 'completions/min_length': 39.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.4682067632675171, 'rewards/meter/std': 0.39334917068481445, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981834888458252, 'rewards/repeat_soft/std': 0.0032929633744060993, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.22965426743030548, 'rewards/total_composite/mean': 0.5504657626152039, 'rewards/total_composite/std': 0.20627839863300323, 'reward': 0.5504657626152039, 'reward_std': 0.20627839863300323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1894393414258957, 'sampling/sampling_logp_difference/max': 2.1364288330078125, 'sampling/importance_sampling_ratio/min': 0.11807575821876526, 'sampling/importance_sampling_ratio/mean': 1.0336065292358398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6671224385499954, 'clip_ratio/low_mean': 0.10553514678031206, 'clip_ratio/low_min': 0.10553514678031206, 'clip_ratio/high_mean': 0.05519563052803278, 'clip_ratio/high_max': 0.05519563052803278, 'clip_ratio/region_mean': 0.16073077730834484, 'reward_total_mean': 0.5504657626152039, 'reward_meter_mean': 0.4682067632675171, 'reward_meter_std': 0.39334917068481445, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981834888458252, 'reward_repeat_soft_std': 0.0032929633744060993, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.22965426743030548, 'reward_total_composite_mean': 0.5504657626152039, 'reward_total_composite_std': 0.20627839863300323, 'epoch': 0.0}

  1%|          | 23/3300 [02:44<6:35:53,  7.25s/it]INFO 04-13 07:23:42 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 24/3300 [02:52<6:47:55,  7.47s/it]2026-04-13 07:23:49,952 | INFO | train_grpo_train | metrics_logged mode=train step=24

                                                   
{'loss': 0.0501, 'grad_norm': 13.180197715759277, 'learning_rate': 9.930303030303031e-06, 'num_tokens': 44607.0, 'completions/mean_length': 62.125, 'completions/min_length': 46.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.125, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.21943849325180054, 'rewards/meter/std': 0.14068876206874847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984592199325562, 'rewards/repeat_soft/std': 0.00178744294680655, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.4418691396713257, 'rewards/total_composite/std': 0.07863625138998032, 'reward': 0.4418691396713257, 'reward_std': 0.07863625138998032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18320536613464355, 'sampling/sampling_logp_difference/max': 1.4752097129821777, 'sampling/importance_sampling_ratio/min': 0.3220268189907074, 'sampling/importance_sampling_ratio/mean': 1.0245449542999268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4479553326964378, 'clip_ratio/low_mean': 0.1257912116125226, 'clip_ratio/low_min': 0.1257912116125226, 'clip_ratio/high_mean': 0.07774893380701542, 'clip_ratio/high_max': 0.07774893380701542, 'clip_ratio/region_mean': 0.20354014541953802, 'reward_total_mean': 0.4418691396713257, 'reward_meter_mean': 0.21943849325180054, 'reward_meter_std': 0.14068876206874847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984592199325562, 'reward_repeat_soft_std': 0.00178744294680655, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.4418691396713257, 'reward_total_composite_std': 0.07863625138998032, 'epoch': 0.0}

  1%|          | 24/3300 [02:52<6:47:55,  7.47s/it]INFO 04-13 07:23:50 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 25/3300 [03:00<7:02:54,  7.75s/it]2026-04-13 07:23:58,346 | INFO | train_grpo_train | metrics_logged mode=train step=25

                                                   
{'loss': 0.0264, 'grad_norm': 8.850423812866211, 'learning_rate': 9.927272727272728e-06, 'num_tokens': 47239.0, 'completions/mean_length': 120.0, 'completions/min_length': 113.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.0, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.3736605644226074, 'rewards/meter/std': 0.2752265930175781, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975237250328064, 'rewards/repeat_soft/std': 0.001548389787785709, 'rewards/judge_quality/mean': 0.809999942779541, 'rewards/judge_quality/std': 0.19146056473255157, 'rewards/total_composite/mean': 0.5476093888282776, 'rewards/total_composite/std': 0.16478323936462402, 'reward': 0.5476093888282776, 'reward_std': 0.16478323936462402, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11544646322727203, 'sampling/sampling_logp_difference/max': 1.9932332038879395, 'sampling/importance_sampling_ratio/min': 0.13625416159629822, 'sampling/importance_sampling_ratio/mean': 1.0179648399353027, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7649307884275913, 'clip_ratio/low_mean': 0.0729277259670198, 'clip_ratio/low_min': 0.0729277259670198, 'clip_ratio/high_mean': 0.0333252027630806, 'clip_ratio/high_max': 0.0333252027630806, 'clip_ratio/region_mean': 0.1062529287301004, 'reward_total_mean': 0.5476093888282776, 'reward_meter_mean': 0.3736605644226074, 'reward_meter_std': 0.2752265930175781, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975237250328064, 'reward_repeat_soft_std': 0.001548389787785709, 'reward_judge_quality_mean': 0.809999942779541, 'reward_judge_quality_std': 0.19146056473255157, 'reward_total_composite_mean': 0.5476093888282776, 'reward_total_composite_std': 0.16478323936462402, 'epoch': 0.0}

  1%|          | 25/3300 [03:00<7:02:54,  7.75s/it]INFO 04-13 07:23:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 26/3300 [03:06<6:34:57,  7.24s/it]2026-04-13 07:24:04,393 | INFO | train_grpo_train | metrics_logged mode=train step=26

                                                   
{'loss': 0.0341, 'grad_norm': 11.169930458068848, 'learning_rate': 9.924242424242425e-06, 'num_tokens': 48753.0, 'completions/mean_length': 32.25, 'completions/min_length': 31.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9985294342041016, 'rewards/meter/std': 0.0007957193301990628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6169735193252563, 'rewards/total_composite/std': 0.00021723672398366034, 'reward': 0.6169735193252563, 'reward_std': 0.0002172399399569258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.044476937502622604, 'sampling/sampling_logp_difference/max': 0.9176238775253296, 'sampling/importance_sampling_ratio/min': 0.3994671106338501, 'sampling/importance_sampling_ratio/mean': 1.0029505491256714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2671094238758087, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.02726221503689885, 'clip_ratio/high_max': 0.02726221503689885, 'clip_ratio/region_mean': 0.0451193586923182, 'reward_total_mean': 0.6169735193252563, 'reward_meter_mean': 0.9985294342041016, 'reward_meter_std': 0.0007957193301990628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6169735193252563, 'reward_total_composite_std': 0.00021723672398366034, 'epoch': 0.0}

  1%|          | 26/3300 [03:06<6:34:57,  7.24s/it]INFO 04-13 07:24:04 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 27/3300 [03:14<6:53:24,  7.58s/it]2026-04-13 07:24:12,766 | INFO | train_grpo_train | metrics_logged mode=train step=27

                                                   
{'loss': -0.0472, 'grad_norm': 8.745513916015625, 'learning_rate': 9.921212121212121e-06, 'num_tokens': 51435.0, 'completions/mean_length': 138.25, 'completions/min_length': 83.0, 'completions/max_length': 191.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 191.0, 'rewards/meter/mean': 0.45913514494895935, 'rewards/meter/std': 0.3512060344219208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.999238133430481, 'rewards/repeat_soft/std': 0.0007623151177540421, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.20469054579734802, 'rewards/total_composite/mean': 0.43480345606803894, 'rewards/total_composite/std': 0.19026042520999908, 'reward': 0.43480345606803894, 'reward_std': 0.1902604103088379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2304992526769638, 'sampling/sampling_logp_difference/max': 1.7399463653564453, 'sampling/importance_sampling_ratio/min': 0.17552980780601501, 'sampling/importance_sampling_ratio/mean': 1.04105544090271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5277575254440308, 'clip_ratio/low_mean': 0.07658866792917252, 'clip_ratio/low_min': 0.07658866792917252, 'clip_ratio/high_mean': 0.15760960429906845, 'clip_ratio/high_max': 0.15760960429906845, 'clip_ratio/region_mean': 0.23419827222824097, 'reward_total_mean': 0.43480345606803894, 'reward_meter_mean': 0.45913514494895935, 'reward_meter_std': 0.3512060344219208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.999238133430481, 'reward_repeat_soft_std': 0.0007623151177540421, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.20469054579734802, 'reward_total_composite_mean': 0.43480345606803894, 'reward_total_composite_std': 0.19026042520999908, 'epoch': 0.0}

  1%|          | 27/3300 [03:14<6:53:24,  7.58s/it]INFO 04-13 07:24:13 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 28/3300 [03:22<6:44:50,  7.42s/it]2026-04-13 07:24:19,830 | INFO | train_grpo_train | metrics_logged mode=train step=28

                                                   
{'loss': 0.0977, 'grad_norm': 23.730607986450195, 'learning_rate': 9.918181818181818e-06, 'num_tokens': 52963.0, 'completions/mean_length': 29.0, 'completions/min_length': 16.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6324816346168518, 'rewards/meter/std': 0.4750705063343048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9585226774215698, 'rewards/repeat_soft/std': 0.011249415576457977, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5238929390907288, 'rewards/total_composite/std': 0.13481268286705017, 'reward': 0.5238929390907288, 'reward_std': 0.13481266796588898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19141219556331635, 'sampling/sampling_logp_difference/max': 1.2272367477416992, 'sampling/importance_sampling_ratio/min': 0.29310137033462524, 'sampling/importance_sampling_ratio/mean': 1.0181347131729126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8611400574445724, 'clip_ratio/low_mean': 0.07508680690079927, 'clip_ratio/low_min': 0.07508680690079927, 'clip_ratio/high_mean': 0.15117423422634602, 'clip_ratio/high_max': 0.15117423422634602, 'clip_ratio/region_mean': 0.2262610411271453, 'reward_total_mean': 0.5238929390907288, 'reward_meter_mean': 0.6324816346168518, 'reward_meter_std': 0.4750705063343048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9585226774215698, 'reward_repeat_soft_std': 0.011249415576457977, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5238929390907288, 'reward_total_composite_std': 0.13481268286705017, 'epoch': 0.0}

  1%|          | 28/3300 [03:22<6:44:50,  7.42s/it]INFO 04-13 07:24:20 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 29/3300 [03:31<7:10:14,  7.89s/it]2026-04-13 07:24:28,815 | INFO | train_grpo_train | metrics_logged mode=train step=29

                                                   
{'loss': 0.062, 'grad_norm': 7.373910427093506, 'learning_rate': 9.915151515151515e-06, 'num_tokens': 56000.0, 'completions/mean_length': 179.625, 'completions/min_length': 150.0, 'completions/max_length': 217.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 179.625, 'completions/min_terminated_length': 150.0, 'completions/max_terminated_length': 217.0, 'rewards/meter/mean': 0.8581138253211975, 'rewards/meter/std': 0.28778037428855896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989545345306396, 'rewards/repeat_soft/std': 0.000916292832698673, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.5919029712677002, 'rewards/total_composite/std': 0.11980368196964264, 'reward': 0.5919029712677002, 'reward_std': 0.11980368942022324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20266062021255493, 'sampling/sampling_logp_difference/max': 1.7890448570251465, 'sampling/importance_sampling_ratio/min': 0.16711971163749695, 'sampling/importance_sampling_ratio/mean': 1.0425152778625488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3875505477190018, 'clip_ratio/low_mean': 0.07503936253488064, 'clip_ratio/low_min': 0.07503936253488064, 'clip_ratio/high_mean': 0.13527059368789196, 'clip_ratio/high_max': 0.13527059368789196, 'clip_ratio/region_mean': 0.2103099562227726, 'reward_total_mean': 0.5919029712677002, 'reward_meter_mean': 0.8581138253211975, 'reward_meter_std': 0.28778037428855896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989545345306396, 'reward_repeat_soft_std': 0.000916292832698673, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.5919029712677002, 'reward_total_composite_std': 0.11980368196964264, 'epoch': 0.0}

  1%|          | 29/3300 [03:31<7:10:14,  7.89s/it]INFO 04-13 07:24:29 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 30/3300 [03:38<7:00:42,  7.72s/it]2026-04-13 07:24:36,131 | INFO | train_grpo_train | metrics_logged mode=train step=30

                                                   
{'loss': -0.0097, 'grad_norm': 12.832189559936523, 'learning_rate': 9.912121212121213e-06, 'num_tokens': 57955.0, 'completions/mean_length': 60.375, 'completions/min_length': 37.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.4923819899559021, 'rewards/meter/std': 0.4194702208042145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.998965859413147, 'rewards/repeat_soft/std': 0.002180325798690319, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.5113567113876343, 'rewards/total_composite/std': 0.15652590990066528, 'reward': 0.5113567113876343, 'reward_std': 0.15652593970298767, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2167627066373825, 'sampling/sampling_logp_difference/max': 1.6749815940856934, 'sampling/importance_sampling_ratio/min': 0.18731161952018738, 'sampling/importance_sampling_ratio/mean': 1.0154328346252441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.061600148677826, 'clip_ratio/low_mean': 0.1053523225709796, 'clip_ratio/low_min': 0.1053523225709796, 'clip_ratio/high_mean': 0.09013216756284237, 'clip_ratio/high_max': 0.09013216756284237, 'clip_ratio/region_mean': 0.19548449013382196, 'reward_total_mean': 0.5113567113876343, 'reward_meter_mean': 0.4923819899559021, 'reward_meter_std': 0.4194702208042145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.998965859413147, 'reward_repeat_soft_std': 0.002180325798690319, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.5113567113876343, 'reward_total_composite_std': 0.15652590990066528, 'epoch': 0.0}

  1%|          | 30/3300 [03:38<7:00:42,  7.72s/it]INFO 04-13 07:24:36 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 31/3300 [03:45<6:55:18,  7.62s/it]2026-04-13 07:24:43,527 | INFO | train_grpo_train | metrics_logged mode=train step=31

                                                   
{'loss': 0.0774, 'grad_norm': 17.354585647583008, 'learning_rate': 9.90909090909091e-06, 'num_tokens': 59958.0, 'completions/mean_length': 78.375, 'completions/min_length': 60.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.375, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.3928065001964569, 'rewards/meter/std': 0.29107925295829773, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9918286800384521, 'rewards/repeat_soft/std': 0.006046677939593792, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.46345409750938416, 'rewards/total_composite/std': 0.07556276023387909, 'reward': 0.46345409750938416, 'reward_std': 0.07556275278329849, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23961949348449707, 'sampling/sampling_logp_difference/max': 2.6145894527435303, 'sampling/importance_sampling_ratio/min': 0.07319783419370651, 'sampling/importance_sampling_ratio/mean': 0.9922016263008118, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2000147253274918, 'clip_ratio/low_mean': 0.09721448086202145, 'clip_ratio/low_min': 0.09721448086202145, 'clip_ratio/high_mean': 0.10816307179629803, 'clip_ratio/high_max': 0.10816307179629803, 'clip_ratio/region_mean': 0.20537755265831947, 'reward_total_mean': 0.46345409750938416, 'reward_meter_mean': 0.3928065001964569, 'reward_meter_std': 0.29107925295829773, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9918286800384521, 'reward_repeat_soft_std': 0.006046677939593792, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.46345409750938416, 'reward_total_composite_std': 0.07556276023387909, 'epoch': 0.0}

  1%|          | 31/3300 [03:45<6:55:18,  7.62s/it]INFO 04-13 07:24:43 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 32/3300 [03:52<6:47:26,  7.48s/it]2026-04-13 07:24:50,677 | INFO | train_grpo_train | metrics_logged mode=train step=32

                                                   
{'loss': 0.0951, 'grad_norm': 33.8635139465332, 'learning_rate': 9.906060606060607e-06, 'num_tokens': 61400.0, 'completions/mean_length': 25.25, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.2765427529811859, 'rewards/meter/std': 0.24548844993114471, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/total_composite/mean': 0.47309568524360657, 'rewards/total_composite/std': 0.12770362198352814, 'reward': 0.47309568524360657, 'reward_std': 0.12770362198352814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24985970556735992, 'sampling/sampling_logp_difference/max': 1.6838467121124268, 'sampling/importance_sampling_ratio/min': 0.18565842509269714, 'sampling/importance_sampling_ratio/mean': 0.9918996691703796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.68735009431839, 'clip_ratio/low_mean': 0.08237721398472786, 'clip_ratio/low_min': 0.08237721398472786, 'clip_ratio/high_mean': 0.10065257363021374, 'clip_ratio/high_max': 0.10065257363021374, 'clip_ratio/region_mean': 0.1830297876149416, 'reward_total_mean': 0.47309568524360657, 'reward_meter_mean': 0.2765427529811859, 'reward_meter_std': 0.24548844993114471, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.47309568524360657, 'reward_total_composite_std': 0.12770362198352814, 'epoch': 0.0}

  1%|          | 32/3300 [03:52<6:47:26,  7.48s/it]INFO 04-13 07:24:50 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 33/3300 [04:00<6:48:14,  7.50s/it]2026-04-13 07:24:58,214 | INFO | train_grpo_train | metrics_logged mode=train step=33

                                                   
{'loss': -0.0659, 'grad_norm': 18.40517234802246, 'learning_rate': 9.903030303030305e-06, 'num_tokens': 63006.0, 'completions/mean_length': 36.75, 'completions/min_length': 25.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.47970691323280334, 'rewards/meter/std': 0.39111989736557007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9984102249145508, 'rewards/repeat_soft/std': 0.0024311586748808622, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.45416176319122314, 'rewards/total_composite/std': 0.22378690540790558, 'reward': 0.45416176319122314, 'reward_std': 0.22378690540790558, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20951895415782928, 'sampling/sampling_logp_difference/max': 1.2337064743041992, 'sampling/importance_sampling_ratio/min': 0.29121121764183044, 'sampling/importance_sampling_ratio/mean': 1.0673010349273682, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.090721681714058, 'clip_ratio/low_mean': 0.07213383819907904, 'clip_ratio/low_min': 0.07213383819907904, 'clip_ratio/high_mean': 0.12102419137954712, 'clip_ratio/high_max': 0.12102419137954712, 'clip_ratio/region_mean': 0.19315802957862616, 'reward_total_mean': 0.45416176319122314, 'reward_meter_mean': 0.47970691323280334, 'reward_meter_std': 0.39111989736557007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9984102249145508, 'reward_repeat_soft_std': 0.0024311586748808622, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.45416176319122314, 'reward_total_composite_std': 0.22378690540790558, 'epoch': 0.0}

  1%|          | 33/3300 [04:00<6:48:14,  7.50s/it]INFO 04-13 07:24:58 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 34/3300 [04:07<6:40:32,  7.36s/it]2026-04-13 07:25:05,247 | INFO | train_grpo_train | metrics_logged mode=train step=34

                                                   
{'loss': 0.0987, 'grad_norm': 16.91385841369629, 'learning_rate': 9.9e-06, 'num_tokens': 64599.0, 'completions/mean_length': 41.125, 'completions/min_length': 27.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.3966813385486603, 'rewards/meter/std': 0.3952701985836029, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9912532567977905, 'rewards/repeat_soft/std': 0.009152887389063835, 'rewards/judge_quality/mean': 0.690000057220459, 'rewards/judge_quality/std': 0.22315914928913116, 'rewards/total_composite/mean': 0.5417426824569702, 'rewards/total_composite/std': 0.21540047228336334, 'reward': 0.5417426824569702, 'reward_std': 0.21540047228336334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2257329225540161, 'sampling/sampling_logp_difference/max': 2.2196807861328125, 'sampling/importance_sampling_ratio/min': 0.1086437851190567, 'sampling/importance_sampling_ratio/mean': 1.0293428897857666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3586771450936794, 'clip_ratio/low_mean': 0.1340062189847231, 'clip_ratio/low_min': 0.1340062189847231, 'clip_ratio/high_mean': 0.07132523134350777, 'clip_ratio/high_max': 0.07132523134350777, 'clip_ratio/region_mean': 0.20533145032823086, 'reward_total_mean': 0.5417426824569702, 'reward_meter_mean': 0.3966813385486603, 'reward_meter_std': 0.3952701985836029, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9912532567977905, 'reward_repeat_soft_std': 0.009152887389063835, 'reward_judge_quality_mean': 0.690000057220459, 'reward_judge_quality_std': 0.22315914928913116, 'reward_total_composite_mean': 0.5417426824569702, 'reward_total_composite_std': 0.21540047228336334, 'epoch': 0.0}

  1%|          | 34/3300 [04:07<6:40:32,  7.36s/it]INFO 04-13 07:25:05 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 35/3300 [04:13<6:24:03,  7.06s/it]2026-04-13 07:25:11,604 | INFO | train_grpo_train | metrics_logged mode=train step=35

                                                   
{'loss': 0.0704, 'grad_norm': 27.953187942504883, 'learning_rate': 9.896969696969699e-06, 'num_tokens': 66168.0, 'completions/mean_length': 31.125, 'completions/min_length': 24.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7837445735931396, 'rewards/meter/std': 0.2611982822418213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9804645776748657, 'rewards/repeat_soft/std': 0.03129582852125168, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.15250293910503387, 'rewards/total_composite/mean': 0.6702450513839722, 'rewards/total_composite/std': 0.1414785236120224, 'reward': 0.6702450513839722, 'reward_std': 0.1414785236120224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1811247318983078, 'sampling/sampling_logp_difference/max': 1.859018325805664, 'sampling/importance_sampling_ratio/min': 0.15582552552223206, 'sampling/importance_sampling_ratio/mean': 0.959926187992096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6075849942862988, 'clip_ratio/low_mean': 0.06071342155337334, 'clip_ratio/low_min': 0.06071342155337334, 'clip_ratio/high_mean': 0.083756223320961, 'clip_ratio/high_max': 0.083756223320961, 'clip_ratio/region_mean': 0.14446964487433434, 'reward_total_mean': 0.6702450513839722, 'reward_meter_mean': 0.7837445735931396, 'reward_meter_std': 0.2611982822418213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9804645776748657, 'reward_repeat_soft_std': 0.03129582852125168, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.15250293910503387, 'reward_total_composite_mean': 0.6702450513839722, 'reward_total_composite_std': 0.1414785236120224, 'epoch': 0.0}

  1%|          | 35/3300 [04:13<6:24:03,  7.06s/it]INFO 04-13 07:25:11 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 36/3300 [04:20<6:14:56,  6.89s/it]2026-04-13 07:25:18,110 | INFO | train_grpo_train | metrics_logged mode=train step=36

                                                   
{'loss': 0.0955, 'grad_norm': 15.208176612854004, 'learning_rate': 9.893939393939395e-06, 'num_tokens': 67682.0, 'completions/mean_length': 40.25, 'completions/min_length': 32.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.4485693871974945, 'rewards/meter/std': 0.3291569948196411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953606724739075, 'rewards/repeat_soft/std': 0.0055753388442099094, 'rewards/judge_quality/mean': 0.7350000143051147, 'rewards/judge_quality/std': 0.15014280378818512, 'rewards/total_composite/mean': 0.5608401298522949, 'rewards/total_composite/std': 0.15952417254447937, 'reward': 0.5608401298522949, 'reward_std': 0.15952417254447937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18473203480243683, 'sampling/sampling_logp_difference/max': 1.8001375198364258, 'sampling/importance_sampling_ratio/min': 0.1652761697769165, 'sampling/importance_sampling_ratio/mean': 1.016972541809082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4089657962322235, 'clip_ratio/low_mean': 0.07581416238099337, 'clip_ratio/low_min': 0.07581416238099337, 'clip_ratio/high_mean': 0.08979000151157379, 'clip_ratio/high_max': 0.08979000151157379, 'clip_ratio/region_mean': 0.16560416389256716, 'reward_total_mean': 0.5608401298522949, 'reward_meter_mean': 0.4485693871974945, 'reward_meter_std': 0.3291569948196411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953606724739075, 'reward_repeat_soft_std': 0.0055753388442099094, 'reward_judge_quality_mean': 0.7350000143051147, 'reward_judge_quality_std': 0.15014280378818512, 'reward_total_composite_mean': 0.5608401298522949, 'reward_total_composite_std': 0.15952417254447937, 'epoch': 0.0}

  1%|          | 36/3300 [04:20<6:14:56,  6.89s/it]INFO 04-13 07:25:18 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 37/3300 [04:26<6:09:29,  6.79s/it]2026-04-13 07:25:24,676 | INFO | train_grpo_train | metrics_logged mode=train step=37

                                                   
{'loss': -0.0345, 'grad_norm': 15.821805000305176, 'learning_rate': 9.890909090909092e-06, 'num_tokens': 69454.0, 'completions/mean_length': 50.5, 'completions/min_length': 34.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7323452234268188, 'rewards/meter/std': 0.3667115867137909, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9898346066474915, 'rewards/repeat_soft/std': 0.013633593916893005, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6005253195762634, 'rewards/total_composite/std': 0.16267496347427368, 'reward': 0.6005253195762634, 'reward_std': 0.16267496347427368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22236742079257965, 'sampling/sampling_logp_difference/max': 1.5800833702087402, 'sampling/importance_sampling_ratio/min': 0.20595793426036835, 'sampling/importance_sampling_ratio/mean': 1.0082014799118042, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.157983884215355, 'clip_ratio/low_mean': 0.10209287982434034, 'clip_ratio/low_min': 0.10209287982434034, 'clip_ratio/high_mean': 0.11077776551246643, 'clip_ratio/high_max': 0.11077776551246643, 'clip_ratio/region_mean': 0.21287064533680677, 'reward_total_mean': 0.6005253195762634, 'reward_meter_mean': 0.7323452234268188, 'reward_meter_std': 0.3667115867137909, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9898346066474915, 'reward_repeat_soft_std': 0.013633593916893005, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6005253195762634, 'reward_total_composite_std': 0.16267496347427368, 'epoch': 0.0}

  1%|          | 37/3300 [04:26<6:09:29,  6.79s/it]INFO 04-13 07:25:24 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 38/3300 [04:34<6:15:48,  6.91s/it]2026-04-13 07:25:31,864 | INFO | train_grpo_train | metrics_logged mode=train step=38

                                                   
{'loss': 0.0838, 'grad_norm': 16.704303741455078, 'learning_rate': 9.887878787878789e-06, 'num_tokens': 71106.0, 'completions/mean_length': 53.5, 'completions/min_length': 33.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5510157942771912, 'rewards/meter/std': 0.41519367694854736, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980113506317139, 'rewards/repeat_soft/std': 0.005624707788228989, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.13979578018188477, 'rewards/total_composite/mean': 0.5337550640106201, 'rewards/total_composite/std': 0.1366468220949173, 'reward': 0.5337550640106201, 'reward_std': 0.1366468071937561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19399528205394745, 'sampling/sampling_logp_difference/max': 1.4106621742248535, 'sampling/importance_sampling_ratio/min': 0.24398167431354523, 'sampling/importance_sampling_ratio/mean': 1.0123108625411987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4564415216445923, 'clip_ratio/low_mean': 0.07825370226055384, 'clip_ratio/low_min': 0.07825370226055384, 'clip_ratio/high_mean': 0.11193818785250187, 'clip_ratio/high_max': 0.11193818785250187, 'clip_ratio/region_mean': 0.1901918901130557, 'reward_total_mean': 0.5337550640106201, 'reward_meter_mean': 0.5510157942771912, 'reward_meter_std': 0.41519367694854736, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980113506317139, 'reward_repeat_soft_std': 0.005624707788228989, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.13979578018188477, 'reward_total_composite_mean': 0.5337550640106201, 'reward_total_composite_std': 0.1366468220949173, 'epoch': 0.0}

  1%|          | 38/3300 [04:34<6:15:48,  6.91s/it]INFO 04-13 07:25:32 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 39/3300 [04:41<6:22:53,  7.04s/it]2026-04-13 07:25:39,218 | INFO | train_grpo_train | metrics_logged mode=train step=39

                                                   
{'loss': 0.1019, 'grad_norm': 35.369117736816406, 'learning_rate': 9.884848484848486e-06, 'num_tokens': 72810.0, 'completions/mean_length': 55.0, 'completions/min_length': 48.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.770334005355835, 'rewards/meter/std': 0.22616279125213623, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9931648969650269, 'rewards/repeat_soft/std': 0.008299742825329304, 'rewards/judge_quality/mean': 0.6524999737739563, 'rewards/judge_quality/std': 0.20190168917179108, 'rewards/total_composite/mean': 0.685139536857605, 'rewards/total_composite/std': 0.15254512429237366, 'reward': 0.685139536857605, 'reward_std': 0.15254512429237366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14268994331359863, 'sampling/sampling_logp_difference/max': 2.045924663543701, 'sampling/importance_sampling_ratio/min': 0.1292606145143509, 'sampling/importance_sampling_ratio/mean': 1.000756025314331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5295688956975937, 'clip_ratio/low_mean': 0.06531935092061758, 'clip_ratio/low_min': 0.06531935092061758, 'clip_ratio/high_mean': 0.04677868401631713, 'clip_ratio/high_max': 0.04677868401631713, 'clip_ratio/region_mean': 0.11209803493693471, 'reward_total_mean': 0.685139536857605, 'reward_meter_mean': 0.770334005355835, 'reward_meter_std': 0.22616279125213623, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9931648969650269, 'reward_repeat_soft_std': 0.008299742825329304, 'reward_judge_quality_mean': 0.6524999737739563, 'reward_judge_quality_std': 0.20190168917179108, 'reward_total_composite_mean': 0.685139536857605, 'reward_total_composite_std': 0.15254512429237366, 'epoch': 0.0}

  1%|          | 39/3300 [04:41<6:22:53,  7.04s/it]INFO 04-13 07:25:39 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 40/3300 [04:48<6:23:55,  7.07s/it]2026-04-13 07:25:46,335 | INFO | train_grpo_train | metrics_logged mode=train step=40

                                                   
{'loss': 0.2088, 'grad_norm': 23.99601173400879, 'learning_rate': 9.881818181818182e-06, 'num_tokens': 74420.0, 'completions/mean_length': 38.25, 'completions/min_length': 28.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6893784403800964, 'rewards/meter/std': 0.4171596169471741, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9735429286956787, 'rewards/repeat_soft/std': 0.04065443575382233, 'rewards/judge_quality/mean': 0.8825000524520874, 'rewards/judge_quality/std': 0.07440238445997238, 'rewards/total_composite/mean': 0.757306694984436, 'rewards/total_composite/std': 0.2491140067577362, 'reward': 0.757306694984436, 'reward_std': 0.24911397695541382, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2222781926393509, 'sampling/sampling_logp_difference/max': 2.1597957611083984, 'sampling/importance_sampling_ratio/min': 0.115348681807518, 'sampling/importance_sampling_ratio/mean': 1.0162009000778198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0633226111531258, 'clip_ratio/low_mean': 0.08311870507895947, 'clip_ratio/low_min': 0.08311870507895947, 'clip_ratio/high_mean': 0.09333791490644217, 'clip_ratio/high_max': 0.09333791490644217, 'clip_ratio/region_mean': 0.17645661998540163, 'reward_total_mean': 0.757306694984436, 'reward_meter_mean': 0.6893784403800964, 'reward_meter_std': 0.4171596169471741, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9735429286956787, 'reward_repeat_soft_std': 0.04065443575382233, 'reward_judge_quality_mean': 0.8825000524520874, 'reward_judge_quality_std': 0.07440238445997238, 'reward_total_composite_mean': 0.757306694984436, 'reward_total_composite_std': 0.2491140067577362, 'epoch': 0.0}

  1%|          | 40/3300 [04:48<6:23:55,  7.07s/it]INFO 04-13 07:25:46 [block_pool.py:378] Successfully reset prefix cache

  1%|          | 41/3300 [04:56<6:39:02,  7.35s/it]2026-04-13 07:25:54,335 | INFO | train_grpo_train | metrics_logged mode=train step=41

                                                   
{'loss': 0.1591, 'grad_norm': 14.506078720092773, 'learning_rate': 9.87878787878788e-06, 'num_tokens': 76186.0, 'completions/mean_length': 60.75, 'completions/min_length': 33.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.6156733632087708, 'rewards/meter/std': 0.4310094714164734, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9923155903816223, 'rewards/repeat_soft/std': 0.013626428321003914, 'rewards/judge_quality/mean': 0.45124998688697815, 'rewards/judge_quality/std': 0.12799973785877228, 'rewards/total_composite/mean': 0.5048567056655884, 'rewards/total_composite/std': 0.14636343717575073, 'reward': 0.5048567056655884, 'reward_std': 0.14636343717575073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20026902854442596, 'sampling/sampling_logp_difference/max': 1.480910301208496, 'sampling/importance_sampling_ratio/min': 0.2274305820465088, 'sampling/importance_sampling_ratio/mean': 1.0431978702545166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1705337911844254, 'clip_ratio/low_mean': 0.06033051759004593, 'clip_ratio/low_min': 0.06033051759004593, 'clip_ratio/high_mean': 0.10448822379112244, 'clip_ratio/high_max': 0.10448822379112244, 'clip_ratio/region_mean': 0.16481874138116837, 'reward_total_mean': 0.5048567056655884, 'reward_meter_mean': 0.6156733632087708, 'reward_meter_std': 0.4310094714164734, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9923155903816223, 'reward_repeat_soft_std': 0.013626428321003914, 'reward_judge_quality_mean': 0.45124998688697815, 'reward_judge_quality_std': 0.12799973785877228, 'reward_total_composite_mean': 0.5048567056655884, 'reward_total_composite_std': 0.14636343717575073, 'epoch': 0.0}

  1%|          | 41/3300 [04:56<6:39:02,  7.35s/it]INFO 04-13 07:25:54 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 42/3300 [05:04<6:41:40,  7.40s/it]2026-04-13 07:26:01,850 | INFO | train_grpo_train | metrics_logged mode=train step=42

                                                   
{'loss': -0.0256, 'grad_norm': 16.84304428100586, 'learning_rate': 9.875757575757576e-06, 'num_tokens': 77725.0, 'completions/mean_length': 43.375, 'completions/min_length': 25.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.582551121711731, 'rewards/meter/std': 0.4320070743560791, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.977229654788971, 'rewards/repeat_soft/std': 0.020586658269166946, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.3022770881652832, 'rewards/total_composite/mean': 0.5598830580711365, 'rewards/total_composite/std': 0.3062882721424103, 'reward': 0.5598830580711365, 'reward_std': 0.3062882423400879, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21116963028907776, 'sampling/sampling_logp_difference/max': 2.439089059829712, 'sampling/importance_sampling_ratio/min': 0.08724027872085571, 'sampling/importance_sampling_ratio/mean': 1.0060738325119019, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6929726600646973, 'clip_ratio/low_mean': 0.0896103922277689, 'clip_ratio/low_min': 0.0896103922277689, 'clip_ratio/high_mean': 0.16354099847376347, 'clip_ratio/high_max': 0.16354099847376347, 'clip_ratio/region_mean': 0.25315139070153236, 'reward_total_mean': 0.5598830580711365, 'reward_meter_mean': 0.582551121711731, 'reward_meter_std': 0.4320070743560791, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.977229654788971, 'reward_repeat_soft_std': 0.020586658269166946, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.3022770881652832, 'reward_total_composite_mean': 0.5598830580711365, 'reward_total_composite_std': 0.3062882721424103, 'epoch': 0.0}

  1%|▏         | 42/3300 [05:04<6:41:40,  7.40s/it]INFO 04-13 07:26:02 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 43/3300 [05:11<6:35:57,  7.29s/it]2026-04-13 07:26:08,904 | INFO | train_grpo_train | metrics_logged mode=train step=43

                                                   
{'loss': -0.0522, 'grad_norm': 17.400814056396484, 'learning_rate': 9.872727272727274e-06, 'num_tokens': 79484.0, 'completions/mean_length': 44.875, 'completions/min_length': 31.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.624340295791626, 'rewards/meter/std': 0.4104929268360138, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.997374415397644, 'rewards/repeat_soft/std': 0.004505294840782881, 'rewards/judge_quality/mean': 0.7224999666213989, 'rewards/judge_quality/std': 0.232854962348938, 'rewards/total_composite/mean': 0.6401236653327942, 'rewards/total_composite/std': 0.20709064602851868, 'reward': 0.6401236653327942, 'reward_std': 0.2070906162261963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22349603474140167, 'sampling/sampling_logp_difference/max': 1.757357120513916, 'sampling/importance_sampling_ratio/min': 0.17250016331672668, 'sampling/importance_sampling_ratio/mean': 1.0200906991958618, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7742891013622284, 'clip_ratio/low_mean': 0.12815146706998348, 'clip_ratio/low_min': 0.12815146706998348, 'clip_ratio/high_mean': 0.08952812105417252, 'clip_ratio/high_max': 0.08952812105417252, 'clip_ratio/region_mean': 0.217679588124156, 'reward_total_mean': 0.6401236653327942, 'reward_meter_mean': 0.624340295791626, 'reward_meter_std': 0.4104929268360138, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.997374415397644, 'reward_repeat_soft_std': 0.004505294840782881, 'reward_judge_quality_mean': 0.7224999666213989, 'reward_judge_quality_std': 0.232854962348938, 'reward_total_composite_mean': 0.6401236653327942, 'reward_total_composite_std': 0.20709064602851868, 'epoch': 0.0}

  1%|▏         | 43/3300 [05:11<6:35:57,  7.29s/it]INFO 04-13 07:26:09 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 44/3300 [05:18<6:35:56,  7.30s/it]2026-04-13 07:26:16,206 | INFO | train_grpo_train | metrics_logged mode=train step=44

                                                   
{'loss': -0.0165, 'grad_norm': 11.687650680541992, 'learning_rate': 9.869696969696971e-06, 'num_tokens': 81723.0, 'completions/mean_length': 99.875, 'completions/min_length': 74.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.875, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.3675351142883301, 'rewards/meter/std': 0.2946922183036804, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9973076581954956, 'rewards/repeat_soft/std': 0.002466563368216157, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.3810186982154846, 'rewards/total_composite/std': 0.1699932962656021, 'reward': 0.3810186982154846, 'reward_std': 0.16999328136444092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20723651349544525, 'sampling/sampling_logp_difference/max': 1.8403129577636719, 'sampling/importance_sampling_ratio/min': 0.1587677299976349, 'sampling/importance_sampling_ratio/mean': 1.039668083190918, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.169490233063698, 'clip_ratio/low_mean': 0.06495124101638794, 'clip_ratio/low_min': 0.06495124101638794, 'clip_ratio/high_mean': 0.11929397657513618, 'clip_ratio/high_max': 0.11929397657513618, 'clip_ratio/region_mean': 0.18424521759152412, 'reward_total_mean': 0.3810186982154846, 'reward_meter_mean': 0.3675351142883301, 'reward_meter_std': 0.2946922183036804, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9973076581954956, 'reward_repeat_soft_std': 0.002466563368216157, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.3810186982154846, 'reward_total_composite_std': 0.1699932962656021, 'epoch': 0.0}

  1%|▏         | 44/3300 [05:18<6:35:56,  7.30s/it]INFO 04-13 07:26:16 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 45/3300 [05:25<6:32:45,  7.24s/it]2026-04-13 07:26:23,313 | INFO | train_grpo_train | metrics_logged mode=train step=45

                                                   
{'loss': 0.1585, 'grad_norm': 14.383469581604004, 'learning_rate': 9.866666666666668e-06, 'num_tokens': 83331.0, 'completions/mean_length': 50.0, 'completions/min_length': 38.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7773149013519287, 'rewards/meter/std': 0.25354528427124023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9865143895149231, 'rewards/repeat_soft/std': 0.02192966639995575, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.5853496193885803, 'rewards/total_composite/std': 0.2865021228790283, 'reward': 0.5853496193885803, 'reward_std': 0.2865021228790283, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20703330636024475, 'sampling/sampling_logp_difference/max': 1.642387866973877, 'sampling/importance_sampling_ratio/min': 0.19351740181446075, 'sampling/importance_sampling_ratio/mean': 1.0344429016113281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8182155936956406, 'clip_ratio/low_mean': 0.08375179395079613, 'clip_ratio/low_min': 0.08375179395079613, 'clip_ratio/high_mean': 0.12188977561891079, 'clip_ratio/high_max': 0.12188977561891079, 'clip_ratio/region_mean': 0.20564156956970692, 'reward_total_mean': 0.5853496193885803, 'reward_meter_mean': 0.7773149013519287, 'reward_meter_std': 0.25354528427124023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9865143895149231, 'reward_repeat_soft_std': 0.02192966639995575, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5853496193885803, 'reward_total_composite_std': 0.2865021228790283, 'epoch': 0.0}

  1%|▏         | 45/3300 [05:25<6:32:45,  7.24s/it]INFO 04-13 07:26:23 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 46/3300 [05:33<6:40:43,  7.39s/it]2026-04-13 07:26:31,050 | INFO | train_grpo_train | metrics_logged mode=train step=46

                                                   
{'loss': -0.0718, 'grad_norm': 15.25479793548584, 'learning_rate': 9.863636363636364e-06, 'num_tokens': 84939.0, 'completions/mean_length': 45.0, 'completions/min_length': 36.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5657721757888794, 'rewards/meter/std': 0.4256070852279663, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984498023986816, 'rewards/repeat_soft/std': 0.004384705331176519, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.6238999366760254, 'rewards/total_composite/std': 0.23371422290802002, 'reward': 0.6238999366760254, 'reward_std': 0.23371422290802002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2331676185131073, 'sampling/sampling_logp_difference/max': 1.3806533813476562, 'sampling/importance_sampling_ratio/min': 0.2514142394065857, 'sampling/importance_sampling_ratio/mean': 1.0120073556900024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.181841418147087, 'clip_ratio/low_mean': 0.08811917342245579, 'clip_ratio/low_min': 0.08811917342245579, 'clip_ratio/high_mean': 0.13466385751962662, 'clip_ratio/high_max': 0.13466385751962662, 'clip_ratio/region_mean': 0.2227830309420824, 'reward_total_mean': 0.6238999366760254, 'reward_meter_mean': 0.5657721757888794, 'reward_meter_std': 0.4256070852279663, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984498023986816, 'reward_repeat_soft_std': 0.004384705331176519, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.6238999366760254, 'reward_total_composite_std': 0.23371422290802002, 'epoch': 0.0}

  1%|▏         | 46/3300 [05:33<6:40:43,  7.39s/it]INFO 04-13 07:26:31 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 47/3300 [05:39<6:29:47,  7.19s/it]2026-04-13 07:26:37,775 | INFO | train_grpo_train | metrics_logged mode=train step=47

                                                   
{'loss': 0.0598, 'grad_norm': 19.374162673950195, 'learning_rate': 9.860606060606061e-06, 'num_tokens': 86499.0, 'completions/mean_length': 46.0, 'completions/min_length': 35.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.26104703545570374, 'rewards/meter/std': 0.2333628535270691, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9847595691680908, 'rewards/repeat_soft/std': 0.011016963981091976, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.44352251291275024, 'rewards/total_composite/std': 0.11558562517166138, 'reward': 0.44352251291275024, 'reward_std': 0.11558562517166138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19819296896457672, 'sampling/sampling_logp_difference/max': 2.590005874633789, 'sampling/importance_sampling_ratio/min': 0.07501959800720215, 'sampling/importance_sampling_ratio/mean': 1.022105097770691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6669579893350601, 'clip_ratio/low_mean': 0.1387309804558754, 'clip_ratio/low_min': 0.1387309804558754, 'clip_ratio/high_mean': 0.04285714402794838, 'clip_ratio/high_max': 0.04285714402794838, 'clip_ratio/region_mean': 0.18158812448382378, 'reward_total_mean': 0.44352251291275024, 'reward_meter_mean': 0.26104703545570374, 'reward_meter_std': 0.2333628535270691, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9847595691680908, 'reward_repeat_soft_std': 0.011016963981091976, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.44352251291275024, 'reward_total_composite_std': 0.11558562517166138, 'epoch': 0.0}

  1%|▏         | 47/3300 [05:40<6:29:47,  7.19s/it]INFO 04-13 07:26:38 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 48/3300 [05:46<6:23:38,  7.08s/it]2026-04-13 07:26:44,594 | INFO | train_grpo_train | metrics_logged mode=train step=48

                                                   
{'loss': 0.1314, 'grad_norm': 28.504535675048828, 'learning_rate': 9.857575757575758e-06, 'num_tokens': 88108.0, 'completions/mean_length': 32.125, 'completions/min_length': 28.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.27756398916244507, 'rewards/meter/std': 0.3191877007484436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9716074466705322, 'rewards/repeat_soft/std': 0.036921996623277664, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2921227812767029, 'rewards/total_composite/mean': 0.4460914433002472, 'rewards/total_composite/std': 0.2519592344760895, 'reward': 0.4460914433002472, 'reward_std': 0.2519592046737671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27054738998413086, 'sampling/sampling_logp_difference/max': 4.1249494552612305, 'sampling/importance_sampling_ratio/min': 0.01616431027650833, 'sampling/importance_sampling_ratio/mean': 1.010880947113037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2761306390166283, 'clip_ratio/low_mean': 0.1557934768497944, 'clip_ratio/low_min': 0.1557934768497944, 'clip_ratio/high_mean': 0.07254464458674192, 'clip_ratio/high_max': 0.07254464458674192, 'clip_ratio/region_mean': 0.2283381214365363, 'reward_total_mean': 0.4460914433002472, 'reward_meter_mean': 0.27756398916244507, 'reward_meter_std': 0.3191877007484436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9716074466705322, 'reward_repeat_soft_std': 0.036921996623277664, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2921227812767029, 'reward_total_composite_mean': 0.4460914433002472, 'reward_total_composite_std': 0.2519592344760895, 'epoch': 0.0}

  1%|▏         | 48/3300 [05:46<6:23:38,  7.08s/it]INFO 04-13 07:26:44 [block_pool.py:378] Successfully reset prefix cache

  1%|▏         | 49/3300 [05:55<6:42:51,  7.43s/it]2026-04-13 07:26:52,860 | INFO | train_grpo_train | metrics_logged mode=train step=49

                                                   
{'loss': 0.0539, 'grad_norm': 19.42587661743164, 'learning_rate': 9.854545454545456e-06, 'num_tokens': 90563.0, 'completions/mean_length': 115.875, 'completions/min_length': 95.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.875, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.30381885170936584, 'rewards/meter/std': 0.18320336937904358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9918557405471802, 'rewards/repeat_soft/std': 0.0065759336575865746, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.3872862756252289, 'rewards/total_composite/std': 0.16922295093536377, 'reward': 0.3872862756252289, 'reward_std': 0.16922295093536377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2825930118560791, 'sampling/sampling_logp_difference/max': 3.059009552001953, 'sampling/importance_sampling_ratio/min': 0.046934157609939575, 'sampling/importance_sampling_ratio/mean': 0.9706795811653137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2838484272360802, 'clip_ratio/low_mean': 0.08543148636817932, 'clip_ratio/low_min': 0.08543148636817932, 'clip_ratio/high_mean': 0.15653525479137897, 'clip_ratio/high_max': 0.15653525479137897, 'clip_ratio/region_mean': 0.2419667411595583, 'reward_total_mean': 0.3872862756252289, 'reward_meter_mean': 0.30381885170936584, 'reward_meter_std': 0.18320336937904358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9918557405471802, 'reward_repeat_soft_std': 0.0065759336575865746, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.3872862756252289, 'reward_total_composite_std': 0.16922295093536377, 'epoch': 0.0}

  1%|▏         | 49/3300 [05:55<6:42:51,  7.43s/it]INFO 04-13 07:26:53 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 50/3300 [06:03<6:56:25,  7.69s/it]2026-04-13 07:27:01,138 | INFO | train_grpo_train | metrics_logged mode=train step=50

                                                   
{'loss': -0.0257, 'grad_norm': 13.771021842956543, 'learning_rate': 9.851515151515151e-06, 'num_tokens': 92565.0, 'completions/mean_length': 88.25, 'completions/min_length': 61.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.329473614692688, 'rewards/meter/std': 0.25273919105529785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9935731887817383, 'rewards/repeat_soft/std': 0.0037070184480398893, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3710108995437622, 'rewards/total_composite/std': 0.15944214165210724, 'reward': 0.3710108995437622, 'reward_std': 0.15944214165210724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2445198893547058, 'sampling/sampling_logp_difference/max': 2.486546516418457, 'sampling/importance_sampling_ratio/min': 0.08319678902626038, 'sampling/importance_sampling_ratio/mean': 1.0304960012435913, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.181565210223198, 'clip_ratio/low_mean': 0.08294316753745079, 'clip_ratio/low_min': 0.08294316753745079, 'clip_ratio/high_mean': 0.164827523753047, 'clip_ratio/high_max': 0.164827523753047, 'clip_ratio/region_mean': 0.24777069129049778, 'reward_total_mean': 0.3710108995437622, 'reward_meter_mean': 0.329473614692688, 'reward_meter_std': 0.25273919105529785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9935731887817383, 'reward_repeat_soft_std': 0.0037070184480398893, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3710108995437622, 'reward_total_composite_std': 0.15944214165210724, 'epoch': 0.01}

  2%|▏         | 50/3300 [06:03<6:56:25,  7.69s/it]INFO 04-13 07:27:01 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:17,  2.20s/it]

 30%|███       | 3/10 [00:13<00:36,  5.23s/it]

 40%|████      | 4/10 [00:18<00:29,  4.93s/it]

 50%|█████     | 5/10 [00:24<00:26,  5.34s/it]

 60%|██████    | 6/10 [00:30<00:22,  5.66s/it]

 70%|███████   | 7/10 [00:36<00:17,  5.71s/it]

 80%|████████  | 8/10 [00:46<00:13,  6.97s/it]

 90%|█████████ | 9/10 [00:54<00:07,  7.31s/it]

100%|██████████| 10/10 [00:59<00:00,  6.52s/it]2026-04-13 07:28:10,359 | INFO | train_grpo_train | metrics_logged mode=eval step=50

                                                   

                                               
{'eval_loss': nan, 'eval_runtime': 69.2141, 'eval_samples_per_second': 1.156, 'eval_steps_per_second': 0.144, 'eval_num_tokens': 92565.0, 'eval_completions/mean_length': 106.85, 'eval_completions/min_length': 31.3, 'eval_completions/max_length': 286.3, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 96.1928581237793, 'eval_completions/min_terminated_length': 31.3, 'eval_completions/max_terminated_length': 221.4, 'eval_rewards/meter/mean': 0.4507799297571182, 'eval_rewards/meter/std': 0.3792104184627533, 'eval_rewards/count_adherence/mean': 0.9110416650772095, 'eval_rewards/count_adherence/std': 0.20400682613253593, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.11700168251991272, 'eval_rewards/repeat_soft/mean': 0.9902224004268646, 'eval_rewards/repeat_soft/std': 0.017161866067908704, 'eval_rewards/judge_quality/mean': 0.5031250029802322, 'eval_rewards/judge_quality/std': 0.18279580399394035, 'eval_rewards/total_composite/mean': 0.4718601256608963, 'eval_rewards/total_composite/std': 0.17194418162107467, 'eval_reward': 0.4718601256608963, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.13539711087942125, 'eval_sampling/sampling_logp_difference/max': 1.1591978073120117, 'eval_sampling/importance_sampling_ratio/min': 0.31711107939481736, 'eval_sampling/importance_sampling_ratio/mean': 1.0372873306274415, 'eval_sampling/importance_sampling_ratio/max': 1.536458969116211, 'eval_entropy': 2.0635273694992065, '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': 0.4718601256608963, 'eval_reward_meter_mean': 0.4507799297571182, 'eval_reward_meter_std': 0.3792104184627533, 'eval_reward_count_adherence_mean': 0.9110416650772095, 'eval_reward_count_adherence_std': 0.20400682613253593, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.11700168251991272, 'eval_reward_repeat_soft_mean': 0.9902224004268646, 'eval_reward_repeat_soft_std': 0.017161866067908704, 'eval_reward_judge_quality_mean': 0.5031250029802322, 'eval_reward_judge_quality_std': 0.18279580399394035, 'eval_reward_total_composite_mean': 0.4718601256608963, 'eval_reward_total_composite_std': 0.17194418162107467, 'epoch': 0.01}

  2%|▏         | 50/3300 [07:12<6:56:25,  7.69s/it]

100%|██████████| 10/10 [00:59<00:00,  6.52s/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(

  2%|▏         | 51/3300 [07:23<26:25:34, 29.28s/it]2026-04-13 07:28:20,805 | INFO | train_grpo_train | metrics_logged mode=train step=51

                                                    
{'loss': 0.1885, 'grad_norm': 15.710267066955566, 'learning_rate': 9.84848484848485e-06, 'num_tokens': 94303.0, 'completions/mean_length': 50.25, 'completions/min_length': 33.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.48370689153671265, 'rewards/meter/std': 0.4339752495288849, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9988887906074524, 'rewards/repeat_soft/std': 0.002272413345053792, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.5073872208595276, 'rewards/total_composite/std': 0.21636413037776947, 'reward': 0.5073872208595276, 'reward_std': 0.21636416018009186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2187386006116867, 'sampling/sampling_logp_difference/max': 1.439584732055664, 'sampling/importance_sampling_ratio/min': 0.2370261698961258, 'sampling/importance_sampling_ratio/mean': 1.015249252319336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4316212087869644, 'clip_ratio/low_mean': 0.11987635493278503, 'clip_ratio/low_min': 0.11987635493278503, 'clip_ratio/high_mean': 0.11116188950836658, 'clip_ratio/high_max': 0.11116188950836658, 'clip_ratio/region_mean': 0.23103824444115162, 'reward_total_mean': 0.5073872208595276, 'reward_meter_mean': 0.48370689153671265, 'reward_meter_std': 0.4339752495288849, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9988887906074524, 'reward_repeat_soft_std': 0.002272413345053792, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.5073872208595276, 'reward_total_composite_std': 0.21636413037776947, 'epoch': 0.01}

  2%|▏         | 51/3300 [07:23<26:25:34, 29.28s/it]INFO 04-13 07:28:21 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 52/3300 [07:36<22:00:37, 24.40s/it]2026-04-13 07:28:33,801 | INFO | train_grpo_train | metrics_logged mode=train step=52

                                                    
{'loss': -0.0651, 'grad_norm': 7.459178924560547, 'learning_rate': 9.845454545454546e-06, 'num_tokens': 96331.0, 'completions/mean_length': 143.5, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 90.85714721679688, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.5149980783462524, 'rewards/meter/std': 0.3162343502044678, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.35197150707244873, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9875726699829102, 'rewards/repeat_soft/std': 0.01723732426762581, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.23793382942676544, 'rewards/total_composite/mean': 0.43574202060699463, 'rewards/total_composite/std': 0.3080213665962219, 'reward': 0.43574202060699463, 'reward_std': 0.3080213963985443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2421015501022339, 'sampling/sampling_logp_difference/max': 1.4144783020019531, 'sampling/importance_sampling_ratio/min': 0.2430523782968521, 'sampling/importance_sampling_ratio/mean': 1.0390788316726685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0853719115257263, 'clip_ratio/low_mean': 0.041592495515942574, 'clip_ratio/low_min': 0.041592495515942574, 'clip_ratio/high_mean': 0.16704162023961544, 'clip_ratio/high_max': 0.16704162023961544, 'clip_ratio/region_mean': 0.20863411575555801, 'reward_total_mean': 0.43574202060699463, 'reward_meter_mean': 0.5149980783462524, 'reward_meter_std': 0.3162343502044678, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.35197150707244873, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9875726699829102, 'reward_repeat_soft_std': 0.01723732426762581, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.23793382942676544, 'reward_total_composite_mean': 0.43574202060699463, 'reward_total_composite_std': 0.3080213665962219, 'epoch': 0.01}

  2%|▏         | 52/3300 [07:36<22:00:37, 24.40s/it]INFO 04-13 07:28:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 53/3300 [07:43<17:18:12, 19.18s/it]2026-04-13 07:28:40,826 | INFO | train_grpo_train | metrics_logged mode=train step=53

                                                    
{'loss': 0.0639, 'grad_norm': 29.835546493530273, 'learning_rate': 9.842424242424243e-06, 'num_tokens': 97818.0, 'completions/mean_length': 32.875, 'completions/min_length': 25.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5110926032066345, 'rewards/meter/std': 0.36550405621528625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9988951683044434, 'rewards/repeat_soft/std': 0.0031248305458575487, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5656921863555908, 'rewards/total_composite/std': 0.3202142119407654, 'reward': 0.5656921863555908, 'reward_std': 0.3202142119407654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20222999155521393, 'sampling/sampling_logp_difference/max': 2.314786434173584, 'sampling/importance_sampling_ratio/min': 0.09878727793693542, 'sampling/importance_sampling_ratio/mean': 0.9813011884689331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6355803459882736, 'clip_ratio/low_mean': 0.10113392397761345, 'clip_ratio/low_min': 0.10113392397761345, 'clip_ratio/high_mean': 0.051487069576978683, 'clip_ratio/high_max': 0.051487069576978683, 'clip_ratio/region_mean': 0.15262099355459213, 'reward_total_mean': 0.5656921863555908, 'reward_meter_mean': 0.5110926032066345, 'reward_meter_std': 0.36550405621528625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9988951683044434, 'reward_repeat_soft_std': 0.0031248305458575487, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5656921863555908, 'reward_total_composite_std': 0.3202142119407654, 'epoch': 0.01}

  2%|▏         | 53/3300 [07:43<17:18:12, 19.18s/it]INFO 04-13 07:28:41 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 54/3300 [07:51<14:26:49, 16.02s/it]2026-04-13 07:28:49,471 | INFO | train_grpo_train | metrics_logged mode=train step=54

                                                    
{'loss': 0.0885, 'grad_norm': 10.542903900146484, 'learning_rate': 9.83939393939394e-06, 'num_tokens': 100231.0, 'completions/mean_length': 124.625, 'completions/min_length': 86.0, 'completions/max_length': 182.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.625, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.2036353051662445, 'rewards/meter/std': 0.22631075978279114, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9939138293266296, 'rewards/repeat_soft/std': 0.007315187249332666, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.38589751720428467, 'rewards/total_composite/std': 0.2055649310350418, 'reward': 0.38589751720428467, 'reward_std': 0.2055649310350418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.228908970952034, 'sampling/sampling_logp_difference/max': 2.3076095581054688, 'sampling/importance_sampling_ratio/min': 0.09949881583452225, 'sampling/importance_sampling_ratio/mean': 1.044135332107544, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1012340635061264, 'clip_ratio/low_mean': 0.13333098776638508, 'clip_ratio/low_min': 0.13333098776638508, 'clip_ratio/high_mean': 0.08108514919877052, 'clip_ratio/high_max': 0.08108514919877052, 'clip_ratio/region_mean': 0.2144161369651556, 'reward_total_mean': 0.38589751720428467, 'reward_meter_mean': 0.2036353051662445, 'reward_meter_std': 0.22631075978279114, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9939138293266296, 'reward_repeat_soft_std': 0.007315187249332666, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.38589751720428467, 'reward_total_composite_std': 0.2055649310350418, 'epoch': 0.01}

  2%|▏         | 54/3300 [07:51<14:26:49, 16.02s/it]INFO 04-13 07:28:49 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 55/3300 [07:58<11:49:05, 13.11s/it]2026-04-13 07:28:55,789 | INFO | train_grpo_train | metrics_logged mode=train step=55

                                                    
{'loss': 0.0139, 'grad_norm': 20.996606826782227, 'learning_rate': 9.836363636363637e-06, 'num_tokens': 101772.0, 'completions/mean_length': 25.625, 'completions/min_length': 18.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.44943761825561523, 'rewards/meter/std': 0.4661872088909149, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9460721015930176, 'rewards/repeat_soft/std': 0.03294894099235535, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.22258226573467255, 'rewards/total_composite/mean': 0.5090354681015015, 'rewards/total_composite/std': 0.20502084493637085, 'reward': 0.5090354681015015, 'reward_std': 0.20502084493637085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19217462837696075, 'sampling/sampling_logp_difference/max': 1.8104991912841797, 'sampling/importance_sampling_ratio/min': 0.16357247531414032, 'sampling/importance_sampling_ratio/mean': 1.0293488502502441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4901997074484825, 'clip_ratio/low_mean': 0.06276315823197365, 'clip_ratio/low_min': 0.06276315823197365, 'clip_ratio/high_mean': 0.08341875951737165, 'clip_ratio/high_max': 0.08341875951737165, 'clip_ratio/region_mean': 0.1461819177493453, 'reward_total_mean': 0.5090354681015015, 'reward_meter_mean': 0.44943761825561523, 'reward_meter_std': 0.4661872088909149, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9460721015930176, 'reward_repeat_soft_std': 0.03294894099235535, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.22258226573467255, 'reward_total_composite_mean': 0.5090354681015015, 'reward_total_composite_std': 0.20502084493637085, 'epoch': 0.01}

  2%|▏         | 55/3300 [07:58<11:49:05, 13.11s/it]INFO 04-13 07:28:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 56/3300 [08:04<9:58:16, 11.07s/it] 2026-04-13 07:29:02,081 | INFO | train_grpo_train | metrics_logged mode=train step=56

                                                   
{'loss': -0.0396, 'grad_norm': 17.0747127532959, 'learning_rate': 9.833333333333333e-06, 'num_tokens': 103282.0, 'completions/mean_length': 30.75, 'completions/min_length': 24.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7303705215454102, 'rewards/meter/std': 0.4463520348072052, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5913689732551575, 'rewards/total_composite/std': 0.18762490153312683, 'reward': 0.5913689732551575, 'reward_std': 0.18762490153312683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18441377580165863, 'sampling/sampling_logp_difference/max': 1.2306923866271973, 'sampling/importance_sampling_ratio/min': 0.2920902669429779, 'sampling/importance_sampling_ratio/mean': 1.0209691524505615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4771666452288628, 'clip_ratio/low_mean': 0.051488833501935005, 'clip_ratio/low_min': 0.051488833501935005, 'clip_ratio/high_mean': 0.12333367951214314, 'clip_ratio/high_max': 0.12333367951214314, 'clip_ratio/region_mean': 0.17482251301407814, 'reward_total_mean': 0.5913689732551575, 'reward_meter_mean': 0.7303705215454102, 'reward_meter_std': 0.4463520348072052, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5913689732551575, 'reward_total_composite_std': 0.18762490153312683, 'epoch': 0.01}

  2%|▏         | 56/3300 [08:04<9:58:16, 11.07s/it]INFO 04-13 07:29:02 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 57/3300 [08:12<9:04:53, 10.08s/it]2026-04-13 07:29:09,866 | INFO | train_grpo_train | metrics_logged mode=train step=57

                                                   
{'loss': 0.0667, 'grad_norm': 9.708298683166504, 'learning_rate': 9.830303030303032e-06, 'num_tokens': 105531.0, 'completions/mean_length': 111.125, 'completions/min_length': 78.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.125, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.425605833530426, 'rewards/meter/std': 0.37622109055519104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9927740693092346, 'rewards/repeat_soft/std': 0.0072191013023257256, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.18539534509181976, 'rewards/total_composite/mean': 0.4965977072715759, 'rewards/total_composite/std': 0.1536581963300705, 'reward': 0.4965977072715759, 'reward_std': 0.1536581963300705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19263683259487152, 'sampling/sampling_logp_difference/max': 1.5666651725769043, 'sampling/importance_sampling_ratio/min': 0.20874013006687164, 'sampling/importance_sampling_ratio/mean': 1.0319360494613647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.05468812584877, 'clip_ratio/low_mean': 0.11199099197983742, 'clip_ratio/low_min': 0.11199099197983742, 'clip_ratio/high_mean': 0.058838874101638794, 'clip_ratio/high_max': 0.058838874101638794, 'clip_ratio/region_mean': 0.1708298660814762, 'reward_total_mean': 0.4965977072715759, 'reward_meter_mean': 0.425605833530426, 'reward_meter_std': 0.37622109055519104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9927740693092346, 'reward_repeat_soft_std': 0.0072191013023257256, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.18539534509181976, 'reward_total_composite_mean': 0.4965977072715759, 'reward_total_composite_std': 0.1536581963300705, 'epoch': 0.01}

  2%|▏         | 57/3300 [08:12<9:04:53, 10.08s/it]INFO 04-13 07:29:10 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 58/3300 [08:19<8:27:02,  9.38s/it]2026-04-13 07:29:17,622 | INFO | train_grpo_train | metrics_logged mode=train step=58

                                                   
{'loss': 0.1245, 'grad_norm': 10.96164608001709, 'learning_rate': 9.827272727272729e-06, 'num_tokens': 107742.0, 'completions/mean_length': 93.375, 'completions/min_length': 58.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.375, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.4870108962059021, 'rewards/meter/std': 0.451114296913147, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.99791419506073, 'rewards/repeat_soft/std': 0.0024445788003504276, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.22385822236537933, 'rewards/total_composite/mean': 0.44787633419036865, 'rewards/total_composite/std': 0.2583901584148407, 'reward': 0.44787633419036865, 'reward_std': 0.2583901882171631, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22290265560150146, 'sampling/sampling_logp_difference/max': 1.734273910522461, 'sampling/importance_sampling_ratio/min': 0.17652831971645355, 'sampling/importance_sampling_ratio/mean': 1.019986867904663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.226851150393486, 'clip_ratio/low_mean': 0.12805911153554916, 'clip_ratio/low_min': 0.12805911153554916, 'clip_ratio/high_mean': 0.08842651732265949, 'clip_ratio/high_max': 0.08842651732265949, 'clip_ratio/region_mean': 0.21648562885820866, 'reward_total_mean': 0.44787633419036865, 'reward_meter_mean': 0.4870108962059021, 'reward_meter_std': 0.451114296913147, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.99791419506073, 'reward_repeat_soft_std': 0.0024445788003504276, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.22385822236537933, 'reward_total_composite_mean': 0.44787633419036865, 'reward_total_composite_std': 0.2583901584148407, 'epoch': 0.01}

  2%|▏         | 58/3300 [08:19<8:27:02,  9.38s/it]INFO 04-13 07:29:17 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 59/3300 [08:29<8:25:26,  9.36s/it]2026-04-13 07:29:26,917 | INFO | train_grpo_train | metrics_logged mode=train step=59

                                                   
{'loss': 0.467, 'grad_norm': 13.283112525939941, 'learning_rate': 9.824242424242425e-06, 'num_tokens': 109807.0, 'completions/mean_length': 79.125, 'completions/min_length': 34.0, 'completions/max_length': 258.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 258.0, 'rewards/meter/mean': 0.4121353030204773, 'rewards/meter/std': 0.47703149914741516, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9690682888031006, 'rewards/repeat_soft/std': 0.02924323081970215, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.5556825399398804, 'rewards/total_composite/std': 0.3225981891155243, 'reward': 0.5556825399398804, 'reward_std': 0.3225981891155243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1868036389350891, 'sampling/sampling_logp_difference/max': 1.6826361417770386, 'sampling/importance_sampling_ratio/min': 0.1858833134174347, 'sampling/importance_sampling_ratio/mean': 1.0397764444351196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.721499502658844, 'clip_ratio/low_mean': 0.12037735246121883, 'clip_ratio/low_min': 0.12037735246121883, 'clip_ratio/high_mean': 0.05528116412460804, 'clip_ratio/high_max': 0.05528116412460804, 'clip_ratio/region_mean': 0.17565851658582687, 'reward_total_mean': 0.5556825399398804, 'reward_meter_mean': 0.4121353030204773, 'reward_meter_std': 0.47703149914741516, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9690682888031006, 'reward_repeat_soft_std': 0.02924323081970215, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.5556825399398804, 'reward_total_composite_std': 0.3225981891155243, 'epoch': 0.01}

  2%|▏         | 59/3300 [08:29<8:25:26,  9.36s/it]INFO 04-13 07:29:27 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 60/3300 [08:35<7:42:12,  8.56s/it]2026-04-13 07:29:33,616 | INFO | train_grpo_train | metrics_logged mode=train step=60

                                                   
{'loss': 0.1308, 'grad_norm': 16.130277633666992, 'learning_rate': 9.821212121212122e-06, 'num_tokens': 111362.0, 'completions/mean_length': 49.375, 'completions/min_length': 31.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6151303052902222, 'rewards/meter/std': 0.3767223656177521, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9902870655059814, 'rewards/repeat_soft/std': 0.013324901461601257, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.22915996611118317, 'rewards/total_composite/mean': 0.6230052709579468, 'rewards/total_composite/std': 0.20295588672161102, 'reward': 0.6230052709579468, 'reward_std': 0.20295588672161102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19441068172454834, 'sampling/sampling_logp_difference/max': 1.3564529418945312, 'sampling/importance_sampling_ratio/min': 0.2575727701187134, 'sampling/importance_sampling_ratio/mean': 1.0002994537353516, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6784673035144806, 'clip_ratio/low_mean': 0.12781431712210178, 'clip_ratio/low_min': 0.12781431712210178, 'clip_ratio/high_mean': 0.07399587146937847, 'clip_ratio/high_max': 0.07399587146937847, 'clip_ratio/region_mean': 0.20181018859148026, 'reward_total_mean': 0.6230052709579468, 'reward_meter_mean': 0.6151303052902222, 'reward_meter_std': 0.3767223656177521, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9902870655059814, 'reward_repeat_soft_std': 0.013324901461601257, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.22915996611118317, 'reward_total_composite_mean': 0.6230052709579468, 'reward_total_composite_std': 0.20295588672161102, 'epoch': 0.01}

  2%|▏         | 60/3300 [08:35<7:42:12,  8.56s/it]INFO 04-13 07:29:33 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 61/3300 [08:42<7:17:32,  8.11s/it]2026-04-13 07:29:40,661 | INFO | train_grpo_train | metrics_logged mode=train step=61

                                                   
{'loss': 0.1142, 'grad_norm': 13.33469295501709, 'learning_rate': 9.81818181818182e-06, 'num_tokens': 113192.0, 'completions/mean_length': 54.75, 'completions/min_length': 38.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.3869086503982544, 'rewards/meter/std': 0.45730486512184143, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9924991130828857, 'rewards/repeat_soft/std': 0.01105050090700388, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.13265827298164368, 'rewards/total_composite/mean': 0.4687749743461609, 'rewards/total_composite/std': 0.16751234233379364, 'reward': 0.4687749743461609, 'reward_std': 0.16751235723495483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23125039041042328, 'sampling/sampling_logp_difference/max': 1.4231446981430054, 'sampling/importance_sampling_ratio/min': 0.24095511436462402, 'sampling/importance_sampling_ratio/mean': 1.0310434103012085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9869754761457443, 'clip_ratio/low_mean': 0.10279995948076248, 'clip_ratio/low_min': 0.10279995948076248, 'clip_ratio/high_mean': 0.08435150422155857, 'clip_ratio/high_max': 0.08435150422155857, 'clip_ratio/region_mean': 0.18715146370232105, 'reward_total_mean': 0.4687749743461609, 'reward_meter_mean': 0.3869086503982544, 'reward_meter_std': 0.45730486512184143, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9924991130828857, 'reward_repeat_soft_std': 0.01105050090700388, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.13265827298164368, 'reward_total_composite_mean': 0.4687749743461609, 'reward_total_composite_std': 0.16751234233379364, 'epoch': 0.01}

  2%|▏         | 61/3300 [08:42<7:17:32,  8.11s/it]INFO 04-13 07:29:40 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 62/3300 [08:54<8:19:16,  9.25s/it]2026-04-13 07:29:52,587 | INFO | train_grpo_train | metrics_logged mode=train step=62

                                                   
{'loss': -0.0608, 'grad_norm': 6.91639518737793, 'learning_rate': 9.815151515151516e-06, 'num_tokens': 116000.0, 'completions/mean_length': 211.0, 'completions/min_length': 140.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 168.0, 'completions/min_terminated_length': 140.0, 'completions/max_terminated_length': 237.0, 'rewards/meter/mean': 0.20621223747730255, 'rewards/meter/std': 0.1470290869474411, 'rewards/count_adherence/mean': 0.824999988079071, 'rewards/count_adherence/std': 0.345377653837204, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9960790872573853, 'rewards/repeat_soft/std': 0.004181248601526022, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.3168066143989563, 'rewards/total_composite/std': 0.15409335494041443, 'reward': 0.3168066143989563, 'reward_std': 0.15409335494041443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22920776903629303, 'sampling/sampling_logp_difference/max': 3.835620164871216, 'sampling/importance_sampling_ratio/min': 0.021587945520877838, 'sampling/importance_sampling_ratio/mean': 1.0088270902633667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5278817415237427, 'clip_ratio/low_mean': 0.028145695105195045, 'clip_ratio/low_min': 0.028145695105195045, 'clip_ratio/high_mean': 0.1687558777630329, 'clip_ratio/high_max': 0.1687558777630329, 'clip_ratio/region_mean': 0.19690157286822796, 'reward_total_mean': 0.3168066143989563, 'reward_meter_mean': 0.20621223747730255, 'reward_meter_std': 0.1470290869474411, 'reward_count_adherence_mean': 0.824999988079071, 'reward_count_adherence_std': 0.345377653837204, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9960790872573853, 'reward_repeat_soft_std': 0.004181248601526022, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.3168066143989563, 'reward_total_composite_std': 0.15409335494041443, 'epoch': 0.01}

  2%|▏         | 62/3300 [08:54<8:19:16,  9.25s/it]INFO 04-13 07:29:52 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 63/3300 [09:02<7:53:10,  8.77s/it]2026-04-13 07:30:00,236 | INFO | train_grpo_train | metrics_logged mode=train step=63

                                                   
{'loss': 0.1668, 'grad_norm': 29.546297073364258, 'learning_rate': 9.812121212121212e-06, 'num_tokens': 117890.0, 'completions/mean_length': 75.25, 'completions/min_length': 55.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.25, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.4134503901004791, 'rewards/meter/std': 0.332876592874527, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9948821067810059, 'rewards/repeat_soft/std': 0.008461784571409225, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.4275481104850769, 'rewards/total_composite/std': 0.1968495398759842, 'reward': 0.4275481104850769, 'reward_std': 0.1968495398759842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22210216522216797, 'sampling/sampling_logp_difference/max': 3.0588936805725098, 'sampling/importance_sampling_ratio/min': 0.04693959653377533, 'sampling/importance_sampling_ratio/mean': 1.0094012022018433, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1671466045081615, 'clip_ratio/low_mean': 0.07250726781785488, 'clip_ratio/low_min': 0.07250726781785488, 'clip_ratio/high_mean': 0.11376019939780235, 'clip_ratio/high_max': 0.11376019939780235, 'clip_ratio/region_mean': 0.18626746721565723, 'reward_total_mean': 0.4275481104850769, 'reward_meter_mean': 0.4134503901004791, 'reward_meter_std': 0.332876592874527, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9948821067810059, 'reward_repeat_soft_std': 0.008461784571409225, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.4275481104850769, 'reward_total_composite_std': 0.1968495398759842, 'epoch': 0.01}

  2%|▏         | 63/3300 [09:02<7:53:10,  8.77s/it]INFO 04-13 07:30:00 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 64/3300 [09:10<7:42:52,  8.58s/it]2026-04-13 07:30:08,379 | INFO | train_grpo_train | metrics_logged mode=train step=64

                                                   
{'loss': 0.051, 'grad_norm': 12.051981925964355, 'learning_rate': 9.809090909090911e-06, 'num_tokens': 120161.0, 'completions/mean_length': 101.875, 'completions/min_length': 83.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.875, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.4644274115562439, 'rewards/meter/std': 0.44741523265838623, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9956849813461304, 'rewards/repeat_soft/std': 0.005578116048127413, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5090116262435913, 'rewards/total_composite/std': 0.1652180701494217, 'reward': 0.5090116262435913, 'reward_std': 0.1652180552482605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22064253687858582, 'sampling/sampling_logp_difference/max': 1.6844162940979004, 'sampling/importance_sampling_ratio/min': 0.1855527013540268, 'sampling/importance_sampling_ratio/mean': 1.0209674835205078, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0903439819812775, 'clip_ratio/low_mean': 0.08959254249930382, 'clip_ratio/low_min': 0.08959254249930382, 'clip_ratio/high_mean': 0.11559330485761166, 'clip_ratio/high_max': 0.11559330485761166, 'clip_ratio/region_mean': 0.20518584735691547, 'reward_total_mean': 0.5090116262435913, 'reward_meter_mean': 0.4644274115562439, 'reward_meter_std': 0.44741523265838623, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9956849813461304, 'reward_repeat_soft_std': 0.005578116048127413, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5090116262435913, 'reward_total_composite_std': 0.1652180701494217, 'epoch': 0.01}

  2%|▏         | 64/3300 [09:10<7:42:52,  8.58s/it]INFO 04-13 07:30:08 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 65/3300 [09:17<7:11:07,  8.00s/it]2026-04-13 07:30:15,008 | INFO | train_grpo_train | metrics_logged mode=train step=65

                                                   
{'loss': 0.1801, 'grad_norm': 24.44002342224121, 'learning_rate': 9.806060606060607e-06, 'num_tokens': 121722.0, 'completions/mean_length': 45.125, 'completions/min_length': 31.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5212327241897583, 'rewards/meter/std': 0.48654648661613464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9990687370300293, 'rewards/repeat_soft/std': 0.0013255642261356115, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.41597363352775574, 'rewards/total_composite/std': 0.2080499529838562, 'reward': 0.41597363352775574, 'reward_std': 0.2080499529838562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.28085407614707947, 'sampling/sampling_logp_difference/max': 2.827544689178467, 'sampling/importance_sampling_ratio/min': 0.059157922863960266, 'sampling/importance_sampling_ratio/mean': 1.0095726251602173, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9724387228488922, 'clip_ratio/low_mean': 0.14638710487633944, 'clip_ratio/low_min': 0.14638710487633944, 'clip_ratio/high_mean': 0.08752177841961384, 'clip_ratio/high_max': 0.08752177841961384, 'clip_ratio/region_mean': 0.23390888329595327, 'reward_total_mean': 0.41597363352775574, 'reward_meter_mean': 0.5212327241897583, 'reward_meter_std': 0.48654648661613464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9990687370300293, 'reward_repeat_soft_std': 0.0013255642261356115, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.41597363352775574, 'reward_total_composite_std': 0.2080499529838562, 'epoch': 0.01}

  2%|▏         | 65/3300 [09:17<7:11:07,  8.00s/it]INFO 04-13 07:30:15 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 66/3300 [09:28<8:09:33,  9.08s/it]2026-04-13 07:30:26,625 | INFO | train_grpo_train | metrics_logged mode=train step=66

                                                   
{'loss': 0.0828, 'grad_norm': 6.262032985687256, 'learning_rate': 9.803030303030304e-06, 'num_tokens': 123325.0, 'completions/mean_length': 120.375, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.42857360839844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.5316421985626221, 'rewards/meter/std': 0.3480125069618225, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.45806270837783813, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9995944499969482, 'rewards/repeat_soft/std': 0.0009662444936111569, 'rewards/judge_quality/mean': 0.4437500238418579, 'rewards/judge_quality/std': 0.19935163855552673, 'rewards/total_composite/mean': 0.4721200466156006, 'rewards/total_composite/std': 0.24633915722370148, 'reward': 0.4721200466156006, 'reward_std': 0.24633914232254028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22367364168167114, 'sampling/sampling_logp_difference/max': 1.7374200820922852, 'sampling/importance_sampling_ratio/min': 0.1759738177061081, 'sampling/importance_sampling_ratio/mean': 1.0281990766525269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8580587804317474, 'clip_ratio/low_mean': 0.08735784329473972, 'clip_ratio/low_min': 0.08735784329473972, 'clip_ratio/high_mean': 0.09469257295131683, 'clip_ratio/high_max': 0.09469257295131683, 'clip_ratio/region_mean': 0.18205041624605656, 'reward_total_mean': 0.4721200466156006, 'reward_meter_mean': 0.5316421985626221, 'reward_meter_std': 0.3480125069618225, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.45806270837783813, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9995944499969482, 'reward_repeat_soft_std': 0.0009662444936111569, 'reward_judge_quality_mean': 0.4437500238418579, 'reward_judge_quality_std': 0.19935163855552673, 'reward_total_composite_mean': 0.4721200466156006, 'reward_total_composite_std': 0.24633915722370148, 'epoch': 0.01}

  2%|▏         | 66/3300 [09:28<8:09:33,  9.08s/it]INFO 04-13 07:30:26 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 67/3300 [09:36<7:42:12,  8.58s/it]2026-04-13 07:30:34,025 | INFO | train_grpo_train | metrics_logged mode=train step=67

                                                   
{'loss': -0.0164, 'grad_norm': 14.978940963745117, 'learning_rate': 9.800000000000001e-06, 'num_tokens': 125078.0, 'completions/mean_length': 60.125, 'completions/min_length': 40.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.5190232396125793, 'rewards/meter/std': 0.41173845529556274, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9963045716285706, 'rewards/repeat_soft/std': 0.006186562590301037, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.19167961180210114, 'rewards/total_composite/mean': 0.41957348585128784, 'rewards/total_composite/std': 0.252689391374588, 'reward': 0.41957348585128784, 'reward_std': 0.252689391374588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21659037470817566, 'sampling/sampling_logp_difference/max': 1.5980007648468018, 'sampling/importance_sampling_ratio/min': 0.202300563454628, 'sampling/importance_sampling_ratio/mean': 1.0314173698425293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.830552101135254, 'clip_ratio/low_mean': 0.1511866319924593, 'clip_ratio/low_min': 0.1511866319924593, 'clip_ratio/high_mean': 0.04352152161300182, 'clip_ratio/high_max': 0.04352152161300182, 'clip_ratio/region_mean': 0.19470815360546112, 'reward_total_mean': 0.41957348585128784, 'reward_meter_mean': 0.5190232396125793, 'reward_meter_std': 0.41173845529556274, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9963045716285706, 'reward_repeat_soft_std': 0.006186562590301037, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.19167961180210114, 'reward_total_composite_mean': 0.41957348585128784, 'reward_total_composite_std': 0.252689391374588, 'epoch': 0.01}

  2%|▏         | 67/3300 [09:36<7:42:12,  8.58s/it]INFO 04-13 07:30:34 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 68/3300 [09:43<7:18:16,  8.14s/it]2026-04-13 07:30:41,131 | INFO | train_grpo_train | metrics_logged mode=train step=68

                                                   
{'loss': 0.1232, 'grad_norm': 23.07606315612793, 'learning_rate': 9.796969696969698e-06, 'num_tokens': 126637.0, 'completions/mean_length': 44.875, 'completions/min_length': 39.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5100425481796265, 'rewards/meter/std': 0.3832434415817261, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.975085973739624, 'rewards/repeat_soft/std': 0.027266472578048706, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.49085086584091187, 'rewards/total_composite/std': 0.10883072763681412, 'reward': 0.49085086584091187, 'reward_std': 0.10883072018623352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21352910995483398, 'sampling/sampling_logp_difference/max': 2.5040464401245117, 'sampling/importance_sampling_ratio/min': 0.08175352215766907, 'sampling/importance_sampling_ratio/mean': 1.0196813344955444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.520227499306202, 'clip_ratio/low_mean': 0.09456020221114159, 'clip_ratio/low_min': 0.09456020221114159, 'clip_ratio/high_mean': 0.09431146644055843, 'clip_ratio/high_max': 0.09431146644055843, 'clip_ratio/region_mean': 0.18887166865170002, 'reward_total_mean': 0.49085086584091187, 'reward_meter_mean': 0.5100425481796265, 'reward_meter_std': 0.3832434415817261, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.975085973739624, 'reward_repeat_soft_std': 0.027266472578048706, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.49085086584091187, 'reward_total_composite_std': 0.10883072763681412, 'epoch': 0.01}

  2%|▏         | 68/3300 [09:43<7:18:16,  8.14s/it]INFO 04-13 07:30:41 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 69/3300 [09:50<7:02:10,  7.84s/it]2026-04-13 07:30:48,280 | INFO | train_grpo_train | metrics_logged mode=train step=69

                                                   
{'loss': 0.0529, 'grad_norm': 22.51520347595215, 'learning_rate': 9.793939393939394e-06, 'num_tokens': 128609.0, 'completions/mean_length': 59.5, 'completions/min_length': 47.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.3263697326183319, 'rewards/meter/std': 0.34783032536506653, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.988911509513855, 'rewards/repeat_soft/std': 0.010033752769231796, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.1524970978498459, 'rewards/total_composite/mean': 0.3768683075904846, 'rewards/total_composite/std': 0.2702862620353699, 'reward': 0.3768683075904846, 'reward_std': 0.2702862620353699, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2645116448402405, 'sampling/sampling_logp_difference/max': 2.039849281311035, 'sampling/importance_sampling_ratio/min': 0.13004830479621887, 'sampling/importance_sampling_ratio/mean': 1.003286361694336, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4887580871582031, 'clip_ratio/low_mean': 0.1026343535631895, 'clip_ratio/low_min': 0.1026343535631895, 'clip_ratio/high_mean': 0.14072966761887074, 'clip_ratio/high_max': 0.14072966761887074, 'clip_ratio/region_mean': 0.24336402118206024, 'reward_total_mean': 0.3768683075904846, 'reward_meter_mean': 0.3263697326183319, 'reward_meter_std': 0.34783032536506653, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.988911509513855, 'reward_repeat_soft_std': 0.010033752769231796, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.1524970978498459, 'reward_total_composite_mean': 0.3768683075904846, 'reward_total_composite_std': 0.2702862620353699, 'epoch': 0.01}

  2%|▏         | 69/3300 [09:50<7:02:10,  7.84s/it]INFO 04-13 07:30:48 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 70/3300 [09:58<7:01:05,  7.82s/it]2026-04-13 07:30:56,061 | INFO | train_grpo_train | metrics_logged mode=train step=70

                                                   
{'loss': -0.0466, 'grad_norm': 12.199811935424805, 'learning_rate': 9.790909090909093e-06, 'num_tokens': 130737.0, 'completions/mean_length': 94.0, 'completions/min_length': 65.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.2902809977531433, 'rewards/meter/std': 0.12799283862113953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939800500869751, 'rewards/repeat_soft/std': 0.004366800654679537, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.0975411981344223, 'rewards/total_composite/mean': 0.4276006817817688, 'rewards/total_composite/std': 0.03887045010924339, 'reward': 0.4276006817817688, 'reward_std': 0.03887045010924339, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.239035964012146, 'sampling/sampling_logp_difference/max': 2.181110382080078, 'sampling/importance_sampling_ratio/min': 0.11291608214378357, 'sampling/importance_sampling_ratio/mean': 1.0222389698028564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2758661061525345, 'clip_ratio/low_mean': 0.10574052855372429, 'clip_ratio/low_min': 0.10574052855372429, 'clip_ratio/high_mean': 0.10558660514652729, 'clip_ratio/high_max': 0.10558660514652729, 'clip_ratio/region_mean': 0.21132713370025158, 'reward_total_mean': 0.4276006817817688, 'reward_meter_mean': 0.2902809977531433, 'reward_meter_std': 0.12799283862113953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939800500869751, 'reward_repeat_soft_std': 0.004366800654679537, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.0975411981344223, 'reward_total_composite_mean': 0.4276006817817688, 'reward_total_composite_std': 0.03887045010924339, 'epoch': 0.01}

  2%|▏         | 70/3300 [09:58<7:01:05,  7.82s/it]INFO 04-13 07:30:56 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 71/3300 [10:04<6:31:31,  7.28s/it]2026-04-13 07:31:02,060 | INFO | train_grpo_train | metrics_logged mode=train step=71

                                                   
{'loss': 0.0143, 'grad_norm': 14.534868240356445, 'learning_rate': 9.787878787878788e-06, 'num_tokens': 132332.0, 'completions/mean_length': 27.375, 'completions/min_length': 26.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.980559229850769, 'rewards/meter/std': 0.01318947970867157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953598380088806, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.9356783628463745, 'rewards/total_composite/std': 0.007609147112816572, 'reward': 0.9356783628463745, 'reward_std': 0.007609153166413307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06976129859685898, 'sampling/sampling_logp_difference/max': 0.8716723918914795, 'sampling/importance_sampling_ratio/min': 0.41825151443481445, 'sampling/importance_sampling_ratio/mean': 1.0073102712631226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42719393596053123, 'clip_ratio/low_mean': 0.022321429569274187, 'clip_ratio/low_min': 0.022321429569274187, 'clip_ratio/high_mean': 0.02252229768782854, 'clip_ratio/high_max': 0.02252229768782854, 'clip_ratio/region_mean': 0.04484372725710273, 'reward_total_mean': 0.9356783628463745, 'reward_meter_mean': 0.980559229850769, 'reward_meter_std': 0.01318947970867157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953598380088806, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.9356783628463745, 'reward_total_composite_std': 0.007609147112816572, 'epoch': 0.01}

  2%|▏         | 71/3300 [10:04<6:31:31,  7.28s/it]INFO 04-13 07:31:02 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 72/3300 [10:11<6:30:02,  7.25s/it]2026-04-13 07:31:09,250 | INFO | train_grpo_train | metrics_logged mode=train step=72

                                                   
{'loss': 0.0149, 'grad_norm': 18.851024627685547, 'learning_rate': 9.784848484848486e-06, 'num_tokens': 134250.0, 'completions/mean_length': 77.75, 'completions/min_length': 65.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5707041025161743, 'rewards/meter/std': 0.3704380989074707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9915953874588013, 'rewards/repeat_soft/std': 0.003521057078614831, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.579657256603241, 'rewards/total_composite/std': 0.16086405515670776, 'reward': 0.579657256603241, 'reward_std': 0.16086405515670776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15077872574329376, 'sampling/sampling_logp_difference/max': 1.8281669616699219, 'sampling/importance_sampling_ratio/min': 0.16070789098739624, 'sampling/importance_sampling_ratio/mean': 1.001753330230713, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7165041640400887, 'clip_ratio/low_mean': 0.077230763155967, 'clip_ratio/low_min': 0.077230763155967, 'clip_ratio/high_mean': 0.03872744762338698, 'clip_ratio/high_max': 0.03872744762338698, 'clip_ratio/region_mean': 0.11595821077935398, 'reward_total_mean': 0.579657256603241, 'reward_meter_mean': 0.5707041025161743, 'reward_meter_std': 0.3704380989074707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9915953874588013, 'reward_repeat_soft_std': 0.003521057078614831, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.579657256603241, 'reward_total_composite_std': 0.16086405515670776, 'epoch': 0.01}

  2%|▏         | 72/3300 [10:11<6:30:02,  7.25s/it]INFO 04-13 07:31:09 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 73/3300 [10:18<6:24:04,  7.14s/it]2026-04-13 07:31:16,138 | INFO | train_grpo_train | metrics_logged mode=train step=73

                                                   
{'loss': 0.0722, 'grad_norm': 16.596925735473633, 'learning_rate': 9.781818181818183e-06, 'num_tokens': 135829.0, 'completions/mean_length': 38.375, 'completions/min_length': 33.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7099935412406921, 'rewards/meter/std': 0.3367968797683716, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981318712234497, 'rewards/repeat_soft/std': 0.005283824168145657, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.22984081506729126, 'rewards/total_composite/mean': 0.6462399959564209, 'rewards/total_composite/std': 0.21094709634780884, 'reward': 0.6462399959564209, 'reward_std': 0.21094708144664764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1958298534154892, 'sampling/sampling_logp_difference/max': 1.196608543395996, 'sampling/importance_sampling_ratio/min': 0.30221742391586304, 'sampling/importance_sampling_ratio/mean': 1.0368590354919434, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8273597061634064, 'clip_ratio/low_mean': 0.13772443495690823, 'clip_ratio/low_min': 0.13772443495690823, 'clip_ratio/high_mean': 0.09337121434509754, 'clip_ratio/high_max': 0.09337121434509754, 'clip_ratio/region_mean': 0.23109564930200577, 'reward_total_mean': 0.6462399959564209, 'reward_meter_mean': 0.7099935412406921, 'reward_meter_std': 0.3367968797683716, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981318712234497, 'reward_repeat_soft_std': 0.005283824168145657, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.22984081506729126, 'reward_total_composite_mean': 0.6462399959564209, 'reward_total_composite_std': 0.21094709634780884, 'epoch': 0.01}

  2%|▏         | 73/3300 [10:18<6:24:04,  7.14s/it]INFO 04-13 07:31:16 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 74/3300 [10:25<6:31:07,  7.27s/it]2026-04-13 07:31:23,724 | INFO | train_grpo_train | metrics_logged mode=train step=74

                                                   
{'loss': 0.0166, 'grad_norm': 16.803701400756836, 'learning_rate': 9.77878787878788e-06, 'num_tokens': 137366.0, 'completions/mean_length': 46.125, 'completions/min_length': 31.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6728395819664001, 'rewards/meter/std': 0.34131065011024475, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953755736351013, 'rewards/repeat_soft/std': 0.004812015686184168, 'rewards/judge_quality/mean': 0.9275000095367432, 'rewards/judge_quality/std': 0.013887288980185986, 'rewards/total_composite/mean': 0.756302535533905, 'rewards/total_composite/std': 0.20829513669013977, 'reward': 0.756302535533905, 'reward_std': 0.20829510688781738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12959113717079163, 'sampling/sampling_logp_difference/max': 1.2154483795166016, 'sampling/importance_sampling_ratio/min': 0.29657700657844543, 'sampling/importance_sampling_ratio/mean': 1.0163564682006836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8411549031734467, 'clip_ratio/low_mean': 0.06045190338045359, 'clip_ratio/low_min': 0.06045190338045359, 'clip_ratio/high_mean': 0.05965705122798681, 'clip_ratio/high_max': 0.05965705122798681, 'clip_ratio/region_mean': 0.1201089546084404, 'reward_total_mean': 0.756302535533905, 'reward_meter_mean': 0.6728395819664001, 'reward_meter_std': 0.34131065011024475, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953755736351013, 'reward_repeat_soft_std': 0.004812015686184168, 'reward_judge_quality_mean': 0.9275000095367432, 'reward_judge_quality_std': 0.013887288980185986, 'reward_total_composite_mean': 0.756302535533905, 'reward_total_composite_std': 0.20829513669013977, 'epoch': 0.01}

  2%|▏         | 74/3300 [10:25<6:31:07,  7.27s/it]INFO 04-13 07:31:23 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 75/3300 [10:34<6:55:34,  7.73s/it]2026-04-13 07:31:32,522 | INFO | train_grpo_train | metrics_logged mode=train step=75

                                                   
{'loss': 0.3489, 'grad_norm': 11.72706127166748, 'learning_rate': 9.775757575757576e-06, 'num_tokens': 139715.0, 'completions/mean_length': 119.625, 'completions/min_length': 79.0, 'completions/max_length': 264.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.625, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 264.0, 'rewards/meter/mean': 0.7541071772575378, 'rewards/meter/std': 0.2662733495235443, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9913018941879272, 'rewards/repeat_soft/std': 0.011900298297405243, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.606920599937439, 'rewards/total_composite/std': 0.20927901566028595, 'reward': 0.606920599937439, 'reward_std': 0.20927901566028595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2122229039669037, 'sampling/sampling_logp_difference/max': 1.480539321899414, 'sampling/importance_sampling_ratio/min': 0.22751495242118835, 'sampling/importance_sampling_ratio/mean': 1.0367655754089355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.362242892384529, 'clip_ratio/low_mean': 0.1122982744127512, 'clip_ratio/low_min': 0.1122982744127512, 'clip_ratio/high_mean': 0.07517952844500542, 'clip_ratio/high_max': 0.07517952844500542, 'clip_ratio/region_mean': 0.18747780285775661, 'reward_total_mean': 0.606920599937439, 'reward_meter_mean': 0.7541071772575378, 'reward_meter_std': 0.2662733495235443, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9913018941879272, 'reward_repeat_soft_std': 0.011900298297405243, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.606920599937439, 'reward_total_composite_std': 0.20927901566028595, 'epoch': 0.01}

  2%|▏         | 75/3300 [10:34<6:55:34,  7.73s/it]INFO 04-13 07:31:32 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 76/3300 [10:41<6:38:56,  7.42s/it]2026-04-13 07:31:39,229 | INFO | train_grpo_train | metrics_logged mode=train step=76

                                                   
{'loss': 0.0372, 'grad_norm': 18.240943908691406, 'learning_rate': 9.772727272727273e-06, 'num_tokens': 141197.0, 'completions/mean_length': 31.25, 'completions/min_length': 18.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6227948069572449, 'rewards/meter/std': 0.4872068166732788, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9694886207580566, 'rewards/repeat_soft/std': 0.019950784742832184, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.2499571591615677, 'rewards/total_composite/mean': 0.5201561450958252, 'rewards/total_composite/std': 0.23341584205627441, 'reward': 0.5201561450958252, 'reward_std': 0.2334158569574356, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19917994737625122, 'sampling/sampling_logp_difference/max': 1.2310771942138672, 'sampling/importance_sampling_ratio/min': 0.2919778823852539, 'sampling/importance_sampling_ratio/mean': 1.0508776903152466, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.957547053694725, 'clip_ratio/low_mean': 0.062437155516818166, 'clip_ratio/low_min': 0.062437155516818166, 'clip_ratio/high_mean': 0.10055694729089737, 'clip_ratio/high_max': 0.10055694729089737, 'clip_ratio/region_mean': 0.16299410280771554, 'reward_total_mean': 0.5201561450958252, 'reward_meter_mean': 0.6227948069572449, 'reward_meter_std': 0.4872068166732788, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9694886207580566, 'reward_repeat_soft_std': 0.019950784742832184, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.2499571591615677, 'reward_total_composite_mean': 0.5201561450958252, 'reward_total_composite_std': 0.23341584205627441, 'epoch': 0.01}

  2%|▏         | 76/3300 [10:41<6:38:56,  7.42s/it]INFO 04-13 07:31:39 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 77/3300 [10:48<6:33:07,  7.32s/it]2026-04-13 07:31:46,302 | INFO | train_grpo_train | metrics_logged mode=train step=77

                                                   
{'loss': 0.132, 'grad_norm': 16.104795455932617, 'learning_rate': 9.76969696969697e-06, 'num_tokens': 142809.0, 'completions/mean_length': 45.5, 'completions/min_length': 32.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5599648952484131, 'rewards/meter/std': 0.4068310260772705, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9915940761566162, 'rewards/repeat_soft/std': 0.012580431066453457, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.22699514031410217, 'rewards/total_composite/mean': 0.5684210658073425, 'rewards/total_composite/std': 0.18888187408447266, 'reward': 0.5684210658073425, 'reward_std': 0.18888187408447266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.200801819562912, 'sampling/sampling_logp_difference/max': 3.130115509033203, 'sampling/importance_sampling_ratio/min': 0.04371274635195732, 'sampling/importance_sampling_ratio/mean': 1.0195224285125732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2177944853901863, 'clip_ratio/low_mean': 0.08189102727919817, 'clip_ratio/low_min': 0.08189102727919817, 'clip_ratio/high_mean': 0.09817468374967575, 'clip_ratio/high_max': 0.09817468374967575, 'clip_ratio/region_mean': 0.18006571102887392, 'reward_total_mean': 0.5684210658073425, 'reward_meter_mean': 0.5599648952484131, 'reward_meter_std': 0.4068310260772705, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9915940761566162, 'reward_repeat_soft_std': 0.012580431066453457, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.22699514031410217, 'reward_total_composite_mean': 0.5684210658073425, 'reward_total_composite_std': 0.18888187408447266, 'epoch': 0.01}

  2%|▏         | 77/3300 [10:48<6:33:07,  7.32s/it]INFO 04-13 07:31:46 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 78/3300 [10:56<6:35:46,  7.37s/it]2026-04-13 07:31:53,792 | INFO | train_grpo_train | metrics_logged mode=train step=78

                                                   
{'loss': 0.1556, 'grad_norm': 9.579455375671387, 'learning_rate': 9.766666666666667e-06, 'num_tokens': 145314.0, 'completions/mean_length': 118.125, 'completions/min_length': 70.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.6192508935928345, 'rewards/meter/std': 0.2793404459953308, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996671199798584, 'rewards/repeat_soft/std': 0.00344731449149549, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.60548335313797, 'rewards/total_composite/std': 0.1651448756456375, 'reward': 0.60548335313797, 'reward_std': 0.1651448756456375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2017233520746231, 'sampling/sampling_logp_difference/max': 2.272388458251953, 'sampling/importance_sampling_ratio/min': 0.10306572169065475, 'sampling/importance_sampling_ratio/mean': 1.0252689123153687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0700202882289886, 'clip_ratio/low_mean': 0.09481169655919075, 'clip_ratio/low_min': 0.09481169655919075, 'clip_ratio/high_mean': 0.10964499041438103, 'clip_ratio/high_max': 0.10964499041438103, 'clip_ratio/region_mean': 0.20445668697357178, 'reward_total_mean': 0.60548335313797, 'reward_meter_mean': 0.6192508935928345, 'reward_meter_std': 0.2793404459953308, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996671199798584, 'reward_repeat_soft_std': 0.00344731449149549, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.60548335313797, 'reward_total_composite_std': 0.1651448756456375, 'epoch': 0.01}

  2%|▏         | 78/3300 [10:56<6:35:46,  7.37s/it]INFO 04-13 07:31:54 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 79/3300 [11:08<7:58:20,  8.91s/it]2026-04-13 07:32:06,295 | INFO | train_grpo_train | metrics_logged mode=train step=79

                                                   
{'loss': -0.1215, 'grad_norm': 5.458466529846191, 'learning_rate': 9.763636363636365e-06, 'num_tokens': 147338.0, 'completions/mean_length': 140.0, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 86.85714721679688, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.2900845408439636, 'rewards/meter/std': 0.2756892442703247, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.3563483655452728, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9936281442642212, 'rewards/repeat_soft/std': 0.0072111692279577255, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.3009004592895508, 'rewards/total_composite/mean': 0.44272592663764954, 'rewards/total_composite/std': 0.18702588975429535, 'reward': 0.44272592663764954, 'reward_std': 0.18702590465545654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20270249247550964, 'sampling/sampling_logp_difference/max': 2.109290599822998, 'sampling/importance_sampling_ratio/min': 0.12132399529218674, 'sampling/importance_sampling_ratio/mean': 1.0167397260665894, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4257397055625916, 'clip_ratio/low_mean': 0.09781302325427532, 'clip_ratio/low_min': 0.09781302325427532, 'clip_ratio/high_mean': 0.07240715436637402, 'clip_ratio/high_max': 0.07240715436637402, 'clip_ratio/region_mean': 0.17022017762064934, 'reward_total_mean': 0.44272592663764954, 'reward_meter_mean': 0.2900845408439636, 'reward_meter_std': 0.2756892442703247, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.3563483655452728, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9936281442642212, 'reward_repeat_soft_std': 0.0072111692279577255, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.3009004592895508, 'reward_total_composite_mean': 0.44272592663764954, 'reward_total_composite_std': 0.18702588975429535, 'epoch': 0.01}

  2%|▏         | 79/3300 [11:08<7:58:20,  8.91s/it]INFO 04-13 07:32:06 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 80/3300 [11:19<8:32:36,  9.55s/it]2026-04-13 07:32:17,345 | INFO | train_grpo_train | metrics_logged mode=train step=80

                                                   
{'loss': 0.6328, 'grad_norm': 9.749089241027832, 'learning_rate': 9.760606060606062e-06, 'num_tokens': 149447.0, 'completions/mean_length': 94.625, 'completions/min_length': 33.0, 'completions/max_length': 422.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 422.0, 'rewards/meter/mean': 0.7609310746192932, 'rewards/meter/std': 0.37621188163757324, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9743838310241699, 'rewards/repeat_soft/std': 0.06604637950658798, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.22965426743030548, 'rewards/total_composite/mean': 0.6037291288375854, 'rewards/total_composite/std': 0.1989206224679947, 'reward': 0.6037291288375854, 'reward_std': 0.19892063736915588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15118443965911865, 'sampling/sampling_logp_difference/max': 1.5827999114990234, 'sampling/importance_sampling_ratio/min': 0.20539918541908264, 'sampling/importance_sampling_ratio/mean': 1.0348566770553589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1621425449848175, 'clip_ratio/low_mean': 0.06869752984493971, 'clip_ratio/low_min': 0.06869752984493971, 'clip_ratio/high_mean': 0.0752419987693429, 'clip_ratio/high_max': 0.0752419987693429, 'clip_ratio/region_mean': 0.1439395286142826, 'reward_total_mean': 0.6037291288375854, 'reward_meter_mean': 0.7609310746192932, 'reward_meter_std': 0.37621188163757324, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9743838310241699, 'reward_repeat_soft_std': 0.06604637950658798, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.22965426743030548, 'reward_total_composite_mean': 0.6037291288375854, 'reward_total_composite_std': 0.1989206224679947, 'epoch': 0.01}

  2%|▏         | 80/3300 [11:19<8:32:36,  9.55s/it]INFO 04-13 07:32:17 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 81/3300 [11:29<8:33:34,  9.57s/it]2026-04-13 07:32:26,966 | INFO | train_grpo_train | metrics_logged mode=train step=81

                                                   
{'loss': 0.2858, 'grad_norm': 9.317183494567871, 'learning_rate': 9.757575757575758e-06, 'num_tokens': 152215.0, 'completions/mean_length': 136.0, 'completions/min_length': 75.0, 'completions/max_length': 226.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.0, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 226.0, 'rewards/meter/mean': 0.8717004060745239, 'rewards/meter/std': 0.18755942583084106, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9913686513900757, 'rewards/repeat_soft/std': 0.01117410883307457, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/total_composite/mean': 0.643903374671936, 'rewards/total_composite/std': 0.1500636488199234, 'reward': 0.643903374671936, 'reward_std': 0.1500636488199234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2290331870317459, 'sampling/sampling_logp_difference/max': 1.9447827339172363, 'sampling/importance_sampling_ratio/min': 0.14301829040050507, 'sampling/importance_sampling_ratio/mean': 1.0455464124679565, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6830307990312576, 'clip_ratio/low_mean': 0.12898291274905205, 'clip_ratio/low_min': 0.12898291274905205, 'clip_ratio/high_mean': 0.0660889595746994, 'clip_ratio/high_max': 0.0660889595746994, 'clip_ratio/region_mean': 0.19507187232375145, 'reward_total_mean': 0.643903374671936, 'reward_meter_mean': 0.8717004060745239, 'reward_meter_std': 0.18755942583084106, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9913686513900757, 'reward_repeat_soft_std': 0.01117410883307457, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.643903374671936, 'reward_total_composite_std': 0.1500636488199234, 'epoch': 0.01}

  2%|▏         | 81/3300 [11:29<8:33:34,  9.57s/it]INFO 04-13 07:32:27 [block_pool.py:378] Successfully reset prefix cache

  2%|▏         | 82/3300 [11:36<7:51:01,  8.78s/it]2026-04-13 07:32:33,903 | INFO | train_grpo_train | metrics_logged mode=train step=82

                                                   
{'loss': 0.2129, 'grad_norm': 18.911800384521484, 'learning_rate': 9.754545454545455e-06, 'num_tokens': 153757.0, 'completions/mean_length': 43.75, 'completions/min_length': 33.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.3292384743690491, 'rewards/meter/std': 0.38146013021469116, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974339008331299, 'rewards/repeat_soft/std': 0.004254767205566168, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.4292713403701782, 'rewards/total_composite/std': 0.1198873519897461, 'reward': 0.4292713403701782, 'reward_std': 0.1198873519897461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22892795503139496, 'sampling/sampling_logp_difference/max': 1.2771120071411133, 'sampling/importance_sampling_ratio/min': 0.27884143590927124, 'sampling/importance_sampling_ratio/mean': 1.0563212633132935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1919937133789062, 'clip_ratio/low_mean': 0.16997694596648216, 'clip_ratio/low_min': 0.16997694596648216, 'clip_ratio/high_mean': 0.048161765560507774, 'clip_ratio/high_max': 0.048161765560507774, 'clip_ratio/region_mean': 0.21813871152698994, 'reward_total_mean': 0.4292713403701782, 'reward_meter_mean': 0.3292384743690491, 'reward_meter_std': 0.38146013021469116, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974339008331299, 'reward_repeat_soft_std': 0.004254767205566168, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.4292713403701782, 'reward_total_composite_std': 0.1198873519897461, 'epoch': 0.01}

  2%|▏         | 82/3300 [11:36<7:51:01,  8.78s/it]INFO 04-13 07:32:34 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 83/3300 [11:42<7:17:35,  8.16s/it]2026-04-13 07:32:40,617 | INFO | train_grpo_train | metrics_logged mode=train step=83

                                                   
{'loss': 0.0794, 'grad_norm': 12.305427551269531, 'learning_rate': 9.751515151515152e-06, 'num_tokens': 155299.0, 'completions/mean_length': 47.75, 'completions/min_length': 37.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7958900928497314, 'rewards/meter/std': 0.3201044499874115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9813162088394165, 'rewards/repeat_soft/std': 0.050224728882312775, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6428062915802002, 'rewards/total_composite/std': 0.19279375672340393, 'reward': 0.6428062915802002, 'reward_std': 0.19279374182224274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18508756160736084, 'sampling/sampling_logp_difference/max': 1.5233378410339355, 'sampling/importance_sampling_ratio/min': 0.21798309683799744, 'sampling/importance_sampling_ratio/mean': 1.0269668102264404, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6632934510707855, 'clip_ratio/low_mean': 0.11185488197952509, 'clip_ratio/low_min': 0.11185488197952509, 'clip_ratio/high_mean': 0.045780474320054054, 'clip_ratio/high_max': 0.045780474320054054, 'clip_ratio/region_mean': 0.15763535629957914, 'reward_total_mean': 0.6428062915802002, 'reward_meter_mean': 0.7958900928497314, 'reward_meter_std': 0.3201044499874115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9813162088394165, 'reward_repeat_soft_std': 0.050224728882312775, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6428062915802002, 'reward_total_composite_std': 0.19279375672340393, 'epoch': 0.01}

  3%|▎         | 83/3300 [11:42<7:17:35,  8.16s/it]INFO 04-13 07:32:40 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 84/3300 [11:50<7:05:29,  7.94s/it]2026-04-13 07:32:48,034 | INFO | train_grpo_train | metrics_logged mode=train step=84

                                                   
{'loss': 0.0305, 'grad_norm': 15.982399940490723, 'learning_rate': 9.74848484848485e-06, 'num_tokens': 157087.0, 'completions/mean_length': 42.5, 'completions/min_length': 34.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.39095914363861084, 'rewards/meter/std': 0.42909476161003113, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967079758644104, 'rewards/repeat_soft/std': 0.008959452621638775, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.4579356908798218, 'rewards/total_composite/std': 0.1210685521364212, 'reward': 0.4579356908798218, 'reward_std': 0.1210685521364212, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24277810752391815, 'sampling/sampling_logp_difference/max': 1.4623289108276367, 'sampling/importance_sampling_ratio/min': 0.23169603943824768, 'sampling/importance_sampling_ratio/mean': 1.0521039962768555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.471090257167816, 'clip_ratio/low_mean': 0.1764454822987318, 'clip_ratio/low_min': 0.1764454822987318, 'clip_ratio/high_mean': 0.08806561306118965, 'clip_ratio/high_max': 0.08806561306118965, 'clip_ratio/region_mean': 0.26451109535992146, 'reward_total_mean': 0.4579356908798218, 'reward_meter_mean': 0.39095914363861084, 'reward_meter_std': 0.42909476161003113, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967079758644104, 'reward_repeat_soft_std': 0.008959452621638775, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.4579356908798218, 'reward_total_composite_std': 0.1210685521364212, 'epoch': 0.01}

  3%|▎         | 84/3300 [11:50<7:05:29,  7.94s/it]INFO 04-13 07:32:48 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 85/3300 [11:57<6:56:25,  7.77s/it]2026-04-13 07:32:55,417 | INFO | train_grpo_train | metrics_logged mode=train step=85

                                                   
{'loss': -0.0612, 'grad_norm': 28.472885131835938, 'learning_rate': 9.745454545454547e-06, 'num_tokens': 158426.0, 'completions/mean_length': 22.375, 'completions/min_length': 17.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6606827974319458, 'rewards/meter/std': 0.4175296425819397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.930449366569519, 'rewards/repeat_soft/std': 0.06305917352437973, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5210764408111572, 'rewards/total_composite/std': 0.11448659002780914, 'reward': 0.5210764408111572, 'reward_std': 0.11448659002780914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22843873500823975, 'sampling/sampling_logp_difference/max': 2.0268335342407227, 'sampling/importance_sampling_ratio/min': 0.13175204396247864, 'sampling/importance_sampling_ratio/mean': 1.0153146982192993, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6977891623973846, 'clip_ratio/low_mean': 0.08766401372849941, 'clip_ratio/low_min': 0.08766401372849941, 'clip_ratio/high_mean': 0.11823012121021748, 'clip_ratio/high_max': 0.11823012121021748, 'clip_ratio/region_mean': 0.2058941349387169, 'reward_total_mean': 0.5210764408111572, 'reward_meter_mean': 0.6606827974319458, 'reward_meter_std': 0.4175296425819397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.930449366569519, 'reward_repeat_soft_std': 0.06305917352437973, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5210764408111572, 'reward_total_composite_std': 0.11448659002780914, 'epoch': 0.01}

  3%|▎         | 85/3300 [11:57<6:56:25,  7.77s/it]INFO 04-13 07:32:55 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 86/3300 [12:05<7:03:33,  7.91s/it]2026-04-13 07:33:03,642 | INFO | train_grpo_train | metrics_logged mode=train step=86

                                                   
{'loss': 0.0581, 'grad_norm': 19.20193862915039, 'learning_rate': 9.742424242424244e-06, 'num_tokens': 160546.0, 'completions/mean_length': 86.0, 'completions/min_length': 67.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.24867114424705505, 'rewards/meter/std': 0.20864476263523102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985886216163635, 'rewards/repeat_soft/std': 0.0018666742835193872, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.41330647468566895, 'rewards/total_composite/std': 0.06062160059809685, 'reward': 0.41330647468566895, 'reward_std': 0.06062160059809685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23355013132095337, 'sampling/sampling_logp_difference/max': 3.66347599029541, 'sampling/importance_sampling_ratio/min': 0.02564322203397751, 'sampling/importance_sampling_ratio/mean': 1.0034799575805664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.218077875673771, 'clip_ratio/low_mean': 0.12993094697594643, 'clip_ratio/low_min': 0.12993094697594643, 'clip_ratio/high_mean': 0.059350667521357536, 'clip_ratio/high_max': 0.059350667521357536, 'clip_ratio/region_mean': 0.18928161449730396, 'reward_total_mean': 0.41330647468566895, 'reward_meter_mean': 0.24867114424705505, 'reward_meter_std': 0.20864476263523102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985886216163635, 'reward_repeat_soft_std': 0.0018666742835193872, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.41330647468566895, 'reward_total_composite_std': 0.06062160059809685, 'epoch': 0.01}

  3%|▎         | 86/3300 [12:05<7:03:33,  7.91s/it]INFO 04-13 07:33:03 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 87/3300 [12:12<6:42:45,  7.52s/it]2026-04-13 07:33:10,260 | INFO | train_grpo_train | metrics_logged mode=train step=87

                                                   
{'loss': 0.145, 'grad_norm': 23.893922805786133, 'learning_rate': 9.739393939393941e-06, 'num_tokens': 161860.0, 'completions/mean_length': 23.25, 'completions/min_length': 15.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.2548510432243347, 'rewards/meter/std': 0.3831472098827362, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9499562978744507, 'rewards/repeat_soft/std': 0.021399270743131638, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.22258226573467255, 'rewards/total_composite/mean': 0.4158652424812317, 'rewards/total_composite/std': 0.10863844305276871, 'reward': 0.4158652424812317, 'reward_std': 0.10863843560218811, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21412859857082367, 'sampling/sampling_logp_difference/max': 1.3826007843017578, 'sampling/importance_sampling_ratio/min': 0.25092509388923645, 'sampling/importance_sampling_ratio/mean': 1.0157396793365479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6046317741274834, 'clip_ratio/low_mean': 0.12450417224317789, 'clip_ratio/low_min': 0.12450417224317789, 'clip_ratio/high_mean': 0.09791667014360428, 'clip_ratio/high_max': 0.09791667014360428, 'clip_ratio/region_mean': 0.22242084238678217, 'reward_total_mean': 0.4158652424812317, 'reward_meter_mean': 0.2548510432243347, 'reward_meter_std': 0.3831472098827362, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9499562978744507, 'reward_repeat_soft_std': 0.021399270743131638, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.22258226573467255, 'reward_total_composite_mean': 0.4158652424812317, 'reward_total_composite_std': 0.10863844305276871, 'epoch': 0.01}

  3%|▎         | 87/3300 [12:12<6:42:45,  7.52s/it]INFO 04-13 07:33:10 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 88/3300 [12:19<6:30:16,  7.29s/it]2026-04-13 07:33:17,013 | INFO | train_grpo_train | metrics_logged mode=train step=88

                                                   
{'loss': -0.112, 'grad_norm': 15.566302299499512, 'learning_rate': 9.736363636363637e-06, 'num_tokens': 163574.0, 'completions/mean_length': 53.25, 'completions/min_length': 38.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6322734951972961, 'rewards/meter/std': 0.4166473150253296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9894940853118896, 'rewards/repeat_soft/std': 0.013136623427271843, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5233603715896606, 'rewards/total_composite/std': 0.11490435898303986, 'reward': 0.5233603715896606, 'reward_std': 0.11490435898303986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22387740015983582, 'sampling/sampling_logp_difference/max': 1.7036323547363281, 'sampling/importance_sampling_ratio/min': 0.1820211559534073, 'sampling/importance_sampling_ratio/mean': 1.0229641199111938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8468092530965805, 'clip_ratio/low_mean': 0.09993897192180157, 'clip_ratio/low_min': 0.09993897192180157, 'clip_ratio/high_mean': 0.1066430639475584, 'clip_ratio/high_max': 0.1066430639475584, 'clip_ratio/region_mean': 0.20658203586935997, 'reward_total_mean': 0.5233603715896606, 'reward_meter_mean': 0.6322734951972961, 'reward_meter_std': 0.4166473150253296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9894940853118896, 'reward_repeat_soft_std': 0.013136623427271843, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5233603715896606, 'reward_total_composite_std': 0.11490435898303986, 'epoch': 0.01}

  3%|▎         | 88/3300 [12:19<6:30:16,  7.29s/it]INFO 04-13 07:33:17 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 89/3300 [12:26<6:25:28,  7.20s/it]2026-04-13 07:33:24,011 | INFO | train_grpo_train | metrics_logged mode=train step=89

                                                   
{'loss': 0.052, 'grad_norm': 22.00596046447754, 'learning_rate': 9.733333333333334e-06, 'num_tokens': 165573.0, 'completions/mean_length': 84.875, 'completions/min_length': 53.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.5694794654846191, 'rewards/meter/std': 0.32996881008148193, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9958401322364807, 'rewards/repeat_soft/std': 0.006230877246707678, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.5533815622329712, 'rewards/total_composite/std': 0.10411175340414047, 'reward': 0.5533815622329712, 'reward_std': 0.10411174595355988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23848699033260345, 'sampling/sampling_logp_difference/max': 3.618814468383789, 'sampling/importance_sampling_ratio/min': 0.026814447715878487, 'sampling/importance_sampling_ratio/mean': 0.9860868453979492, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9971832558512688, 'clip_ratio/low_mean': 0.07648433186113834, 'clip_ratio/low_min': 0.07648433186113834, 'clip_ratio/high_mean': 0.09676007181406021, 'clip_ratio/high_max': 0.09676007181406021, 'clip_ratio/region_mean': 0.17324440367519855, 'reward_total_mean': 0.5533815622329712, 'reward_meter_mean': 0.5694794654846191, 'reward_meter_std': 0.32996881008148193, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9958401322364807, 'reward_repeat_soft_std': 0.006230877246707678, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.5533815622329712, 'reward_total_composite_std': 0.10411175340414047, 'epoch': 0.01}

  3%|▎         | 89/3300 [12:26<6:25:28,  7.20s/it]INFO 04-13 07:33:24 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 90/3300 [12:32<6:08:10,  6.88s/it]2026-04-13 07:33:30,143 | INFO | train_grpo_train | metrics_logged mode=train step=90

                                                   
{'loss': 0.0951, 'grad_norm': 23.296001434326172, 'learning_rate': 9.730303030303031e-06, 'num_tokens': 167021.0, 'completions/mean_length': 24.0, 'completions/min_length': 21.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.36750417947769165, 'rewards/meter/std': 0.3164251148700714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948339462280273, 'rewards/repeat_soft/std': 0.01454270351678133, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.21866071224212646, 'rewards/total_composite/mean': 0.5101072788238525, 'rewards/total_composite/std': 0.18115530908107758, 'reward': 0.5101072788238525, 'reward_std': 0.18115530908107758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21593332290649414, 'sampling/sampling_logp_difference/max': 1.3680660724639893, 'sampling/importance_sampling_ratio/min': 0.25459885597229004, 'sampling/importance_sampling_ratio/mean': 1.0594345331192017, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3456864058971405, 'clip_ratio/low_mean': 0.16109457332640886, 'clip_ratio/low_min': 0.16109457332640886, 'clip_ratio/high_mean': 0.02095238072797656, 'clip_ratio/high_max': 0.02095238072797656, 'clip_ratio/region_mean': 0.18204695405438542, 'reward_total_mean': 0.5101072788238525, 'reward_meter_mean': 0.36750417947769165, 'reward_meter_std': 0.3164251148700714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948339462280273, 'reward_repeat_soft_std': 0.01454270351678133, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.21866071224212646, 'reward_total_composite_mean': 0.5101072788238525, 'reward_total_composite_std': 0.18115530908107758, 'epoch': 0.01}

  3%|▎         | 90/3300 [12:32<6:08:10,  6.88s/it]INFO 04-13 07:33:30 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 91/3300 [12:40<6:33:37,  7.36s/it]2026-04-13 07:33:38,619 | INFO | train_grpo_train | metrics_logged mode=train step=91

                                                   
{'loss': -0.0065, 'grad_norm': 11.63347053527832, 'learning_rate': 9.727272727272728e-06, 'num_tokens': 169080.0, 'completions/mean_length': 84.375, 'completions/min_length': 64.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.4681720435619354, 'rewards/meter/std': 0.38532817363739014, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995926022529602, 'rewards/repeat_soft/std': 0.003701572772115469, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.19190309941768646, 'rewards/total_composite/mean': 0.525649905204773, 'rewards/total_composite/std': 0.17196428775787354, 'reward': 0.525649905204773, 'reward_std': 0.17196428775787354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2211720496416092, 'sampling/sampling_logp_difference/max': 1.5955607891082764, 'sampling/importance_sampling_ratio/min': 0.20279477536678314, 'sampling/importance_sampling_ratio/mean': 1.041763186454773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2681550085544586, 'clip_ratio/low_mean': 0.09809909947216511, 'clip_ratio/low_min': 0.09809909947216511, 'clip_ratio/high_mean': 0.09866714477539062, 'clip_ratio/high_max': 0.09866714477539062, 'clip_ratio/region_mean': 0.19676624424755573, 'reward_total_mean': 0.525649905204773, 'reward_meter_mean': 0.4681720435619354, 'reward_meter_std': 0.38532817363739014, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995926022529602, 'reward_repeat_soft_std': 0.003701572772115469, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.19190309941768646, 'reward_total_composite_mean': 0.525649905204773, 'reward_total_composite_std': 0.17196428775787354, 'epoch': 0.01}

  3%|▎         | 91/3300 [12:40<6:33:37,  7.36s/it]INFO 04-13 07:33:38 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 92/3300 [12:47<6:30:17,  7.30s/it]2026-04-13 07:33:45,779 | INFO | train_grpo_train | metrics_logged mode=train step=92

                                                   
{'loss': 0.0501, 'grad_norm': 17.996219635009766, 'learning_rate': 9.724242424242426e-06, 'num_tokens': 170756.0, 'completions/mean_length': 41.5, 'completions/min_length': 32.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.2540503144264221, 'rewards/meter/std': 0.31625130772590637, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9889193773269653, 'rewards/repeat_soft/std': 0.01017341110855341, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.44203996658325195, 'rewards/total_composite/std': 0.11635930091142654, 'reward': 0.44203996658325195, 'reward_std': 0.11635930836200714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23884731531143188, 'sampling/sampling_logp_difference/max': 2.4119834899902344, 'sampling/importance_sampling_ratio/min': 0.08963732421398163, 'sampling/importance_sampling_ratio/mean': 1.0279514789581299, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8763217777013779, 'clip_ratio/low_mean': 0.12165019661188126, 'clip_ratio/low_min': 0.12165019661188126, 'clip_ratio/high_mean': 0.07047158665955067, 'clip_ratio/high_max': 0.07047158665955067, 'clip_ratio/region_mean': 0.19212178327143192, 'reward_total_mean': 0.44203996658325195, 'reward_meter_mean': 0.2540503144264221, 'reward_meter_std': 0.31625130772590637, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9889193773269653, 'reward_repeat_soft_std': 0.01017341110855341, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.44203996658325195, 'reward_total_composite_std': 0.11635930091142654, 'epoch': 0.01}

  3%|▎         | 92/3300 [12:48<6:30:17,  7.30s/it]INFO 04-13 07:33:46 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 93/3300 [12:55<6:28:32,  7.27s/it]2026-04-13 07:33:52,978 | INFO | train_grpo_train | metrics_logged mode=train step=93

                                                   
{'loss': -0.004, 'grad_norm': 12.886481285095215, 'learning_rate': 9.721212121212123e-06, 'num_tokens': 172367.0, 'completions/mean_length': 54.375, 'completions/min_length': 44.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.667952299118042, 'rewards/meter/std': 0.42871978878974915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986583590507507, 'rewards/repeat_soft/std': 0.0020897281356155872, 'rewards/judge_quality/mean': 0.39249998331069946, 'rewards/judge_quality/std': 0.08892211318016052, 'rewards/total_composite/mean': 0.5257940292358398, 'rewards/total_composite/std': 0.11939812451601028, 'reward': 0.5257940292358398, 'reward_std': 0.11939812451601028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21974819898605347, 'sampling/sampling_logp_difference/max': 2.408405303955078, 'sampling/importance_sampling_ratio/min': 0.08995863795280457, 'sampling/importance_sampling_ratio/mean': 1.0207183361053467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9088242799043655, 'clip_ratio/low_mean': 0.09619488101452589, 'clip_ratio/low_min': 0.09619488101452589, 'clip_ratio/high_mean': 0.11377151682972908, 'clip_ratio/high_max': 0.11377151682972908, 'clip_ratio/region_mean': 0.20996639784425497, 'reward_total_mean': 0.5257940292358398, 'reward_meter_mean': 0.667952299118042, 'reward_meter_std': 0.42871978878974915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986583590507507, 'reward_repeat_soft_std': 0.0020897281356155872, 'reward_judge_quality_mean': 0.39249998331069946, 'reward_judge_quality_std': 0.08892211318016052, 'reward_total_composite_mean': 0.5257940292358398, 'reward_total_composite_std': 0.11939812451601028, 'epoch': 0.01}

  3%|▎         | 93/3300 [12:55<6:28:32,  7.27s/it]INFO 04-13 07:33:53 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 94/3300 [13:02<6:21:07,  7.13s/it]2026-04-13 07:33:59,792 | INFO | train_grpo_train | metrics_logged mode=train step=94

                                                   
{'loss': 0.066, 'grad_norm': 72.35273742675781, 'learning_rate': 9.718181818181818e-06, 'num_tokens': 174097.0, 'completions/mean_length': 38.25, 'completions/min_length': 32.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.3880566954612732, 'rewards/meter/std': 0.3312193751335144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9880231618881226, 'rewards/repeat_soft/std': 0.011508532799780369, 'rewards/judge_quality/mean': 0.6812499761581421, 'rewards/judge_quality/std': 0.25542333722114563, 'rewards/total_composite/mean': 0.5185790061950684, 'rewards/total_composite/std': 0.18502680957317352, 'reward': 0.5185790061950684, 'reward_std': 0.18502680957317352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20673184096813202, 'sampling/sampling_logp_difference/max': 1.6290805339813232, 'sampling/importance_sampling_ratio/min': 0.196109801530838, 'sampling/importance_sampling_ratio/mean': 1.0550040006637573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.473860390484333, 'clip_ratio/low_mean': 0.11757584288716316, 'clip_ratio/low_min': 0.11757584288716316, 'clip_ratio/high_mean': 0.061890242621302605, 'clip_ratio/high_max': 0.061890242621302605, 'clip_ratio/region_mean': 0.17946608550846577, 'reward_total_mean': 0.5185790061950684, 'reward_meter_mean': 0.3880566954612732, 'reward_meter_std': 0.3312193751335144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9880231618881226, 'reward_repeat_soft_std': 0.011508532799780369, 'reward_judge_quality_mean': 0.6812499761581421, 'reward_judge_quality_std': 0.25542333722114563, 'reward_total_composite_mean': 0.5185790061950684, 'reward_total_composite_std': 0.18502680957317352, 'epoch': 0.01}

  3%|▎         | 94/3300 [13:02<6:21:07,  7.13s/it]INFO 04-13 07:34:00 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 95/3300 [13:09<6:26:00,  7.23s/it]2026-04-13 07:34:07,237 | INFO | train_grpo_train | metrics_logged mode=train step=95

                                                   
{'loss': 0.0334, 'grad_norm': 10.50815486907959, 'learning_rate': 9.715151515151516e-06, 'num_tokens': 176393.0, 'completions/mean_length': 104.0, 'completions/min_length': 75.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.0, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.5511190891265869, 'rewards/meter/std': 0.3720203936100006, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9935624003410339, 'rewards/repeat_soft/std': 0.009157720021903515, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.5496877431869507, 'rewards/total_composite/std': 0.179593026638031, 'reward': 0.5496877431869507, 'reward_std': 0.1795930117368698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23180857300758362, 'sampling/sampling_logp_difference/max': 1.9675331115722656, 'sampling/importance_sampling_ratio/min': 0.13980130851268768, 'sampling/importance_sampling_ratio/mean': 1.0309317111968994, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4507804214954376, 'clip_ratio/low_mean': 0.12515215389430523, 'clip_ratio/low_min': 0.12515215389430523, 'clip_ratio/high_mean': 0.07687505520880222, 'clip_ratio/high_max': 0.07687505520880222, 'clip_ratio/region_mean': 0.20202720910310745, 'reward_total_mean': 0.5496877431869507, 'reward_meter_mean': 0.5511190891265869, 'reward_meter_std': 0.3720203936100006, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9935624003410339, 'reward_repeat_soft_std': 0.009157720021903515, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.5496877431869507, 'reward_total_composite_std': 0.179593026638031, 'epoch': 0.01}

  3%|▎         | 95/3300 [13:09<6:26:00,  7.23s/it]INFO 04-13 07:34:07 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 96/3300 [13:16<6:28:18,  7.27s/it]2026-04-13 07:34:14,616 | INFO | train_grpo_train | metrics_logged mode=train step=96

                                                   
{'loss': -0.002, 'grad_norm': 18.13115692138672, 'learning_rate': 9.712121212121213e-06, 'num_tokens': 177986.0, 'completions/mean_length': 44.125, 'completions/min_length': 32.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.35852187871932983, 'rewards/meter/std': 0.3847596347332001, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928138256072998, 'rewards/repeat_soft/std': 0.008599132299423218, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955308735370636, 'rewards/total_composite/mean': 0.45549121499061584, 'rewards/total_composite/std': 0.09961428493261337, 'reward': 0.45549121499061584, 'reward_std': 0.09961428493261337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21455226838588715, 'sampling/sampling_logp_difference/max': 1.6288232803344727, 'sampling/importance_sampling_ratio/min': 0.19616025686264038, 'sampling/importance_sampling_ratio/mean': 1.0390883684158325, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0448380559682846, 'clip_ratio/low_mean': 0.12832126952707767, 'clip_ratio/low_min': 0.12832126952707767, 'clip_ratio/high_mean': 0.0959168542176485, 'clip_ratio/high_max': 0.0959168542176485, 'clip_ratio/region_mean': 0.22423812374472618, 'reward_total_mean': 0.45549121499061584, 'reward_meter_mean': 0.35852187871932983, 'reward_meter_std': 0.3847596347332001, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928138256072998, 'reward_repeat_soft_std': 0.008599132299423218, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955308735370636, 'reward_total_composite_mean': 0.45549121499061584, 'reward_total_composite_std': 0.09961428493261337, 'epoch': 0.01}

  3%|▎         | 96/3300 [13:16<6:28:18,  7.27s/it]INFO 04-13 07:34:14 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 97/3300 [13:24<6:31:29,  7.33s/it]2026-04-13 07:34:22,091 | INFO | train_grpo_train | metrics_logged mode=train step=97

                                                   
{'loss': 0.1103, 'grad_norm': 17.4005126953125, 'learning_rate': 9.70909090909091e-06, 'num_tokens': 179656.0, 'completions/mean_length': 43.75, 'completions/min_length': 38.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4310287833213806, 'rewards/meter/std': 0.37155377864837646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9705358147621155, 'rewards/repeat_soft/std': 0.03972465544939041, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.4592880606651306, 'rewards/total_composite/std': 0.10607890784740448, 'reward': 0.4592880606651306, 'reward_std': 0.10607889294624329, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22239214181900024, 'sampling/sampling_logp_difference/max': 1.114335536956787, 'sampling/importance_sampling_ratio/min': 0.328133225440979, 'sampling/importance_sampling_ratio/mean': 1.0250064134597778, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.259575054049492, 'clip_ratio/low_mean': 0.11555228382349014, 'clip_ratio/low_min': 0.11555228382349014, 'clip_ratio/high_mean': 0.06651191785931587, 'clip_ratio/high_max': 0.06651191785931587, 'clip_ratio/region_mean': 0.18206420168280602, 'reward_total_mean': 0.4592880606651306, 'reward_meter_mean': 0.4310287833213806, 'reward_meter_std': 0.37155377864837646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9705358147621155, 'reward_repeat_soft_std': 0.03972465544939041, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.4592880606651306, 'reward_total_composite_std': 0.10607890784740448, 'epoch': 0.01}

  3%|▎         | 97/3300 [13:24<6:31:29,  7.33s/it]INFO 04-13 07:34:22 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 98/3300 [13:32<6:43:25,  7.56s/it]2026-04-13 07:34:30,179 | INFO | train_grpo_train | metrics_logged mode=train step=98

                                                   
{'loss': 0.183, 'grad_norm': 18.236595153808594, 'learning_rate': 9.706060606060606e-06, 'num_tokens': 181422.0, 'completions/mean_length': 56.75, 'completions/min_length': 42.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.43153274059295654, 'rewards/meter/std': 0.4027426838874817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9987303614616394, 'rewards/repeat_soft/std': 0.0030256821773946285, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.19255799055099487, 'rewards/total_composite/mean': 0.4896277189254761, 'rewards/total_composite/std': 0.12527494132518768, 'reward': 0.4896277189254761, 'reward_std': 0.12527494132518768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19202686846256256, 'sampling/sampling_logp_difference/max': 1.3069257736206055, 'sampling/importance_sampling_ratio/min': 0.27065083384513855, 'sampling/importance_sampling_ratio/mean': 1.0375019311904907, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.928257331252098, 'clip_ratio/low_mean': 0.08768315427005291, 'clip_ratio/low_min': 0.08768315427005291, 'clip_ratio/high_mean': 0.08970235660672188, 'clip_ratio/high_max': 0.08970235660672188, 'clip_ratio/region_mean': 0.1773855108767748, 'reward_total_mean': 0.4896277189254761, 'reward_meter_mean': 0.43153274059295654, 'reward_meter_std': 0.4027426838874817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9987303614616394, 'reward_repeat_soft_std': 0.0030256821773946285, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.19255799055099487, 'reward_total_composite_mean': 0.4896277189254761, 'reward_total_composite_std': 0.12527494132518768, 'epoch': 0.01}

  3%|▎         | 98/3300 [13:32<6:43:25,  7.56s/it]INFO 04-13 07:34:30 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 99/3300 [13:38<6:26:20,  7.24s/it]2026-04-13 07:34:36,679 | INFO | train_grpo_train | metrics_logged mode=train step=99

                                                   
{'loss': -0.0123, 'grad_norm': 20.899450302124023, 'learning_rate': 9.703030303030305e-06, 'num_tokens': 182801.0, 'completions/mean_length': 28.375, 'completions/min_length': 19.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.24775895476341248, 'rewards/meter/std': 0.35025811195373535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9559658765792847, 'rewards/repeat_soft/std': 0.01644347980618477, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.3981969356536865, 'rewards/total_composite/std': 0.06609706580638885, 'reward': 0.3981969356536865, 'reward_std': 0.06609705835580826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17724387347698212, 'sampling/sampling_logp_difference/max': 1.2565131187438965, 'sampling/importance_sampling_ratio/min': 0.28464481234550476, 'sampling/importance_sampling_ratio/mean': 1.0068453550338745, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.540700025856495, 'clip_ratio/low_mean': 0.08678907621651888, 'clip_ratio/low_min': 0.08678907621651888, 'clip_ratio/high_mean': 0.061506545171141624, 'clip_ratio/high_max': 0.061506545171141624, 'clip_ratio/region_mean': 0.1482956213876605, 'reward_total_mean': 0.3981969356536865, 'reward_meter_mean': 0.24775895476341248, 'reward_meter_std': 0.35025811195373535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9559658765792847, 'reward_repeat_soft_std': 0.01644347980618477, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.3981969356536865, 'reward_total_composite_std': 0.06609706580638885, 'epoch': 0.01}

  3%|▎         | 99/3300 [13:38<6:26:20,  7.24s/it]INFO 04-13 07:34:36 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 100/3300 [13:46<6:28:10,  7.28s/it]2026-04-13 07:34:44,044 | INFO | train_grpo_train | metrics_logged mode=train step=100

                                                    
{'loss': 0.1025, 'grad_norm': 15.105440139770508, 'learning_rate': 9.7e-06, 'num_tokens': 184736.0, 'completions/mean_length': 67.875, 'completions/min_length': 54.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.5348031520843506, 'rewards/meter/std': 0.3817266523838043, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9921558499336243, 'rewards/repeat_soft/std': 0.0050133876502513885, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1440981924533844, 'rewards/total_composite/mean': 0.4639267325401306, 'rewards/total_composite/std': 0.20549696683883667, 'reward': 0.4639267325401306, 'reward_std': 0.20549695193767548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21573399007320404, 'sampling/sampling_logp_difference/max': 1.7488975524902344, 'sampling/importance_sampling_ratio/min': 0.17396561801433563, 'sampling/importance_sampling_ratio/mean': 1.0013856887817383, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9593983590602875, 'clip_ratio/low_mean': 0.06655851006507874, 'clip_ratio/low_min': 0.06655851006507874, 'clip_ratio/high_mean': 0.13855982199311256, 'clip_ratio/high_max': 0.13855982199311256, 'clip_ratio/region_mean': 0.2051183320581913, 'reward_total_mean': 0.4639267325401306, 'reward_meter_mean': 0.5348031520843506, 'reward_meter_std': 0.3817266523838043, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9921558499336243, 'reward_repeat_soft_std': 0.0050133876502513885, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1440981924533844, 'reward_total_composite_mean': 0.4639267325401306, 'reward_total_composite_std': 0.20549696683883667, 'epoch': 0.01}

  3%|▎         | 100/3300 [13:46<6:28:10,  7.28s/it]INFO 04-13 07:34:44 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:05<00:20,  2.54s/it]

 30%|███       | 3/10 [00:10<00:25,  3.60s/it]

 40%|████      | 4/10 [00:14<00:22,  3.78s/it]

 50%|█████     | 5/10 [00:20<00:22,  4.60s/it]

 60%|██████    | 6/10 [00:26<00:19,  4.97s/it]

 70%|███████   | 7/10 [00:31<00:15,  5.09s/it]

 80%|████████  | 8/10 [00:35<00:09,  4.87s/it]

 90%|█████████ | 9/10 [00:40<00:04,  4.94s/it]

100%|██████████| 10/10 [00:45<00:00,  4.77s/it]2026-04-13 07:35:34,748 | INFO | train_grpo_train | metrics_logged mode=eval step=100

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 50.6952, 'eval_samples_per_second': 1.578, 'eval_steps_per_second': 0.197, 'eval_num_tokens': 184736.0, 'eval_completions/mean_length': 79.6375, 'eval_completions/min_length': 32.0, 'eval_completions/max_length': 143.1, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 79.6375, 'eval_completions/min_terminated_length': 32.0, 'eval_completions/max_terminated_length': 143.1, 'eval_rewards/meter/mean': 0.4947461932897568, 'eval_rewards/meter/std': 0.3523126423358917, 'eval_rewards/count_adherence/mean': 0.9681249976158142, 'eval_rewards/count_adherence/std': 0.09015611484646797, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.1767766922712326, 'eval_rewards/repeat_soft/mean': 0.991607666015625, 'eval_rewards/repeat_soft/std': 0.011974084354005755, 'eval_rewards/judge_quality/mean': 0.47200000286102295, 'eval_rewards/judge_quality/std': 0.1661804124712944, 'eval_rewards/total_composite/mean': 0.471698522567749, 'eval_rewards/total_composite/std': 0.17879592925310134, 'eval_reward': 0.471698522567749, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.14445659518241882, 'eval_sampling/sampling_logp_difference/max': 1.225595474243164, 'eval_sampling/importance_sampling_ratio/min': 0.2977974995970726, 'eval_sampling/importance_sampling_ratio/mean': 1.0436278223991393, 'eval_sampling/importance_sampling_ratio/max': 1.5806872248649597, 'eval_entropy': 2.184915065765381, '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': 0.471698522567749, 'eval_reward_meter_mean': 0.4947461932897568, 'eval_reward_meter_std': 0.3523126423358917, 'eval_reward_count_adherence_mean': 0.9681249976158142, 'eval_reward_count_adherence_std': 0.09015611484646797, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.1767766922712326, 'eval_reward_repeat_soft_mean': 0.991607666015625, 'eval_reward_repeat_soft_std': 0.011974084354005755, 'eval_reward_judge_quality_mean': 0.47200000286102295, 'eval_reward_judge_quality_std': 0.1661804124712944, 'eval_reward_total_composite_mean': 0.471698522567749, 'eval_reward_total_composite_std': 0.17879592925310134, 'epoch': 0.01}

  3%|▎         | 100/3300 [14:36<6:28:10,  7.28s/it]

100%|██████████| 10/10 [00:45<00:00,  4.77s/it]

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

  3%|▎         | 101/3300 [14:48<21:00:09, 23.64s/it]2026-04-13 07:35:45,845 | INFO | train_grpo_train | metrics_logged mode=train step=101

                                                     
{'loss': 0.0118, 'grad_norm': 16.414012908935547, 'learning_rate': 9.696969696969698e-06, 'num_tokens': 186614.0, 'completions/mean_length': 58.75, 'completions/min_length': 45.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.31307607889175415, 'rewards/meter/std': 0.37023577094078064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9960094690322876, 'rewards/repeat_soft/std': 0.003800690406933427, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.1566559225320816, 'rewards/total_composite/mean': 0.41067010164260864, 'rewards/total_composite/std': 0.2034657597541809, 'reward': 0.41067010164260864, 'reward_std': 0.2034657597541809, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23996204137802124, 'sampling/sampling_logp_difference/max': 1.938732624053955, 'sampling/importance_sampling_ratio/min': 0.17366401851177216, 'sampling/importance_sampling_ratio/mean': 1.0507534742355347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.916314333677292, 'clip_ratio/low_mean': 0.09495464991778135, 'clip_ratio/low_min': 0.09495464991778135, 'clip_ratio/high_mean': 0.1007721908390522, 'clip_ratio/high_max': 0.1007721908390522, 'clip_ratio/region_mean': 0.19572684075683355, 'reward_total_mean': 0.41067010164260864, 'reward_meter_mean': 0.31307607889175415, 'reward_meter_std': 0.37023577094078064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9960094690322876, 'reward_repeat_soft_std': 0.003800690406933427, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.1566559225320816, 'reward_total_composite_mean': 0.41067010164260864, 'reward_total_composite_std': 0.2034657597541809, 'epoch': 0.01}

  3%|▎         | 101/3300 [14:48<21:00:09, 23.64s/it]INFO 04-13 07:35:46 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 102/3300 [14:56<16:49:32, 18.94s/it]2026-04-13 07:35:53,832 | INFO | train_grpo_train | metrics_logged mode=train step=102

                                                     
{'loss': 0.0199, 'grad_norm': 13.145352363586426, 'learning_rate': 9.693939393939395e-06, 'num_tokens': 188740.0, 'completions/mean_length': 86.75, 'completions/min_length': 79.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.75, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.32552945613861084, 'rewards/meter/std': 0.29995009303092957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9968774914741516, 'rewards/repeat_soft/std': 0.002367391949519515, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43840116262435913, 'rewards/total_composite/std': 0.0820188969373703, 'reward': 0.43840116262435913, 'reward_std': 0.08201887458562851, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21868892014026642, 'sampling/sampling_logp_difference/max': 2.1058950424194336, 'sampling/importance_sampling_ratio/min': 0.12173666059970856, 'sampling/importance_sampling_ratio/mean': 1.0365042686462402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.044892191886902, 'clip_ratio/low_mean': 0.11399291269481182, 'clip_ratio/low_min': 0.11399291269481182, 'clip_ratio/high_mean': 0.07613622397184372, 'clip_ratio/high_max': 0.07613622397184372, 'clip_ratio/region_mean': 0.19012913666665554, 'reward_total_mean': 0.43840116262435913, 'reward_meter_mean': 0.32552945613861084, 'reward_meter_std': 0.29995009303092957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9968774914741516, 'reward_repeat_soft_std': 0.002367391949519515, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43840116262435913, 'reward_total_composite_std': 0.0820188969373703, 'epoch': 0.01}

  3%|▎         | 102/3300 [14:56<16:49:32, 18.94s/it]INFO 04-13 07:35:54 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 103/3300 [15:03<13:46:25, 15.51s/it]2026-04-13 07:36:01,337 | INFO | train_grpo_train | metrics_logged mode=train step=103

                                                     
{'loss': 0.0483, 'grad_norm': 13.322657585144043, 'learning_rate': 9.690909090909092e-06, 'num_tokens': 190428.0, 'completions/mean_length': 54.0, 'completions/min_length': 41.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.4907858967781067, 'rewards/meter/std': 0.4426969289779663, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974976778030396, 'rewards/repeat_soft/std': 0.0037490674294531345, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404788017273, 'rewards/total_composite/mean': 0.5294002294540405, 'rewards/total_composite/std': 0.14981630444526672, 'reward': 0.5294002294540405, 'reward_std': 0.14981631934642792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1901864856481552, 'sampling/sampling_logp_difference/max': 1.7494490146636963, 'sampling/importance_sampling_ratio/min': 0.17386971414089203, 'sampling/importance_sampling_ratio/mean': 1.0129625797271729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3739588409662247, 'clip_ratio/low_mean': 0.06334241945296526, 'clip_ratio/low_min': 0.06334241945296526, 'clip_ratio/high_mean': 0.09503825195133686, 'clip_ratio/high_max': 0.09503825195133686, 'clip_ratio/region_mean': 0.15838067140430212, 'reward_total_mean': 0.5294002294540405, 'reward_meter_mean': 0.4907858967781067, 'reward_meter_std': 0.4426969289779663, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974976778030396, 'reward_repeat_soft_std': 0.0037490674294531345, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404788017273, 'reward_total_composite_mean': 0.5294002294540405, 'reward_total_composite_std': 0.14981630444526672, 'epoch': 0.01}

  3%|▎         | 103/3300 [15:03<13:46:25, 15.51s/it]INFO 04-13 07:36:01 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 104/3300 [15:10<11:35:07, 13.05s/it]2026-04-13 07:36:08,645 | INFO | train_grpo_train | metrics_logged mode=train step=104

                                                     
{'loss': 0.067, 'grad_norm': 15.824663162231445, 'learning_rate': 9.687878787878788e-06, 'num_tokens': 192042.0, 'completions/mean_length': 47.75, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.4273160696029663, 'rewards/meter/std': 0.3913634121417999, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948374032974243, 'rewards/repeat_soft/std': 0.011940538883209229, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2418234497308731, 'rewards/total_composite/mean': 0.5368877649307251, 'rewards/total_composite/std': 0.18454858660697937, 'reward': 0.5368877649307251, 'reward_std': 0.18454860150814056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2321016639471054, 'sampling/sampling_logp_difference/max': 2.423257827758789, 'sampling/importance_sampling_ratio/min': 0.08863239735364914, 'sampling/importance_sampling_ratio/mean': 1.0372896194458008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.88126140832901, 'clip_ratio/low_mean': 0.10984764993190765, 'clip_ratio/low_min': 0.10984764993190765, 'clip_ratio/high_mean': 0.08070892840623856, 'clip_ratio/high_max': 0.08070892840623856, 'clip_ratio/region_mean': 0.1905565783381462, 'reward_total_mean': 0.5368877649307251, 'reward_meter_mean': 0.4273160696029663, 'reward_meter_std': 0.3913634121417999, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948374032974243, 'reward_repeat_soft_std': 0.011940538883209229, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2418234497308731, 'reward_total_composite_mean': 0.5368877649307251, 'reward_total_composite_std': 0.18454858660697937, 'epoch': 0.01}

  3%|▎         | 104/3300 [15:10<11:35:07, 13.05s/it]INFO 04-13 07:36:08 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 105/3300 [15:18<10:02:37, 11.32s/it]2026-04-13 07:36:15,919 | INFO | train_grpo_train | metrics_logged mode=train step=105

                                                     
{'loss': -0.0929, 'grad_norm': 10.761244773864746, 'learning_rate': 9.684848484848487e-06, 'num_tokens': 194281.0, 'completions/mean_length': 95.875, 'completions/min_length': 68.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.3006367087364197, 'rewards/meter/std': 0.3225688934326172, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9966142773628235, 'rewards/repeat_soft/std': 0.004816299770027399, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.14574319124221802, 'rewards/total_composite/mean': 0.45476049184799194, 'rewards/total_composite/std': 0.15194669365882874, 'reward': 0.45476049184799194, 'reward_std': 0.15194669365882874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2330923229455948, 'sampling/sampling_logp_difference/max': 1.461348533630371, 'sampling/importance_sampling_ratio/min': 0.23192331194877625, 'sampling/importance_sampling_ratio/mean': 1.0213702917099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5699221789836884, 'clip_ratio/low_mean': 0.17626286298036575, 'clip_ratio/low_min': 0.17626286298036575, 'clip_ratio/high_mean': 0.048399388790130615, 'clip_ratio/high_max': 0.048399388790130615, 'clip_ratio/region_mean': 0.22466225177049637, 'reward_total_mean': 0.45476049184799194, 'reward_meter_mean': 0.3006367087364197, 'reward_meter_std': 0.3225688934326172, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9966142773628235, 'reward_repeat_soft_std': 0.004816299770027399, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.14574319124221802, 'reward_total_composite_mean': 0.45476049184799194, 'reward_total_composite_std': 0.15194669365882874, 'epoch': 0.01}

  3%|▎         | 105/3300 [15:18<10:02:37, 11.32s/it]INFO 04-13 07:36:16 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 106/3300 [15:25<8:57:40, 10.10s/it] 2026-04-13 07:36:23,181 | INFO | train_grpo_train | metrics_logged mode=train step=106

                                                    
{'loss': 0.045, 'grad_norm': 32.41508102416992, 'learning_rate': 9.681818181818182e-06, 'num_tokens': 195731.0, 'completions/mean_length': 29.25, 'completions/min_length': 26.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.41725194454193115, 'rewards/meter/std': 0.2775874435901642, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9898457527160645, 'rewards/repeat_soft/std': 0.005739502143114805, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.1954299360513687, 'rewards/total_composite/mean': 0.5062490105628967, 'rewards/total_composite/std': 0.13873079419136047, 'reward': 0.5062490105628967, 'reward_std': 0.13873079419136047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19110196828842163, 'sampling/sampling_logp_difference/max': 1.9213972091674805, 'sampling/importance_sampling_ratio/min': 0.1464022547006607, 'sampling/importance_sampling_ratio/mean': 1.0243908166885376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9342062026262283, 'clip_ratio/low_mean': 0.1261307019740343, 'clip_ratio/low_min': 0.1261307019740343, 'clip_ratio/high_mean': 0.04126082360744476, 'clip_ratio/high_max': 0.04126082360744476, 'clip_ratio/region_mean': 0.16739152558147907, 'reward_total_mean': 0.5062490105628967, 'reward_meter_mean': 0.41725194454193115, 'reward_meter_std': 0.2775874435901642, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9898457527160645, 'reward_repeat_soft_std': 0.005739502143114805, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.1954299360513687, 'reward_total_composite_mean': 0.5062490105628967, 'reward_total_composite_std': 0.13873079419136047, 'epoch': 0.01}

  3%|▎         | 106/3300 [15:25<8:57:40, 10.10s/it]INFO 04-13 07:36:23 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 107/3300 [15:33<8:21:01,  9.41s/it]2026-04-13 07:36:30,996 | INFO | train_grpo_train | metrics_logged mode=train step=107

                                                    
{'loss': 0.2057, 'grad_norm': 20.082307815551758, 'learning_rate': 9.67878787878788e-06, 'num_tokens': 197443.0, 'completions/mean_length': 51.0, 'completions/min_length': 31.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7254915833473206, 'rewards/meter/std': 0.40892869234085083, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9917322993278503, 'rewards/repeat_soft/std': 0.010131089948117733, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.7275059223175049, 'rewards/total_composite/std': 0.27270805835723877, 'reward': 0.7275059223175049, 'reward_std': 0.27270805835723877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17575828731060028, 'sampling/sampling_logp_difference/max': 1.2130756378173828, 'sampling/importance_sampling_ratio/min': 0.29728156328201294, 'sampling/importance_sampling_ratio/mean': 1.0152204036712646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3037724867463112, 'clip_ratio/low_mean': 0.04798919567838311, 'clip_ratio/low_min': 0.04798919567838311, 'clip_ratio/high_mean': 0.12696912698447704, 'clip_ratio/high_max': 0.12696912698447704, 'clip_ratio/region_mean': 0.17495832266286016, 'reward_total_mean': 0.7275059223175049, 'reward_meter_mean': 0.7254915833473206, 'reward_meter_std': 0.40892869234085083, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9917322993278503, 'reward_repeat_soft_std': 0.010131089948117733, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.7275059223175049, 'reward_total_composite_std': 0.27270805835723877, 'epoch': 0.01}

  3%|▎         | 107/3300 [15:33<8:21:01,  9.41s/it]INFO 04-13 07:36:31 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 108/3300 [15:40<7:50:33,  8.85s/it]2026-04-13 07:36:38,512 | INFO | train_grpo_train | metrics_logged mode=train step=108

                                                    
{'loss': -0.0728, 'grad_norm': 31.107200622558594, 'learning_rate': 9.675757575757577e-06, 'num_tokens': 198743.0, 'completions/mean_length': 27.5, 'completions/min_length': 21.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.17419485747814178, 'rewards/meter/std': 0.17303234338760376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976261854171753, 'rewards/repeat_soft/std': 0.003726511262357235, 'rewards/judge_quality/mean': 0.5350000262260437, 'rewards/judge_quality/std': 0.1911618709564209, 'rewards/total_composite/mean': 0.40953513979911804, 'rewards/total_composite/std': 0.061801183968782425, 'reward': 0.40953513979911804, 'reward_std': 0.061801180243492126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.31676381826400757, 'sampling/sampling_logp_difference/max': 2.9225621223449707, 'sampling/importance_sampling_ratio/min': 0.05379568040370941, 'sampling/importance_sampling_ratio/mean': 1.0342049598693848, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0506207197904587, 'clip_ratio/low_mean': 0.15618032403290272, 'clip_ratio/low_min': 0.15618032403290272, 'clip_ratio/high_mean': 0.10346023179590702, 'clip_ratio/high_max': 0.10346023179590702, 'clip_ratio/region_mean': 0.25964055582880974, 'reward_total_mean': 0.40953513979911804, 'reward_meter_mean': 0.17419485747814178, 'reward_meter_std': 0.17303234338760376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976261854171753, 'reward_repeat_soft_std': 0.003726511262357235, 'reward_judge_quality_mean': 0.5350000262260437, 'reward_judge_quality_std': 0.1911618709564209, 'reward_total_composite_mean': 0.40953513979911804, 'reward_total_composite_std': 0.061801183968782425, 'epoch': 0.01}

  3%|▎         | 108/3300 [15:40<7:50:33,  8.85s/it]INFO 04-13 07:36:38 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 109/3300 [15:47<7:15:03,  8.18s/it]2026-04-13 07:36:45,141 | INFO | train_grpo_train | metrics_logged mode=train step=109

                                                    
{'loss': 0.0143, 'grad_norm': 13.207579612731934, 'learning_rate': 9.672727272727274e-06, 'num_tokens': 200420.0, 'completions/mean_length': 49.625, 'completions/min_length': 33.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9199807643890381, 'rewards/meter/std': 0.18390671908855438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9963053464889526, 'rewards/repeat_soft/std': 0.004458059091120958, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5862847566604614, 'rewards/total_composite/std': 0.26184654235839844, 'reward': 0.5862847566604614, 'reward_std': 0.26184654235839844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22561202943325043, 'sampling/sampling_logp_difference/max': 1.3455278873443604, 'sampling/importance_sampling_ratio/min': 0.26040223240852356, 'sampling/importance_sampling_ratio/mean': 1.032443881034851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4901363253593445, 'clip_ratio/low_mean': 0.02556818164885044, 'clip_ratio/low_min': 0.02556818164885044, 'clip_ratio/high_mean': 0.20468470267951488, 'clip_ratio/high_max': 0.20468470267951488, 'clip_ratio/region_mean': 0.23025288432836533, 'reward_total_mean': 0.5862847566604614, 'reward_meter_mean': 0.9199807643890381, 'reward_meter_std': 0.18390671908855438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9963053464889526, 'reward_repeat_soft_std': 0.004458059091120958, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5862847566604614, 'reward_total_composite_std': 0.26184654235839844, 'epoch': 0.01}

  3%|▎         | 109/3300 [15:47<7:15:03,  8.18s/it]INFO 04-13 07:36:45 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 110/3300 [15:54<6:51:43,  7.74s/it]2026-04-13 07:36:51,868 | INFO | train_grpo_train | metrics_logged mode=train step=110

                                                    
{'loss': 0.0721, 'grad_norm': 13.896294593811035, 'learning_rate': 9.66969696969697e-06, 'num_tokens': 202156.0, 'completions/mean_length': 48.0, 'completions/min_length': 37.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.892151951789856, 'rewards/meter/std': 0.26865240931510925, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9856277704238892, 'rewards/repeat_soft/std': 0.017225177958607674, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.220891073346138, 'rewards/total_composite/mean': 0.6859848499298096, 'rewards/total_composite/std': 0.16468684375286102, 'reward': 0.6859848499298096, 'reward_std': 0.16468684375286102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2160102128982544, 'sampling/sampling_logp_difference/max': 1.8444128036499023, 'sampling/importance_sampling_ratio/min': 0.1581181436777115, 'sampling/importance_sampling_ratio/mean': 1.036145806312561, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3849086612462997, 'clip_ratio/low_mean': 0.11906185001134872, 'clip_ratio/low_min': 0.11906185001134872, 'clip_ratio/high_mean': 0.048634571954607964, 'clip_ratio/high_max': 0.048634571954607964, 'clip_ratio/region_mean': 0.1676964219659567, 'reward_total_mean': 0.6859848499298096, 'reward_meter_mean': 0.892151951789856, 'reward_meter_std': 0.26865240931510925, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9856277704238892, 'reward_repeat_soft_std': 0.017225177958607674, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.220891073346138, 'reward_total_composite_mean': 0.6859848499298096, 'reward_total_composite_std': 0.16468684375286102, 'epoch': 0.01}

  3%|▎         | 110/3300 [15:54<6:51:43,  7.74s/it]INFO 04-13 07:36:52 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 111/3300 [16:01<6:47:33,  7.67s/it]2026-04-13 07:36:59,357 | INFO | train_grpo_train | metrics_logged mode=train step=111

                                                    
{'loss': -0.0285, 'grad_norm': 10.08155632019043, 'learning_rate': 9.666666666666667e-06, 'num_tokens': 204200.0, 'completions/mean_length': 91.5, 'completions/min_length': 69.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.6964003443717957, 'rewards/meter/std': 0.3417958915233612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986571073532104, 'rewards/repeat_soft/std': 0.0013369194930419326, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/total_composite/mean': 0.5833487510681152, 'rewards/total_composite/std': 0.14542630314826965, 'reward': 0.5833487510681152, 'reward_std': 0.14542630314826965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22398188710212708, 'sampling/sampling_logp_difference/max': 1.491401195526123, 'sampling/importance_sampling_ratio/min': 0.22505709528923035, 'sampling/importance_sampling_ratio/mean': 1.0558103322982788, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0628055930137634, 'clip_ratio/low_mean': 0.09790320880711079, 'clip_ratio/low_min': 0.09790320880711079, 'clip_ratio/high_mean': 0.12199797295033932, 'clip_ratio/high_max': 0.12199797295033932, 'clip_ratio/region_mean': 0.2199011817574501, 'reward_total_mean': 0.5833487510681152, 'reward_meter_mean': 0.6964003443717957, 'reward_meter_std': 0.3417958915233612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986571073532104, 'reward_repeat_soft_std': 0.0013369194930419326, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.5833487510681152, 'reward_total_composite_std': 0.14542630314826965, 'epoch': 0.01}

  3%|▎         | 111/3300 [16:01<6:47:33,  7.67s/it]INFO 04-13 07:36:59 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 112/3300 [16:08<6:37:06,  7.47s/it]2026-04-13 07:37:06,378 | INFO | train_grpo_train | metrics_logged mode=train step=112

                                                    
{'loss': 0.1087, 'grad_norm': 20.6037540435791, 'learning_rate': 9.663636363636364e-06, 'num_tokens': 205731.0, 'completions/mean_length': 26.375, 'completions/min_length': 16.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5949082970619202, 'rewards/meter/std': 0.494957834482193, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9459680318832397, 'rewards/repeat_soft/std': 0.04101009666919708, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.22258225083351135, 'rewards/total_composite/mean': 0.5486740469932556, 'rewards/total_composite/std': 0.20209264755249023, 'reward': 0.5486740469932556, 'reward_std': 0.20209261775016785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18613271415233612, 'sampling/sampling_logp_difference/max': 1.6638908386230469, 'sampling/importance_sampling_ratio/min': 0.18940061330795288, 'sampling/importance_sampling_ratio/mean': 1.0181546211242676, 'sampling/importance_sampling_ratio/max': 1.8430614471435547, 'entropy': 1.4548956751823425, 'clip_ratio/low_mean': 0.06781775318086147, 'clip_ratio/low_min': 0.06781775318086147, 'clip_ratio/high_mean': 0.10600329004228115, 'clip_ratio/high_max': 0.10600329004228115, 'clip_ratio/region_mean': 0.17382104322314262, 'reward_total_mean': 0.5486740469932556, 'reward_meter_mean': 0.5949082970619202, 'reward_meter_std': 0.494957834482193, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9459680318832397, 'reward_repeat_soft_std': 0.04101009666919708, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.22258225083351135, 'reward_total_composite_mean': 0.5486740469932556, 'reward_total_composite_std': 0.20209264755249023, 'epoch': 0.01}

  3%|▎         | 112/3300 [16:08<6:37:06,  7.47s/it]INFO 04-13 07:37:06 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 113/3300 [16:16<6:39:12,  7.52s/it]2026-04-13 07:37:13,992 | INFO | train_grpo_train | metrics_logged mode=train step=113

                                                    
{'loss': 0.0605, 'grad_norm': 11.709718704223633, 'learning_rate': 9.660606060606061e-06, 'num_tokens': 207592.0, 'completions/mean_length': 76.625, 'completions/min_length': 64.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.5275776386260986, 'rewards/meter/std': 0.32127389311790466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9927095770835876, 'rewards/repeat_soft/std': 0.007404064293950796, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5425221920013428, 'rewards/total_composite/std': 0.16789241135120392, 'reward': 0.5425221920013428, 'reward_std': 0.16789241135120392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2142861932516098, 'sampling/sampling_logp_difference/max': 2.0963268280029297, 'sampling/importance_sampling_ratio/min': 0.12290704995393753, 'sampling/importance_sampling_ratio/mean': 1.0310522317886353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.829342469573021, 'clip_ratio/low_mean': 0.09125383384525776, 'clip_ratio/low_min': 0.09125383384525776, 'clip_ratio/high_mean': 0.07866137474775314, 'clip_ratio/high_max': 0.07866137474775314, 'clip_ratio/region_mean': 0.1699152085930109, 'reward_total_mean': 0.5425221920013428, 'reward_meter_mean': 0.5275776386260986, 'reward_meter_std': 0.32127389311790466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9927095770835876, 'reward_repeat_soft_std': 0.007404064293950796, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5425221920013428, 'reward_total_composite_std': 0.16789241135120392, 'epoch': 0.01}

  3%|▎         | 113/3300 [16:16<6:39:12,  7.52s/it]INFO 04-13 07:37:14 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 114/3300 [16:22<6:25:58,  7.27s/it]2026-04-13 07:37:20,686 | INFO | train_grpo_train | metrics_logged mode=train step=114

                                                    
{'loss': -0.0008, 'grad_norm': 15.91455078125, 'learning_rate': 9.657575757575758e-06, 'num_tokens': 209166.0, 'completions/mean_length': 42.75, 'completions/min_length': 32.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.41170966625213623, 'rewards/meter/std': 0.39287254214286804, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9963371753692627, 'rewards/repeat_soft/std': 0.0069928220473229885, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.3950405418872833, 'rewards/total_composite/std': 0.20517849922180176, 'reward': 0.3950405418872833, 'reward_std': 0.20517848432064056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23370611667633057, 'sampling/sampling_logp_difference/max': 1.5829181671142578, 'sampling/importance_sampling_ratio/min': 0.20537489652633667, 'sampling/importance_sampling_ratio/mean': 1.0398354530334473, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.164140671491623, 'clip_ratio/low_mean': 0.11847196333110332, 'clip_ratio/low_min': 0.11847196333110332, 'clip_ratio/high_mean': 0.08506148867309093, 'clip_ratio/high_max': 0.08506148867309093, 'clip_ratio/region_mean': 0.20353345200419426, 'reward_total_mean': 0.3950405418872833, 'reward_meter_mean': 0.41170966625213623, 'reward_meter_std': 0.39287254214286804, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9963371753692627, 'reward_repeat_soft_std': 0.0069928220473229885, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.3950405418872833, 'reward_total_composite_std': 0.20517849922180176, 'epoch': 0.01}

  3%|▎         | 114/3300 [16:22<6:25:58,  7.27s/it]INFO 04-13 07:37:20 [block_pool.py:378] Successfully reset prefix cache

  3%|▎         | 115/3300 [16:30<6:38:06,  7.50s/it]2026-04-13 07:37:28,723 | INFO | train_grpo_train | metrics_logged mode=train step=115

                                                    
{'loss': -0.1652, 'grad_norm': 10.372628211975098, 'learning_rate': 9.654545454545456e-06, 'num_tokens': 211540.0, 'completions/mean_length': 108.75, 'completions/min_length': 62.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.3633410930633545, 'rewards/meter/std': 0.29584023356437683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.997858464717865, 'rewards/repeat_soft/std': 0.0016820140881463885, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.41038256883621216, 'rewards/total_composite/std': 0.1801348179578781, 'reward': 0.41038256883621216, 'reward_std': 0.1801348179578781, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2164894938468933, 'sampling/sampling_logp_difference/max': 1.768528938293457, 'sampling/importance_sampling_ratio/min': 0.17058373987674713, 'sampling/importance_sampling_ratio/mean': 1.0498042106628418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7530274242162704, 'clip_ratio/low_mean': 0.051655348390340805, 'clip_ratio/low_min': 0.051655348390340805, 'clip_ratio/high_mean': 0.16114997677505016, 'clip_ratio/high_max': 0.16114997677505016, 'clip_ratio/region_mean': 0.21280532516539097, 'reward_total_mean': 0.41038256883621216, 'reward_meter_mean': 0.3633410930633545, 'reward_meter_std': 0.29584023356437683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.997858464717865, 'reward_repeat_soft_std': 0.0016820140881463885, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.41038256883621216, 'reward_total_composite_std': 0.1801348179578781, 'epoch': 0.01}

  3%|▎         | 115/3300 [16:30<6:38:06,  7.50s/it]INFO 04-13 07:37:28 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 116/3300 [16:38<6:31:06,  7.37s/it]2026-04-13 07:37:35,791 | INFO | train_grpo_train | metrics_logged mode=train step=116

                                                    
{'loss': 0.0623, 'grad_norm': 15.952447891235352, 'learning_rate': 9.651515151515153e-06, 'num_tokens': 213129.0, 'completions/mean_length': 43.625, 'completions/min_length': 30.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8196585178375244, 'rewards/meter/std': 0.2894671559333801, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9935790300369263, 'rewards/repeat_soft/std': 0.007324355188757181, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.615305483341217, 'rewards/total_composite/std': 0.15007352828979492, 'reward': 0.615305483341217, 'reward_std': 0.15007352828979492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2312517613172531, 'sampling/sampling_logp_difference/max': 2.045539140701294, 'sampling/importance_sampling_ratio/min': 0.12931044399738312, 'sampling/importance_sampling_ratio/mean': 1.0233547687530518, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1754041761159897, 'clip_ratio/low_mean': 0.08337220642715693, 'clip_ratio/low_min': 0.08337220642715693, 'clip_ratio/high_mean': 0.12287236005067825, 'clip_ratio/high_max': 0.12287236005067825, 'clip_ratio/region_mean': 0.20624456647783518, 'reward_total_mean': 0.615305483341217, 'reward_meter_mean': 0.8196585178375244, 'reward_meter_std': 0.2894671559333801, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9935790300369263, 'reward_repeat_soft_std': 0.007324355188757181, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.615305483341217, 'reward_total_composite_std': 0.15007352828979492, 'epoch': 0.01}

  4%|▎         | 116/3300 [16:38<6:31:06,  7.37s/it]INFO 04-13 07:37:36 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 117/3300 [16:44<6:21:24,  7.19s/it]2026-04-13 07:37:42,560 | INFO | train_grpo_train | metrics_logged mode=train step=117

                                                    
{'loss': 0.0196, 'grad_norm': 11.291123390197754, 'learning_rate': 9.648484848484849e-06, 'num_tokens': 214683.0, 'completions/mean_length': 43.25, 'completions/min_length': 38.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9655974507331848, 'rewards/meter/std': 0.01152831595391035, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.989479660987854, 'rewards/repeat_soft/std': 0.006363891065120697, 'rewards/judge_quality/mean': 0.9099999666213989, 'rewards/judge_quality/std': 0.018516412004828453, 'rewards/total_composite/mean': 0.9195917248725891, 'rewards/total_composite/std': 0.014142941683530807, 'reward': 0.9195917248725891, 'reward_std': 0.014142945408821106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07092545926570892, 'sampling/sampling_logp_difference/max': 1.4284789562225342, 'sampling/importance_sampling_ratio/min': 0.3051346242427826, 'sampling/importance_sampling_ratio/mean': 1.006468653678894, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.338104959577322, 'clip_ratio/low_mean': 0.012290635146200657, 'clip_ratio/low_min': 0.012290635146200657, 'clip_ratio/high_mean': 0.02475078939460218, 'clip_ratio/high_max': 0.02475078939460218, 'clip_ratio/region_mean': 0.037041424540802836, 'reward_total_mean': 0.9195917248725891, 'reward_meter_mean': 0.9655974507331848, 'reward_meter_std': 0.01152831595391035, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.989479660987854, 'reward_repeat_soft_std': 0.006363891065120697, 'reward_judge_quality_mean': 0.9099999666213989, 'reward_judge_quality_std': 0.018516412004828453, 'reward_total_composite_mean': 0.9195917248725891, 'reward_total_composite_std': 0.014142941683530807, 'epoch': 0.01}

  4%|▎         | 117/3300 [16:44<6:21:24,  7.19s/it]INFO 04-13 07:37:42 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 118/3300 [16:52<6:36:37,  7.48s/it]2026-04-13 07:37:50,714 | INFO | train_grpo_train | metrics_logged mode=train step=118

                                                    
{'loss': -0.0076, 'grad_norm': 7.859068393707275, 'learning_rate': 9.645454545454548e-06, 'num_tokens': 217324.0, 'completions/mean_length': 130.125, 'completions/min_length': 86.0, 'completions/max_length': 184.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.125, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.5821857452392578, 'rewards/meter/std': 0.3516339659690857, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9981836080551147, 'rewards/repeat_soft/std': 0.0015714566688984632, 'rewards/judge_quality/mean': 0.42250001430511475, 'rewards/judge_quality/std': 0.14616528153419495, 'rewards/total_composite/mean': 0.3869349956512451, 'rewards/total_composite/std': 0.2658952474594116, 'reward': 0.3869349956512451, 'reward_std': 0.2658952474594116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23853881657123566, 'sampling/sampling_logp_difference/max': 1.860891342163086, 'sampling/importance_sampling_ratio/min': 0.15553393959999084, 'sampling/importance_sampling_ratio/mean': 1.0505948066711426, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.004988729953766, 'clip_ratio/low_mean': 0.03663211967796087, 'clip_ratio/low_min': 0.03663211967796087, 'clip_ratio/high_mean': 0.15321064181625843, 'clip_ratio/high_max': 0.15321064181625843, 'clip_ratio/region_mean': 0.1898427614942193, 'reward_total_mean': 0.3869349956512451, 'reward_meter_mean': 0.5821857452392578, 'reward_meter_std': 0.3516339659690857, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9981836080551147, 'reward_repeat_soft_std': 0.0015714566688984632, 'reward_judge_quality_mean': 0.42250001430511475, 'reward_judge_quality_std': 0.14616528153419495, 'reward_total_composite_mean': 0.3869349956512451, 'reward_total_composite_std': 0.2658952474594116, 'epoch': 0.01}

  4%|▎         | 118/3300 [16:52<6:36:37,  7.48s/it]INFO 04-13 07:37:50 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 119/3300 [17:00<6:34:16,  7.44s/it]2026-04-13 07:37:58,052 | INFO | train_grpo_train | metrics_logged mode=train step=119

                                                    
{'loss': 0.2217, 'grad_norm': 18.535816192626953, 'learning_rate': 9.642424242424243e-06, 'num_tokens': 219114.0, 'completions/mean_length': 57.75, 'completions/min_length': 35.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.403515487909317, 'rewards/meter/std': 0.4413924217224121, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9988487362861633, 'rewards/repeat_soft/std': 0.0016559003852307796, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.12772038578987122, 'rewards/total_composite/mean': 0.4148494601249695, 'rewards/total_composite/std': 0.202467143535614, 'reward': 0.4148494601249695, 'reward_std': 0.20246712863445282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21171781420707703, 'sampling/sampling_logp_difference/max': 1.7705450057983398, 'sampling/importance_sampling_ratio/min': 0.17024017870426178, 'sampling/importance_sampling_ratio/mean': 1.0579570531845093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.507441684603691, 'clip_ratio/low_mean': 0.09634115733206272, 'clip_ratio/low_min': 0.09634115733206272, 'clip_ratio/high_mean': 0.07929547782987356, 'clip_ratio/high_max': 0.07929547782987356, 'clip_ratio/region_mean': 0.17563663516193628, 'reward_total_mean': 0.4148494601249695, 'reward_meter_mean': 0.403515487909317, 'reward_meter_std': 0.4413924217224121, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9988487362861633, 'reward_repeat_soft_std': 0.0016559003852307796, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.12772038578987122, 'reward_total_composite_mean': 0.4148494601249695, 'reward_total_composite_std': 0.202467143535614, 'epoch': 0.01}

  4%|▎         | 119/3300 [17:00<6:34:16,  7.44s/it]INFO 04-13 07:37:58 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 120/3300 [17:07<6:38:11,  7.51s/it]2026-04-13 07:38:05,742 | INFO | train_grpo_train | metrics_logged mode=train step=120

                                                    
{'loss': 0.0885, 'grad_norm': 11.159661293029785, 'learning_rate': 9.63939393939394e-06, 'num_tokens': 221565.0, 'completions/mean_length': 112.375, 'completions/min_length': 63.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.4886738955974579, 'rewards/meter/std': 0.4274240732192993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9982452988624573, 'rewards/repeat_soft/std': 0.0026251429226249456, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4151157736778259, 'rewards/total_composite/std': 0.20270900428295135, 'reward': 0.4151157736778259, 'reward_std': 0.20270900428295135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2313220500946045, 'sampling/sampling_logp_difference/max': 1.854095458984375, 'sampling/importance_sampling_ratio/min': 0.15659452974796295, 'sampling/importance_sampling_ratio/mean': 1.014096736907959, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4326031506061554, 'clip_ratio/low_mean': 0.11750435456633568, 'clip_ratio/low_min': 0.11750435456633568, 'clip_ratio/high_mean': 0.1027077529579401, 'clip_ratio/high_max': 0.1027077529579401, 'clip_ratio/region_mean': 0.22021210752427578, 'reward_total_mean': 0.4151157736778259, 'reward_meter_mean': 0.4886738955974579, 'reward_meter_std': 0.4274240732192993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9982452988624573, 'reward_repeat_soft_std': 0.0026251429226249456, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4151157736778259, 'reward_total_composite_std': 0.20270900428295135, 'epoch': 0.01}

  4%|▎         | 120/3300 [17:07<6:38:11,  7.51s/it]INFO 04-13 07:38:05 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 121/3300 [17:15<6:44:36,  7.64s/it]2026-04-13 07:38:13,669 | INFO | train_grpo_train | metrics_logged mode=train step=121

                                                    
{'loss': 0.0284, 'grad_norm': 9.35704231262207, 'learning_rate': 9.636363636363638e-06, 'num_tokens': 224368.0, 'completions/mean_length': 150.375, 'completions/min_length': 101.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 150.375, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.30038923025131226, 'rewards/meter/std': 0.26126354932785034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967457056045532, 'rewards/repeat_soft/std': 0.0029344093054533005, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4315181374549866, 'rewards/total_composite/std': 0.07135103642940521, 'reward': 0.4315181374549866, 'reward_std': 0.07135103642940521, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21734307706356049, 'sampling/sampling_logp_difference/max': 1.5603547096252441, 'sampling/importance_sampling_ratio/min': 0.21006155014038086, 'sampling/importance_sampling_ratio/mean': 1.047584891319275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4637745916843414, 'clip_ratio/low_mean': 0.10287107527256012, 'clip_ratio/low_min': 0.10287107527256012, 'clip_ratio/high_mean': 0.11393089592456818, 'clip_ratio/high_max': 0.11393089592456818, 'clip_ratio/region_mean': 0.2168019711971283, 'reward_total_mean': 0.4315181374549866, 'reward_meter_mean': 0.30038923025131226, 'reward_meter_std': 0.26126354932785034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967457056045532, 'reward_repeat_soft_std': 0.0029344093054533005, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4315181374549866, 'reward_total_composite_std': 0.07135103642940521, 'epoch': 0.01}

  4%|▎         | 121/3300 [17:15<6:44:36,  7.64s/it]INFO 04-13 07:38:13 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 122/3300 [17:23<6:51:16,  7.76s/it]2026-04-13 07:38:21,732 | INFO | train_grpo_train | metrics_logged mode=train step=122

                                                    
{'loss': 0.4034, 'grad_norm': 11.443964004516602, 'learning_rate': 9.633333333333335e-06, 'num_tokens': 226239.0, 'completions/mean_length': 75.875, 'completions/min_length': 38.0, 'completions/max_length': 175.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 175.0, 'rewards/meter/mean': 0.24370630085468292, 'rewards/meter/std': 0.3257603347301483, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9960834383964539, 'rewards/repeat_soft/std': 0.005711476318538189, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.16446886956691742, 'rewards/total_composite/mean': 0.4244243800640106, 'rewards/total_composite/std': 0.19604572653770447, 'reward': 0.4244243800640106, 'reward_std': 0.19604569673538208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20754988491535187, 'sampling/sampling_logp_difference/max': 1.750192642211914, 'sampling/importance_sampling_ratio/min': 0.1737404763698578, 'sampling/importance_sampling_ratio/mean': 1.04361891746521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.415245831012726, 'clip_ratio/low_mean': 0.09463843237608671, 'clip_ratio/low_min': 0.09463843237608671, 'clip_ratio/high_mean': 0.07074670866131783, 'clip_ratio/high_max': 0.07074670866131783, 'clip_ratio/region_mean': 0.16538514103740454, 'reward_total_mean': 0.4244243800640106, 'reward_meter_mean': 0.24370630085468292, 'reward_meter_std': 0.3257603347301483, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9960834383964539, 'reward_repeat_soft_std': 0.005711476318538189, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.16446886956691742, 'reward_total_composite_mean': 0.4244243800640106, 'reward_total_composite_std': 0.19604572653770447, 'epoch': 0.01}

  4%|▎         | 122/3300 [17:23<6:51:16,  7.76s/it]INFO 04-13 07:38:21 [block_pool.py:378] Successfully reset prefix cache

  4%|▎         | 123/3300 [17:30<6:32:32,  7.41s/it]2026-04-13 07:38:28,325 | INFO | train_grpo_train | metrics_logged mode=train step=123

                                                    
{'loss': 0.0613, 'grad_norm': 18.49869155883789, 'learning_rate': 9.63030303030303e-06, 'num_tokens': 227650.0, 'completions/mean_length': 28.375, 'completions/min_length': 24.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6262902021408081, 'rewards/meter/std': 0.3038237392902374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953283071517944, 'rewards/repeat_soft/std': 0.0074143316596746445, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2921227812767029, 'rewards/total_composite/mean': 0.6145520210266113, 'rewards/total_composite/std': 0.19700059294700623, 'reward': 0.6145520210266113, 'reward_std': 0.19700059294700623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21750284731388092, 'sampling/sampling_logp_difference/max': 2.3387670516967773, 'sampling/importance_sampling_ratio/min': 0.09644648432731628, 'sampling/importance_sampling_ratio/mean': 0.9930812120437622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5528540685772896, 'clip_ratio/low_mean': 0.14069444499909878, 'clip_ratio/low_min': 0.14069444499909878, 'clip_ratio/high_mean': 0.08239214681088924, 'clip_ratio/high_max': 0.08239214681088924, 'clip_ratio/region_mean': 0.22308659180998802, 'reward_total_mean': 0.6145520210266113, 'reward_meter_mean': 0.6262902021408081, 'reward_meter_std': 0.3038237392902374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953283071517944, 'reward_repeat_soft_std': 0.0074143316596746445, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2921227812767029, 'reward_total_composite_mean': 0.6145520210266113, 'reward_total_composite_std': 0.19700059294700623, 'epoch': 0.01}

  4%|▎         | 123/3300 [17:30<6:32:32,  7.41s/it]INFO 04-13 07:38:28 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 124/3300 [17:37<6:20:21,  7.19s/it]2026-04-13 07:38:34,980 | INFO | train_grpo_train | metrics_logged mode=train step=124

                                                    
{'loss': 0.2265, 'grad_norm': 35.97916030883789, 'learning_rate': 9.627272727272728e-06, 'num_tokens': 229010.0, 'completions/mean_length': 21.0, 'completions/min_length': 14.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8410463333129883, 'rewards/meter/std': 0.26125726103782654, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9526292085647583, 'rewards/repeat_soft/std': 0.03851768746972084, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.549912691116333, 'rewards/total_composite/std': 0.12555654346942902, 'reward': 0.549912691116333, 'reward_std': 0.12555654346942902, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16508203744888306, 'sampling/sampling_logp_difference/max': 0.9596343040466309, 'sampling/importance_sampling_ratio/min': 0.419406920671463, 'sampling/importance_sampling_ratio/mean': 1.040798306465149, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4324931278824806, 'clip_ratio/low_mean': 0.04584703966975212, 'clip_ratio/low_min': 0.04584703966975212, 'clip_ratio/high_mean': 0.09017446916550398, 'clip_ratio/high_max': 0.09017446916550398, 'clip_ratio/region_mean': 0.1360215088352561, 'reward_total_mean': 0.549912691116333, 'reward_meter_mean': 0.8410463333129883, 'reward_meter_std': 0.26125726103782654, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9526292085647583, 'reward_repeat_soft_std': 0.03851768746972084, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.549912691116333, 'reward_total_composite_std': 0.12555654346942902, 'epoch': 0.01}

  4%|▍         | 124/3300 [17:37<6:20:21,  7.19s/it]INFO 04-13 07:38:35 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 125/3300 [17:44<6:20:01,  7.18s/it]2026-04-13 07:38:42,152 | INFO | train_grpo_train | metrics_logged mode=train step=125

                                                    
{'loss': 0.0966, 'grad_norm': 17.960975646972656, 'learning_rate': 9.624242424242425e-06, 'num_tokens': 230885.0, 'completions/mean_length': 65.375, 'completions/min_length': 52.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.31092584133148193, 'rewards/meter/std': 0.3357381820678711, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9908581972122192, 'rewards/repeat_soft/std': 0.009348123334348202, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.45793652534484863, 'rewards/total_composite/std': 0.1227300614118576, 'reward': 0.45793652534484863, 'reward_std': 0.1227300614118576, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27371904253959656, 'sampling/sampling_logp_difference/max': 2.2968673706054688, 'sampling/importance_sampling_ratio/min': 0.10057341307401657, 'sampling/importance_sampling_ratio/mean': 0.9990410804748535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9411287307739258, 'clip_ratio/low_mean': 0.1552787609398365, 'clip_ratio/low_min': 0.1552787609398365, 'clip_ratio/high_mean': 0.09186454489827156, 'clip_ratio/high_max': 0.09186454489827156, 'clip_ratio/region_mean': 0.24714330583810806, 'reward_total_mean': 0.45793652534484863, 'reward_meter_mean': 0.31092584133148193, 'reward_meter_std': 0.3357381820678711, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9908581972122192, 'reward_repeat_soft_std': 0.009348123334348202, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.45793652534484863, 'reward_total_composite_std': 0.1227300614118576, 'epoch': 0.01}

  4%|▍         | 125/3300 [17:44<6:20:01,  7.18s/it]INFO 04-13 07:38:42 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 126/3300 [17:53<6:57:12,  7.89s/it]2026-04-13 07:38:51,685 | INFO | train_grpo_train | metrics_logged mode=train step=126

                                                    
{'loss': 0.134, 'grad_norm': 24.076974868774414, 'learning_rate': 9.621212121212122e-06, 'num_tokens': 232246.0, 'completions/mean_length': 26.125, 'completions/min_length': 19.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7440218925476074, 'rewards/meter/std': 0.45724499225616455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9560876488685608, 'rewards/repeat_soft/std': 0.018136821687221527, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.25709572434425354, 'rewards/total_composite/mean': 0.508181095123291, 'rewards/total_composite/std': 0.13012929260730743, 'reward': 0.508181095123291, 'reward_std': 0.13012929260730743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21434858441352844, 'sampling/sampling_logp_difference/max': 1.1329364776611328, 'sampling/importance_sampling_ratio/min': 0.32208606600761414, 'sampling/importance_sampling_ratio/mean': 1.0196175575256348, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0628183633089066, 'clip_ratio/low_mean': 0.0923573998734355, 'clip_ratio/low_min': 0.0923573998734355, 'clip_ratio/high_mean': 0.06937899254262447, 'clip_ratio/high_max': 0.06937899254262447, 'clip_ratio/region_mean': 0.16173639241605997, 'reward_total_mean': 0.508181095123291, 'reward_meter_mean': 0.7440218925476074, 'reward_meter_std': 0.45724499225616455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9560876488685608, 'reward_repeat_soft_std': 0.018136821687221527, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.25709572434425354, 'reward_total_composite_mean': 0.508181095123291, 'reward_total_composite_std': 0.13012929260730743, 'epoch': 0.01}

  4%|▍         | 126/3300 [17:53<6:57:12,  7.89s/it]INFO 04-13 07:38:51 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 127/3300 [18:01<6:46:14,  7.68s/it]2026-04-13 07:38:58,887 | INFO | train_grpo_train | metrics_logged mode=train step=127

                                                    
{'loss': 0.1658, 'grad_norm': 15.085627555847168, 'learning_rate': 9.61818181818182e-06, 'num_tokens': 233806.0, 'completions/mean_length': 49.0, 'completions/min_length': 36.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5365213751792908, 'rewards/meter/std': 0.4541638195514679, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.993897557258606, 'rewards/repeat_soft/std': 0.009357492439448833, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.4945714473724365, 'rewards/total_composite/std': 0.14548209309577942, 'reward': 0.4945714473724365, 'reward_std': 0.14548209309577942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21035538613796234, 'sampling/sampling_logp_difference/max': 1.697453498840332, 'sampling/importance_sampling_ratio/min': 0.1831493079662323, 'sampling/importance_sampling_ratio/mean': 1.0626816749572754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.212236076593399, 'clip_ratio/low_mean': 0.07253067381680012, 'clip_ratio/low_min': 0.07253067381680012, 'clip_ratio/high_mean': 0.10060209594666958, 'clip_ratio/high_max': 0.10060209594666958, 'clip_ratio/region_mean': 0.1731327697634697, 'reward_total_mean': 0.4945714473724365, 'reward_meter_mean': 0.5365213751792908, 'reward_meter_std': 0.4541638195514679, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.993897557258606, 'reward_repeat_soft_std': 0.009357492439448833, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.4945714473724365, 'reward_total_composite_std': 0.14548209309577942, 'epoch': 0.01}

  4%|▍         | 127/3300 [18:01<6:46:14,  7.68s/it]INFO 04-13 07:38:59 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 128/3300 [18:08<6:45:14,  7.67s/it]2026-04-13 07:39:06,515 | INFO | train_grpo_train | metrics_logged mode=train step=128

                                                    
{'loss': -0.0173, 'grad_norm': 11.304403305053711, 'learning_rate': 9.615151515151517e-06, 'num_tokens': 235866.0, 'completions/mean_length': 84.5, 'completions/min_length': 61.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.42456066608428955, 'rewards/meter/std': 0.3047213554382324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.993170440196991, 'rewards/repeat_soft/std': 0.009959491901099682, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.4069315791130066, 'rewards/total_composite/std': 0.17672067880630493, 'reward': 0.4069315791130066, 'reward_std': 0.17672067880630493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2357020229101181, 'sampling/sampling_logp_difference/max': 1.8493890762329102, 'sampling/importance_sampling_ratio/min': 0.15733325481414795, 'sampling/importance_sampling_ratio/mean': 1.0373672246932983, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6177062541246414, 'clip_ratio/low_mean': 0.05069653131067753, 'clip_ratio/low_min': 0.05069653131067753, 'clip_ratio/high_mean': 0.17823937721550465, 'clip_ratio/high_max': 0.17823937721550465, 'clip_ratio/region_mean': 0.22893590852618217, 'reward_total_mean': 0.4069315791130066, 'reward_meter_mean': 0.42456066608428955, 'reward_meter_std': 0.3047213554382324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.993170440196991, 'reward_repeat_soft_std': 0.009959491901099682, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.4069315791130066, 'reward_total_composite_std': 0.17672067880630493, 'epoch': 0.01}

  4%|▍         | 128/3300 [18:08<6:45:14,  7.67s/it]INFO 04-13 07:39:06 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 129/3300 [18:16<6:45:56,  7.68s/it]2026-04-13 07:39:14,232 | INFO | train_grpo_train | metrics_logged mode=train step=129

                                                    
{'loss': 0.071, 'grad_norm': 11.652441024780273, 'learning_rate': 9.612121212121212e-06, 'num_tokens': 237693.0, 'completions/mean_length': 75.375, 'completions/min_length': 55.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.375, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.3495359718799591, 'rewards/meter/std': 0.26254764199256897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9941217303276062, 'rewards/repeat_soft/std': 0.014426924288272858, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.40984025597572327, 'rewards/total_composite/std': 0.19771966338157654, 'reward': 0.40984025597572327, 'reward_std': 0.19771966338157654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23216703534126282, 'sampling/sampling_logp_difference/max': 1.5401573181152344, 'sampling/importance_sampling_ratio/min': 0.21434737741947174, 'sampling/importance_sampling_ratio/mean': 1.0318115949630737, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.789889633655548, 'clip_ratio/low_mean': 0.10355722904205322, 'clip_ratio/low_min': 0.10355722904205322, 'clip_ratio/high_mean': 0.10569983534514904, 'clip_ratio/high_max': 0.10569983534514904, 'clip_ratio/region_mean': 0.20925706438720226, 'reward_total_mean': 0.40984025597572327, 'reward_meter_mean': 0.3495359718799591, 'reward_meter_std': 0.26254764199256897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9941217303276062, 'reward_repeat_soft_std': 0.014426924288272858, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.40984025597572327, 'reward_total_composite_std': 0.19771966338157654, 'epoch': 0.01}

  4%|▍         | 129/3300 [18:16<6:45:56,  7.68s/it]INFO 04-13 07:39:14 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 130/3300 [18:25<7:14:03,  8.22s/it]2026-04-13 07:39:23,696 | INFO | train_grpo_train | metrics_logged mode=train step=130

                                                    
{'loss': 0.0814, 'grad_norm': 8.03388500213623, 'learning_rate': 9.60909090909091e-06, 'num_tokens': 240318.0, 'completions/mean_length': 151.125, 'completions/min_length': 109.0, 'completions/max_length': 209.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 151.125, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 209.0, 'rewards/meter/mean': 0.4799237847328186, 'rewards/meter/std': 0.3395111560821533, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976893663406372, 'rewards/repeat_soft/std': 0.00220813718624413, 'rewards/judge_quality/mean': 0.2887499928474426, 'rewards/judge_quality/std': 0.11630470305681229, 'rewards/total_composite/mean': 0.42769867181777954, 'rewards/total_composite/std': 0.09508629143238068, 'reward': 0.42769867181777954, 'reward_std': 0.09508628398180008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2239220142364502, 'sampling/sampling_logp_difference/max': 1.2956266403198242, 'sampling/importance_sampling_ratio/min': 0.2811238467693329, 'sampling/importance_sampling_ratio/mean': 1.0429434776306152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.850192368030548, 'clip_ratio/low_mean': 0.09250323288142681, 'clip_ratio/low_min': 0.09250323288142681, 'clip_ratio/high_mean': 0.10934280045330524, 'clip_ratio/high_max': 0.10934280045330524, 'clip_ratio/region_mean': 0.20184603333473206, 'reward_total_mean': 0.42769867181777954, 'reward_meter_mean': 0.4799237847328186, 'reward_meter_std': 0.3395111560821533, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976893663406372, 'reward_repeat_soft_std': 0.00220813718624413, 'reward_judge_quality_mean': 0.2887499928474426, 'reward_judge_quality_std': 0.11630470305681229, 'reward_total_composite_mean': 0.42769867181777954, 'reward_total_composite_std': 0.09508629143238068, 'epoch': 0.01}

  4%|▍         | 130/3300 [18:25<7:14:03,  8.22s/it]INFO 04-13 07:39:23 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 131/3300 [18:35<7:31:27,  8.55s/it]2026-04-13 07:39:33,018 | INFO | train_grpo_train | metrics_logged mode=train step=131

                                                    
{'loss': 0.0504, 'grad_norm': 23.772933959960938, 'learning_rate': 9.606060606060607e-06, 'num_tokens': 241857.0, 'completions/mean_length': 34.375, 'completions/min_length': 26.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5936883091926575, 'rewards/meter/std': 0.38848671317100525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9913143515586853, 'rewards/repeat_soft/std': 0.009590408764779568, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5737825632095337, 'rewards/total_composite/std': 0.18015919625759125, 'reward': 0.5737825632095337, 'reward_std': 0.18015921115875244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2230602502822876, 'sampling/sampling_logp_difference/max': 1.8174285888671875, 'sampling/importance_sampling_ratio/min': 0.16244292259216309, 'sampling/importance_sampling_ratio/mean': 1.016609787940979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.539472445845604, 'clip_ratio/low_mean': 0.0953866122290492, 'clip_ratio/low_min': 0.0953866122290492, 'clip_ratio/high_mean': 0.13186106085777283, 'clip_ratio/high_max': 0.13186106085777283, 'clip_ratio/region_mean': 0.22724767308682203, 'reward_total_mean': 0.5737825632095337, 'reward_meter_mean': 0.5936883091926575, 'reward_meter_std': 0.38848671317100525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9913143515586853, 'reward_repeat_soft_std': 0.009590408764779568, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5737825632095337, 'reward_total_composite_std': 0.18015919625759125, 'epoch': 0.01}

  4%|▍         | 131/3300 [18:35<7:31:27,  8.55s/it]INFO 04-13 07:39:33 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 132/3300 [18:43<7:25:38,  8.44s/it]2026-04-13 07:39:41,207 | INFO | train_grpo_train | metrics_logged mode=train step=132

                                                    
{'loss': 0.105, 'grad_norm': 10.050190925598145, 'learning_rate': 9.603030303030304e-06, 'num_tokens': 244452.0, 'completions/mean_length': 121.375, 'completions/min_length': 90.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.375, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.4083891212940216, 'rewards/meter/std': 0.3109019994735718, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9966987371444702, 'rewards/repeat_soft/std': 0.0018282209057360888, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.445995032787323, 'rewards/total_composite/std': 0.09745653718709946, 'reward': 0.445995032787323, 'reward_std': 0.09745653718709946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21541813015937805, 'sampling/sampling_logp_difference/max': 1.6822853088378906, 'sampling/importance_sampling_ratio/min': 0.18594853579998016, 'sampling/importance_sampling_ratio/mean': 1.0369393825531006, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3659809827804565, 'clip_ratio/low_mean': 0.09735992550849915, 'clip_ratio/low_min': 0.09735992550849915, 'clip_ratio/high_mean': 0.1149887666106224, 'clip_ratio/high_max': 0.1149887666106224, 'clip_ratio/region_mean': 0.21234869211912155, 'reward_total_mean': 0.445995032787323, 'reward_meter_mean': 0.4083891212940216, 'reward_meter_std': 0.3109019994735718, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9966987371444702, 'reward_repeat_soft_std': 0.0018282209057360888, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.445995032787323, 'reward_total_composite_std': 0.09745653718709946, 'epoch': 0.01}

  4%|▍         | 132/3300 [18:43<7:25:38,  8.44s/it]INFO 04-13 07:39:41 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 133/3300 [18:51<7:22:04,  8.38s/it]2026-04-13 07:39:49,432 | INFO | train_grpo_train | metrics_logged mode=train step=133

                                                    
{'loss': 0.1965, 'grad_norm': 11.466776847839355, 'learning_rate': 9.600000000000001e-06, 'num_tokens': 246683.0, 'completions/mean_length': 79.875, 'completions/min_length': 45.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.5764611959457397, 'rewards/meter/std': 0.38859090209007263, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995432436466217, 'rewards/repeat_soft/std': 0.006603735499083996, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5122246742248535, 'rewards/total_composite/std': 0.2017168402671814, 'reward': 0.5122246742248535, 'reward_std': 0.2017168253660202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2326447069644928, 'sampling/sampling_logp_difference/max': 2.0507612228393555, 'sampling/importance_sampling_ratio/min': 0.12863695621490479, 'sampling/importance_sampling_ratio/mean': 1.0416052341461182, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.39143967628479, 'clip_ratio/low_mean': 0.11269427556544542, 'clip_ratio/low_min': 0.11269427556544542, 'clip_ratio/high_mean': 0.09823915176093578, 'clip_ratio/high_max': 0.09823915176093578, 'clip_ratio/region_mean': 0.2109334273263812, 'reward_total_mean': 0.5122246742248535, 'reward_meter_mean': 0.5764611959457397, 'reward_meter_std': 0.38859090209007263, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995432436466217, 'reward_repeat_soft_std': 0.006603735499083996, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5122246742248535, 'reward_total_composite_std': 0.2017168402671814, 'epoch': 0.01}

  4%|▍         | 133/3300 [18:51<7:22:04,  8.38s/it]INFO 04-13 07:39:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 134/3300 [19:00<7:23:58,  8.41s/it]2026-04-13 07:39:57,936 | INFO | train_grpo_train | metrics_logged mode=train step=134

                                                    
{'loss': 0.1005, 'grad_norm': 9.484580993652344, 'learning_rate': 9.596969696969699e-06, 'num_tokens': 249180.0, 'completions/mean_length': 119.125, 'completions/min_length': 94.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.125, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.889968991279602, 'rewards/meter/std': 0.2062651365995407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984543919563293, 'rewards/repeat_soft/std': 0.001125063979998231, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/total_composite/mean': 0.6544889807701111, 'rewards/total_composite/std': 0.12818054854869843, 'reward': 0.6544889807701111, 'reward_std': 0.12818053364753723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24180176854133606, 'sampling/sampling_logp_difference/max': 1.5038063526153564, 'sampling/importance_sampling_ratio/min': 0.22228246927261353, 'sampling/importance_sampling_ratio/mean': 1.0537770986557007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.092375487089157, 'clip_ratio/low_mean': 0.12049133330583572, 'clip_ratio/low_min': 0.12049133330583572, 'clip_ratio/high_mean': 0.08003133162856102, 'clip_ratio/high_max': 0.08003133162856102, 'clip_ratio/region_mean': 0.20052266493439674, 'reward_total_mean': 0.6544889807701111, 'reward_meter_mean': 0.889968991279602, 'reward_meter_std': 0.2062651365995407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984543919563293, 'reward_repeat_soft_std': 0.001125063979998231, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.6544889807701111, 'reward_total_composite_std': 0.12818054854869843, 'epoch': 0.01}

  4%|▍         | 134/3300 [19:00<7:23:58,  8.41s/it]INFO 04-13 07:39:58 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 135/3300 [19:08<7:20:54,  8.36s/it]2026-04-13 07:40:06,165 | INFO | train_grpo_train | metrics_logged mode=train step=135

                                                    
{'loss': 0.0019, 'grad_norm': 9.059500694274902, 'learning_rate': 9.593939393939394e-06, 'num_tokens': 251497.0, 'completions/mean_length': 123.625, 'completions/min_length': 87.0, 'completions/max_length': 165.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.625, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 165.0, 'rewards/meter/mean': 0.3144863247871399, 'rewards/meter/std': 0.28785303235054016, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9965611100196838, 'rewards/repeat_soft/std': 0.003245075698941946, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4088372588157654, 'rewards/total_composite/std': 0.08868332952260971, 'reward': 0.4088372588157654, 'reward_std': 0.08868333697319031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22240547835826874, 'sampling/sampling_logp_difference/max': 1.379908561706543, 'sampling/importance_sampling_ratio/min': 0.25160157680511475, 'sampling/importance_sampling_ratio/mean': 1.0371609926223755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.719191998243332, 'clip_ratio/low_mean': 0.1372978687286377, 'clip_ratio/low_min': 0.1372978687286377, 'clip_ratio/high_mean': 0.05146290548145771, 'clip_ratio/high_max': 0.05146290548145771, 'clip_ratio/region_mean': 0.1887607742100954, 'reward_total_mean': 0.4088372588157654, 'reward_meter_mean': 0.3144863247871399, 'reward_meter_std': 0.28785303235054016, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9965611100196838, 'reward_repeat_soft_std': 0.003245075698941946, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4088372588157654, 'reward_total_composite_std': 0.08868332952260971, 'epoch': 0.01}

  4%|▍         | 135/3300 [19:08<7:20:54,  8.36s/it]INFO 04-13 07:40:06 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 136/3300 [19:15<6:59:50,  7.96s/it]2026-04-13 07:40:13,200 | INFO | train_grpo_train | metrics_logged mode=train step=136

                                                    
{'loss': 0.0238, 'grad_norm': 11.203330993652344, 'learning_rate': 9.590909090909091e-06, 'num_tokens': 253602.0, 'completions/mean_length': 83.125, 'completions/min_length': 70.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.4576031267642975, 'rewards/meter/std': 0.43045464158058167, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996076226234436, 'rewards/repeat_soft/std': 0.004953309893608093, 'rewards/judge_quality/mean': 0.6325000524520874, 'rewards/judge_quality/std': 0.18850921094417572, 'rewards/total_composite/mean': 0.5137057304382324, 'rewards/total_composite/std': 0.15345917642116547, 'reward': 0.5137057304382324, 'reward_std': 0.15345917642116547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20802229642868042, 'sampling/sampling_logp_difference/max': 1.8483209609985352, 'sampling/importance_sampling_ratio/min': 0.15750139951705933, 'sampling/importance_sampling_ratio/mean': 1.0360591411590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.956837147474289, 'clip_ratio/low_mean': 0.10638447385281324, 'clip_ratio/low_min': 0.10638447385281324, 'clip_ratio/high_mean': 0.06874127313494682, 'clip_ratio/high_max': 0.06874127313494682, 'clip_ratio/region_mean': 0.17512574698776007, 'reward_total_mean': 0.5137057304382324, 'reward_meter_mean': 0.4576031267642975, 'reward_meter_std': 0.43045464158058167, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996076226234436, 'reward_repeat_soft_std': 0.004953309893608093, 'reward_judge_quality_mean': 0.6325000524520874, 'reward_judge_quality_std': 0.18850921094417572, 'reward_total_composite_mean': 0.5137057304382324, 'reward_total_composite_std': 0.15345917642116547, 'epoch': 0.01}

  4%|▍         | 136/3300 [19:15<6:59:50,  7.96s/it]INFO 04-13 07:40:13 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 137/3300 [19:21<6:35:13,  7.50s/it]2026-04-13 07:40:19,613 | INFO | train_grpo_train | metrics_logged mode=train step=137

                                                    
{'loss': -0.0257, 'grad_norm': 18.757707595825195, 'learning_rate': 9.587878787878789e-06, 'num_tokens': 255110.0, 'completions/mean_length': 34.5, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.42969757318496704, 'rewards/meter/std': 0.3924821615219116, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.997260332107544, 'rewards/repeat_soft/std': 0.004701441153883934, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5069931745529175, 'rewards/total_composite/std': 0.19512850046157837, 'reward': 0.5069931745529175, 'reward_std': 0.19512848556041718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19561952352523804, 'sampling/sampling_logp_difference/max': 1.0618122816085815, 'sampling/importance_sampling_ratio/min': 0.34582850337028503, 'sampling/importance_sampling_ratio/mean': 1.0478719472885132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3865096867084503, 'clip_ratio/low_mean': 0.11821861192584038, 'clip_ratio/low_min': 0.11821861192584038, 'clip_ratio/high_mean': 0.08298234827816486, 'clip_ratio/high_max': 0.08298234827816486, 'clip_ratio/region_mean': 0.20120096020400524, 'reward_total_mean': 0.5069931745529175, 'reward_meter_mean': 0.42969757318496704, 'reward_meter_std': 0.3924821615219116, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.997260332107544, 'reward_repeat_soft_std': 0.004701441153883934, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5069931745529175, 'reward_total_composite_std': 0.19512850046157837, 'epoch': 0.01}

  4%|▍         | 137/3300 [19:21<6:35:13,  7.50s/it]INFO 04-13 07:40:19 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 138/3300 [19:29<6:31:40,  7.43s/it]2026-04-13 07:40:26,894 | INFO | train_grpo_train | metrics_logged mode=train step=138

                                                    
{'loss': 0.2324, 'grad_norm': 19.762535095214844, 'learning_rate': 9.584848484848486e-06, 'num_tokens': 256848.0, 'completions/mean_length': 48.25, 'completions/min_length': 33.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.22157180309295654, 'rewards/meter/std': 0.23181016743183136, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9917092323303223, 'rewards/repeat_soft/std': 0.012141245417296886, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.42576566338539124, 'rewards/total_composite/std': 0.1713506579399109, 'reward': 0.42576566338539124, 'reward_std': 0.17135067284107208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27887052297592163, 'sampling/sampling_logp_difference/max': 2.26450514793396, 'sampling/importance_sampling_ratio/min': 0.10388142615556717, 'sampling/importance_sampling_ratio/mean': 1.038446307182312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.365349918603897, 'clip_ratio/low_mean': 0.13626768812537193, 'clip_ratio/low_min': 0.13626768812537193, 'clip_ratio/high_mean': 0.0834330152720213, 'clip_ratio/high_max': 0.0834330152720213, 'clip_ratio/region_mean': 0.21970070339739323, 'reward_total_mean': 0.42576566338539124, 'reward_meter_mean': 0.22157180309295654, 'reward_meter_std': 0.23181016743183136, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9917092323303223, 'reward_repeat_soft_std': 0.012141245417296886, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.42576566338539124, 'reward_total_composite_std': 0.1713506579399109, 'epoch': 0.01}

  4%|▍         | 138/3300 [19:29<6:31:40,  7.43s/it]INFO 04-13 07:40:27 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 139/3300 [19:35<6:20:00,  7.21s/it]2026-04-13 07:40:33,596 | INFO | train_grpo_train | metrics_logged mode=train step=139

                                                    
{'loss': 0.0099, 'grad_norm': 15.954649925231934, 'learning_rate': 9.581818181818181e-06, 'num_tokens': 258427.0, 'completions/mean_length': 46.375, 'completions/min_length': 37.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4192202091217041, 'rewards/meter/std': 0.4678597152233124, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9996633529663086, 'rewards/repeat_soft/std': 0.0009521605097688735, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.46494266390800476, 'rewards/total_composite/std': 0.14755646884441376, 'reward': 0.46494266390800476, 'reward_std': 0.14755645394325256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2324858456850052, 'sampling/sampling_logp_difference/max': 1.674997329711914, 'sampling/importance_sampling_ratio/min': 0.1873086839914322, 'sampling/importance_sampling_ratio/mean': 1.0463426113128662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6474010050296783, 'clip_ratio/low_mean': 0.10105736553668976, 'clip_ratio/low_min': 0.10105736553668976, 'clip_ratio/high_mean': 0.11296234279870987, 'clip_ratio/high_max': 0.11296234279870987, 'clip_ratio/region_mean': 0.21401970833539963, 'reward_total_mean': 0.46494266390800476, 'reward_meter_mean': 0.4192202091217041, 'reward_meter_std': 0.4678597152233124, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9996633529663086, 'reward_repeat_soft_std': 0.0009521605097688735, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.46494266390800476, 'reward_total_composite_std': 0.14755646884441376, 'epoch': 0.01}

  4%|▍         | 139/3300 [19:35<6:20:00,  7.21s/it]INFO 04-13 07:40:33 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 140/3300 [19:43<6:23:32,  7.28s/it]2026-04-13 07:40:41,040 | INFO | train_grpo_train | metrics_logged mode=train step=140

                                                    
{'loss': 0.1831, 'grad_norm': 20.739917755126953, 'learning_rate': 9.57878787878788e-06, 'num_tokens': 260043.0, 'completions/mean_length': 44.0, 'completions/min_length': 23.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.4250248372554779, 'rewards/meter/std': 0.34178006649017334, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9861264824867249, 'rewards/repeat_soft/std': 0.023440638557076454, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.2125651240348816, 'rewards/total_composite/mean': 0.4422690272331238, 'rewards/total_composite/std': 0.09150385111570358, 'reward': 0.4422690272331238, 'reward_std': 0.09150385111570358, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23629844188690186, 'sampling/sampling_logp_difference/max': 1.9918994903564453, 'sampling/importance_sampling_ratio/min': 0.13643601536750793, 'sampling/importance_sampling_ratio/mean': 0.9938191771507263, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3093688562512398, 'clip_ratio/low_mean': 0.11244106199592352, 'clip_ratio/low_min': 0.11244106199592352, 'clip_ratio/high_mean': 0.09376941062510014, 'clip_ratio/high_max': 0.09376941062510014, 'clip_ratio/region_mean': 0.20621047262102365, 'reward_total_mean': 0.4422690272331238, 'reward_meter_mean': 0.4250248372554779, 'reward_meter_std': 0.34178006649017334, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9861264824867249, 'reward_repeat_soft_std': 0.023440638557076454, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.2125651240348816, 'reward_total_composite_mean': 0.4422690272331238, 'reward_total_composite_std': 0.09150385111570358, 'epoch': 0.01}

  4%|▍         | 140/3300 [19:43<6:23:32,  7.28s/it]INFO 04-13 07:40:41 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 141/3300 [19:51<6:45:35,  7.70s/it]2026-04-13 07:40:49,726 | INFO | train_grpo_train | metrics_logged mode=train step=141

                                                    
{'loss': -0.0717, 'grad_norm': 7.565834999084473, 'learning_rate': 9.575757575757576e-06, 'num_tokens': 263212.0, 'completions/mean_length': 190.125, 'completions/min_length': 153.0, 'completions/max_length': 233.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 190.125, 'completions/min_terminated_length': 153.0, 'completions/max_terminated_length': 233.0, 'rewards/meter/mean': 0.45693451166152954, 'rewards/meter/std': 0.2026921510696411, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9970571994781494, 'rewards/repeat_soft/std': 0.002340377075597644, 'rewards/judge_quality/mean': 0.32249999046325684, 'rewards/judge_quality/std': 0.10925068706274033, 'rewards/total_composite/mean': 0.35164132714271545, 'rewards/total_composite/std': 0.22201529145240784, 'reward': 0.35164132714271545, 'reward_std': 0.22201529145240784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21230649948120117, 'sampling/sampling_logp_difference/max': 1.8322811126708984, 'sampling/importance_sampling_ratio/min': 0.16004806756973267, 'sampling/importance_sampling_ratio/mean': 1.0359517335891724, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3753039687871933, 'clip_ratio/low_mean': 0.057767802849411964, 'clip_ratio/low_min': 0.057767802849411964, 'clip_ratio/high_mean': 0.15601530484855175, 'clip_ratio/high_max': 0.15601530484855175, 'clip_ratio/region_mean': 0.21378310769796371, 'reward_total_mean': 0.35164132714271545, 'reward_meter_mean': 0.45693451166152954, 'reward_meter_std': 0.2026921510696411, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9970571994781494, 'reward_repeat_soft_std': 0.002340377075597644, 'reward_judge_quality_mean': 0.32249999046325684, 'reward_judge_quality_std': 0.10925068706274033, 'reward_total_composite_mean': 0.35164132714271545, 'reward_total_composite_std': 0.22201529145240784, 'epoch': 0.01}

  4%|▍         | 141/3300 [19:51<6:45:35,  7.70s/it]INFO 04-13 07:40:49 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 142/3300 [19:58<6:27:35,  7.36s/it]2026-04-13 07:40:56,297 | INFO | train_grpo_train | metrics_logged mode=train step=142

                                                    
{'loss': -0.0096, 'grad_norm': 17.13520622253418, 'learning_rate': 9.572727272727273e-06, 'num_tokens': 264821.0, 'completions/mean_length': 42.125, 'completions/min_length': 37.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6349161863327026, 'rewards/meter/std': 0.3797772526741028, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980885982513428, 'rewards/repeat_soft/std': 0.0032931563910096884, 'rewards/judge_quality/mean': 0.5987499952316284, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.6001944541931152, 'rewards/total_composite/std': 0.17118269205093384, 'reward': 0.6001944541931152, 'reward_std': 0.17118267714977264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2538120746612549, 'sampling/sampling_logp_difference/max': 2.163132667541504, 'sampling/importance_sampling_ratio/min': 0.11496441066265106, 'sampling/importance_sampling_ratio/mean': 1.057732343673706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7320705354213715, 'clip_ratio/low_mean': 0.0996027011424303, 'clip_ratio/low_min': 0.0996027011424303, 'clip_ratio/high_mean': 0.0937653873115778, 'clip_ratio/high_max': 0.0937653873115778, 'clip_ratio/region_mean': 0.1933680884540081, 'reward_total_mean': 0.6001944541931152, 'reward_meter_mean': 0.6349161863327026, 'reward_meter_std': 0.3797772526741028, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980885982513428, 'reward_repeat_soft_std': 0.0032931563910096884, 'reward_judge_quality_mean': 0.5987499952316284, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.6001944541931152, 'reward_total_composite_std': 0.17118269205093384, 'epoch': 0.01}

  4%|▍         | 142/3300 [19:58<6:27:35,  7.36s/it]INFO 04-13 07:40:56 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 143/3300 [20:06<6:39:25,  7.59s/it]2026-04-13 07:41:04,419 | INFO | train_grpo_train | metrics_logged mode=train step=143

                                                    
{'loss': 0.022, 'grad_norm': 7.532354354858398, 'learning_rate': 9.56969696969697e-06, 'num_tokens': 267745.0, 'completions/mean_length': 178.5, 'completions/min_length': 127.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 178.5, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.6103168725967407, 'rewards/meter/std': 0.23132774233818054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9940958023071289, 'rewards/repeat_soft/std': 0.006265062373131514, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.5102753043174744, 'rewards/total_composite/std': 0.07604330033063889, 'reward': 0.5102753043174744, 'reward_std': 0.07604330778121948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19552329182624817, 'sampling/sampling_logp_difference/max': 1.7654690742492676, 'sampling/importance_sampling_ratio/min': 0.1711065024137497, 'sampling/importance_sampling_ratio/mean': 1.0306285619735718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4039943516254425, 'clip_ratio/low_mean': 0.12090220488607883, 'clip_ratio/low_min': 0.12090220488607883, 'clip_ratio/high_mean': 0.0771543812006712, 'clip_ratio/high_max': 0.0771543812006712, 'clip_ratio/region_mean': 0.19805658608675003, 'reward_total_mean': 0.5102753043174744, 'reward_meter_mean': 0.6103168725967407, 'reward_meter_std': 0.23132774233818054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9940958023071289, 'reward_repeat_soft_std': 0.006265062373131514, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.5102753043174744, 'reward_total_composite_std': 0.07604330033063889, 'epoch': 0.01}

  4%|▍         | 143/3300 [20:06<6:39:25,  7.59s/it]INFO 04-13 07:41:04 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 144/3300 [20:13<6:32:37,  7.46s/it]2026-04-13 07:41:11,587 | INFO | train_grpo_train | metrics_logged mode=train step=144

                                                    
{'loss': 0.0646, 'grad_norm': 16.326160430908203, 'learning_rate': 9.566666666666668e-06, 'num_tokens': 269457.0, 'completions/mean_length': 42.0, 'completions/min_length': 36.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8048985004425049, 'rewards/meter/std': 0.31660687923431396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9979851245880127, 'rewards/repeat_soft/std': 0.0019468758255243301, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.6515523195266724, 'rewards/total_composite/std': 0.18092253804206848, 'reward': 0.6515523195266724, 'reward_std': 0.1809225231409073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2573746144771576, 'sampling/sampling_logp_difference/max': 1.4676711559295654, 'sampling/importance_sampling_ratio/min': 0.23046156764030457, 'sampling/importance_sampling_ratio/mean': 1.0187652111053467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5685728192329407, 'clip_ratio/low_mean': 0.14593254402279854, 'clip_ratio/low_min': 0.14593254402279854, 'clip_ratio/high_mean': 0.06349206529557705, 'clip_ratio/high_max': 0.06349206529557705, 'clip_ratio/region_mean': 0.2094246093183756, 'reward_total_mean': 0.6515523195266724, 'reward_meter_mean': 0.8048985004425049, 'reward_meter_std': 0.31660687923431396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9979851245880127, 'reward_repeat_soft_std': 0.0019468758255243301, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.6515523195266724, 'reward_total_composite_std': 0.18092253804206848, 'epoch': 0.01}

  4%|▍         | 144/3300 [20:13<6:32:37,  7.46s/it]INFO 04-13 07:41:11 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 145/3300 [20:20<6:19:40,  7.22s/it]2026-04-13 07:41:18,239 | INFO | train_grpo_train | metrics_logged mode=train step=145

                                                    
{'loss': -0.0136, 'grad_norm': 13.721324920654297, 'learning_rate': 9.563636363636365e-06, 'num_tokens': 271000.0, 'completions/mean_length': 37.875, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.41020357608795166, 'rewards/meter/std': 0.468443363904953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9902426600456238, 'rewards/repeat_soft/std': 0.014534646645188332, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.4963862895965576, 'rewards/total_composite/std': 0.16369476914405823, 'reward': 0.4963862895965576, 'reward_std': 0.16369476914405823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25721344351768494, 'sampling/sampling_logp_difference/max': 1.5890684127807617, 'sampling/importance_sampling_ratio/min': 0.20411567389965057, 'sampling/importance_sampling_ratio/mean': 1.0349671840667725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1548607498407364, 'clip_ratio/low_mean': 0.1378176361322403, 'clip_ratio/low_min': 0.1378176361322403, 'clip_ratio/high_mean': 0.08839784376323223, 'clip_ratio/high_max': 0.08839784376323223, 'clip_ratio/region_mean': 0.22621547989547253, 'reward_total_mean': 0.4963862895965576, 'reward_meter_mean': 0.41020357608795166, 'reward_meter_std': 0.468443363904953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9902426600456238, 'reward_repeat_soft_std': 0.014534646645188332, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.4963862895965576, 'reward_total_composite_std': 0.16369476914405823, 'epoch': 0.01}

  4%|▍         | 145/3300 [20:20<6:19:40,  7.22s/it]INFO 04-13 07:41:18 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 146/3300 [20:27<6:19:07,  7.21s/it]2026-04-13 07:41:25,434 | INFO | train_grpo_train | metrics_logged mode=train step=146

                                                    
{'loss': 0.3095, 'grad_norm': 20.189346313476562, 'learning_rate': 9.56060606060606e-06, 'num_tokens': 272617.0, 'completions/mean_length': 33.125, 'completions/min_length': 23.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.31975626945495605, 'rewards/meter/std': 0.41625145077705383, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9912452697753906, 'rewards/repeat_soft/std': 0.007847433909773827, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.3928127586841583, 'rewards/total_composite/std': 0.19192352890968323, 'reward': 0.3928127586841583, 'reward_std': 0.19192351400852203, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24665847420692444, 'sampling/sampling_logp_difference/max': 1.5526294708251953, 'sampling/importance_sampling_ratio/min': 0.21169060468673706, 'sampling/importance_sampling_ratio/mean': 1.0148438215255737, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.123029351234436, 'clip_ratio/low_mean': 0.10523463413119316, 'clip_ratio/low_min': 0.10523463413119316, 'clip_ratio/high_mean': 0.13094667345285416, 'clip_ratio/high_max': 0.13094667345285416, 'clip_ratio/region_mean': 0.23618130758404732, 'reward_total_mean': 0.3928127586841583, 'reward_meter_mean': 0.31975626945495605, 'reward_meter_std': 0.41625145077705383, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9912452697753906, 'reward_repeat_soft_std': 0.007847433909773827, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.3928127586841583, 'reward_total_composite_std': 0.19192352890968323, 'epoch': 0.01}

  4%|▍         | 146/3300 [20:27<6:19:07,  7.21s/it]INFO 04-13 07:41:25 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 147/3300 [20:34<6:08:14,  7.01s/it]2026-04-13 07:41:31,964 | INFO | train_grpo_train | metrics_logged mode=train step=147

                                                    
{'loss': -0.007, 'grad_norm': 22.95297622680664, 'learning_rate': 9.55757575757576e-06, 'num_tokens': 274100.0, 'completions/mean_length': 28.375, 'completions/min_length': 25.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7506793141365051, 'rewards/meter/std': 0.35030412673950195, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9963032007217407, 'rewards/repeat_soft/std': 0.007101293187588453, 'rewards/judge_quality/mean': 0.48125001788139343, 'rewards/judge_quality/std': 0.22949868440628052, 'rewards/total_composite/mean': 0.5625472068786621, 'rewards/total_composite/std': 0.27261069416999817, 'reward': 0.5625472068786621, 'reward_std': 0.27261069416999817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2450028359889984, 'sampling/sampling_logp_difference/max': 1.6066884994506836, 'sampling/importance_sampling_ratio/min': 0.2005506455898285, 'sampling/importance_sampling_ratio/mean': 0.9905889630317688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7993857562541962, 'clip_ratio/low_mean': 0.05597480107098818, 'clip_ratio/low_min': 0.05597480107098818, 'clip_ratio/high_mean': 0.15836711786687374, 'clip_ratio/high_max': 0.15836711786687374, 'clip_ratio/region_mean': 0.21434191893786192, 'reward_total_mean': 0.5625472068786621, 'reward_meter_mean': 0.7506793141365051, 'reward_meter_std': 0.35030412673950195, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9963032007217407, 'reward_repeat_soft_std': 0.007101293187588453, 'reward_judge_quality_mean': 0.48125001788139343, 'reward_judge_quality_std': 0.22949868440628052, 'reward_total_composite_mean': 0.5625472068786621, 'reward_total_composite_std': 0.27261069416999817, 'epoch': 0.01}

  4%|▍         | 147/3300 [20:34<6:08:14,  7.01s/it]INFO 04-13 07:41:32 [block_pool.py:378] Successfully reset prefix cache

  4%|▍         | 148/3300 [20:42<6:35:41,  7.53s/it]2026-04-13 07:41:40,720 | INFO | train_grpo_train | metrics_logged mode=train step=148

                                                    
{'loss': -0.0024, 'grad_norm': 11.477031707763672, 'learning_rate': 9.554545454545455e-06, 'num_tokens': 276455.0, 'completions/mean_length': 105.375, 'completions/min_length': 86.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.375, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.7320038080215454, 'rewards/meter/std': 0.3469865918159485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950052499771118, 'rewards/repeat_soft/std': 0.003556472947821021, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5532875061035156, 'rewards/total_composite/std': 0.08732420206069946, 'reward': 0.5532875061035156, 'reward_std': 0.08732419461011887, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21940535306930542, 'sampling/sampling_logp_difference/max': 2.2431559562683105, 'sampling/importance_sampling_ratio/min': 0.10612305998802185, 'sampling/importance_sampling_ratio/mean': 1.0478465557098389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.533340662717819, 'clip_ratio/low_mean': 0.07008165493607521, 'clip_ratio/low_min': 0.07008165493607521, 'clip_ratio/high_mean': 0.15172532014548779, 'clip_ratio/high_max': 0.15172532014548779, 'clip_ratio/region_mean': 0.221806975081563, 'reward_total_mean': 0.5532875061035156, 'reward_meter_mean': 0.7320038080215454, 'reward_meter_std': 0.3469865918159485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950052499771118, 'reward_repeat_soft_std': 0.003556472947821021, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5532875061035156, 'reward_total_composite_std': 0.08732420206069946, 'epoch': 0.01}

  4%|▍         | 148/3300 [20:42<6:35:41,  7.53s/it]INFO 04-13 07:41:40 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 149/3300 [20:51<6:53:39,  7.88s/it]2026-04-13 07:41:49,400 | INFO | train_grpo_train | metrics_logged mode=train step=149

                                                    
{'loss': 0.114, 'grad_norm': 7.869264602661133, 'learning_rate': 9.551515151515152e-06, 'num_tokens': 279217.0, 'completions/mean_length': 154.25, 'completions/min_length': 118.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.25, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.9051262140274048, 'rewards/meter/std': 0.10090532898902893, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9958059787750244, 'rewards/repeat_soft/std': 0.0055671576410532, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5140024423599243, 'rewards/total_composite/std': 0.21084751188755035, 'reward': 0.5140024423599243, 'reward_std': 0.21084751188755035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22774766385555267, 'sampling/sampling_logp_difference/max': 1.7449188232421875, 'sampling/importance_sampling_ratio/min': 0.17465916275978088, 'sampling/importance_sampling_ratio/mean': 1.0631458759307861, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3837957084178925, 'clip_ratio/low_mean': 0.03640520200133324, 'clip_ratio/low_min': 0.03640520200133324, 'clip_ratio/high_mean': 0.1736424770206213, 'clip_ratio/high_max': 0.1736424770206213, 'clip_ratio/region_mean': 0.21004767902195454, 'reward_total_mean': 0.5140024423599243, 'reward_meter_mean': 0.9051262140274048, 'reward_meter_std': 0.10090532898902893, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9958059787750244, 'reward_repeat_soft_std': 0.0055671576410532, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5140024423599243, 'reward_total_composite_std': 0.21084751188755035, 'epoch': 0.01}

  5%|▍         | 149/3300 [20:51<6:53:39,  7.88s/it]INFO 04-13 07:41:49 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 150/3300 [20:59<6:52:33,  7.86s/it]2026-04-13 07:41:57,215 | INFO | train_grpo_train | metrics_logged mode=train step=150

                                                    
{'loss': 0.0883, 'grad_norm': 12.227827072143555, 'learning_rate': 9.54848484848485e-06, 'num_tokens': 281338.0, 'completions/mean_length': 95.125, 'completions/min_length': 70.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.3991343379020691, 'rewards/meter/std': 0.34587615728378296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.991938591003418, 'rewards/repeat_soft/std': 0.006831905338913202, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4577544927597046, 'rewards/total_composite/std': 0.09414715319871902, 'reward': 0.4577544927597046, 'reward_std': 0.09414716064929962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2518637478351593, 'sampling/sampling_logp_difference/max': 2.145904541015625, 'sampling/importance_sampling_ratio/min': 0.11696219444274902, 'sampling/importance_sampling_ratio/mean': 1.0258044004440308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.635319620370865, 'clip_ratio/low_mean': 0.13855895400047302, 'clip_ratio/low_min': 0.13855895400047302, 'clip_ratio/high_mean': 0.09044339880347252, 'clip_ratio/high_max': 0.09044339880347252, 'clip_ratio/region_mean': 0.22900235280394554, 'reward_total_mean': 0.4577544927597046, 'reward_meter_mean': 0.3991343379020691, 'reward_meter_std': 0.34587615728378296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.991938591003418, 'reward_repeat_soft_std': 0.006831905338913202, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4577544927597046, 'reward_total_composite_std': 0.09414715319871902, 'epoch': 0.02}

  5%|▍         | 150/3300 [20:59<6:52:33,  7.86s/it]INFO 04-13 07:41:57 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:17,  2.19s/it]

 30%|███       | 3/10 [00:10<00:25,  3.63s/it]

 40%|████      | 4/10 [00:13<00:22,  3.67s/it]

 50%|█████     | 5/10 [00:19<00:21,  4.33s/it]

 60%|██████    | 6/10 [00:24<00:17,  4.46s/it]

 70%|███████   | 7/10 [00:29<00:14,  4.83s/it]

 80%|████████  | 8/10 [00:34<00:09,  4.93s/it]

 90%|█████████ | 9/10 [00:39<00:04,  4.92s/it]

100%|██████████| 10/10 [00:43<00:00,  4.69s/it]2026-04-13 07:42:46,815 | INFO | train_grpo_train | metrics_logged mode=eval step=150

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 49.5893, 'eval_samples_per_second': 1.613, 'eval_steps_per_second': 0.202, 'eval_num_tokens': 281338.0, 'eval_completions/mean_length': 72.475, 'eval_completions/min_length': 28.6, 'eval_completions/max_length': 138.9, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 72.475, 'eval_completions/min_terminated_length': 28.6, 'eval_completions/max_terminated_length': 138.9, 'eval_rewards/meter/mean': 0.5251416712999344, 'eval_rewards/meter/std': 0.39672465324401857, 'eval_rewards/count_adherence/mean': 0.9891666531562805, 'eval_rewards/count_adherence/std': 0.03064129650592804, 'eval_rewards/hard_gate/mean': 0.9625, 'eval_rewards/hard_gate/std': 0.0816463440656662, 'eval_rewards/repeat_soft/mean': 0.9936495900154114, 'eval_rewards/repeat_soft/std': 0.00908708639908582, 'eval_rewards/judge_quality/mean': 0.4742499977350235, 'eval_rewards/judge_quality/std': 0.17209785655140877, 'eval_rewards/total_composite/mean': 0.5050697594881057, 'eval_rewards/total_composite/std': 0.17620564475655556, 'eval_reward': 0.5050697594881057, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1579168274998665, 'eval_sampling/sampling_logp_difference/max': 1.197763156890869, 'eval_sampling/importance_sampling_ratio/min': 0.3089745670557022, 'eval_sampling/importance_sampling_ratio/mean': 1.0503769397735596, 'eval_sampling/importance_sampling_ratio/max': 1.5593770742416382, 'eval_entropy': 2.493303906917572, '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': 0.5050697594881057, 'eval_reward_meter_mean': 0.5251416712999344, 'eval_reward_meter_std': 0.39672465324401857, 'eval_reward_count_adherence_mean': 0.9891666531562805, 'eval_reward_count_adherence_std': 0.03064129650592804, 'eval_reward_hard_gate_mean': 0.9625, 'eval_reward_hard_gate_std': 0.0816463440656662, 'eval_reward_repeat_soft_mean': 0.9936495900154114, 'eval_reward_repeat_soft_std': 0.00908708639908582, 'eval_reward_judge_quality_mean': 0.4742499977350235, 'eval_reward_judge_quality_std': 0.17209785655140877, 'eval_reward_total_composite_mean': 0.5050697594881057, 'eval_reward_total_composite_std': 0.17620564475655556, 'epoch': 0.02}

  5%|▍         | 150/3300 [21:49<6:52:33,  7.86s/it]

100%|██████████| 10/10 [00:43<00:00,  4.69s/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(

  5%|▍         | 151/3300 [21:57<20:07:45, 23.01s/it]2026-04-13 07:42:55,587 | INFO | train_grpo_train | metrics_logged mode=train step=151

                                                     
{'loss': -0.0139, 'grad_norm': 19.790287017822266, 'learning_rate': 9.545454545454547e-06, 'num_tokens': 282777.0, 'completions/mean_length': 27.875, 'completions/min_length': 20.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6628537178039551, 'rewards/meter/std': 0.2736438810825348, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9575977325439453, 'rewards/repeat_soft/std': 0.013865554705262184, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.5267788767814636, 'rewards/total_composite/std': 0.08878375589847565, 'reward': 0.5267788767814636, 'reward_std': 0.08878375589847565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19810548424720764, 'sampling/sampling_logp_difference/max': 1.0578088760375977, 'sampling/importance_sampling_ratio/min': 0.34721577167510986, 'sampling/importance_sampling_ratio/mean': 1.0431057214736938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.909252405166626, 'clip_ratio/low_mean': 0.08024779241532087, 'clip_ratio/low_min': 0.08024779241532087, 'clip_ratio/high_mean': 0.08133874274790287, 'clip_ratio/high_max': 0.08133874274790287, 'clip_ratio/region_mean': 0.16158653516322374, 'reward_total_mean': 0.5267788767814636, 'reward_meter_mean': 0.6628537178039551, 'reward_meter_std': 0.2736438810825348, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9575977325439453, 'reward_repeat_soft_std': 0.013865554705262184, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.5267788767814636, 'reward_total_composite_std': 0.08878375589847565, 'epoch': 0.02}

  5%|▍         | 151/3300 [21:57<20:07:45, 23.01s/it]INFO 04-13 07:42:55 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 152/3300 [22:06<16:19:32, 18.67s/it]2026-04-13 07:43:04,124 | INFO | train_grpo_train | metrics_logged mode=train step=152

                                                     
{'loss': -0.0552, 'grad_norm': 9.100715637207031, 'learning_rate': 9.542424242424242e-06, 'num_tokens': 284947.0, 'completions/mean_length': 108.25, 'completions/min_length': 66.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.696089506149292, 'rewards/meter/std': 0.368598073720932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.996252179145813, 'rewards/repeat_soft/std': 0.0039284187369048595, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.39426562190055847, 'rewards/total_composite/std': 0.25745531916618347, 'reward': 0.39426562190055847, 'reward_std': 0.2574552893638611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23037481307983398, 'sampling/sampling_logp_difference/max': 1.9209213256835938, 'sampling/importance_sampling_ratio/min': 0.14647194743156433, 'sampling/importance_sampling_ratio/mean': 1.0479086637496948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9769923985004425, 'clip_ratio/low_mean': 0.07217737101018429, 'clip_ratio/low_min': 0.07217737101018429, 'clip_ratio/high_mean': 0.13639610446989536, 'clip_ratio/high_max': 0.13639610446989536, 'clip_ratio/region_mean': 0.20857347548007965, 'reward_total_mean': 0.39426562190055847, 'reward_meter_mean': 0.696089506149292, 'reward_meter_std': 0.368598073720932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.996252179145813, 'reward_repeat_soft_std': 0.0039284187369048595, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.39426562190055847, 'reward_total_composite_std': 0.25745531916618347, 'epoch': 0.02}

  5%|▍         | 152/3300 [22:06<16:19:32, 18.67s/it]INFO 04-13 07:43:04 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 153/3300 [22:13<13:17:11, 15.20s/it]2026-04-13 07:43:11,225 | INFO | train_grpo_train | metrics_logged mode=train step=153

                                                     
{'loss': 0.0375, 'grad_norm': 22.370010375976562, 'learning_rate': 9.539393939393941e-06, 'num_tokens': 286482.0, 'completions/mean_length': 45.875, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7622062563896179, 'rewards/meter/std': 0.33447155356407166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981516599655151, 'rewards/repeat_soft/std': 0.0038459571078419685, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5700531005859375, 'rewards/total_composite/std': 0.09727941453456879, 'reward': 0.5700531005859375, 'reward_std': 0.09727941453456879, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21602395176887512, 'sampling/sampling_logp_difference/max': 1.35379958152771, 'sampling/importance_sampling_ratio/min': 0.258257120847702, 'sampling/importance_sampling_ratio/mean': 1.035374402999878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.340231239795685, 'clip_ratio/low_mean': 0.06541666761040688, 'clip_ratio/low_min': 0.06541666761040688, 'clip_ratio/high_mean': 0.10533494502305984, 'clip_ratio/high_max': 0.10533494502305984, 'clip_ratio/region_mean': 0.17075161263346672, 'reward_total_mean': 0.5700531005859375, 'reward_meter_mean': 0.7622062563896179, 'reward_meter_std': 0.33447155356407166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981516599655151, 'reward_repeat_soft_std': 0.0038459571078419685, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5700531005859375, 'reward_total_composite_std': 0.09727941453456879, 'epoch': 0.02}

  5%|▍         | 153/3300 [22:13<13:17:11, 15.20s/it]INFO 04-13 07:43:11 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 154/3300 [22:20<11:06:49, 12.72s/it]2026-04-13 07:43:18,153 | INFO | train_grpo_train | metrics_logged mode=train step=154

                                                     
{'loss': -0.0714, 'grad_norm': 16.674734115600586, 'learning_rate': 9.536363636363637e-06, 'num_tokens': 288210.0, 'completions/mean_length': 44.0, 'completions/min_length': 26.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.37179723381996155, 'rewards/meter/std': 0.3616820275783539, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980150461196899, 'rewards/repeat_soft/std': 0.004647395573556423, 'rewards/judge_quality/mean': 0.35499998927116394, 'rewards/judge_quality/std': 0.11928357183933258, 'rewards/total_composite/mean': 0.4344622492790222, 'rewards/total_composite/std': 0.09572197496891022, 'reward': 0.4344622492790222, 'reward_std': 0.09572196751832962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.247511625289917, 'sampling/sampling_logp_difference/max': 1.6972360610961914, 'sampling/importance_sampling_ratio/min': 0.18318915367126465, 'sampling/importance_sampling_ratio/mean': 1.0838520526885986, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3497215807437897, 'clip_ratio/low_mean': 0.12646591290831566, 'clip_ratio/low_min': 0.12646591290831566, 'clip_ratio/high_mean': 0.0763119999319315, 'clip_ratio/high_max': 0.0763119999319315, 'clip_ratio/region_mean': 0.20277791284024715, 'reward_total_mean': 0.4344622492790222, 'reward_meter_mean': 0.37179723381996155, 'reward_meter_std': 0.3616820275783539, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980150461196899, 'reward_repeat_soft_std': 0.004647395573556423, 'reward_judge_quality_mean': 0.35499998927116394, 'reward_judge_quality_std': 0.11928357183933258, 'reward_total_composite_mean': 0.4344622492790222, 'reward_total_composite_std': 0.09572197496891022, 'epoch': 0.02}

  5%|▍         | 154/3300 [22:20<11:06:49, 12.72s/it]INFO 04-13 07:43:18 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 155/3300 [22:27<9:31:12, 10.90s/it] 2026-04-13 07:43:24,803 | INFO | train_grpo_train | metrics_logged mode=train step=155

                                                    
{'loss': 0.1263, 'grad_norm': 13.679117202758789, 'learning_rate': 9.533333333333334e-06, 'num_tokens': 290021.0, 'completions/mean_length': 50.375, 'completions/min_length': 33.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.3696790039539337, 'rewards/meter/std': 0.3822711706161499, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9926376342773438, 'rewards/repeat_soft/std': 0.012575971893966198, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.4402022957801819, 'rewards/total_composite/std': 0.12922914326190948, 'reward': 0.4402022957801819, 'reward_std': 0.12922914326190948, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22571733593940735, 'sampling/sampling_logp_difference/max': 1.248558521270752, 'sampling/importance_sampling_ratio/min': 0.28691810369491577, 'sampling/importance_sampling_ratio/mean': 1.0440168380737305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9790842533111572, 'clip_ratio/low_mean': 0.09155844338238239, 'clip_ratio/low_min': 0.09155844338238239, 'clip_ratio/high_mean': 0.09837852418422699, 'clip_ratio/high_max': 0.09837852418422699, 'clip_ratio/region_mean': 0.18993696756660938, 'reward_total_mean': 0.4402022957801819, 'reward_meter_mean': 0.3696790039539337, 'reward_meter_std': 0.3822711706161499, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9926376342773438, 'reward_repeat_soft_std': 0.012575971893966198, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.4402022957801819, 'reward_total_composite_std': 0.12922914326190948, 'epoch': 0.02}

  5%|▍         | 155/3300 [22:27<9:31:12, 10.90s/it]INFO 04-13 07:43:25 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 156/3300 [22:34<8:32:01,  9.77s/it]2026-04-13 07:43:31,947 | INFO | train_grpo_train | metrics_logged mode=train step=156

                                                    
{'loss': -0.0267, 'grad_norm': 19.237661361694336, 'learning_rate': 9.530303030303031e-06, 'num_tokens': 291601.0, 'completions/mean_length': 34.5, 'completions/min_length': 30.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7581612467765808, 'rewards/meter/std': 0.3878549337387085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9873794317245483, 'rewards/repeat_soft/std': 0.028771191835403442, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.562809944152832, 'rewards/total_composite/std': 0.10876806825399399, 'reward': 0.562809944152832, 'reward_std': 0.10876806825399399, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21949085593223572, 'sampling/sampling_logp_difference/max': 1.4842844009399414, 'sampling/importance_sampling_ratio/min': 0.2266644984483719, 'sampling/importance_sampling_ratio/mean': 1.0325417518615723, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2725380212068558, 'clip_ratio/low_mean': 0.05151515267789364, 'clip_ratio/low_min': 0.05151515267789364, 'clip_ratio/high_mean': 0.14134271629154682, 'clip_ratio/high_max': 0.14134271629154682, 'clip_ratio/region_mean': 0.19285786896944046, 'reward_total_mean': 0.562809944152832, 'reward_meter_mean': 0.7581612467765808, 'reward_meter_std': 0.3878549337387085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9873794317245483, 'reward_repeat_soft_std': 0.028771191835403442, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.562809944152832, 'reward_total_composite_std': 0.10876806825399399, 'epoch': 0.02}

  5%|▍         | 156/3300 [22:34<8:32:01,  9.77s/it]INFO 04-13 07:43:32 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 157/3300 [22:40<7:43:07,  8.84s/it]2026-04-13 07:43:38,618 | INFO | train_grpo_train | metrics_logged mode=train step=157

                                                    
{'loss': 0.0124, 'grad_norm': 18.231548309326172, 'learning_rate': 9.527272727272729e-06, 'num_tokens': 293024.0, 'completions/mean_length': 24.875, 'completions/min_length': 17.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.484775185585022, 'rewards/meter/std': 0.4647696018218994, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.3512499928474426, 'rewards/judge_quality/std': 0.11630471050739288, 'rewards/total_composite/mean': 0.4474857449531555, 'rewards/total_composite/std': 0.11991287022829056, 'reward': 0.4474857449531555, 'reward_std': 0.11991287022829056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2207426279783249, 'sampling/sampling_logp_difference/max': 1.7150442600250244, 'sampling/importance_sampling_ratio/min': 0.17995575070381165, 'sampling/importance_sampling_ratio/mean': 1.051077961921692, 'sampling/importance_sampling_ratio/max': 1.9085829257965088, 'entropy': 2.074438989162445, 'clip_ratio/low_mean': 0.117631989531219, 'clip_ratio/low_min': 0.117631989531219, 'clip_ratio/high_mean': 0.09234270825982094, 'clip_ratio/high_max': 0.09234270825982094, 'clip_ratio/region_mean': 0.20997469779103994, 'reward_total_mean': 0.4474857449531555, 'reward_meter_mean': 0.484775185585022, 'reward_meter_std': 0.4647696018218994, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.3512499928474426, 'reward_judge_quality_std': 0.11630471050739288, 'reward_total_composite_mean': 0.4474857449531555, 'reward_total_composite_std': 0.11991287022829056, 'epoch': 0.02}

  5%|▍         | 157/3300 [22:40<7:43:07,  8.84s/it]INFO 04-13 07:43:38 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 158/3300 [22:48<7:26:04,  8.52s/it]2026-04-13 07:43:46,383 | INFO | train_grpo_train | metrics_logged mode=train step=158

                                                    
{'loss': 0.3052, 'grad_norm': 15.065874099731445, 'learning_rate': 9.524242424242424e-06, 'num_tokens': 294583.0, 'completions/mean_length': 46.875, 'completions/min_length': 33.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.8994266986846924, 'rewards/meter/std': 0.15715087950229645, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950315952301025, 'rewards/repeat_soft/std': 0.006452750880271196, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5939333438873291, 'rewards/total_composite/std': 0.11897853761911392, 'reward': 0.5939333438873291, 'reward_std': 0.11897853016853333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2199499011039734, 'sampling/sampling_logp_difference/max': 1.2576894760131836, 'sampling/importance_sampling_ratio/min': 0.28431016206741333, 'sampling/importance_sampling_ratio/mean': 1.0469788312911987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.943353235721588, 'clip_ratio/low_mean': 0.07501471135765314, 'clip_ratio/low_min': 0.07501471135765314, 'clip_ratio/high_mean': 0.14738421700894833, 'clip_ratio/high_max': 0.14738421700894833, 'clip_ratio/region_mean': 0.22239892836660147, 'reward_total_mean': 0.5939333438873291, 'reward_meter_mean': 0.8994266986846924, 'reward_meter_std': 0.15715087950229645, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950315952301025, 'reward_repeat_soft_std': 0.006452750880271196, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5939333438873291, 'reward_total_composite_std': 0.11897853761911392, 'epoch': 0.02}

  5%|▍         | 158/3300 [22:48<7:26:04,  8.52s/it]INFO 04-13 07:43:46 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 159/3300 [22:55<7:00:49,  8.04s/it]2026-04-13 07:43:53,303 | INFO | train_grpo_train | metrics_logged mode=train step=159

                                                    
{'loss': 0.0037, 'grad_norm': 13.692737579345703, 'learning_rate': 9.521212121212121e-06, 'num_tokens': 296129.0, 'completions/mean_length': 40.25, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8731591701507568, 'rewards/meter/std': 0.2492521107196808, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9962665438652039, 'rewards/repeat_soft/std': 0.004987684544175863, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.6194087266921997, 'rewards/total_composite/std': 0.10308452695608139, 'reward': 0.6194087266921997, 'reward_std': 0.10308452695608139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21989428997039795, 'sampling/sampling_logp_difference/max': 2.073795795440674, 'sampling/importance_sampling_ratio/min': 0.1257077157497406, 'sampling/importance_sampling_ratio/mean': 1.0411407947540283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5392094254493713, 'clip_ratio/low_mean': 0.07449252251535654, 'clip_ratio/low_min': 0.07449252251535654, 'clip_ratio/high_mean': 0.13062131218612194, 'clip_ratio/high_max': 0.13062131218612194, 'clip_ratio/region_mean': 0.20511383470147848, 'reward_total_mean': 0.6194087266921997, 'reward_meter_mean': 0.8731591701507568, 'reward_meter_std': 0.2492521107196808, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9962665438652039, 'reward_repeat_soft_std': 0.004987684544175863, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.6194087266921997, 'reward_total_composite_std': 0.10308452695608139, 'epoch': 0.02}

  5%|▍         | 159/3300 [22:55<7:00:49,  8.04s/it]INFO 04-13 07:43:53 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 160/3300 [23:01<6:36:07,  7.57s/it]2026-04-13 07:43:59,777 | INFO | train_grpo_train | metrics_logged mode=train step=160

                                                    
{'loss': -0.0457, 'grad_norm': 20.655656814575195, 'learning_rate': 9.518181818181819e-06, 'num_tokens': 297670.0, 'completions/mean_length': 35.625, 'completions/min_length': 27.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.730107843875885, 'rewards/meter/std': 0.3464305102825165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9918180704116821, 'rewards/repeat_soft/std': 0.02118588052690029, 'rewards/judge_quality/mean': 0.70250004529953, 'rewards/judge_quality/std': 0.17862571775913239, 'rewards/total_composite/mean': 0.6042086482048035, 'rewards/total_composite/std': 0.3024638891220093, 'reward': 0.6042086482048035, 'reward_std': 0.3024638593196869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20736154913902283, 'sampling/sampling_logp_difference/max': 3.032355308532715, 'sampling/importance_sampling_ratio/min': 0.04820197448134422, 'sampling/importance_sampling_ratio/mean': 1.0262166261672974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.698737993836403, 'clip_ratio/low_mean': 0.07620247639715672, 'clip_ratio/low_min': 0.07620247639715672, 'clip_ratio/high_mean': 0.10640341974794865, 'clip_ratio/high_max': 0.10640341974794865, 'clip_ratio/region_mean': 0.18260589614510536, 'reward_total_mean': 0.6042086482048035, 'reward_meter_mean': 0.730107843875885, 'reward_meter_std': 0.3464305102825165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9918180704116821, 'reward_repeat_soft_std': 0.02118588052690029, 'reward_judge_quality_mean': 0.70250004529953, 'reward_judge_quality_std': 0.17862571775913239, 'reward_total_composite_mean': 0.6042086482048035, 'reward_total_composite_std': 0.3024638891220093, 'epoch': 0.02}

  5%|▍         | 160/3300 [23:02<6:36:07,  7.57s/it]INFO 04-13 07:44:00 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 161/3300 [23:09<6:29:00,  7.44s/it]2026-04-13 07:44:06,901 | INFO | train_grpo_train | metrics_logged mode=train step=161

                                                    
{'loss': 0.0961, 'grad_norm': 20.642906188964844, 'learning_rate': 9.515151515151516e-06, 'num_tokens': 299034.0, 'completions/mean_length': 24.5, 'completions/min_length': 19.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.8733356595039368, 'rewards/meter/std': 0.15982545912265778, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9569480419158936, 'rewards/repeat_soft/std': 0.010414325632154942, 'rewards/judge_quality/mean': 0.5724999904632568, 'rewards/judge_quality/std': 0.38425251841545105, 'rewards/total_composite/mean': 0.6833223104476929, 'rewards/total_composite/std': 0.2551179528236389, 'reward': 0.6833223104476929, 'reward_std': 0.2551179528236389, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21364142000675201, 'sampling/sampling_logp_difference/max': 1.458907127380371, 'sampling/importance_sampling_ratio/min': 0.23249022662639618, 'sampling/importance_sampling_ratio/mean': 1.0699632167816162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2969636619091034, 'clip_ratio/low_mean': 0.08534770319238305, 'clip_ratio/low_min': 0.08534770319238305, 'clip_ratio/high_mean': 0.07862903131172061, 'clip_ratio/high_max': 0.07862903131172061, 'clip_ratio/region_mean': 0.16397673450410366, 'reward_total_mean': 0.6833223104476929, 'reward_meter_mean': 0.8733356595039368, 'reward_meter_std': 0.15982545912265778, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9569480419158936, 'reward_repeat_soft_std': 0.010414325632154942, 'reward_judge_quality_mean': 0.5724999904632568, 'reward_judge_quality_std': 0.38425251841545105, 'reward_total_composite_mean': 0.6833223104476929, 'reward_total_composite_std': 0.2551179528236389, 'epoch': 0.02}

  5%|▍         | 161/3300 [23:09<6:29:00,  7.44s/it]INFO 04-13 07:44:07 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 162/3300 [23:15<6:15:19,  7.18s/it]2026-04-13 07:44:13,472 | INFO | train_grpo_train | metrics_logged mode=train step=162

                                                    
{'loss': -0.0122, 'grad_norm': 25.833316802978516, 'learning_rate': 9.512121212121213e-06, 'num_tokens': 300545.0, 'completions/mean_length': 28.875, 'completions/min_length': 22.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.5683460235595703, 'rewards/meter/std': 0.4032154679298401, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980363845825195, 'rewards/repeat_soft/std': 0.004261462949216366, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.509473979473114, 'rewards/total_composite/std': 0.11501184850931168, 'reward': 0.509473979473114, 'reward_std': 0.11501184850931168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2363111972808838, 'sampling/sampling_logp_difference/max': 2.019106864929199, 'sampling/importance_sampling_ratio/min': 0.1327739953994751, 'sampling/importance_sampling_ratio/mean': 1.0371768474578857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1397273391485214, 'clip_ratio/low_mean': 0.09411422163248062, 'clip_ratio/low_min': 0.09411422163248062, 'clip_ratio/high_mean': 0.13323844783008099, 'clip_ratio/high_max': 0.13323844783008099, 'clip_ratio/region_mean': 0.2273526694625616, 'reward_total_mean': 0.509473979473114, 'reward_meter_mean': 0.5683460235595703, 'reward_meter_std': 0.4032154679298401, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980363845825195, 'reward_repeat_soft_std': 0.004261462949216366, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.509473979473114, 'reward_total_composite_std': 0.11501184850931168, 'epoch': 0.02}

  5%|▍         | 162/3300 [23:15<6:15:19,  7.18s/it]INFO 04-13 07:44:13 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 163/3300 [23:22<6:10:04,  7.08s/it]2026-04-13 07:44:20,322 | INFO | train_grpo_train | metrics_logged mode=train step=163

                                                    
{'loss': 0.167, 'grad_norm': 16.014175415039062, 'learning_rate': 9.50909090909091e-06, 'num_tokens': 302176.0, 'completions/mean_length': 43.875, 'completions/min_length': 28.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7519853115081787, 'rewards/meter/std': 0.37167438864707947, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9982995986938477, 'rewards/repeat_soft/std': 0.0027445803862065077, 'rewards/judge_quality/mean': 0.5712500214576721, 'rewards/judge_quality/std': 0.24468272924423218, 'rewards/total_composite/mean': 0.5745153427124023, 'rewards/total_composite/std': 0.3193117678165436, 'reward': 0.5745153427124023, 'reward_std': 0.3193117678165436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2562037706375122, 'sampling/sampling_logp_difference/max': 2.644084930419922, 'sampling/importance_sampling_ratio/min': 0.0710703581571579, 'sampling/importance_sampling_ratio/mean': 1.032947301864624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6925185918807983, 'clip_ratio/low_mean': 0.09780711121857166, 'clip_ratio/low_min': 0.09780711121857166, 'clip_ratio/high_mean': 0.10800338350236416, 'clip_ratio/high_max': 0.10800338350236416, 'clip_ratio/region_mean': 0.20581049472093582, 'reward_total_mean': 0.5745153427124023, 'reward_meter_mean': 0.7519853115081787, 'reward_meter_std': 0.37167438864707947, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9982995986938477, 'reward_repeat_soft_std': 0.0027445803862065077, 'reward_judge_quality_mean': 0.5712500214576721, 'reward_judge_quality_std': 0.24468272924423218, 'reward_total_composite_mean': 0.5745153427124023, 'reward_total_composite_std': 0.3193117678165436, 'epoch': 0.02}

  5%|▍         | 163/3300 [23:22<6:10:04,  7.08s/it]INFO 04-13 07:44:20 [block_pool.py:378] Successfully reset prefix cache

  5%|▍         | 164/3300 [23:29<6:14:39,  7.17s/it]2026-04-13 07:44:27,701 | INFO | train_grpo_train | metrics_logged mode=train step=164

                                                    
{'loss': 0.006, 'grad_norm': 10.29527473449707, 'learning_rate': 9.506060606060606e-06, 'num_tokens': 304486.0, 'completions/mean_length': 101.75, 'completions/min_length': 73.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.75, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.5555546879768372, 'rewards/meter/std': 0.39578020572662354, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981604814529419, 'rewards/repeat_soft/std': 0.001627032645046711, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.17266297340393066, 'rewards/total_composite/mean': 0.5006445646286011, 'rewards/total_composite/std': 0.1136239543557167, 'reward': 0.5006445646286011, 'reward_std': 0.1136239543557167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23732200264930725, 'sampling/sampling_logp_difference/max': 1.4951133728027344, 'sampling/importance_sampling_ratio/min': 0.22422318160533905, 'sampling/importance_sampling_ratio/mean': 1.0576616525650024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6199953258037567, 'clip_ratio/low_mean': 0.08103533275425434, 'clip_ratio/low_min': 0.08103533275425434, 'clip_ratio/high_mean': 0.14940478280186653, 'clip_ratio/high_max': 0.14940478280186653, 'clip_ratio/region_mean': 0.23044011555612087, 'reward_total_mean': 0.5006445646286011, 'reward_meter_mean': 0.5555546879768372, 'reward_meter_std': 0.39578020572662354, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981604814529419, 'reward_repeat_soft_std': 0.001627032645046711, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.17266297340393066, 'reward_total_composite_mean': 0.5006445646286011, 'reward_total_composite_std': 0.1136239543557167, 'epoch': 0.02}

  5%|▍         | 164/3300 [23:29<6:14:39,  7.17s/it]INFO 04-13 07:44:27 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 165/3300 [23:37<6:15:14,  7.18s/it]2026-04-13 07:44:34,913 | INFO | train_grpo_train | metrics_logged mode=train step=165

                                                    
{'loss': 0.2365, 'grad_norm': 22.70522117614746, 'learning_rate': 9.503030303030303e-06, 'num_tokens': 306042.0, 'completions/mean_length': 36.5, 'completions/min_length': 23.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.49694645404815674, 'rewards/meter/std': 0.3599587380886078, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9899628162384033, 'rewards/repeat_soft/std': 0.015568912029266357, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.5811421275138855, 'rewards/total_composite/std': 0.21729491651058197, 'reward': 0.5811421275138855, 'reward_std': 0.21729491651058197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2418246865272522, 'sampling/sampling_logp_difference/max': 3.4209160804748535, 'sampling/importance_sampling_ratio/min': 0.03268248215317726, 'sampling/importance_sampling_ratio/mean': 0.9846088886260986, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0012380257248878, 'clip_ratio/low_mean': 0.11417735926806927, 'clip_ratio/low_min': 0.11417735926806927, 'clip_ratio/high_mean': 0.09063545241951942, 'clip_ratio/high_max': 0.09063545241951942, 'clip_ratio/region_mean': 0.2048128116875887, 'reward_total_mean': 0.5811421275138855, 'reward_meter_mean': 0.49694645404815674, 'reward_meter_std': 0.3599587380886078, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9899628162384033, 'reward_repeat_soft_std': 0.015568912029266357, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.5811421275138855, 'reward_total_composite_std': 0.21729491651058197, 'epoch': 0.02}

  5%|▌         | 165/3300 [23:37<6:15:14,  7.18s/it]INFO 04-13 07:44:35 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 166/3300 [23:44<6:17:34,  7.23s/it]2026-04-13 07:44:42,251 | INFO | train_grpo_train | metrics_logged mode=train step=166

                                                    
{'loss': 0.1229, 'grad_norm': 14.87144947052002, 'learning_rate': 9.5e-06, 'num_tokens': 307629.0, 'completions/mean_length': 34.375, 'completions/min_length': 24.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5064627528190613, 'rewards/meter/std': 0.37647876143455505, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9944350123405457, 'rewards/repeat_soft/std': 0.004966493230313063, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5132936239242554, 'rewards/total_composite/std': 0.15036630630493164, 'reward': 0.5132936239242554, 'reward_std': 0.15036627650260925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22471682727336884, 'sampling/sampling_logp_difference/max': 1.9718409776687622, 'sampling/importance_sampling_ratio/min': 0.1392003446817398, 'sampling/importance_sampling_ratio/mean': 1.0255013704299927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8131614699959755, 'clip_ratio/low_mean': 0.11302317772060633, 'clip_ratio/low_min': 0.11302317772060633, 'clip_ratio/high_mean': 0.08667140640318394, 'clip_ratio/high_max': 0.08667140640318394, 'clip_ratio/region_mean': 0.19969458412379026, 'reward_total_mean': 0.5132936239242554, 'reward_meter_mean': 0.5064627528190613, 'reward_meter_std': 0.37647876143455505, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9944350123405457, 'reward_repeat_soft_std': 0.004966493230313063, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5132936239242554, 'reward_total_composite_std': 0.15036630630493164, 'epoch': 0.02}

  5%|▌         | 166/3300 [23:44<6:17:34,  7.23s/it]INFO 04-13 07:44:42 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 167/3300 [23:51<6:14:10,  7.17s/it]2026-04-13 07:44:49,271 | INFO | train_grpo_train | metrics_logged mode=train step=167

                                                    
{'loss': 0.0906, 'grad_norm': 14.257298469543457, 'learning_rate': 9.496969696969698e-06, 'num_tokens': 309198.0, 'completions/mean_length': 49.125, 'completions/min_length': 34.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.6849068999290466, 'rewards/meter/std': 0.4111604392528534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.998062789440155, 'rewards/repeat_soft/std': 0.00547928037121892, 'rewards/judge_quality/mean': 0.42249998450279236, 'rewards/judge_quality/std': 0.15471172332763672, 'rewards/total_composite/mean': 0.46477779746055603, 'rewards/total_composite/std': 0.30610400438308716, 'reward': 0.46477779746055603, 'reward_std': 0.30610397458076477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22429433465003967, 'sampling/sampling_logp_difference/max': 1.3679618835449219, 'sampling/importance_sampling_ratio/min': 0.2546253800392151, 'sampling/importance_sampling_ratio/mean': 1.047497272491455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0990798473358154, 'clip_ratio/low_mean': 0.07017498183995485, 'clip_ratio/low_min': 0.07017498183995485, 'clip_ratio/high_mean': 0.12275388929992914, 'clip_ratio/high_max': 0.12275388929992914, 'clip_ratio/region_mean': 0.192928871139884, 'reward_total_mean': 0.46477779746055603, 'reward_meter_mean': 0.6849068999290466, 'reward_meter_std': 0.4111604392528534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.998062789440155, 'reward_repeat_soft_std': 0.00547928037121892, 'reward_judge_quality_mean': 0.42249998450279236, 'reward_judge_quality_std': 0.15471172332763672, 'reward_total_composite_mean': 0.46477779746055603, 'reward_total_composite_std': 0.30610400438308716, 'epoch': 0.02}

  5%|▌         | 167/3300 [23:51<6:14:10,  7.17s/it]INFO 04-13 07:44:49 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 168/3300 [23:58<6:15:02,  7.18s/it]2026-04-13 07:44:56,499 | INFO | train_grpo_train | metrics_logged mode=train step=168

                                                    
{'loss': 0.016, 'grad_norm': 24.050216674804688, 'learning_rate': 9.493939393939395e-06, 'num_tokens': 310708.0, 'completions/mean_length': 23.75, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7325787544250488, 'rewards/meter/std': 0.39579838514328003, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.36374998092651367, 'rewards/judge_quality/std': 0.13265825808048248, 'rewards/total_composite/mean': 0.4768216609954834, 'rewards/total_composite/std': 0.21741239726543427, 'reward': 0.4768216609954834, 'reward_std': 0.21741239726543427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22347064316272736, 'sampling/sampling_logp_difference/max': 1.902097225189209, 'sampling/importance_sampling_ratio/min': 0.14925527572631836, 'sampling/importance_sampling_ratio/mean': 1.030205249786377, 'sampling/importance_sampling_ratio/max': 1.782066822052002, 'entropy': 1.9697849974036217, 'clip_ratio/low_mean': 0.09212963096797466, 'clip_ratio/low_min': 0.09212963096797466, 'clip_ratio/high_mean': 0.10793161019682884, 'clip_ratio/high_max': 0.10793161019682884, 'clip_ratio/region_mean': 0.2000612411648035, 'reward_total_mean': 0.4768216609954834, 'reward_meter_mean': 0.7325787544250488, 'reward_meter_std': 0.39579838514328003, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.36374998092651367, 'reward_judge_quality_std': 0.13265825808048248, 'reward_total_composite_mean': 0.4768216609954834, 'reward_total_composite_std': 0.21741239726543427, 'epoch': 0.02}

  5%|▌         | 168/3300 [23:58<6:15:02,  7.18s/it]INFO 04-13 07:44:56 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 169/3300 [24:05<6:02:54,  6.95s/it]2026-04-13 07:45:02,916 | INFO | train_grpo_train | metrics_logged mode=train step=169

                                                    
{'loss': 0.0113, 'grad_norm': 21.640043258666992, 'learning_rate': 9.490909090909092e-06, 'num_tokens': 312307.0, 'completions/mean_length': 38.875, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7079165577888489, 'rewards/meter/std': 0.40104302763938904, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9991671442985535, 'rewards/repeat_soft/std': 0.0023556507658213377, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.2728651165962219, 'rewards/total_composite/mean': 0.6812629699707031, 'rewards/total_composite/std': 0.23844361305236816, 'reward': 0.6812629699707031, 'reward_std': 0.23844359815120697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22272655367851257, 'sampling/sampling_logp_difference/max': 1.4653444290161133, 'sampling/importance_sampling_ratio/min': 0.2309984266757965, 'sampling/importance_sampling_ratio/mean': 1.0448509454727173, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.164816826581955, 'clip_ratio/low_mean': 0.0858415924012661, 'clip_ratio/low_min': 0.0858415924012661, 'clip_ratio/high_mean': 0.09395995084196329, 'clip_ratio/high_max': 0.09395995084196329, 'clip_ratio/region_mean': 0.1798015432432294, 'reward_total_mean': 0.6812629699707031, 'reward_meter_mean': 0.7079165577888489, 'reward_meter_std': 0.40104302763938904, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9991671442985535, 'reward_repeat_soft_std': 0.0023556507658213377, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.2728651165962219, 'reward_total_composite_mean': 0.6812629699707031, 'reward_total_composite_std': 0.23844361305236816, 'epoch': 0.02}

  5%|▌         | 169/3300 [24:05<6:02:54,  6.95s/it]INFO 04-13 07:45:03 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 170/3300 [24:11<6:00:31,  6.91s/it]2026-04-13 07:45:09,725 | INFO | train_grpo_train | metrics_logged mode=train step=170

                                                    
{'loss': 0.0214, 'grad_norm': 14.737093925476074, 'learning_rate': 9.487878787878788e-06, 'num_tokens': 313806.0, 'completions/mean_length': 40.375, 'completions/min_length': 35.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8741497993469238, 'rewards/meter/std': 0.3062879741191864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9916309118270874, 'rewards/repeat_soft/std': 0.010093681514263153, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.22696760296821594, 'rewards/total_composite/mean': 0.6944446563720703, 'rewards/total_composite/std': 0.19079254567623138, 'reward': 0.6944446563720703, 'reward_std': 0.1907925307750702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21184924244880676, 'sampling/sampling_logp_difference/max': 1.7176275253295898, 'sampling/importance_sampling_ratio/min': 0.17949149012565613, 'sampling/importance_sampling_ratio/mean': 1.075750470161438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6572673618793488, 'clip_ratio/low_mean': 0.15218430012464523, 'clip_ratio/low_min': 0.15218430012464523, 'clip_ratio/high_mean': 0.07383041083812714, 'clip_ratio/high_max': 0.07383041083812714, 'clip_ratio/region_mean': 0.22601471096277237, 'reward_total_mean': 0.6944446563720703, 'reward_meter_mean': 0.8741497993469238, 'reward_meter_std': 0.3062879741191864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9916309118270874, 'reward_repeat_soft_std': 0.010093681514263153, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.22696760296821594, 'reward_total_composite_mean': 0.6944446563720703, 'reward_total_composite_std': 0.19079254567623138, 'epoch': 0.02}

  5%|▌         | 170/3300 [24:11<6:00:31,  6.91s/it]INFO 04-13 07:45:09 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 171/3300 [24:19<6:03:09,  6.96s/it]2026-04-13 07:45:16,814 | INFO | train_grpo_train | metrics_logged mode=train step=171

                                                    
{'loss': 0.1001, 'grad_norm': 15.172825813293457, 'learning_rate': 9.484848484848485e-06, 'num_tokens': 315546.0, 'completions/mean_length': 51.5, 'completions/min_length': 32.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7064719200134277, 'rewards/meter/std': 0.3983866274356842, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9972226023674011, 'rewards/repeat_soft/std': 0.007504946552217007, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/total_composite/mean': 0.6153678297996521, 'rewards/total_composite/std': 0.17465214431285858, 'reward': 0.6153678297996521, 'reward_std': 0.17465214431285858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2158244103193283, 'sampling/sampling_logp_difference/max': 1.5815305709838867, 'sampling/importance_sampling_ratio/min': 0.20566007494926453, 'sampling/importance_sampling_ratio/mean': 1.0184203386306763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.242932051420212, 'clip_ratio/low_mean': 0.0644509494304657, 'clip_ratio/low_min': 0.0644509494304657, 'clip_ratio/high_mean': 0.1464561503380537, 'clip_ratio/high_max': 0.1464561503380537, 'clip_ratio/region_mean': 0.2109070997685194, 'reward_total_mean': 0.6153678297996521, 'reward_meter_mean': 0.7064719200134277, 'reward_meter_std': 0.3983866274356842, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9972226023674011, 'reward_repeat_soft_std': 0.007504946552217007, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.6153678297996521, 'reward_total_composite_std': 0.17465214431285858, 'epoch': 0.02}

  5%|▌         | 171/3300 [24:19<6:03:09,  6.96s/it]INFO 04-13 07:45:17 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 172/3300 [24:25<5:55:36,  6.82s/it]2026-04-13 07:45:23,301 | INFO | train_grpo_train | metrics_logged mode=train step=172

                                                    
{'loss': -0.1089, 'grad_norm': 19.612377166748047, 'learning_rate': 9.481818181818182e-06, 'num_tokens': 317117.0, 'completions/mean_length': 33.375, 'completions/min_length': 24.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5997594594955444, 'rewards/meter/std': 0.3780299425125122, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9781632423400879, 'rewards/repeat_soft/std': 0.02724776230752468, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.5745712518692017, 'rewards/total_composite/std': 0.17878694832324982, 'reward': 0.5745712518692017, 'reward_std': 0.17878693342208862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2273896038532257, 'sampling/sampling_logp_difference/max': 1.4090204238891602, 'sampling/importance_sampling_ratio/min': 0.2443825751543045, 'sampling/importance_sampling_ratio/mean': 1.0568751096725464, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2210111021995544, 'clip_ratio/low_mean': 0.07622195687144995, 'clip_ratio/low_min': 0.07622195687144995, 'clip_ratio/high_mean': 0.11674907617270947, 'clip_ratio/high_max': 0.11674907617270947, 'clip_ratio/region_mean': 0.1929710330441594, 'reward_total_mean': 0.5745712518692017, 'reward_meter_mean': 0.5997594594955444, 'reward_meter_std': 0.3780299425125122, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9781632423400879, 'reward_repeat_soft_std': 0.02724776230752468, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5745712518692017, 'reward_total_composite_std': 0.17878694832324982, 'epoch': 0.02}

  5%|▌         | 172/3300 [24:25<5:55:36,  6.82s/it]INFO 04-13 07:45:23 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 173/3300 [24:32<5:56:47,  6.85s/it]2026-04-13 07:45:30,206 | INFO | train_grpo_train | metrics_logged mode=train step=173

                                                    
{'loss': -0.0445, 'grad_norm': 16.774700164794922, 'learning_rate': 9.47878787878788e-06, 'num_tokens': 318678.0, 'completions/mean_length': 31.125, 'completions/min_length': 24.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5051141381263733, 'rewards/meter/std': 0.3768838942050934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.951198160648346, 'rewards/repeat_soft/std': 0.023612266406416893, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.3564407229423523, 'rewards/total_composite/mean': 0.5817102193832397, 'rewards/total_composite/std': 0.2026408612728119, 'reward': 0.5817102193832397, 'reward_std': 0.20264087617397308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19832772016525269, 'sampling/sampling_logp_difference/max': 1.467167854309082, 'sampling/importance_sampling_ratio/min': 0.23057760298252106, 'sampling/importance_sampling_ratio/mean': 1.015305519104004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5307548269629478, 'clip_ratio/low_mean': 0.08864182699471712, 'clip_ratio/low_min': 0.08864182699471712, 'clip_ratio/high_mean': 0.11992127541452646, 'clip_ratio/high_max': 0.11992127541452646, 'clip_ratio/region_mean': 0.20856310240924358, 'reward_total_mean': 0.5817102193832397, 'reward_meter_mean': 0.5051141381263733, 'reward_meter_std': 0.3768838942050934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.951198160648346, 'reward_repeat_soft_std': 0.023612266406416893, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.3564407229423523, 'reward_total_composite_mean': 0.5817102193832397, 'reward_total_composite_std': 0.2026408612728119, 'epoch': 0.02}

  5%|▌         | 173/3300 [24:32<5:56:47,  6.85s/it]INFO 04-13 07:45:30 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 174/3300 [24:39<5:58:27,  6.88s/it]2026-04-13 07:45:37,165 | INFO | train_grpo_train | metrics_logged mode=train step=174

                                                    
{'loss': 0.0959, 'grad_norm': 10.710179328918457, 'learning_rate': 9.475757575757577e-06, 'num_tokens': 320342.0, 'completions/mean_length': 51.0, 'completions/min_length': 38.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.15989206731319427, 'rewards/meter/std': 0.31738677620887756, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953173398971558, 'rewards/repeat_soft/std': 0.007823141291737556, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.39320605993270874, 'rewards/total_composite/std': 0.08556443452835083, 'reward': 0.39320605993270874, 'reward_std': 0.08556442707777023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24069498479366302, 'sampling/sampling_logp_difference/max': 1.5249691009521484, 'sampling/importance_sampling_ratio/min': 0.21762779355049133, 'sampling/importance_sampling_ratio/mean': 1.0592713356018066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0124190747737885, 'clip_ratio/low_mean': 0.15173242706805468, 'clip_ratio/low_min': 0.15173242706805468, 'clip_ratio/high_mean': 0.037790697067976, 'clip_ratio/high_max': 0.037790697067976, 'clip_ratio/region_mean': 0.18952312413603067, 'reward_total_mean': 0.39320605993270874, 'reward_meter_mean': 0.15989206731319427, 'reward_meter_std': 0.31738677620887756, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953173398971558, 'reward_repeat_soft_std': 0.007823141291737556, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.39320605993270874, 'reward_total_composite_std': 0.08556443452835083, 'epoch': 0.02}

  5%|▌         | 174/3300 [24:39<5:58:27,  6.88s/it]INFO 04-13 07:45:37 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 175/3300 [24:47<6:13:19,  7.17s/it]2026-04-13 07:45:45,005 | INFO | train_grpo_train | metrics_logged mode=train step=175

                                                    
{'loss': 0.1424, 'grad_norm': 16.74935531616211, 'learning_rate': 9.472727272727274e-06, 'num_tokens': 321934.0, 'completions/mean_length': 44.0, 'completions/min_length': 33.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8714777827262878, 'rewards/meter/std': 0.27093425393104553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9930651187896729, 'rewards/repeat_soft/std': 0.00821753591299057, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.17210878431797028, 'rewards/total_composite/mean': 0.7505009174346924, 'rewards/total_composite/std': 0.1681082397699356, 'reward': 0.7505009174346924, 'reward_std': 0.16810822486877441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16729728877544403, 'sampling/sampling_logp_difference/max': 1.657027244567871, 'sampling/importance_sampling_ratio/min': 0.1907050460577011, 'sampling/importance_sampling_ratio/mean': 1.0261846780776978, 'sampling/importance_sampling_ratio/max': 1.8706589937210083, 'entropy': 1.6106469556689262, 'clip_ratio/low_mean': 0.039352625608444214, 'clip_ratio/low_min': 0.039352625608444214, 'clip_ratio/high_mean': 0.11767790373414755, 'clip_ratio/high_max': 0.11767790373414755, 'clip_ratio/region_mean': 0.15703052934259176, 'reward_total_mean': 0.7505009174346924, 'reward_meter_mean': 0.8714777827262878, 'reward_meter_std': 0.27093425393104553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9930651187896729, 'reward_repeat_soft_std': 0.00821753591299057, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.17210878431797028, 'reward_total_composite_mean': 0.7505009174346924, 'reward_total_composite_std': 0.1681082397699356, 'epoch': 0.02}

  5%|▌         | 175/3300 [24:47<6:13:19,  7.17s/it]INFO 04-13 07:45:45 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 176/3300 [24:53<6:03:45,  6.99s/it]2026-04-13 07:45:51,568 | INFO | train_grpo_train | metrics_logged mode=train step=176

                                                    
{'loss': 0.067, 'grad_norm': 23.75896453857422, 'learning_rate': 9.469696969696971e-06, 'num_tokens': 323519.0, 'completions/mean_length': 40.125, 'completions/min_length': 34.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5985807180404663, 'rewards/meter/std': 0.43099549412727356, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.992364764213562, 'rewards/repeat_soft/std': 0.007917179726064205, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.587218165397644, 'rewards/total_composite/std': 0.3418886363506317, 'reward': 0.587218165397644, 'reward_std': 0.3418886363506317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21122638881206512, 'sampling/sampling_logp_difference/max': 2.2637810707092285, 'sampling/importance_sampling_ratio/min': 0.1039566695690155, 'sampling/importance_sampling_ratio/mean': 1.0143921375274658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0883154422044754, 'clip_ratio/low_mean': 0.122177729383111, 'clip_ratio/low_min': 0.122177729383111, 'clip_ratio/high_mean': 0.07360660284757614, 'clip_ratio/high_max': 0.07360660284757614, 'clip_ratio/region_mean': 0.19578433223068714, 'reward_total_mean': 0.587218165397644, 'reward_meter_mean': 0.5985807180404663, 'reward_meter_std': 0.43099549412727356, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.992364764213562, 'reward_repeat_soft_std': 0.007917179726064205, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.587218165397644, 'reward_total_composite_std': 0.3418886363506317, 'epoch': 0.02}

  5%|▌         | 176/3300 [24:53<6:03:45,  6.99s/it]INFO 04-13 07:45:51 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 177/3300 [25:01<6:18:15,  7.27s/it]2026-04-13 07:45:59,489 | INFO | train_grpo_train | metrics_logged mode=train step=177

                                                    
{'loss': 0.0198, 'grad_norm': 13.65584659576416, 'learning_rate': 9.466666666666667e-06, 'num_tokens': 325161.0, 'completions/mean_length': 54.25, 'completions/min_length': 36.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.3896534740924835, 'rewards/meter/std': 0.318272203207016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989882707595825, 'rewards/repeat_soft/std': 0.002861686749383807, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.481770396232605, 'rewards/total_composite/std': 0.10922771692276001, 'reward': 0.481770396232605, 'reward_std': 0.10922770202159882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22601468861103058, 'sampling/sampling_logp_difference/max': 1.635115623474121, 'sampling/importance_sampling_ratio/min': 0.1949298232793808, 'sampling/importance_sampling_ratio/mean': 1.0414087772369385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6458567082881927, 'clip_ratio/low_mean': 0.0945226140320301, 'clip_ratio/low_min': 0.0945226140320301, 'clip_ratio/high_mean': 0.08183415047824383, 'clip_ratio/high_max': 0.08183415047824383, 'clip_ratio/region_mean': 0.17635676451027393, 'reward_total_mean': 0.481770396232605, 'reward_meter_mean': 0.3896534740924835, 'reward_meter_std': 0.318272203207016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989882707595825, 'reward_repeat_soft_std': 0.002861686749383807, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.481770396232605, 'reward_total_composite_std': 0.10922771692276001, 'epoch': 0.02}

  5%|▌         | 177/3300 [25:01<6:18:15,  7.27s/it]INFO 04-13 07:45:59 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 178/3300 [25:10<6:34:27,  7.58s/it]2026-04-13 07:46:07,802 | INFO | train_grpo_train | metrics_logged mode=train step=178

                                                    
{'loss': 0.0195, 'grad_norm': 8.766780853271484, 'learning_rate': 9.463636363636364e-06, 'num_tokens': 327668.0, 'completions/mean_length': 115.375, 'completions/min_length': 95.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.375, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.7110608816146851, 'rewards/meter/std': 0.3456106185913086, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9969117641448975, 'rewards/repeat_soft/std': 0.0029427569825202227, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.43055757880210876, 'rewards/total_composite/std': 0.20038874447345734, 'reward': 0.43055757880210876, 'reward_std': 0.20038872957229614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2415567934513092, 'sampling/sampling_logp_difference/max': 1.5120110511779785, 'sampling/importance_sampling_ratio/min': 0.22046616673469543, 'sampling/importance_sampling_ratio/mean': 1.0475821495056152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4164766371250153, 'clip_ratio/low_mean': 0.07863680832087994, 'clip_ratio/low_min': 0.07863680832087994, 'clip_ratio/high_mean': 0.1258973926305771, 'clip_ratio/high_max': 0.1258973926305771, 'clip_ratio/region_mean': 0.20453420095145702, 'reward_total_mean': 0.43055757880210876, 'reward_meter_mean': 0.7110608816146851, 'reward_meter_std': 0.3456106185913086, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9969117641448975, 'reward_repeat_soft_std': 0.0029427569825202227, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.43055757880210876, 'reward_total_composite_std': 0.20038874447345734, 'epoch': 0.02}

  5%|▌         | 178/3300 [25:10<6:34:27,  7.58s/it]INFO 04-13 07:46:08 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 179/3300 [25:17<6:27:00,  7.44s/it]2026-04-13 07:46:14,914 | INFO | train_grpo_train | metrics_logged mode=train step=179

                                                    
{'loss': 0.0365, 'grad_norm': 11.071876525878906, 'learning_rate': 9.460606060606061e-06, 'num_tokens': 329759.0, 'completions/mean_length': 85.375, 'completions/min_length': 73.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.4272414445877075, 'rewards/meter/std': 0.25779473781585693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996740996837616, 'rewards/repeat_soft/std': 0.0038489243015646935, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.47163498401641846, 'rewards/total_composite/std': 0.0906149297952652, 'reward': 0.47163498401641846, 'reward_std': 0.0906149372458458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24609750509262085, 'sampling/sampling_logp_difference/max': 1.5515518188476562, 'sampling/importance_sampling_ratio/min': 0.21191886067390442, 'sampling/importance_sampling_ratio/mean': 1.0612713098526, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.618737071752548, 'clip_ratio/low_mean': 0.10393072944134474, 'clip_ratio/low_min': 0.10393072944134474, 'clip_ratio/high_mean': 0.07959592528641224, 'clip_ratio/high_max': 0.07959592528641224, 'clip_ratio/region_mean': 0.18352665472775698, 'reward_total_mean': 0.47163498401641846, 'reward_meter_mean': 0.4272414445877075, 'reward_meter_std': 0.25779473781585693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996740996837616, 'reward_repeat_soft_std': 0.0038489243015646935, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.47163498401641846, 'reward_total_composite_std': 0.0906149297952652, 'epoch': 0.02}

  5%|▌         | 179/3300 [25:17<6:27:00,  7.44s/it]INFO 04-13 07:46:15 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 180/3300 [25:24<6:31:19,  7.53s/it]2026-04-13 07:46:22,641 | INFO | train_grpo_train | metrics_logged mode=train step=180

                                                    
{'loss': 0.0235, 'grad_norm': 11.487159729003906, 'learning_rate': 9.457575757575759e-06, 'num_tokens': 331979.0, 'completions/mean_length': 79.5, 'completions/min_length': 61.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.31564873456954956, 'rewards/meter/std': 0.30156204104423523, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9972239136695862, 'rewards/repeat_soft/std': 0.0030039141420274973, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.414653480052948, 'rewards/total_composite/std': 0.08931313455104828, 'reward': 0.414653480052948, 'reward_std': 0.08931312710046768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2503102719783783, 'sampling/sampling_logp_difference/max': 1.6738219261169434, 'sampling/importance_sampling_ratio/min': 0.18752898275852203, 'sampling/importance_sampling_ratio/mean': 1.042333960533142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0768066942691803, 'clip_ratio/low_mean': 0.13392947241663933, 'clip_ratio/low_min': 0.13392947241663933, 'clip_ratio/high_mean': 0.0785590372979641, 'clip_ratio/high_max': 0.0785590372979641, 'clip_ratio/region_mean': 0.21248850971460342, 'reward_total_mean': 0.414653480052948, 'reward_meter_mean': 0.31564873456954956, 'reward_meter_std': 0.30156204104423523, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9972239136695862, 'reward_repeat_soft_std': 0.0030039141420274973, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.414653480052948, 'reward_total_composite_std': 0.08931313455104828, 'epoch': 0.02}

  5%|▌         | 180/3300 [25:24<6:31:19,  7.53s/it]INFO 04-13 07:46:22 [block_pool.py:378] Successfully reset prefix cache

  5%|▌         | 181/3300 [25:32<6:32:24,  7.55s/it]2026-04-13 07:46:30,242 | INFO | train_grpo_train | metrics_logged mode=train step=181

                                                    
{'loss': 0.0879, 'grad_norm': 12.22265338897705, 'learning_rate': 9.454545454545456e-06, 'num_tokens': 333734.0, 'completions/mean_length': 59.375, 'completions/min_length': 46.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.4035565257072449, 'rewards/meter/std': 0.4075183570384979, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969116449356079, 'rewards/repeat_soft/std': 0.0037362936418503523, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.2327437400817871, 'rewards/total_composite/mean': 0.48305100202560425, 'rewards/total_composite/std': 0.20548434555530548, 'reward': 0.48305100202560425, 'reward_std': 0.20548434555530548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2156466543674469, 'sampling/sampling_logp_difference/max': 1.5678300857543945, 'sampling/importance_sampling_ratio/min': 0.20849710702896118, 'sampling/importance_sampling_ratio/mean': 1.0440673828125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9228354692459106, 'clip_ratio/low_mean': 0.1170395202934742, 'clip_ratio/low_min': 0.1170395202934742, 'clip_ratio/high_mean': 0.06735390983521938, 'clip_ratio/high_max': 0.06735390983521938, 'clip_ratio/region_mean': 0.18439343012869358, 'reward_total_mean': 0.48305100202560425, 'reward_meter_mean': 0.4035565257072449, 'reward_meter_std': 0.4075183570384979, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969116449356079, 'reward_repeat_soft_std': 0.0037362936418503523, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.2327437400817871, 'reward_total_composite_mean': 0.48305100202560425, 'reward_total_composite_std': 0.20548434555530548, 'epoch': 0.02}

  5%|▌         | 181/3300 [25:32<6:32:24,  7.55s/it]INFO 04-13 07:46:30 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 182/3300 [25:38<6:15:20,  7.22s/it]2026-04-13 07:46:36,705 | INFO | train_grpo_train | metrics_logged mode=train step=182

                                                    
{'loss': 0.2062, 'grad_norm': 26.969940185546875, 'learning_rate': 9.451515151515153e-06, 'num_tokens': 335325.0, 'completions/mean_length': 34.875, 'completions/min_length': 24.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5274512767791748, 'rewards/meter/std': 0.4633592963218689, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.977503776550293, 'rewards/repeat_soft/std': 0.020300719887018204, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5399748086929321, 'rewards/total_composite/std': 0.1919652372598648, 'reward': 0.5399748086929321, 'reward_std': 0.1919652372598648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23039452731609344, 'sampling/sampling_logp_difference/max': 2.461104393005371, 'sampling/importance_sampling_ratio/min': 0.08534064888954163, 'sampling/importance_sampling_ratio/mean': 0.9769449234008789, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1152411550283432, 'clip_ratio/low_mean': 0.07630632631480694, 'clip_ratio/low_min': 0.07630632631480694, 'clip_ratio/high_mean': 0.14935007132589817, 'clip_ratio/high_max': 0.14935007132589817, 'clip_ratio/region_mean': 0.2256563976407051, 'reward_total_mean': 0.5399748086929321, 'reward_meter_mean': 0.5274512767791748, 'reward_meter_std': 0.4633592963218689, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.977503776550293, 'reward_repeat_soft_std': 0.020300719887018204, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5399748086929321, 'reward_total_composite_std': 0.1919652372598648, 'epoch': 0.02}

  6%|▌         | 182/3300 [25:38<6:15:20,  7.22s/it]INFO 04-13 07:46:36 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 183/3300 [25:45<6:00:20,  6.94s/it]2026-04-13 07:46:42,972 | INFO | train_grpo_train | metrics_logged mode=train step=183

                                                    
{'loss': 0.047, 'grad_norm': 23.409812927246094, 'learning_rate': 9.448484848484849e-06, 'num_tokens': 336977.0, 'completions/mean_length': 33.5, 'completions/min_length': 30.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8624000549316406, 'rewards/meter/std': 0.3189335763454437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9998470544815063, 'rewards/repeat_soft/std': 0.00043267954606562853, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.2512255907058716, 'rewards/total_composite/mean': 0.7162233591079712, 'rewards/total_composite/std': 0.19540704786777496, 'reward': 0.7162233591079712, 'reward_std': 0.19540704786777496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14572322368621826, 'sampling/sampling_logp_difference/max': 1.710498332977295, 'sampling/importance_sampling_ratio/min': 0.1807756870985031, 'sampling/importance_sampling_ratio/mean': 1.0187908411026, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.661162231117487, 'clip_ratio/low_mean': 0.06761159980669618, 'clip_ratio/low_min': 0.06761159980669618, 'clip_ratio/high_mean': 0.0531698577105999, 'clip_ratio/high_max': 0.0531698577105999, 'clip_ratio/region_mean': 0.12078145751729608, 'reward_total_mean': 0.7162233591079712, 'reward_meter_mean': 0.8624000549316406, 'reward_meter_std': 0.3189335763454437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9998470544815063, 'reward_repeat_soft_std': 0.00043267954606562853, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.2512255907058716, 'reward_total_composite_mean': 0.7162233591079712, 'reward_total_composite_std': 0.19540704786777496, 'epoch': 0.02}

  6%|▌         | 183/3300 [25:45<6:00:20,  6.94s/it]INFO 04-13 07:46:43 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 184/3300 [25:51<5:55:39,  6.85s/it]2026-04-13 07:46:49,616 | INFO | train_grpo_train | metrics_logged mode=train step=184

                                                    
{'loss': 0.0993, 'grad_norm': 15.767330169677734, 'learning_rate': 9.445454545454546e-06, 'num_tokens': 338608.0, 'completions/mean_length': 47.875, 'completions/min_length': 37.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5631375908851624, 'rewards/meter/std': 0.3976224958896637, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984055757522583, 'rewards/repeat_soft/std': 0.0025164084509015083, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.09754122048616409, 'rewards/total_composite/mean': 0.5147405862808228, 'rewards/total_composite/std': 0.11590421199798584, 'reward': 0.5147405862808228, 'reward_std': 0.11590423434972763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22204120457172394, 'sampling/sampling_logp_difference/max': 1.1929655075073242, 'sampling/importance_sampling_ratio/min': 0.30332043766975403, 'sampling/importance_sampling_ratio/mean': 1.0571694374084473, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.678367853164673, 'clip_ratio/low_mean': 0.0853744950145483, 'clip_ratio/low_min': 0.0853744950145483, 'clip_ratio/high_mean': 0.12821896374225616, 'clip_ratio/high_max': 0.12821896374225616, 'clip_ratio/region_mean': 0.21359345875680447, 'reward_total_mean': 0.5147405862808228, 'reward_meter_mean': 0.5631375908851624, 'reward_meter_std': 0.3976224958896637, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984055757522583, 'reward_repeat_soft_std': 0.0025164084509015083, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.09754122048616409, 'reward_total_composite_mean': 0.5147405862808228, 'reward_total_composite_std': 0.11590421199798584, 'epoch': 0.02}

  6%|▌         | 184/3300 [25:51<5:55:39,  6.85s/it]INFO 04-13 07:46:49 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 185/3300 [25:58<5:50:06,  6.74s/it]2026-04-13 07:46:56,116 | INFO | train_grpo_train | metrics_logged mode=train step=185

                                                    
{'loss': -0.0371, 'grad_norm': 17.47907257080078, 'learning_rate': 9.442424242424243e-06, 'num_tokens': 340039.0, 'completions/mean_length': 26.875, 'completions/min_length': 19.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.549547553062439, 'rewards/meter/std': 0.46725091338157654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9615821838378906, 'rewards/repeat_soft/std': 0.002596014179289341, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5350043773651123, 'rewards/total_composite/std': 0.19496038556098938, 'reward': 0.5350043773651123, 'reward_std': 0.19496038556098938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21452338993549347, 'sampling/sampling_logp_difference/max': 1.4326086044311523, 'sampling/importance_sampling_ratio/min': 0.2386854887008667, 'sampling/importance_sampling_ratio/mean': 1.0544756650924683, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4443026930093765, 'clip_ratio/low_mean': 0.11357326246798038, 'clip_ratio/low_min': 0.11357326246798038, 'clip_ratio/high_mean': 0.12992424704134464, 'clip_ratio/high_max': 0.12992424704134464, 'clip_ratio/region_mean': 0.24349750950932503, 'reward_total_mean': 0.5350043773651123, 'reward_meter_mean': 0.549547553062439, 'reward_meter_std': 0.46725091338157654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9615821838378906, 'reward_repeat_soft_std': 0.002596014179289341, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5350043773651123, 'reward_total_composite_std': 0.19496038556098938, 'epoch': 0.02}

  6%|▌         | 185/3300 [25:58<5:50:06,  6.74s/it]INFO 04-13 07:46:56 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 186/3300 [26:05<5:51:49,  6.78s/it]2026-04-13 07:47:02,976 | INFO | train_grpo_train | metrics_logged mode=train step=186

                                                    
{'loss': 0.0235, 'grad_norm': 17.311351776123047, 'learning_rate': 9.43939393939394e-06, 'num_tokens': 342081.0, 'completions/mean_length': 72.25, 'completions/min_length': 54.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.3997536301612854, 'rewards/meter/std': 0.37315547466278076, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9957786798477173, 'rewards/repeat_soft/std': 0.005814037751406431, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.398880273103714, 'rewards/total_composite/std': 0.1924026906490326, 'reward': 0.398880273103714, 'reward_std': 0.1924026757478714, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.262917697429657, 'sampling/sampling_logp_difference/max': 3.0831079483032227, 'sampling/importance_sampling_ratio/min': 0.04581664130091667, 'sampling/importance_sampling_ratio/mean': 1.00346040725708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.003795847296715, 'clip_ratio/low_mean': 0.12207727693021297, 'clip_ratio/low_min': 0.12207727693021297, 'clip_ratio/high_mean': 0.082015885040164, 'clip_ratio/high_max': 0.082015885040164, 'clip_ratio/region_mean': 0.20409316197037697, 'reward_total_mean': 0.398880273103714, 'reward_meter_mean': 0.3997536301612854, 'reward_meter_std': 0.37315547466278076, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9957786798477173, 'reward_repeat_soft_std': 0.005814037751406431, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.398880273103714, 'reward_total_composite_std': 0.1924026906490326, 'epoch': 0.02}

  6%|▌         | 186/3300 [26:05<5:51:49,  6.78s/it]INFO 04-13 07:47:03 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 187/3300 [26:13<6:12:30,  7.18s/it]2026-04-13 07:47:11,091 | INFO | train_grpo_train | metrics_logged mode=train step=187

                                                    
{'loss': 0.0127, 'grad_norm': 7.8910722732543945, 'learning_rate': 9.436363636363636e-06, 'num_tokens': 344835.0, 'completions/mean_length': 146.25, 'completions/min_length': 104.0, 'completions/max_length': 181.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.25, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.5559006929397583, 'rewards/meter/std': 0.28145334124565125, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9929571151733398, 'rewards/repeat_soft/std': 0.007836537435650826, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4760981798171997, 'rewards/total_composite/std': 0.09441790729761124, 'reward': 0.4760981798171997, 'reward_std': 0.09441790729761124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22341546416282654, 'sampling/sampling_logp_difference/max': 1.3534693717956543, 'sampling/importance_sampling_ratio/min': 0.2583424150943756, 'sampling/importance_sampling_ratio/mean': 1.062045693397522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1189622282981873, 'clip_ratio/low_mean': 0.0684958714991808, 'clip_ratio/low_min': 0.0684958714991808, 'clip_ratio/high_mean': 0.1289503499865532, 'clip_ratio/high_max': 0.1289503499865532, 'clip_ratio/region_mean': 0.19744622148573399, 'reward_total_mean': 0.4760981798171997, 'reward_meter_mean': 0.5559006929397583, 'reward_meter_std': 0.28145334124565125, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9929571151733398, 'reward_repeat_soft_std': 0.007836537435650826, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4760981798171997, 'reward_total_composite_std': 0.09441790729761124, 'epoch': 0.02}

  6%|▌         | 187/3300 [26:13<6:12:30,  7.18s/it]INFO 04-13 07:47:11 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 188/3300 [26:20<6:05:36,  7.05s/it]2026-04-13 07:47:17,835 | INFO | train_grpo_train | metrics_logged mode=train step=188

                                                    
{'loss': -0.1187, 'grad_norm': 15.400172233581543, 'learning_rate': 9.433333333333335e-06, 'num_tokens': 346495.0, 'completions/mean_length': 50.5, 'completions/min_length': 39.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8723356127738953, 'rewards/meter/std': 0.2369154691696167, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9981918334960938, 'rewards/repeat_soft/std': 0.004290272947400808, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.6304661631584167, 'rewards/total_composite/std': 0.14154669642448425, 'reward': 0.6304661631584167, 'reward_std': 0.14154671132564545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23977959156036377, 'sampling/sampling_logp_difference/max': 1.277449607849121, 'sampling/importance_sampling_ratio/min': 0.2787473201751709, 'sampling/importance_sampling_ratio/mean': 1.0627295970916748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.198775589466095, 'clip_ratio/low_mean': 0.1782637145370245, 'clip_ratio/low_min': 0.1782637145370245, 'clip_ratio/high_mean': 0.045219458639621735, 'clip_ratio/high_max': 0.045219458639621735, 'clip_ratio/region_mean': 0.22348317317664623, 'reward_total_mean': 0.6304661631584167, 'reward_meter_mean': 0.8723356127738953, 'reward_meter_std': 0.2369154691696167, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9981918334960938, 'reward_repeat_soft_std': 0.004290272947400808, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.6304661631584167, 'reward_total_composite_std': 0.14154669642448425, 'epoch': 0.02}

  6%|▌         | 188/3300 [26:20<6:05:36,  7.05s/it]INFO 04-13 07:47:18 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 189/3300 [26:27<6:06:10,  7.06s/it]2026-04-13 07:47:24,937 | INFO | train_grpo_train | metrics_logged mode=train step=189

                                                    
{'loss': 0.1108, 'grad_norm': 35.58707809448242, 'learning_rate': 9.43030303030303e-06, 'num_tokens': 348008.0, 'completions/mean_length': 33.125, 'completions/min_length': 29.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6826807260513306, 'rewards/meter/std': 0.306144118309021, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9814399480819702, 'rewards/repeat_soft/std': 0.027393506839871407, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6602128744125366, 'rewards/total_composite/std': 0.3298247158527374, 'reward': 0.6602128744125366, 'reward_std': 0.32982468605041504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23146715760231018, 'sampling/sampling_logp_difference/max': 2.0614590644836426, 'sampling/importance_sampling_ratio/min': 0.12726813554763794, 'sampling/importance_sampling_ratio/mean': 1.003521203994751, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.97225097194314, 'clip_ratio/low_mean': 0.07323569525033236, 'clip_ratio/low_min': 0.07323569525033236, 'clip_ratio/high_mean': 0.10484400764107704, 'clip_ratio/high_max': 0.10484400764107704, 'clip_ratio/region_mean': 0.1780797028914094, 'reward_total_mean': 0.6602128744125366, 'reward_meter_mean': 0.6826807260513306, 'reward_meter_std': 0.306144118309021, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9814399480819702, 'reward_repeat_soft_std': 0.027393506839871407, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6602128744125366, 'reward_total_composite_std': 0.3298247158527374, 'epoch': 0.02}

  6%|▌         | 189/3300 [26:27<6:06:10,  7.06s/it]INFO 04-13 07:47:25 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 190/3300 [26:34<6:14:29,  7.22s/it]2026-04-13 07:47:32,541 | INFO | train_grpo_train | metrics_logged mode=train step=190

                                                    
{'loss': 0.068, 'grad_norm': 20.739145278930664, 'learning_rate': 9.427272727272728e-06, 'num_tokens': 349685.0, 'completions/mean_length': 40.625, 'completions/min_length': 16.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.3190385401248932, 'rewards/meter/std': 0.38571831583976746, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9899997115135193, 'rewards/repeat_soft/std': 0.01640298217535019, 'rewards/judge_quality/mean': 0.45124998688697815, 'rewards/judge_quality/std': 0.13767845928668976, 'rewards/total_composite/mean': 0.38480043411254883, 'rewards/total_composite/std': 0.1965453326702118, 'reward': 0.38480043411254883, 'reward_std': 0.1965453326702118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25145721435546875, 'sampling/sampling_logp_difference/max': 1.5635175704956055, 'sampling/importance_sampling_ratio/min': 0.20939819514751434, 'sampling/importance_sampling_ratio/mean': 1.0337738990783691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.717229127883911, 'clip_ratio/low_mean': 0.11086361482739449, 'clip_ratio/low_min': 0.11086361482739449, 'clip_ratio/high_mean': 0.1254812590777874, 'clip_ratio/high_max': 0.1254812590777874, 'clip_ratio/region_mean': 0.23634487390518188, 'reward_total_mean': 0.38480043411254883, 'reward_meter_mean': 0.3190385401248932, 'reward_meter_std': 0.38571831583976746, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9899997115135193, 'reward_repeat_soft_std': 0.01640298217535019, 'reward_judge_quality_mean': 0.45124998688697815, 'reward_judge_quality_std': 0.13767845928668976, 'reward_total_composite_mean': 0.38480043411254883, 'reward_total_composite_std': 0.1965453326702118, 'epoch': 0.02}

  6%|▌         | 190/3300 [26:34<6:14:29,  7.22s/it]INFO 04-13 07:47:32 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 191/3300 [26:44<6:50:19,  7.92s/it]2026-04-13 07:47:42,078 | INFO | train_grpo_train | metrics_logged mode=train step=191

                                                    
{'loss': 0.0167, 'grad_norm': 7.381566524505615, 'learning_rate': 9.424242424242425e-06, 'num_tokens': 352401.0, 'completions/mean_length': 154.5, 'completions/min_length': 111.0, 'completions/max_length': 185.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.5, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.8647301197052002, 'rewards/meter/std': 0.28420892357826233, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9960473775863647, 'rewards/repeat_soft/std': 0.004961573984473944, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.48406609892845154, 'rewards/total_composite/std': 0.21130195260047913, 'reward': 0.48406609892845154, 'reward_std': 0.21130193769931793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20887885987758636, 'sampling/sampling_logp_difference/max': 2.524847984313965, 'sampling/importance_sampling_ratio/min': 0.08007048070430756, 'sampling/importance_sampling_ratio/mean': 1.0513540506362915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.996433287858963, 'clip_ratio/low_mean': 0.0464102178812027, 'clip_ratio/low_min': 0.0464102178812027, 'clip_ratio/high_mean': 0.1237588468939066, 'clip_ratio/high_max': 0.1237588468939066, 'clip_ratio/region_mean': 0.1701690647751093, 'reward_total_mean': 0.48406609892845154, 'reward_meter_mean': 0.8647301197052002, 'reward_meter_std': 0.28420892357826233, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9960473775863647, 'reward_repeat_soft_std': 0.004961573984473944, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.48406609892845154, 'reward_total_composite_std': 0.21130195260047913, 'epoch': 0.02}

  6%|▌         | 191/3300 [26:44<6:50:19,  7.92s/it]INFO 04-13 07:47:42 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 192/3300 [26:51<6:43:18,  7.79s/it]2026-04-13 07:47:49,555 | INFO | train_grpo_train | metrics_logged mode=train step=192

                                                    
{'loss': -0.0564, 'grad_norm': 21.528614044189453, 'learning_rate': 9.421212121212122e-06, 'num_tokens': 354050.0, 'completions/mean_length': 35.125, 'completions/min_length': 22.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.4956907033920288, 'rewards/meter/std': 0.35053014755249023, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9955316781997681, 'rewards/repeat_soft/std': 0.006094147451221943, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.48836007714271545, 'rewards/total_composite/std': 0.09807529300451279, 'reward': 0.48836007714271545, 'reward_std': 0.09807528555393219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2227940410375595, 'sampling/sampling_logp_difference/max': 1.2411231994628906, 'sampling/importance_sampling_ratio/min': 0.2890593707561493, 'sampling/importance_sampling_ratio/mean': 1.0087708234786987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9255905747413635, 'clip_ratio/low_mean': 0.12833823263645172, 'clip_ratio/low_min': 0.12833823263645172, 'clip_ratio/high_mean': 0.08295800909399986, 'clip_ratio/high_max': 0.08295800909399986, 'clip_ratio/region_mean': 0.21129624173045158, 'reward_total_mean': 0.48836007714271545, 'reward_meter_mean': 0.4956907033920288, 'reward_meter_std': 0.35053014755249023, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9955316781997681, 'reward_repeat_soft_std': 0.006094147451221943, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.48836007714271545, 'reward_total_composite_std': 0.09807529300451279, 'epoch': 0.02}

  6%|▌         | 192/3300 [26:51<6:43:18,  7.79s/it]INFO 04-13 07:47:49 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 193/3300 [26:59<6:39:15,  7.71s/it]2026-04-13 07:47:57,080 | INFO | train_grpo_train | metrics_logged mode=train step=193

                                                    
{'loss': -0.1608, 'grad_norm': 13.815486907958984, 'learning_rate': 9.418181818181818e-06, 'num_tokens': 355910.0, 'completions/mean_length': 54.5, 'completions/min_length': 17.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6727335453033447, 'rewards/meter/std': 0.29759544134140015, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9905915260314941, 'rewards/repeat_soft/std': 0.013315845280885696, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.4954334795475006, 'rewards/total_composite/std': 0.09651092439889908, 'reward': 0.4954334795475006, 'reward_std': 0.09651093184947968, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24432682991027832, 'sampling/sampling_logp_difference/max': 1.2650394439697266, 'sampling/importance_sampling_ratio/min': 0.28222814202308655, 'sampling/importance_sampling_ratio/mean': 1.0481576919555664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2915859818458557, 'clip_ratio/low_mean': 0.10347050474956632, 'clip_ratio/low_min': 0.10347050474956632, 'clip_ratio/high_mean': 0.08240798488259315, 'clip_ratio/high_max': 0.08240798488259315, 'clip_ratio/region_mean': 0.18587848963215947, 'reward_total_mean': 0.4954334795475006, 'reward_meter_mean': 0.6727335453033447, 'reward_meter_std': 0.29759544134140015, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9905915260314941, 'reward_repeat_soft_std': 0.013315845280885696, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.4954334795475006, 'reward_total_composite_std': 0.09651092439889908, 'epoch': 0.02}

  6%|▌         | 193/3300 [26:59<6:39:15,  7.71s/it]INFO 04-13 07:47:57 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 194/3300 [27:07<6:42:21,  7.77s/it]2026-04-13 07:48:05,006 | INFO | train_grpo_train | metrics_logged mode=train step=194

                                                    
{'loss': 0.2029, 'grad_norm': 12.876761436462402, 'learning_rate': 9.415151515151515e-06, 'num_tokens': 357585.0, 'completions/mean_length': 52.375, 'completions/min_length': 39.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.7496733665466309, 'rewards/meter/std': 0.36694076657295227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9982771873474121, 'rewards/repeat_soft/std': 0.004632346797734499, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.21387162804603577, 'rewards/total_composite/mean': 0.5776740908622742, 'rewards/total_composite/std': 0.1355331391096115, 'reward': 0.5776740908622742, 'reward_std': 0.13553312420845032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20649906992912292, 'sampling/sampling_logp_difference/max': 1.2183427810668945, 'sampling/importance_sampling_ratio/min': 0.29571983218193054, 'sampling/importance_sampling_ratio/mean': 1.0395499467849731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.77356718480587, 'clip_ratio/low_mean': 0.06917979381978512, 'clip_ratio/low_min': 0.06917979381978512, 'clip_ratio/high_mean': 0.12279230542480946, 'clip_ratio/high_max': 0.12279230542480946, 'clip_ratio/region_mean': 0.19197209924459457, 'reward_total_mean': 0.5776740908622742, 'reward_meter_mean': 0.7496733665466309, 'reward_meter_std': 0.36694076657295227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9982771873474121, 'reward_repeat_soft_std': 0.004632346797734499, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.21387162804603577, 'reward_total_composite_mean': 0.5776740908622742, 'reward_total_composite_std': 0.1355331391096115, 'epoch': 0.02}

  6%|▌         | 194/3300 [27:07<6:42:21,  7.77s/it]INFO 04-13 07:48:05 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 195/3300 [27:14<6:36:25,  7.66s/it]2026-04-13 07:48:12,397 | INFO | train_grpo_train | metrics_logged mode=train step=195

                                                    
{'loss': -0.0291, 'grad_norm': 12.050610542297363, 'learning_rate': 9.412121212121212e-06, 'num_tokens': 359238.0, 'completions/mean_length': 47.625, 'completions/min_length': 35.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.2897016704082489, 'rewards/meter/std': 0.3196074366569519, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.998386025428772, 'rewards/repeat_soft/std': 0.0037448338698595762, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.34002625942230225, 'rewards/total_composite/std': 0.2229786217212677, 'reward': 0.34002625942230225, 'reward_std': 0.2229786068201065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23713691532611847, 'sampling/sampling_logp_difference/max': 1.4660263061523438, 'sampling/importance_sampling_ratio/min': 0.23084096610546112, 'sampling/importance_sampling_ratio/mean': 1.0835492610931396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.733154684305191, 'clip_ratio/low_mean': 0.0436507947742939, 'clip_ratio/low_min': 0.0436507947742939, 'clip_ratio/high_mean': 0.16215167567133904, 'clip_ratio/high_max': 0.16215167567133904, 'clip_ratio/region_mean': 0.20580247044563293, 'reward_total_mean': 0.34002625942230225, 'reward_meter_mean': 0.2897016704082489, 'reward_meter_std': 0.3196074366569519, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.998386025428772, 'reward_repeat_soft_std': 0.0037448338698595762, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.34002625942230225, 'reward_total_composite_std': 0.2229786217212677, 'epoch': 0.02}

  6%|▌         | 195/3300 [27:14<6:36:25,  7.66s/it]INFO 04-13 07:48:12 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 196/3300 [27:20<6:15:15,  7.25s/it]2026-04-13 07:48:18,703 | INFO | train_grpo_train | metrics_logged mode=train step=196

                                                    
{'loss': 0.002, 'grad_norm': 21.040252685546875, 'learning_rate': 9.40909090909091e-06, 'num_tokens': 360614.0, 'completions/mean_length': 23.0, 'completions/min_length': 19.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6609889268875122, 'rewards/meter/std': 0.37576085329055786, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9615821838378906, 'rewards/repeat_soft/std': 0.002596014179289341, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.2499571591615677, 'rewards/total_composite/mean': 0.5649632215499878, 'rewards/total_composite/std': 0.1667466014623642, 'reward': 0.5649632215499878, 'reward_std': 0.1667466014623642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18898744881153107, 'sampling/sampling_logp_difference/max': 1.400949478149414, 'sampling/importance_sampling_ratio/min': 0.24636293947696686, 'sampling/importance_sampling_ratio/mean': 1.0581945180892944, 'sampling/importance_sampling_ratio/max': 1.849361538887024, 'entropy': 1.814273864030838, 'clip_ratio/low_mean': 0.06732456292957067, 'clip_ratio/low_min': 0.06732456292957067, 'clip_ratio/high_mean': 0.06317934766411781, 'clip_ratio/high_max': 0.06317934766411781, 'clip_ratio/region_mean': 0.1305039105936885, 'reward_total_mean': 0.5649632215499878, 'reward_meter_mean': 0.6609889268875122, 'reward_meter_std': 0.37576085329055786, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9615821838378906, 'reward_repeat_soft_std': 0.002596014179289341, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.2499571591615677, 'reward_total_composite_mean': 0.5649632215499878, 'reward_total_composite_std': 0.1667466014623642, 'epoch': 0.02}

  6%|▌         | 196/3300 [27:20<6:15:15,  7.25s/it]INFO 04-13 07:48:18 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 197/3300 [27:27<5:57:49,  6.92s/it]2026-04-13 07:48:24,841 | INFO | train_grpo_train | metrics_logged mode=train step=197

                                                    
{'loss': 0.0144, 'grad_norm': 14.87818717956543, 'learning_rate': 9.406060606060607e-06, 'num_tokens': 362018.0, 'completions/mean_length': 28.5, 'completions/min_length': 22.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7591187953948975, 'rewards/meter/std': 0.3778837025165558, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9472886323928833, 'rewards/repeat_soft/std': 0.033697858452796936, 'rewards/judge_quality/mean': 0.71875, 'rewards/judge_quality/std': 0.2992102801799774, 'rewards/total_composite/mean': 0.7138622999191284, 'rewards/total_composite/std': 0.24715933203697205, 'reward': 0.7138622999191284, 'reward_std': 0.24715931713581085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17814406752586365, 'sampling/sampling_logp_difference/max': 1.0547475814819336, 'sampling/importance_sampling_ratio/min': 0.34828031063079834, 'sampling/importance_sampling_ratio/mean': 1.0341427326202393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9291309788823128, 'clip_ratio/low_mean': 0.08132582809776068, 'clip_ratio/low_min': 0.08132582809776068, 'clip_ratio/high_mean': 0.10434755031019449, 'clip_ratio/high_max': 0.10434755031019449, 'clip_ratio/region_mean': 0.18567337840795517, 'reward_total_mean': 0.7138622999191284, 'reward_meter_mean': 0.7591187953948975, 'reward_meter_std': 0.3778837025165558, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9472886323928833, 'reward_repeat_soft_std': 0.033697858452796936, 'reward_judge_quality_mean': 0.71875, 'reward_judge_quality_std': 0.2992102801799774, 'reward_total_composite_mean': 0.7138622999191284, 'reward_total_composite_std': 0.24715933203697205, 'epoch': 0.02}

  6%|▌         | 197/3300 [27:27<5:57:49,  6.92s/it]INFO 04-13 07:48:25 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 198/3300 [27:34<6:00:24,  6.97s/it]2026-04-13 07:48:31,934 | INFO | train_grpo_train | metrics_logged mode=train step=198

                                                    
{'loss': 0.0672, 'grad_norm': 15.537421226501465, 'learning_rate': 9.403030303030304e-06, 'num_tokens': 363963.0, 'completions/mean_length': 60.125, 'completions/min_length': 51.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5443116426467896, 'rewards/meter/std': 0.37749719619750977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9726038575172424, 'rewards/repeat_soft/std': 0.05000071972608566, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.5973585247993469, 'rewards/total_composite/std': 0.20080992579460144, 'reward': 0.5973585247993469, 'reward_std': 0.20080992579460144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18356436491012573, 'sampling/sampling_logp_difference/max': 2.705568313598633, 'sampling/importance_sampling_ratio/min': 0.06683232635259628, 'sampling/importance_sampling_ratio/mean': 1.002237319946289, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0010015740990639, 'clip_ratio/low_mean': 0.11279523558914661, 'clip_ratio/low_min': 0.11279523558914661, 'clip_ratio/high_mean': 0.06206165812909603, 'clip_ratio/high_max': 0.06206165812909603, 'clip_ratio/region_mean': 0.17485689371824265, 'reward_total_mean': 0.5973585247993469, 'reward_meter_mean': 0.5443116426467896, 'reward_meter_std': 0.37749719619750977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9726038575172424, 'reward_repeat_soft_std': 0.05000071972608566, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.5973585247993469, 'reward_total_composite_std': 0.20080992579460144, 'epoch': 0.02}

  6%|▌         | 198/3300 [27:34<6:00:24,  6.97s/it]INFO 04-13 07:48:32 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 199/3300 [27:40<5:55:22,  6.88s/it]2026-04-13 07:48:38,586 | INFO | train_grpo_train | metrics_logged mode=train step=199

                                                    
{'loss': 0.0029, 'grad_norm': 23.622743606567383, 'learning_rate': 9.4e-06, 'num_tokens': 365314.0, 'completions/mean_length': 23.875, 'completions/min_length': 18.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5365124940872192, 'rewards/meter/std': 0.39426350593566895, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.11501552164554596, 'rewards/total_composite/mean': 0.40696874260902405, 'rewards/total_composite/std': 0.1946464478969574, 'reward': 0.40696874260902405, 'reward_std': 0.1946464478969574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23545973002910614, 'sampling/sampling_logp_difference/max': 1.2556962966918945, 'sampling/importance_sampling_ratio/min': 0.2848774194717407, 'sampling/importance_sampling_ratio/mean': 1.0553559064865112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6063058376312256, 'clip_ratio/low_mean': 0.16020440869033337, 'clip_ratio/low_min': 0.16020440869033337, 'clip_ratio/high_mean': 0.09125663712620735, 'clip_ratio/high_max': 0.09125663712620735, 'clip_ratio/region_mean': 0.2514610458165407, 'reward_total_mean': 0.40696874260902405, 'reward_meter_mean': 0.5365124940872192, 'reward_meter_std': 0.39426350593566895, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.11501552164554596, 'reward_total_composite_mean': 0.40696874260902405, 'reward_total_composite_std': 0.1946464478969574, 'epoch': 0.02}

  6%|▌         | 199/3300 [27:40<5:55:22,  6.88s/it]INFO 04-13 07:48:38 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 200/3300 [27:48<6:13:59,  7.24s/it]2026-04-13 07:48:46,671 | INFO | train_grpo_train | metrics_logged mode=train step=200

                                                    
{'loss': 0.0264, 'grad_norm': 6.864881992340088, 'learning_rate': 9.396969696969697e-06, 'num_tokens': 367566.0, 'completions/mean_length': 107.5, 'completions/min_length': 93.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.8985656499862671, 'rewards/meter/std': 0.15440396964550018, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986597299575806, 'rewards/repeat_soft/std': 0.0012854236410930753, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.6393650770187378, 'rewards/total_composite/std': 0.13178274035453796, 'reward': 0.6393650770187378, 'reward_std': 0.13178272545337677, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2367682307958603, 'sampling/sampling_logp_difference/max': 1.4644136428833008, 'sampling/importance_sampling_ratio/min': 0.2312135398387909, 'sampling/importance_sampling_ratio/mean': 1.041608214378357, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4034773111343384, 'clip_ratio/low_mean': 0.1424694685265422, 'clip_ratio/low_min': 0.1424694685265422, 'clip_ratio/high_mean': 0.027499999850988388, 'clip_ratio/high_max': 0.027499999850988388, 'clip_ratio/region_mean': 0.16996946837753057, 'reward_total_mean': 0.6393650770187378, 'reward_meter_mean': 0.8985656499862671, 'reward_meter_std': 0.15440396964550018, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986597299575806, 'reward_repeat_soft_std': 0.0012854236410930753, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.6393650770187378, 'reward_total_composite_std': 0.13178274035453796, 'epoch': 0.02}

  6%|▌         | 200/3300 [27:48<6:13:59,  7.24s/it]INFO 04-13 07:48:46 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:19,  2.43s/it]

 30%|███       | 3/10 [00:09<00:22,  3.28s/it]

 40%|████      | 4/10 [00:13<00:21,  3.61s/it]

 50%|█████     | 5/10 [00:18<00:20,  4.16s/it]

 60%|██████    | 6/10 [00:24<00:19,  4.78s/it]

 70%|███████   | 7/10 [00:30<00:14,  4.96s/it]

 80%|████████  | 8/10 [00:35<00:09,  4.98s/it]

 90%|█████████ | 9/10 [00:39<00:04,  4.91s/it]

100%|██████████| 10/10 [00:43<00:00,  4.61s/it]2026-04-13 07:49:35,968 | INFO | train_grpo_train | metrics_logged mode=eval step=200

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 49.2826, 'eval_samples_per_second': 1.623, 'eval_steps_per_second': 0.203, 'eval_num_tokens': 367566.0, 'eval_completions/mean_length': 73.6125, 'eval_completions/min_length': 29.2, 'eval_completions/max_length': 134.9, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 73.6125, 'eval_completions/min_terminated_length': 29.2, 'eval_completions/max_terminated_length': 134.9, 'eval_rewards/meter/mean': 0.5391432881355286, 'eval_rewards/meter/std': 0.38314679861068723, 'eval_rewards/count_adherence/mean': 0.970000010728836, 'eval_rewards/count_adherence/std': 0.0788449503481388, 'eval_rewards/hard_gate/mean': 0.925, 'eval_rewards/hard_gate/std': 0.18771235942840575, 'eval_rewards/repeat_soft/mean': 0.9907893419265748, 'eval_rewards/repeat_soft/std': 0.012920224107801914, 'eval_rewards/judge_quality/mean': 0.4497499972581863, 'eval_rewards/judge_quality/std': 0.15556477615609765, 'eval_rewards/total_composite/mean': 0.4672420650720596, 'eval_rewards/total_composite/std': 0.18860110267996788, 'eval_reward': 0.4672420650720596, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.16620644852519034, 'eval_sampling/sampling_logp_difference/max': 1.0829068183898927, 'eval_sampling/importance_sampling_ratio/min': 0.34725173711776736, 'eval_sampling/importance_sampling_ratio/mean': 1.0519151449203492, 'eval_sampling/importance_sampling_ratio/max': 1.6473710417747498, 'eval_entropy': 2.6686283588409423, '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': 0.4672420650720596, 'eval_reward_meter_mean': 0.5391432881355286, 'eval_reward_meter_std': 0.38314679861068723, 'eval_reward_count_adherence_mean': 0.970000010728836, 'eval_reward_count_adherence_std': 0.0788449503481388, 'eval_reward_hard_gate_mean': 0.925, 'eval_reward_hard_gate_std': 0.18771235942840575, 'eval_reward_repeat_soft_mean': 0.9907893419265748, 'eval_reward_repeat_soft_std': 0.012920224107801914, 'eval_reward_judge_quality_mean': 0.4497499972581863, 'eval_reward_judge_quality_std': 0.15556477615609765, 'eval_reward_total_composite_mean': 0.4672420650720596, 'eval_reward_total_composite_std': 0.18860110267996788, 'epoch': 0.02}

  6%|▌         | 200/3300 [28:38<6:13:59,  7.24s/it]

100%|██████████| 10/10 [00:43<00:00,  4.61s/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 [28:48<19:41:53, 22.88s/it]2026-04-13 07:49:46,059 | INFO | train_grpo_train | metrics_logged mode=train step=201

                                                     
{'loss': -0.1181, 'grad_norm': 10.919523239135742, 'learning_rate': 9.393939393939396e-06, 'num_tokens': 369668.0, 'completions/mean_length': 73.75, 'completions/min_length': 35.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.3700682520866394, 'rewards/meter/std': 0.3486845791339874, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.997421145439148, 'rewards/repeat_soft/std': 0.0027841662522405386, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.4321563243865967, 'rewards/total_composite/std': 0.10706186294555664, 'reward': 0.4321563243865967, 'reward_std': 0.10706186294555664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21561789512634277, 'sampling/sampling_logp_difference/max': 1.2610321044921875, 'sampling/importance_sampling_ratio/min': 0.28336140513420105, 'sampling/importance_sampling_ratio/mean': 1.0400891304016113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.015211820602417, 'clip_ratio/low_mean': 0.138279652222991, 'clip_ratio/low_min': 0.138279652222991, 'clip_ratio/high_mean': 0.07148056291043758, 'clip_ratio/high_max': 0.07148056291043758, 'clip_ratio/region_mean': 0.20976021513342857, 'reward_total_mean': 0.4321563243865967, 'reward_meter_mean': 0.3700682520866394, 'reward_meter_std': 0.3486845791339874, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.997421145439148, 'reward_repeat_soft_std': 0.0027841662522405386, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.4321563243865967, 'reward_total_composite_std': 0.10706186294555664, 'epoch': 0.02}

  6%|▌         | 201/3300 [28:48<19:41:53, 22.88s/it]INFO 04-13 07:49:46 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 202/3300 [28:56<15:51:41, 18.43s/it]2026-04-13 07:49:54,103 | INFO | train_grpo_train | metrics_logged mode=train step=202

                                                     
{'loss': 0.0434, 'grad_norm': 11.686323165893555, 'learning_rate': 9.390909090909092e-06, 'num_tokens': 371570.0, 'completions/mean_length': 65.75, 'completions/min_length': 57.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9234566688537598, 'rewards/meter/std': 0.1139838844537735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9957336783409119, 'rewards/repeat_soft/std': 0.005028849001973867, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.5512813329696655, 'rewards/total_composite/std': 0.23752687871456146, 'reward': 0.5512813329696655, 'reward_std': 0.23752686381340027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2418219894170761, 'sampling/sampling_logp_difference/max': 1.222823143005371, 'sampling/importance_sampling_ratio/min': 0.29439786076545715, 'sampling/importance_sampling_ratio/mean': 1.0712027549743652, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4567994475364685, 'clip_ratio/low_mean': 0.03870224021375179, 'clip_ratio/low_min': 0.03870224021375179, 'clip_ratio/high_mean': 0.13332573510706425, 'clip_ratio/high_max': 0.13332573510706425, 'clip_ratio/region_mean': 0.17202797532081604, 'reward_total_mean': 0.5512813329696655, 'reward_meter_mean': 0.9234566688537598, 'reward_meter_std': 0.1139838844537735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9957336783409119, 'reward_repeat_soft_std': 0.005028849001973867, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.5512813329696655, 'reward_total_composite_std': 0.23752687871456146, 'epoch': 0.02}

  6%|▌         | 202/3300 [28:56<15:51:41, 18.43s/it]INFO 04-13 07:49:54 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 203/3300 [29:05<13:23:54, 15.57s/it]2026-04-13 07:50:03,013 | INFO | train_grpo_train | metrics_logged mode=train step=203

                                                     
{'loss': -0.0037, 'grad_norm': 9.007547378540039, 'learning_rate': 9.387878787878789e-06, 'num_tokens': 374317.0, 'completions/mean_length': 136.375, 'completions/min_length': 109.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.375, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.6907858848571777, 'rewards/meter/std': 0.3659006953239441, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.999170184135437, 'rewards/repeat_soft/std': 0.0015115885762497783, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.4157761335372925, 'rewards/total_composite/std': 0.18969878554344177, 'reward': 0.4157761335372925, 'reward_std': 0.18969878554344177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22914695739746094, 'sampling/sampling_logp_difference/max': 1.5648012161254883, 'sampling/importance_sampling_ratio/min': 0.20912957191467285, 'sampling/importance_sampling_ratio/mean': 1.0604259967803955, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.437682628631592, 'clip_ratio/low_mean': 0.07899650000035763, 'clip_ratio/low_min': 0.07899650000035763, 'clip_ratio/high_mean': 0.11292179115116596, 'clip_ratio/high_max': 0.11292179115116596, 'clip_ratio/region_mean': 0.1919182911515236, 'reward_total_mean': 0.4157761335372925, 'reward_meter_mean': 0.6907858848571777, 'reward_meter_std': 0.3659006953239441, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.999170184135437, 'reward_repeat_soft_std': 0.0015115885762497783, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.4157761335372925, 'reward_total_composite_std': 0.18969878554344177, 'epoch': 0.02}

  6%|▌         | 203/3300 [29:05<13:23:54, 15.57s/it]INFO 04-13 07:50:03 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 204/3300 [29:12<11:11:25, 13.01s/it]2026-04-13 07:50:10,046 | INFO | train_grpo_train | metrics_logged mode=train step=204

                                                     
{'loss': 0.1175, 'grad_norm': 18.759410858154297, 'learning_rate': 9.384848484848486e-06, 'num_tokens': 376023.0, 'completions/mean_length': 42.25, 'completions/min_length': 34.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7116919755935669, 'rewards/meter/std': 0.38788291811943054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996666431427002, 'rewards/repeat_soft/std': 0.005294432397931814, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.2568177580833435, 'rewards/total_composite/mean': 0.6003146171569824, 'rewards/total_composite/std': 0.18185506761074066, 'reward': 0.6003146171569824, 'reward_std': 0.18185506761074066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25824519991874695, 'sampling/sampling_logp_difference/max': 1.8854713439941406, 'sampling/importance_sampling_ratio/min': 0.15175750851631165, 'sampling/importance_sampling_ratio/mean': 1.0785658359527588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1049128472805023, 'clip_ratio/low_mean': 0.10476029291749, 'clip_ratio/low_min': 0.10476029291749, 'clip_ratio/high_mean': 0.11218062974512577, 'clip_ratio/high_max': 0.11218062974512577, 'clip_ratio/region_mean': 0.21694092266261578, 'reward_total_mean': 0.6003146171569824, 'reward_meter_mean': 0.7116919755935669, 'reward_meter_std': 0.38788291811943054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996666431427002, 'reward_repeat_soft_std': 0.005294432397931814, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.2568177580833435, 'reward_total_composite_mean': 0.6003146171569824, 'reward_total_composite_std': 0.18185506761074066, 'epoch': 0.02}

  6%|▌         | 204/3300 [29:12<11:11:25, 13.01s/it]INFO 04-13 07:50:10 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 205/3300 [29:19<9:45:54, 11.36s/it] 2026-04-13 07:50:17,545 | INFO | train_grpo_train | metrics_logged mode=train step=205

                                                    
{'loss': 0.1287, 'grad_norm': 9.911831855773926, 'learning_rate': 9.381818181818183e-06, 'num_tokens': 378179.0, 'completions/mean_length': 96.5, 'completions/min_length': 76.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.8738713264465332, 'rewards/meter/std': 0.12859904766082764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973481893539429, 'rewards/repeat_soft/std': 0.00395160960033536, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.6006252765655518, 'rewards/total_composite/std': 0.08549527823925018, 'reward': 0.6006252765655518, 'reward_std': 0.08549528568983078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2232775241136551, 'sampling/sampling_logp_difference/max': 1.452596664428711, 'sampling/importance_sampling_ratio/min': 0.23396198451519012, 'sampling/importance_sampling_ratio/mean': 1.062524437904358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4127989411354065, 'clip_ratio/low_mean': 0.0587850222364068, 'clip_ratio/low_min': 0.0587850222364068, 'clip_ratio/high_mean': 0.13522934913635254, 'clip_ratio/high_max': 0.13522934913635254, 'clip_ratio/region_mean': 0.19401437137275934, 'reward_total_mean': 0.6006252765655518, 'reward_meter_mean': 0.8738713264465332, 'reward_meter_std': 0.12859904766082764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973481893539429, 'reward_repeat_soft_std': 0.00395160960033536, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.6006252765655518, 'reward_total_composite_std': 0.08549527823925018, 'epoch': 0.02}

  6%|▌         | 205/3300 [29:19<9:45:54, 11.36s/it]INFO 04-13 07:50:17 [block_pool.py:378] Successfully reset prefix cache

  6%|▌         | 206/3300 [29:26<8:37:33, 10.04s/it]2026-04-13 07:50:24,500 | INFO | train_grpo_train | metrics_logged mode=train step=206

                                                    
{'loss': 0.1188, 'grad_norm': 14.01938247680664, 'learning_rate': 9.378787878787879e-06, 'num_tokens': 379858.0, 'completions/mean_length': 49.875, 'completions/min_length': 35.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.3101171553134918, 'rewards/meter/std': 0.19971781969070435, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9820109605789185, 'rewards/repeat_soft/std': 0.016810083761811256, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.21224987506866455, 'rewards/total_composite/mean': 0.46407532691955566, 'rewards/total_composite/std': 0.10568965971469879, 'reward': 0.46407532691955566, 'reward_std': 0.10568965971469879, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2285691648721695, 'sampling/sampling_logp_difference/max': 1.982824683189392, 'sampling/importance_sampling_ratio/min': 0.137679785490036, 'sampling/importance_sampling_ratio/mean': 1.0584670305252075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4913979172706604, 'clip_ratio/low_mean': 0.10394087061285973, 'clip_ratio/low_min': 0.10394087061285973, 'clip_ratio/high_mean': 0.08304168842732906, 'clip_ratio/high_max': 0.08304168842732906, 'clip_ratio/region_mean': 0.1869825590401888, 'reward_total_mean': 0.46407532691955566, 'reward_meter_mean': 0.3101171553134918, 'reward_meter_std': 0.19971781969070435, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9820109605789185, 'reward_repeat_soft_std': 0.016810083761811256, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.21224987506866455, 'reward_total_composite_mean': 0.46407532691955566, 'reward_total_composite_std': 0.10568965971469879, 'epoch': 0.02}

  6%|▌         | 206/3300 [29:26<8:37:33, 10.04s/it]INFO 04-13 07:50:24 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 207/3300 [29:34<8:00:04,  9.31s/it]2026-04-13 07:50:32,121 | INFO | train_grpo_train | metrics_logged mode=train step=207

                                                    
{'loss': 0.1399, 'grad_norm': 16.71153450012207, 'learning_rate': 9.375757575757576e-06, 'num_tokens': 381676.0, 'completions/mean_length': 58.25, 'completions/min_length': 50.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.4424096941947937, 'rewards/meter/std': 0.311807781457901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986618161201477, 'rewards/repeat_soft/std': 0.0023097151424735785, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/total_composite/mean': 0.5327742099761963, 'rewards/total_composite/std': 0.18730784952640533, 'reward': 0.5327742099761963, 'reward_std': 0.18730784952640533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25211960077285767, 'sampling/sampling_logp_difference/max': 2.9538486003875732, 'sampling/importance_sampling_ratio/min': 0.05213866010308266, 'sampling/importance_sampling_ratio/mean': 0.9958014488220215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1974932253360748, 'clip_ratio/low_mean': 0.13629101030528545, 'clip_ratio/low_min': 0.13629101030528545, 'clip_ratio/high_mean': 0.04890734329819679, 'clip_ratio/high_max': 0.04890734329819679, 'clip_ratio/region_mean': 0.18519835360348225, 'reward_total_mean': 0.5327742099761963, 'reward_meter_mean': 0.4424096941947937, 'reward_meter_std': 0.311807781457901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986618161201477, 'reward_repeat_soft_std': 0.0023097151424735785, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.5327742099761963, 'reward_total_composite_std': 0.18730784952640533, 'epoch': 0.02}

  6%|▋         | 207/3300 [29:34<8:00:04,  9.31s/it]INFO 04-13 07:50:32 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 208/3300 [29:40<7:12:24,  8.39s/it]2026-04-13 07:50:38,366 | INFO | train_grpo_train | metrics_logged mode=train step=208

                                                    
{'loss': -0.1236, 'grad_norm': 19.451038360595703, 'learning_rate': 9.372727272727273e-06, 'num_tokens': 383076.0, 'completions/mean_length': 26.0, 'completions/min_length': 18.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6434499025344849, 'rewards/meter/std': 0.40879666805267334, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.952542781829834, 'rewards/repeat_soft/std': 0.011327398009598255, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.10507649928331375, 'rewards/total_composite/mean': 0.5123113393783569, 'rewards/total_composite/std': 0.12489007413387299, 'reward': 0.5123113393783569, 'reward_std': 0.12489006668329239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1586938351392746, 'sampling/sampling_logp_difference/max': 1.4763026237487793, 'sampling/importance_sampling_ratio/min': 0.22848092019557953, 'sampling/importance_sampling_ratio/mean': 1.011419653892517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8310238495469093, 'clip_ratio/low_mean': 0.09997650468721986, 'clip_ratio/low_min': 0.09997650468721986, 'clip_ratio/high_mean': 0.05958451982587576, 'clip_ratio/high_max': 0.05958451982587576, 'clip_ratio/region_mean': 0.15956102451309562, 'reward_total_mean': 0.5123113393783569, 'reward_meter_mean': 0.6434499025344849, 'reward_meter_std': 0.40879666805267334, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.952542781829834, 'reward_repeat_soft_std': 0.011327398009598255, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.10507649928331375, 'reward_total_composite_mean': 0.5123113393783569, 'reward_total_composite_std': 0.12489007413387299, 'epoch': 0.02}

  6%|▋         | 208/3300 [29:40<7:12:24,  8.39s/it]INFO 04-13 07:50:38 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 209/3300 [29:47<6:48:38,  7.93s/it]2026-04-13 07:50:45,222 | INFO | train_grpo_train | metrics_logged mode=train step=209

                                                    
{'loss': 0.0732, 'grad_norm': 11.658206939697266, 'learning_rate': 9.36969696969697e-06, 'num_tokens': 384645.0, 'completions/mean_length': 46.125, 'completions/min_length': 37.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.31836074590682983, 'rewards/meter/std': 0.3821393847465515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9982406497001648, 'rewards/repeat_soft/std': 0.004306571092456579, 'rewards/judge_quality/mean': 0.6325000524520874, 'rewards/judge_quality/std': 0.2474873960018158, 'rewards/total_composite/mean': 0.49626511335372925, 'rewards/total_composite/std': 0.19499576091766357, 'reward': 0.49626511335372925, 'reward_std': 0.19499576091766357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19146274030208588, 'sampling/sampling_logp_difference/max': 1.2691144943237305, 'sampling/importance_sampling_ratio/min': 0.281080424785614, 'sampling/importance_sampling_ratio/mean': 1.0481432676315308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.180514022707939, 'clip_ratio/low_mean': 0.07780514191836119, 'clip_ratio/low_min': 0.07780514191836119, 'clip_ratio/high_mean': 0.06334707047790289, 'clip_ratio/high_max': 0.06334707047790289, 'clip_ratio/region_mean': 0.14115221239626408, 'reward_total_mean': 0.49626511335372925, 'reward_meter_mean': 0.31836074590682983, 'reward_meter_std': 0.3821393847465515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9982406497001648, 'reward_repeat_soft_std': 0.004306571092456579, 'reward_judge_quality_mean': 0.6325000524520874, 'reward_judge_quality_std': 0.2474873960018158, 'reward_total_composite_mean': 0.49626511335372925, 'reward_total_composite_std': 0.19499576091766357, 'epoch': 0.02}

  6%|▋         | 209/3300 [29:47<6:48:38,  7.93s/it]INFO 04-13 07:50:45 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 210/3300 [29:54<6:40:45,  7.78s/it]2026-04-13 07:50:52,653 | INFO | train_grpo_train | metrics_logged mode=train step=210

                                                    
{'loss': 0.1079, 'grad_norm': 8.990679740905762, 'learning_rate': 9.366666666666668e-06, 'num_tokens': 386866.0, 'completions/mean_length': 112.625, 'completions/min_length': 88.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.625, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.3633342683315277, 'rewards/meter/std': 0.30514103174209595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9887441396713257, 'rewards/repeat_soft/std': 0.02644232101738453, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.3910030722618103, 'rewards/total_composite/std': 0.1865619570016861, 'reward': 0.3910030722618103, 'reward_std': 0.1865619570016861, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2292454093694687, 'sampling/sampling_logp_difference/max': 1.7594289779663086, 'sampling/importance_sampling_ratio/min': 0.1721431314945221, 'sampling/importance_sampling_ratio/mean': 1.053958535194397, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2477812469005585, 'clip_ratio/low_mean': 0.0655823927372694, 'clip_ratio/low_min': 0.0655823927372694, 'clip_ratio/high_mean': 0.12628458999097347, 'clip_ratio/high_max': 0.12628458999097347, 'clip_ratio/region_mean': 0.19186698272824287, 'reward_total_mean': 0.3910030722618103, 'reward_meter_mean': 0.3633342683315277, 'reward_meter_std': 0.30514103174209595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9887441396713257, 'reward_repeat_soft_std': 0.02644232101738453, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.3910030722618103, 'reward_total_composite_std': 0.1865619570016861, 'epoch': 0.02}

  6%|▋         | 210/3300 [29:54<6:40:45,  7.78s/it]INFO 04-13 07:50:52 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 211/3300 [30:01<6:15:20,  7.29s/it]2026-04-13 07:50:58,797 | INFO | train_grpo_train | metrics_logged mode=train step=211

                                                    
{'loss': 0.0086, 'grad_norm': 21.03252601623535, 'learning_rate': 9.363636363636365e-06, 'num_tokens': 388355.0, 'completions/mean_length': 26.125, 'completions/min_length': 18.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.7306721210479736, 'rewards/meter/std': 0.43190711736679077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42624998092651367, 'rewards/judge_quality/std': 0.21980105340480804, 'rewards/total_composite/mean': 0.5577149987220764, 'rewards/total_composite/std': 0.1788819283246994, 'reward': 0.5577149987220764, 'reward_std': 0.17888189852237701, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20815829932689667, 'sampling/sampling_logp_difference/max': 1.2290782928466797, 'sampling/importance_sampling_ratio/min': 0.2925620973110199, 'sampling/importance_sampling_ratio/mean': 1.0683070421218872, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.57223704457283, 'clip_ratio/low_mean': 0.12673186231404543, 'clip_ratio/low_min': 0.12673186231404543, 'clip_ratio/high_mean': 0.1146390400826931, 'clip_ratio/high_max': 0.1146390400826931, 'clip_ratio/region_mean': 0.24137090239673853, 'reward_total_mean': 0.5577149987220764, 'reward_meter_mean': 0.7306721210479736, 'reward_meter_std': 0.43190711736679077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42624998092651367, 'reward_judge_quality_std': 0.21980105340480804, 'reward_total_composite_mean': 0.5577149987220764, 'reward_total_composite_std': 0.1788819283246994, 'epoch': 0.02}

  6%|▋         | 211/3300 [30:01<6:15:20,  7.29s/it]INFO 04-13 07:50:59 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 212/3300 [30:08<6:23:39,  7.45s/it]2026-04-13 07:51:06,636 | INFO | train_grpo_train | metrics_logged mode=train step=212

                                                    
{'loss': 0.2297, 'grad_norm': 13.52635669708252, 'learning_rate': 9.36060606060606e-06, 'num_tokens': 390055.0, 'completions/mean_length': 54.5, 'completions/min_length': 38.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9868708848953247, 'rewards/meter/std': 0.004795430693775415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9918505549430847, 'rewards/repeat_soft/std': 0.020045360550284386, 'rewards/judge_quality/mean': 0.3474999964237213, 'rewards/judge_quality/std': 0.11310549825429916, 'rewards/total_composite/mean': 0.5717463493347168, 'rewards/total_composite/std': 0.07369498163461685, 'reward': 0.5717463493347168, 'reward_std': 0.07369498163461685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24047841131687164, 'sampling/sampling_logp_difference/max': 1.0447254180908203, 'sampling/importance_sampling_ratio/min': 0.35178840160369873, 'sampling/importance_sampling_ratio/mean': 1.0468429327011108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4133982360363007, 'clip_ratio/low_mean': 0.06530501134693623, 'clip_ratio/low_min': 0.06530501134693623, 'clip_ratio/high_mean': 0.15588182397186756, 'clip_ratio/high_max': 0.15588182397186756, 'clip_ratio/region_mean': 0.2211868353188038, 'reward_total_mean': 0.5717463493347168, 'reward_meter_mean': 0.9868708848953247, 'reward_meter_std': 0.004795430693775415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9918505549430847, 'reward_repeat_soft_std': 0.020045360550284386, 'reward_judge_quality_mean': 0.3474999964237213, 'reward_judge_quality_std': 0.11310549825429916, 'reward_total_composite_mean': 0.5717463493347168, 'reward_total_composite_std': 0.07369498163461685, 'epoch': 0.02}

  6%|▋         | 212/3300 [30:08<6:23:39,  7.45s/it]INFO 04-13 07:51:06 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 213/3300 [30:16<6:33:39,  7.65s/it]2026-04-13 07:51:14,747 | INFO | train_grpo_train | metrics_logged mode=train step=213

                                                    
{'loss': 0.0765, 'grad_norm': 8.80113697052002, 'learning_rate': 9.357575757575758e-06, 'num_tokens': 392499.0, 'completions/mean_length': 121.5, 'completions/min_length': 62.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.5, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.32365942001342773, 'rewards/meter/std': 0.33367863297462463, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969816207885742, 'rewards/repeat_soft/std': 0.002814301522448659, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.18100909888744354, 'rewards/total_composite/mean': 0.45135462284088135, 'rewards/total_composite/std': 0.12482835352420807, 'reward': 0.45135462284088135, 'reward_std': 0.12482836097478867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21992932260036469, 'sampling/sampling_logp_difference/max': 1.7447614669799805, 'sampling/importance_sampling_ratio/min': 0.17468665540218353, 'sampling/importance_sampling_ratio/mean': 1.0516859292984009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.712333858013153, 'clip_ratio/low_mean': 0.15063858963549137, 'clip_ratio/low_min': 0.15063858963549137, 'clip_ratio/high_mean': 0.09270507097244263, 'clip_ratio/high_max': 0.09270507097244263, 'clip_ratio/region_mean': 0.243343660607934, 'reward_total_mean': 0.45135462284088135, 'reward_meter_mean': 0.32365942001342773, 'reward_meter_std': 0.33367863297462463, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969816207885742, 'reward_repeat_soft_std': 0.002814301522448659, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.18100909888744354, 'reward_total_composite_mean': 0.45135462284088135, 'reward_total_composite_std': 0.12482835352420807, 'epoch': 0.02}

  6%|▋         | 213/3300 [30:16<6:33:39,  7.65s/it]INFO 04-13 07:51:14 [block_pool.py:378] Successfully reset prefix cache

  6%|▋         | 214/3300 [30:23<6:22:47,  7.44s/it]2026-04-13 07:51:21,703 | INFO | train_grpo_train | metrics_logged mode=train step=214

                                                    
{'loss': -0.0382, 'grad_norm': 15.697773933410645, 'learning_rate': 9.354545454545455e-06, 'num_tokens': 394082.0, 'completions/mean_length': 38.875, 'completions/min_length': 33.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.3175961375236511, 'rewards/meter/std': 0.38890060782432556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995468020439148, 'rewards/repeat_soft/std': 0.0074559105560183525, 'rewards/judge_quality/mean': 0.627500057220459, 'rewards/judge_quality/std': 0.21952873468399048, 'rewards/total_composite/mean': 0.47513413429260254, 'rewards/total_composite/std': 0.17617429792881012, 'reward': 0.47513413429260254, 'reward_std': 0.1761743128299713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17709468305110931, 'sampling/sampling_logp_difference/max': 1.4810428619384766, 'sampling/importance_sampling_ratio/min': 0.22740042209625244, 'sampling/importance_sampling_ratio/mean': 1.0321656465530396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6610290557146072, 'clip_ratio/low_mean': 0.09056520229205489, 'clip_ratio/low_min': 0.09056520229205489, 'clip_ratio/high_mean': 0.05239104572683573, 'clip_ratio/high_max': 0.05239104572683573, 'clip_ratio/region_mean': 0.14295624801889062, 'reward_total_mean': 0.47513413429260254, 'reward_meter_mean': 0.3175961375236511, 'reward_meter_std': 0.38890060782432556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995468020439148, 'reward_repeat_soft_std': 0.0074559105560183525, 'reward_judge_quality_mean': 0.627500057220459, 'reward_judge_quality_std': 0.21952873468399048, 'reward_total_composite_mean': 0.47513413429260254, 'reward_total_composite_std': 0.17617429792881012, 'epoch': 0.02}

  6%|▋         | 214/3300 [30:23<6:22:47,  7.44s/it]INFO 04-13 07:51:21 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 215/3300 [30:31<6:18:40,  7.36s/it]2026-04-13 07:51:28,885 | INFO | train_grpo_train | metrics_logged mode=train step=215

                                                    
{'loss': 0.1524, 'grad_norm': 13.002318382263184, 'learning_rate': 9.351515151515152e-06, 'num_tokens': 396112.0, 'completions/mean_length': 77.75, 'completions/min_length': 58.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.75, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.3979058265686035, 'rewards/meter/std': 0.42231327295303345, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9979252219200134, 'rewards/repeat_soft/std': 0.001515165320597589, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.39308273792266846, 'rewards/total_composite/std': 0.19107317924499512, 'reward': 0.39308273792266846, 'reward_std': 0.19107317924499512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25099608302116394, 'sampling/sampling_logp_difference/max': 1.6520500183105469, 'sampling/importance_sampling_ratio/min': 0.19165661931037903, 'sampling/importance_sampling_ratio/mean': 1.0474159717559814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.917648583650589, 'clip_ratio/low_mean': 0.10196277312934399, 'clip_ratio/low_min': 0.10196277312934399, 'clip_ratio/high_mean': 0.10599080473184586, 'clip_ratio/high_max': 0.10599080473184586, 'clip_ratio/region_mean': 0.20795357786118984, 'reward_total_mean': 0.39308273792266846, 'reward_meter_mean': 0.3979058265686035, 'reward_meter_std': 0.42231327295303345, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9979252219200134, 'reward_repeat_soft_std': 0.001515165320597589, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.39308273792266846, 'reward_total_composite_std': 0.19107317924499512, 'epoch': 0.02}

  7%|▋         | 215/3300 [30:31<6:18:40,  7.36s/it]INFO 04-13 07:51:29 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 216/3300 [30:38<6:22:10,  7.44s/it]2026-04-13 07:51:36,486 | INFO | train_grpo_train | metrics_logged mode=train step=216

                                                    
{'loss': 0.053, 'grad_norm': 13.459993362426758, 'learning_rate': 9.34848484848485e-06, 'num_tokens': 398528.0, 'completions/mean_length': 88.0, 'completions/min_length': 69.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.2462480366230011, 'rewards/meter/std': 0.1796661615371704, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9924381375312805, 'rewards/repeat_soft/std': 0.0044843521900475025, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4137886166572571, 'rewards/total_composite/std': 0.050832126289606094, 'reward': 0.4137886166572571, 'reward_std': 0.0508321188390255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20468120276927948, 'sampling/sampling_logp_difference/max': 2.7741663455963135, 'sampling/importance_sampling_ratio/min': 0.06240147724747658, 'sampling/importance_sampling_ratio/mean': 1.0105061531066895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4371421933174133, 'clip_ratio/low_mean': 0.0938820093870163, 'clip_ratio/low_min': 0.0938820093870163, 'clip_ratio/high_mean': 0.09630271978676319, 'clip_ratio/high_max': 0.09630271978676319, 'clip_ratio/region_mean': 0.1901847291737795, 'reward_total_mean': 0.4137886166572571, 'reward_meter_mean': 0.2462480366230011, 'reward_meter_std': 0.1796661615371704, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9924381375312805, 'reward_repeat_soft_std': 0.0044843521900475025, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4137886166572571, 'reward_total_composite_std': 0.050832126289606094, 'epoch': 0.02}

  7%|▋         | 216/3300 [30:38<6:22:10,  7.44s/it]INFO 04-13 07:51:36 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 217/3300 [30:46<6:25:23,  7.50s/it]2026-04-13 07:51:44,138 | INFO | train_grpo_train | metrics_logged mode=train step=217

                                                    
{'loss': 0.0276, 'grad_norm': 9.43420124053955, 'learning_rate': 9.345454545454547e-06, 'num_tokens': 401007.0, 'completions/mean_length': 110.875, 'completions/min_length': 50.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.8257700204849243, 'rewards/meter/std': 0.24290962517261505, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.2121320217847824, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9925748109817505, 'rewards/repeat_soft/std': 0.004607753828167915, 'rewards/judge_quality/mean': 0.3137499690055847, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.4005802571773529, 'rewards/total_composite/std': 0.2571200430393219, 'reward': 0.4005802571773529, 'reward_std': 0.2571200430393219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25337132811546326, 'sampling/sampling_logp_difference/max': 1.4746055603027344, 'sampling/importance_sampling_ratio/min': 0.2288689911365509, 'sampling/importance_sampling_ratio/mean': 1.0746126174926758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7519125640392303, 'clip_ratio/low_mean': 0.06241595558822155, 'clip_ratio/low_min': 0.06241595558822155, 'clip_ratio/high_mean': 0.12228629365563393, 'clip_ratio/high_max': 0.12228629365563393, 'clip_ratio/region_mean': 0.18470224924385548, 'reward_total_mean': 0.4005802571773529, 'reward_meter_mean': 0.8257700204849243, 'reward_meter_std': 0.24290962517261505, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.2121320217847824, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9925748109817505, 'reward_repeat_soft_std': 0.004607753828167915, 'reward_judge_quality_mean': 0.3137499690055847, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.4005802571773529, 'reward_total_composite_std': 0.2571200430393219, 'epoch': 0.02}

  7%|▋         | 217/3300 [30:46<6:25:23,  7.50s/it]INFO 04-13 07:51:44 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 218/3300 [30:54<6:32:28,  7.64s/it]2026-04-13 07:51:52,104 | INFO | train_grpo_train | metrics_logged mode=train step=218

                                                    
{'loss': 0.0009, 'grad_norm': 16.42949676513672, 'learning_rate': 9.342424242424243e-06, 'num_tokens': 402693.0, 'completions/mean_length': 42.75, 'completions/min_length': 35.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5975445508956909, 'rewards/meter/std': 0.41694375872612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9960700869560242, 'rewards/repeat_soft/std': 0.007433554623275995, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5205863118171692, 'rewards/total_composite/std': 0.10966836661100388, 'reward': 0.5205863118171692, 'reward_std': 0.10966836661100388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22321748733520508, 'sampling/sampling_logp_difference/max': 1.2820940017700195, 'sampling/importance_sampling_ratio/min': 0.2774556875228882, 'sampling/importance_sampling_ratio/mean': 1.075124740600586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6578336507081985, 'clip_ratio/low_mean': 0.08264311123639345, 'clip_ratio/low_min': 0.08264311123639345, 'clip_ratio/high_mean': 0.09011243376880884, 'clip_ratio/high_max': 0.09011243376880884, 'clip_ratio/region_mean': 0.1727555450052023, 'reward_total_mean': 0.5205863118171692, 'reward_meter_mean': 0.5975445508956909, 'reward_meter_std': 0.41694375872612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9960700869560242, 'reward_repeat_soft_std': 0.007433554623275995, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5205863118171692, 'reward_total_composite_std': 0.10966836661100388, 'epoch': 0.02}

  7%|▋         | 218/3300 [30:54<6:32:28,  7.64s/it]INFO 04-13 07:51:52 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 219/3300 [31:01<6:23:46,  7.47s/it]2026-04-13 07:51:59,189 | INFO | train_grpo_train | metrics_logged mode=train step=219

                                                    
{'loss': 0.0003, 'grad_norm': 13.988005638122559, 'learning_rate': 9.33939393939394e-06, 'num_tokens': 404229.0, 'completions/mean_length': 44.0, 'completions/min_length': 40.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.22564613819122314, 'rewards/meter/std': 0.2446339875459671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9994621872901917, 'rewards/repeat_soft/std': 0.0015211430145427585, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/total_composite/mean': 0.46914806962013245, 'rewards/total_composite/std': 0.14648151397705078, 'reward': 0.46914806962013245, 'reward_std': 0.14648151397705078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18220719695091248, 'sampling/sampling_logp_difference/max': 1.2400398254394531, 'sampling/importance_sampling_ratio/min': 0.2893727123737335, 'sampling/importance_sampling_ratio/mean': 1.0694383382797241, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2795404344797134, 'clip_ratio/low_mean': 0.11006227973848581, 'clip_ratio/low_min': 0.11006227973848581, 'clip_ratio/high_mean': 0.06450820062309504, 'clip_ratio/high_max': 0.06450820062309504, 'clip_ratio/region_mean': 0.17457048036158085, 'reward_total_mean': 0.46914806962013245, 'reward_meter_mean': 0.22564613819122314, 'reward_meter_std': 0.2446339875459671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9994621872901917, 'reward_repeat_soft_std': 0.0015211430145427585, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.46914806962013245, 'reward_total_composite_std': 0.14648151397705078, 'epoch': 0.02}

  7%|▋         | 219/3300 [31:01<6:23:46,  7.47s/it]INFO 04-13 07:51:59 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 220/3300 [31:08<6:17:59,  7.36s/it]2026-04-13 07:52:06,295 | INFO | train_grpo_train | metrics_logged mode=train step=220

                                                    
{'loss': 0.0078, 'grad_norm': 13.550875663757324, 'learning_rate': 9.336363636363637e-06, 'num_tokens': 405869.0, 'completions/mean_length': 40.0, 'completions/min_length': 34.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.46394291520118713, 'rewards/meter/std': 0.2540203332901001, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9941651821136475, 'rewards/repeat_soft/std': 0.007878880016505718, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.24663449823856354, 'rewards/total_composite/mean': 0.468347430229187, 'rewards/total_composite/std': 0.2445186972618103, 'reward': 0.468347430229187, 'reward_std': 0.2445186972618103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20536713302135468, 'sampling/sampling_logp_difference/max': 1.563028335571289, 'sampling/importance_sampling_ratio/min': 0.20950067043304443, 'sampling/importance_sampling_ratio/mean': 1.0111911296844482, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3682081010192633, 'clip_ratio/low_mean': 0.12227142415940762, 'clip_ratio/low_min': 0.12227142415940762, 'clip_ratio/high_mean': 0.06669955886900425, 'clip_ratio/high_max': 0.06669955886900425, 'clip_ratio/region_mean': 0.18897098302841187, 'reward_total_mean': 0.468347430229187, 'reward_meter_mean': 0.46394291520118713, 'reward_meter_std': 0.2540203332901001, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9941651821136475, 'reward_repeat_soft_std': 0.007878880016505718, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.24663449823856354, 'reward_total_composite_mean': 0.468347430229187, 'reward_total_composite_std': 0.2445186972618103, 'epoch': 0.02}

  7%|▋         | 220/3300 [31:08<6:17:59,  7.36s/it]INFO 04-13 07:52:06 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 221/3300 [31:14<6:02:11,  7.06s/it]2026-04-13 07:52:12,640 | INFO | train_grpo_train | metrics_logged mode=train step=221

                                                    
{'loss': -0.0787, 'grad_norm': 22.784671783447266, 'learning_rate': 9.333333333333334e-06, 'num_tokens': 407234.0, 'completions/mean_length': 20.625, 'completions/min_length': 15.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7635542154312134, 'rewards/meter/std': 0.41205763816833496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8938847780227661, 'rewards/repeat_soft/std': 0.16115181148052216, 'rewards/judge_quality/mean': 0.4012500047683716, 'rewards/judge_quality/std': 0.10260012745857239, 'rewards/total_composite/mean': 0.5158623456954956, 'rewards/total_composite/std': 0.22194378077983856, 'reward': 0.5158623456954956, 'reward_std': 0.22194376587867737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1971156895160675, 'sampling/sampling_logp_difference/max': 1.424391746520996, 'sampling/importance_sampling_ratio/min': 0.2406548112630844, 'sampling/importance_sampling_ratio/mean': 1.0366919040679932, 'sampling/importance_sampling_ratio/max': 1.914661169052124, 'entropy': 1.931114301085472, 'clip_ratio/low_mean': 0.04880952462553978, 'clip_ratio/low_min': 0.04880952462553978, 'clip_ratio/high_mean': 0.14553836593404412, 'clip_ratio/high_max': 0.14553836593404412, 'clip_ratio/region_mean': 0.1943478905595839, 'reward_total_mean': 0.5158623456954956, 'reward_meter_mean': 0.7635542154312134, 'reward_meter_std': 0.41205763816833496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8938847780227661, 'reward_repeat_soft_std': 0.16115181148052216, 'reward_judge_quality_mean': 0.4012500047683716, 'reward_judge_quality_std': 0.10260012745857239, 'reward_total_composite_mean': 0.5158623456954956, 'reward_total_composite_std': 0.22194378077983856, 'epoch': 0.02}

  7%|▋         | 221/3300 [31:14<6:02:11,  7.06s/it]INFO 04-13 07:52:12 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 222/3300 [31:21<6:03:02,  7.08s/it]2026-04-13 07:52:19,762 | INFO | train_grpo_train | metrics_logged mode=train step=222

                                                    
{'loss': 0.0696, 'grad_norm': 12.268951416015625, 'learning_rate': 9.33030303030303e-06, 'num_tokens': 408868.0, 'completions/mean_length': 41.25, 'completions/min_length': 37.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.24934370815753937, 'rewards/meter/std': 0.352595716714859, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9918965101242065, 'rewards/repeat_soft/std': 0.008309373632073402, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.41893401741981506, 'rewards/total_composite/std': 0.09815803915262222, 'reward': 0.41893401741981506, 'reward_std': 0.09815803915262222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19416873157024384, 'sampling/sampling_logp_difference/max': 1.2836970090866089, 'sampling/importance_sampling_ratio/min': 0.33600616455078125, 'sampling/importance_sampling_ratio/mean': 1.062048077583313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3454811573028564, 'clip_ratio/low_mean': 0.13969633169472218, 'clip_ratio/low_min': 0.13969633169472218, 'clip_ratio/high_mean': 0.04002079088240862, 'clip_ratio/high_max': 0.04002079088240862, 'clip_ratio/region_mean': 0.1797171225771308, 'reward_total_mean': 0.41893401741981506, 'reward_meter_mean': 0.24934370815753937, 'reward_meter_std': 0.352595716714859, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9918965101242065, 'reward_repeat_soft_std': 0.008309373632073402, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.41893401741981506, 'reward_total_composite_std': 0.09815803915262222, 'epoch': 0.02}

  7%|▋         | 222/3300 [31:21<6:03:02,  7.08s/it]INFO 04-13 07:52:19 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 223/3300 [31:30<6:17:56,  7.37s/it]2026-04-13 07:52:27,813 | INFO | train_grpo_train | metrics_logged mode=train step=223

                                                    
{'loss': -0.0104, 'grad_norm': 8.55893611907959, 'learning_rate': 9.327272727272729e-06, 'num_tokens': 411085.0, 'completions/mean_length': 105.125, 'completions/min_length': 88.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.125, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.7916437983512878, 'rewards/meter/std': 0.24882195889949799, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9991984367370605, 'rewards/repeat_soft/std': 0.001545512699522078, 'rewards/judge_quality/mean': 0.23749999701976776, 'rewards/judge_quality/std': 0.09910311549901962, 'rewards/total_composite/mean': 0.40125560760498047, 'rewards/total_composite/std': 0.16618593037128448, 'reward': 0.40125560760498047, 'reward_std': 0.16618593037128448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2330099493265152, 'sampling/sampling_logp_difference/max': 1.4484367370605469, 'sampling/importance_sampling_ratio/min': 0.23493726551532745, 'sampling/importance_sampling_ratio/mean': 1.07094407081604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.7301617562770844, 'clip_ratio/low_mean': 0.025815216824412346, 'clip_ratio/low_min': 0.025815216824412346, 'clip_ratio/high_mean': 0.16274195350706577, 'clip_ratio/high_max': 0.16274195350706577, 'clip_ratio/region_mean': 0.18855717033147812, 'reward_total_mean': 0.40125560760498047, 'reward_meter_mean': 0.7916437983512878, 'reward_meter_std': 0.24882195889949799, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9991984367370605, 'reward_repeat_soft_std': 0.001545512699522078, 'reward_judge_quality_mean': 0.23749999701976776, 'reward_judge_quality_std': 0.09910311549901962, 'reward_total_composite_mean': 0.40125560760498047, 'reward_total_composite_std': 0.16618593037128448, 'epoch': 0.02}

  7%|▋         | 223/3300 [31:30<6:17:56,  7.37s/it]INFO 04-13 07:52:28 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 224/3300 [31:37<6:17:03,  7.35s/it]2026-04-13 07:52:35,133 | INFO | train_grpo_train | metrics_logged mode=train step=224

                                                    
{'loss': 0.1158, 'grad_norm': 13.292996406555176, 'learning_rate': 9.324242424242424e-06, 'num_tokens': 412778.0, 'completions/mean_length': 58.625, 'completions/min_length': 48.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.5885939598083496, 'rewards/meter/std': 0.3498261570930481, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996599018573761, 'rewards/repeat_soft/std': 0.0025883002672344446, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.1566559225320816, 'rewards/total_composite/mean': 0.5756773948669434, 'rewards/total_composite/std': 0.1631508320569992, 'reward': 0.5756773948669434, 'reward_std': 0.1631508320569992, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2243303507566452, 'sampling/sampling_logp_difference/max': 1.7579689025878906, 'sampling/importance_sampling_ratio/min': 0.17239466309547424, 'sampling/importance_sampling_ratio/mean': 1.0411030054092407, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9070996046066284, 'clip_ratio/low_mean': 0.09940476529300213, 'clip_ratio/low_min': 0.09940476529300213, 'clip_ratio/high_mean': 0.1176231075078249, 'clip_ratio/high_max': 0.1176231075078249, 'clip_ratio/region_mean': 0.21702787280082703, 'reward_total_mean': 0.5756773948669434, 'reward_meter_mean': 0.5885939598083496, 'reward_meter_std': 0.3498261570930481, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996599018573761, 'reward_repeat_soft_std': 0.0025883002672344446, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.1566559225320816, 'reward_total_composite_mean': 0.5756773948669434, 'reward_total_composite_std': 0.1631508320569992, 'epoch': 0.02}

  7%|▋         | 224/3300 [31:37<6:17:03,  7.35s/it]INFO 04-13 07:52:35 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 225/3300 [31:44<6:21:29,  7.44s/it]2026-04-13 07:52:42,786 | INFO | train_grpo_train | metrics_logged mode=train step=225

                                                    
{'loss': 0.0863, 'grad_norm': 9.034035682678223, 'learning_rate': 9.321212121212122e-06, 'num_tokens': 415296.0, 'completions/mean_length': 114.75, 'completions/min_length': 93.0, 'completions/max_length': 146.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.132351815700531, 'rewards/meter/std': 0.15599699318408966, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9927990436553955, 'rewards/repeat_soft/std': 0.009575942531228065, 'rewards/judge_quality/mean': 0.26749998331069946, 'rewards/judge_quality/std': 0.10375107079744339, 'rewards/total_composite/mean': 0.371489018201828, 'rewards/total_composite/std': 0.026845812797546387, 'reward': 0.371489018201828, 'reward_std': 0.026845814660191536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24750882387161255, 'sampling/sampling_logp_difference/max': 1.3614006042480469, 'sampling/importance_sampling_ratio/min': 0.25630155205726624, 'sampling/importance_sampling_ratio/mean': 1.0550618171691895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.477659285068512, 'clip_ratio/low_mean': 0.12866520509123802, 'clip_ratio/low_min': 0.12866520509123802, 'clip_ratio/high_mean': 0.04819867201149464, 'clip_ratio/high_max': 0.04819867201149464, 'clip_ratio/region_mean': 0.17686387710273266, 'reward_total_mean': 0.371489018201828, 'reward_meter_mean': 0.132351815700531, 'reward_meter_std': 0.15599699318408966, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9927990436553955, 'reward_repeat_soft_std': 0.009575942531228065, 'reward_judge_quality_mean': 0.26749998331069946, 'reward_judge_quality_std': 0.10375107079744339, 'reward_total_composite_mean': 0.371489018201828, 'reward_total_composite_std': 0.026845812797546387, 'epoch': 0.02}

  7%|▋         | 225/3300 [31:45<6:21:29,  7.44s/it]INFO 04-13 07:52:43 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 226/3300 [31:52<6:15:06,  7.32s/it]2026-04-13 07:52:49,823 | INFO | train_grpo_train | metrics_logged mode=train step=226

                                                    
{'loss': -0.085, 'grad_norm': 15.010398864746094, 'learning_rate': 9.318181818181819e-06, 'num_tokens': 416665.0, 'completions/mean_length': 21.125, 'completions/min_length': 16.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.4514320194721222, 'rewards/meter/std': 0.3760293424129486, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.4698364734649658, 'rewards/total_composite/std': 0.11148180067539215, 'reward': 0.4698364734649658, 'reward_std': 0.11148180812597275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14198869466781616, 'sampling/sampling_logp_difference/max': 0.9297459125518799, 'sampling/importance_sampling_ratio/min': 0.39465397596359253, 'sampling/importance_sampling_ratio/mean': 1.024768590927124, 'sampling/importance_sampling_ratio/max': 1.9370837211608887, 'entropy': 0.9989406615495682, 'clip_ratio/low_mean': 0.06531892158091068, 'clip_ratio/low_min': 0.06531892158091068, 'clip_ratio/high_mean': 0.06238553300499916, 'clip_ratio/high_max': 0.06238553300499916, 'clip_ratio/region_mean': 0.12770445458590984, 'reward_total_mean': 0.4698364734649658, 'reward_meter_mean': 0.4514320194721222, 'reward_meter_std': 0.3760293424129486, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.4698364734649658, 'reward_total_composite_std': 0.11148180067539215, 'epoch': 0.02}

  7%|▋         | 226/3300 [31:52<6:15:06,  7.32s/it]INFO 04-13 07:52:50 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 227/3300 [31:59<6:19:23,  7.41s/it]2026-04-13 07:52:57,430 | INFO | train_grpo_train | metrics_logged mode=train step=227

                                                    
{'loss': 0.1018, 'grad_norm': 9.615615844726562, 'learning_rate': 9.315151515151516e-06, 'num_tokens': 418843.0, 'completions/mean_length': 96.25, 'completions/min_length': 74.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.25, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.4061456024646759, 'rewards/meter/std': 0.3346509337425232, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9986236095428467, 'rewards/repeat_soft/std': 0.0010029288241639733, 'rewards/judge_quality/mean': 0.2849999964237213, 'rewards/judge_quality/std': 0.12177261710166931, 'rewards/total_composite/mean': 0.30179980397224426, 'rewards/total_composite/std': 0.20693248510360718, 'reward': 0.30179980397224426, 'reward_std': 0.20693247020244598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24260981380939484, 'sampling/sampling_logp_difference/max': 1.4735374450683594, 'sampling/importance_sampling_ratio/min': 0.22911356389522552, 'sampling/importance_sampling_ratio/mean': 1.0716917514801025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.097844690084457, 'clip_ratio/low_mean': 0.044188953936100006, 'clip_ratio/low_min': 0.044188953936100006, 'clip_ratio/high_mean': 0.14237085357308388, 'clip_ratio/high_max': 0.14237085357308388, 'clip_ratio/region_mean': 0.18655980750918388, 'reward_total_mean': 0.30179980397224426, 'reward_meter_mean': 0.4061456024646759, 'reward_meter_std': 0.3346509337425232, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9986236095428467, 'reward_repeat_soft_std': 0.0010029288241639733, 'reward_judge_quality_mean': 0.2849999964237213, 'reward_judge_quality_std': 0.12177261710166931, 'reward_total_composite_mean': 0.30179980397224426, 'reward_total_composite_std': 0.20693248510360718, 'epoch': 0.02}

  7%|▋         | 227/3300 [31:59<6:19:23,  7.41s/it]INFO 04-13 07:52:57 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 228/3300 [32:07<6:22:08,  7.46s/it]2026-04-13 07:53:05,026 | INFO | train_grpo_train | metrics_logged mode=train step=228

                                                    
{'loss': 0.0147, 'grad_norm': 11.655343055725098, 'learning_rate': 9.312121212121212e-06, 'num_tokens': 420670.0, 'completions/mean_length': 58.375, 'completions/min_length': 46.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.5829670429229736, 'rewards/meter/std': 0.42181527614593506, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9993855953216553, 'rewards/repeat_soft/std': 0.0017378831980749965, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.10690449178218842, 'rewards/total_composite/mean': 0.44242537021636963, 'rewards/total_composite/std': 0.21695834398269653, 'reward': 0.44242537021636963, 'reward_std': 0.21695834398269653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2133873850107193, 'sampling/sampling_logp_difference/max': 1.4677143096923828, 'sampling/importance_sampling_ratio/min': 0.23045162856578827, 'sampling/importance_sampling_ratio/mean': 1.0543785095214844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.69685398042202, 'clip_ratio/low_mean': 0.06226163078099489, 'clip_ratio/low_min': 0.06226163078099489, 'clip_ratio/high_mean': 0.13080392964184284, 'clip_ratio/high_max': 0.13080392964184284, 'clip_ratio/region_mean': 0.19306556042283773, 'reward_total_mean': 0.44242537021636963, 'reward_meter_mean': 0.5829670429229736, 'reward_meter_std': 0.42181527614593506, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9993855953216553, 'reward_repeat_soft_std': 0.0017378831980749965, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.10690449178218842, 'reward_total_composite_mean': 0.44242537021636963, 'reward_total_composite_std': 0.21695834398269653, 'epoch': 0.02}

  7%|▋         | 228/3300 [32:07<6:22:08,  7.46s/it]INFO 04-13 07:53:05 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 229/3300 [32:13<6:06:36,  7.16s/it]2026-04-13 07:53:11,486 | INFO | train_grpo_train | metrics_logged mode=train step=229

                                                    
{'loss': 0.1245, 'grad_norm': 17.736083984375, 'learning_rate': 9.30909090909091e-06, 'num_tokens': 422246.0, 'completions/mean_length': 41.0, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9324971437454224, 'rewards/meter/std': 0.10176718235015869, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9925810694694519, 'rewards/repeat_soft/std': 0.007585481274873018, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6034588813781738, 'rewards/total_composite/std': 0.0284835547208786, 'reward': 0.6034588813781738, 'reward_std': 0.028483537957072258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22982312738895416, 'sampling/sampling_logp_difference/max': 1.3124065399169922, 'sampling/importance_sampling_ratio/min': 0.26917150616645813, 'sampling/importance_sampling_ratio/mean': 1.020029067993164, 'sampling/importance_sampling_ratio/max': 1.9736405611038208, 'entropy': 2.5963164418935776, 'clip_ratio/low_mean': 0.04195804335176945, 'clip_ratio/low_min': 0.04195804335176945, 'clip_ratio/high_mean': 0.18593565560877323, 'clip_ratio/high_max': 0.18593565560877323, 'clip_ratio/region_mean': 0.22789369896054268, 'reward_total_mean': 0.6034588813781738, 'reward_meter_mean': 0.9324971437454224, 'reward_meter_std': 0.10176718235015869, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9925810694694519, 'reward_repeat_soft_std': 0.007585481274873018, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6034588813781738, 'reward_total_composite_std': 0.0284835547208786, 'epoch': 0.02}

  7%|▋         | 229/3300 [32:13<6:06:36,  7.16s/it]INFO 04-13 07:53:11 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 230/3300 [32:20<5:57:35,  6.99s/it]2026-04-13 07:53:18,068 | INFO | train_grpo_train | metrics_logged mode=train step=230

                                                    
{'loss': 0.0795, 'grad_norm': 20.825172424316406, 'learning_rate': 9.306060606060608e-06, 'num_tokens': 424107.0, 'completions/mean_length': 48.625, 'completions/min_length': 44.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.4202841818332672, 'rewards/meter/std': 0.26213112473487854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9920446872711182, 'rewards/repeat_soft/std': 0.006530150305479765, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.41952887177467346, 'rewards/total_composite/std': 0.17820438742637634, 'reward': 0.41952887177467346, 'reward_std': 0.17820440232753754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23581646382808685, 'sampling/sampling_logp_difference/max': 1.6607871055603027, 'sampling/importance_sampling_ratio/min': 0.1899893879890442, 'sampling/importance_sampling_ratio/mean': 1.0112229585647583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.193833500146866, 'clip_ratio/low_mean': 0.01886792480945587, 'clip_ratio/low_min': 0.01886792480945587, 'clip_ratio/high_mean': 0.22404972184449434, 'clip_ratio/high_max': 0.22404972184449434, 'clip_ratio/region_mean': 0.24291764665395021, 'reward_total_mean': 0.41952887177467346, 'reward_meter_mean': 0.4202841818332672, 'reward_meter_std': 0.26213112473487854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9920446872711182, 'reward_repeat_soft_std': 0.006530150305479765, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.41952887177467346, 'reward_total_composite_std': 0.17820438742637634, 'epoch': 0.02}

  7%|▋         | 230/3300 [32:20<5:57:35,  6.99s/it]INFO 04-13 07:53:18 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 231/3300 [32:27<5:53:47,  6.92s/it]2026-04-13 07:53:24,817 | INFO | train_grpo_train | metrics_logged mode=train step=231

                                                    
{'loss': 0.1162, 'grad_norm': 9.56534194946289, 'learning_rate': 9.303030303030303e-06, 'num_tokens': 425901.0, 'completions/mean_length': 60.25, 'completions/min_length': 42.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9549270868301392, 'rewards/meter/std': 0.06464817374944687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9862501621246338, 'rewards/repeat_soft/std': 0.023244677111506462, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.20860078930854797, 'rewards/total_composite/mean': 0.6222413182258606, 'rewards/total_composite/std': 0.13817496597766876, 'reward': 0.6222413182258606, 'reward_std': 0.13817498087882996, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21060773730278015, 'sampling/sampling_logp_difference/max': 1.1233088970184326, 'sampling/importance_sampling_ratio/min': 0.3354046046733856, 'sampling/importance_sampling_ratio/mean': 1.0613038539886475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8028874695301056, 'clip_ratio/low_mean': 0.14069972466677427, 'clip_ratio/low_min': 0.14069972466677427, 'clip_ratio/high_mean': 0.032738097012043, 'clip_ratio/high_max': 0.032738097012043, 'clip_ratio/region_mean': 0.17343782167881727, 'reward_total_mean': 0.6222413182258606, 'reward_meter_mean': 0.9549270868301392, 'reward_meter_std': 0.06464817374944687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9862501621246338, 'reward_repeat_soft_std': 0.023244677111506462, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.20860078930854797, 'reward_total_composite_mean': 0.6222413182258606, 'reward_total_composite_std': 0.13817496597766876, 'epoch': 0.02}

  7%|▋         | 231/3300 [32:27<5:53:47,  6.92s/it]INFO 04-13 07:53:25 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 232/3300 [32:34<5:58:44,  7.02s/it]2026-04-13 07:53:32,064 | INFO | train_grpo_train | metrics_logged mode=train step=232

                                                    
{'loss': 0.0053, 'grad_norm': 10.48536491394043, 'learning_rate': 9.3e-06, 'num_tokens': 427731.0, 'completions/mean_length': 62.75, 'completions/min_length': 53.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7720727920532227, 'rewards/meter/std': 0.38212233781814575, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976580142974854, 'rewards/repeat_soft/std': 0.006362478248775005, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7141427397727966, 'rewards/total_composite/std': 0.2544417083263397, 'reward': 0.7141427397727966, 'reward_std': 0.2544417381286621, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19832520186901093, 'sampling/sampling_logp_difference/max': 1.1894540786743164, 'sampling/importance_sampling_ratio/min': 0.3043873906135559, 'sampling/importance_sampling_ratio/mean': 1.0476423501968384, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.552804797887802, 'clip_ratio/low_mean': 0.08357260189950466, 'clip_ratio/low_min': 0.08357260189950466, 'clip_ratio/high_mean': 0.09559883549809456, 'clip_ratio/high_max': 0.09559883549809456, 'clip_ratio/region_mean': 0.17917143739759922, 'reward_total_mean': 0.7141427397727966, 'reward_meter_mean': 0.7720727920532227, 'reward_meter_std': 0.38212233781814575, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976580142974854, 'reward_repeat_soft_std': 0.006362478248775005, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7141427397727966, 'reward_total_composite_std': 0.2544417083263397, 'epoch': 0.02}

  7%|▋         | 232/3300 [32:34<5:58:44,  7.02s/it]INFO 04-13 07:53:32 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 233/3300 [32:43<6:33:58,  7.71s/it]2026-04-13 07:53:41,385 | INFO | train_grpo_train | metrics_logged mode=train step=233

                                                    
{'loss': 0.0208, 'grad_norm': 7.633944511413574, 'learning_rate': 9.296969696969698e-06, 'num_tokens': 430685.0, 'completions/mean_length': 177.25, 'completions/min_length': 134.0, 'completions/max_length': 232.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.25, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 232.0, 'rewards/meter/mean': 0.1322435736656189, 'rewards/meter/std': 0.123989999294281, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.0890870913863182, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.998924970626831, 'rewards/repeat_soft/std': 0.0005966781172901392, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.18431341648101807, 'rewards/total_composite/mean': 0.32089686393737793, 'rewards/total_composite/std': 0.13752084970474243, 'reward': 0.32089686393737793, 'reward_std': 0.13752086460590363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23580877482891083, 'sampling/sampling_logp_difference/max': 1.2670583724975586, 'sampling/importance_sampling_ratio/min': 0.28165894746780396, 'sampling/importance_sampling_ratio/mean': 1.0716066360473633, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9186656773090363, 'clip_ratio/low_mean': 0.03377334773540497, 'clip_ratio/low_min': 0.03377334773540497, 'clip_ratio/high_mean': 0.15041503123939037, 'clip_ratio/high_max': 0.15041503123939037, 'clip_ratio/region_mean': 0.18418837897479534, 'reward_total_mean': 0.32089686393737793, 'reward_meter_mean': 0.1322435736656189, 'reward_meter_std': 0.123989999294281, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.0890870913863182, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.998924970626831, 'reward_repeat_soft_std': 0.0005966781172901392, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.18431341648101807, 'reward_total_composite_mean': 0.32089686393737793, 'reward_total_composite_std': 0.13752084970474243, 'epoch': 0.02}

  7%|▋         | 233/3300 [32:43<6:33:58,  7.71s/it]INFO 04-13 07:53:41 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 234/3300 [32:50<6:20:02,  7.44s/it]2026-04-13 07:53:48,191 | INFO | train_grpo_train | metrics_logged mode=train step=234

                                                    
{'loss': -0.0942, 'grad_norm': 17.617843627929688, 'learning_rate': 9.293939393939395e-06, 'num_tokens': 432140.0, 'completions/mean_length': 25.875, 'completions/min_length': 18.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.315799742937088, 'rewards/meter/std': 0.3497828245162964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9580180644989014, 'rewards/repeat_soft/std': 0.009412449784576893, 'rewards/judge_quality/mean': 0.5437500476837158, 'rewards/judge_quality/std': 0.3325631320476532, 'rewards/total_composite/mean': 0.48225101828575134, 'rewards/total_composite/std': 0.19423244893550873, 'reward': 0.48225101828575134, 'reward_std': 0.19423244893550873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20272737741470337, 'sampling/sampling_logp_difference/max': 1.5750455856323242, 'sampling/importance_sampling_ratio/min': 0.20699812471866608, 'sampling/importance_sampling_ratio/mean': 1.0383530855178833, 'sampling/importance_sampling_ratio/max': 1.8340553045272827, 'entropy': 1.7450096309185028, 'clip_ratio/low_mean': 0.07225429499521852, 'clip_ratio/low_min': 0.07225429499521852, 'clip_ratio/high_mean': 0.09486280009150505, 'clip_ratio/high_max': 0.09486280009150505, 'clip_ratio/region_mean': 0.16711709508672357, 'reward_total_mean': 0.48225101828575134, 'reward_meter_mean': 0.315799742937088, 'reward_meter_std': 0.3497828245162964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9580180644989014, 'reward_repeat_soft_std': 0.009412449784576893, 'reward_judge_quality_mean': 0.5437500476837158, 'reward_judge_quality_std': 0.3325631320476532, 'reward_total_composite_mean': 0.48225101828575134, 'reward_total_composite_std': 0.19423244893550873, 'epoch': 0.02}

  7%|▋         | 234/3300 [32:50<6:20:02,  7.44s/it]INFO 04-13 07:53:48 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 235/3300 [33:00<6:53:48,  8.10s/it]2026-04-13 07:53:57,842 | INFO | train_grpo_train | metrics_logged mode=train step=235

                                                    
{'loss': 0.0866, 'grad_norm': 7.5676469802856445, 'learning_rate': 9.29090909090909e-06, 'num_tokens': 434776.0, 'completions/mean_length': 134.5, 'completions/min_length': 82.0, 'completions/max_length': 332.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 134.5, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 332.0, 'rewards/meter/mean': 0.5199936628341675, 'rewards/meter/std': 0.3432442545890808, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9915440678596497, 'rewards/repeat_soft/std': 0.013796639628708363, 'rewards/judge_quality/mean': 0.26874998211860657, 'rewards/judge_quality/std': 0.13505950570106506, 'rewards/total_composite/mean': 0.33347609639167786, 'rewards/total_composite/std': 0.22276601195335388, 'reward': 0.33347609639167786, 'reward_std': 0.22276602685451508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24162688851356506, 'sampling/sampling_logp_difference/max': 1.4961953163146973, 'sampling/importance_sampling_ratio/min': 0.22398072481155396, 'sampling/importance_sampling_ratio/mean': 1.0915111303329468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.0247719287872314, 'clip_ratio/low_mean': 0.043165650218725204, 'clip_ratio/low_min': 0.043165650218725204, 'clip_ratio/high_mean': 0.14359846152365208, 'clip_ratio/high_max': 0.14359846152365208, 'clip_ratio/region_mean': 0.18676411174237728, 'reward_total_mean': 0.33347609639167786, 'reward_meter_mean': 0.5199936628341675, 'reward_meter_std': 0.3432442545890808, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9915440678596497, 'reward_repeat_soft_std': 0.013796639628708363, 'reward_judge_quality_mean': 0.26874998211860657, 'reward_judge_quality_std': 0.13505950570106506, 'reward_total_composite_mean': 0.33347609639167786, 'reward_total_composite_std': 0.22276601195335388, 'epoch': 0.02}

  7%|▋         | 235/3300 [33:00<6:53:48,  8.10s/it]INFO 04-13 07:53:58 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 236/3300 [33:08<6:59:49,  8.22s/it]2026-04-13 07:54:06,342 | INFO | train_grpo_train | metrics_logged mode=train step=236

                                                    
{'loss': 0.0838, 'grad_norm': 11.439743041992188, 'learning_rate': 9.28787878787879e-06, 'num_tokens': 437304.0, 'completions/mean_length': 122.0, 'completions/min_length': 104.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.0, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.6040682792663574, 'rewards/meter/std': 0.22797472774982452, 'rewards/count_adherence/mean': 0.9166666269302368, 'rewards/count_adherence/std': 0.0890870913863182, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9847686886787415, 'rewards/repeat_soft/std': 0.017375530675053596, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5345677137374878, 'rewards/total_composite/std': 0.10127469152212143, 'reward': 0.5345677137374878, 'reward_std': 0.10127468407154083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23363074660301208, 'sampling/sampling_logp_difference/max': 1.8927171230316162, 'sampling/importance_sampling_ratio/min': 0.1506618857383728, 'sampling/importance_sampling_ratio/mean': 1.0151104927062988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7888659238815308, 'clip_ratio/low_mean': 0.11487327516078949, 'clip_ratio/low_min': 0.11487327516078949, 'clip_ratio/high_mean': 0.12245976366102695, 'clip_ratio/high_max': 0.12245976366102695, 'clip_ratio/region_mean': 0.23733303882181644, 'reward_total_mean': 0.5345677137374878, 'reward_meter_mean': 0.6040682792663574, 'reward_meter_std': 0.22797472774982452, 'reward_count_adherence_mean': 0.9166666269302368, 'reward_count_adherence_std': 0.0890870913863182, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9847686886787415, 'reward_repeat_soft_std': 0.017375530675053596, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5345677137374878, 'reward_total_composite_std': 0.10127469152212143, 'epoch': 0.02}

  7%|▋         | 236/3300 [33:08<6:59:49,  8.22s/it]INFO 04-13 07:54:06 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 237/3300 [33:20<7:59:24,  9.39s/it]2026-04-13 07:54:18,463 | INFO | train_grpo_train | metrics_logged mode=train step=237

                                                    
{'loss': -0.1479, 'grad_norm': 3.540306568145752, 'learning_rate': 9.284848484848485e-06, 'num_tokens': 439032.0, 'completions/mean_length': 126.0, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.85714721679688, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.33497804403305054, 'rewards/meter/std': 0.40851643681526184, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9910533428192139, 'rewards/repeat_soft/std': 0.014115079306066036, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.1345893144607544, 'rewards/total_composite/mean': 0.38116806745529175, 'rewards/total_composite/std': 0.17786787450313568, 'reward': 0.38116806745529175, 'reward_std': 0.17786787450313568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23217308521270752, 'sampling/sampling_logp_difference/max': 1.4035301208496094, 'sampling/importance_sampling_ratio/min': 0.24572797119617462, 'sampling/importance_sampling_ratio/mean': 1.089635968208313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.430901527404785, 'clip_ratio/low_mean': 0.055441899225115776, 'clip_ratio/low_min': 0.055441899225115776, 'clip_ratio/high_mean': 0.11968373786658049, 'clip_ratio/high_max': 0.11968373786658049, 'clip_ratio/region_mean': 0.17512563709169626, 'reward_total_mean': 0.38116806745529175, 'reward_meter_mean': 0.33497804403305054, 'reward_meter_std': 0.40851643681526184, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9910533428192139, 'reward_repeat_soft_std': 0.014115079306066036, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.1345893144607544, 'reward_total_composite_mean': 0.38116806745529175, 'reward_total_composite_std': 0.17786787450313568, 'epoch': 0.02}

  7%|▋         | 237/3300 [33:20<7:59:24,  9.39s/it]INFO 04-13 07:54:18 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 238/3300 [33:28<7:42:29,  9.06s/it]2026-04-13 07:54:26,762 | INFO | train_grpo_train | metrics_logged mode=train step=238

                                                    
{'loss': 0.0075, 'grad_norm': 10.821329116821289, 'learning_rate': 9.281818181818183e-06, 'num_tokens': 440710.0, 'completions/mean_length': 52.75, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9594722390174866, 'rewards/meter/std': 0.06740762293338776, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9954468607902527, 'rewards/repeat_soft/std': 0.00810985453426838, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.1440672129392624, 'rewards/total_composite/mean': 0.6222023367881775, 'rewards/total_composite/std': 0.10977651923894882, 'reward': 0.6222023367881775, 'reward_std': 0.10977650433778763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21899792551994324, 'sampling/sampling_logp_difference/max': 1.0820188522338867, 'sampling/importance_sampling_ratio/min': 0.33891063928604126, 'sampling/importance_sampling_ratio/mean': 1.0673742294311523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.817643642425537, 'clip_ratio/low_mean': 0.12432807870209217, 'clip_ratio/low_min': 0.12432807870209217, 'clip_ratio/high_mean': 0.03415032755583525, 'clip_ratio/high_max': 0.03415032755583525, 'clip_ratio/region_mean': 0.15847840625792742, 'reward_total_mean': 0.6222023367881775, 'reward_meter_mean': 0.9594722390174866, 'reward_meter_std': 0.06740762293338776, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9954468607902527, 'reward_repeat_soft_std': 0.00810985453426838, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.1440672129392624, 'reward_total_composite_mean': 0.6222023367881775, 'reward_total_composite_std': 0.10977651923894882, 'epoch': 0.02}

  7%|▋         | 238/3300 [33:28<7:42:29,  9.06s/it]INFO 04-13 07:54:26 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 239/3300 [33:35<7:10:27,  8.44s/it]2026-04-13 07:54:33,740 | INFO | train_grpo_train | metrics_logged mode=train step=239

                                                    
{'loss': 0.1266, 'grad_norm': 16.885149002075195, 'learning_rate': 9.27878787878788e-06, 'num_tokens': 442355.0, 'completions/mean_length': 42.625, 'completions/min_length': 29.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5577026605606079, 'rewards/meter/std': 0.41305047273635864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9940649271011353, 'rewards/repeat_soft/std': 0.015189046040177345, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.4251362085342407, 'rewards/total_composite/std': 0.199260875582695, 'reward': 0.4251362085342407, 'reward_std': 0.1992608904838562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23422786593437195, 'sampling/sampling_logp_difference/max': 1.3673014640808105, 'sampling/importance_sampling_ratio/min': 0.25479358434677124, 'sampling/importance_sampling_ratio/mean': 1.0716608762741089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5329028964042664, 'clip_ratio/low_mean': 0.08765532821416855, 'clip_ratio/low_min': 0.08765532821416855, 'clip_ratio/high_mean': 0.08749033231288195, 'clip_ratio/high_max': 0.08749033231288195, 'clip_ratio/region_mean': 0.1751456605270505, 'reward_total_mean': 0.4251362085342407, 'reward_meter_mean': 0.5577026605606079, 'reward_meter_std': 0.41305047273635864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9940649271011353, 'reward_repeat_soft_std': 0.015189046040177345, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.4251362085342407, 'reward_total_composite_std': 0.199260875582695, 'epoch': 0.02}

  7%|▋         | 239/3300 [33:35<7:10:27,  8.44s/it]INFO 04-13 07:54:33 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 240/3300 [33:43<6:53:30,  8.11s/it]2026-04-13 07:54:41,079 | INFO | train_grpo_train | metrics_logged mode=train step=240

                                                    
{'loss': 0.1303, 'grad_norm': 11.940001487731934, 'learning_rate': 9.275757575757577e-06, 'num_tokens': 444059.0, 'completions/mean_length': 56.0, 'completions/min_length': 45.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.4955418109893799, 'rewards/meter/std': 0.45711708068847656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9967243671417236, 'rewards/repeat_soft/std': 0.005859177093952894, 'rewards/judge_quality/mean': 0.33125001192092896, 'rewards/judge_quality/std': 0.13611313700675964, 'rewards/total_composite/mean': 0.3668709099292755, 'rewards/total_composite/std': 0.25199586153030396, 'reward': 0.3668709099292755, 'reward_std': 0.25199583172798157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.26978856325149536, 'sampling/sampling_logp_difference/max': 1.5231103897094727, 'sampling/importance_sampling_ratio/min': 0.21803268790245056, 'sampling/importance_sampling_ratio/mean': 1.0623533725738525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.923114836215973, 'clip_ratio/low_mean': 0.046531341038644314, 'clip_ratio/low_min': 0.046531341038644314, 'clip_ratio/high_mean': 0.12861622124910355, 'clip_ratio/high_max': 0.12861622124910355, 'clip_ratio/region_mean': 0.17514756228774786, 'reward_total_mean': 0.3668709099292755, 'reward_meter_mean': 0.4955418109893799, 'reward_meter_std': 0.45711708068847656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9967243671417236, 'reward_repeat_soft_std': 0.005859177093952894, 'reward_judge_quality_mean': 0.33125001192092896, 'reward_judge_quality_std': 0.13611313700675964, 'reward_total_composite_mean': 0.3668709099292755, 'reward_total_composite_std': 0.25199586153030396, 'epoch': 0.02}

  7%|▋         | 240/3300 [33:43<6:53:30,  8.11s/it]INFO 04-13 07:54:41 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 241/3300 [33:57<8:21:38,  9.84s/it]2026-04-13 07:54:54,957 | INFO | train_grpo_train | metrics_logged mode=train step=241

                                                    
{'loss': 0.0658, 'grad_norm': 12.133151054382324, 'learning_rate': 9.272727272727273e-06, 'num_tokens': 445973.0, 'completions/mean_length': 76.25, 'completions/min_length': 54.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.19163626432418823, 'rewards/meter/std': 0.26619383692741394, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9973886609077454, 'rewards/repeat_soft/std': 0.0026149640325456858, 'rewards/judge_quality/mean': 0.6025000214576721, 'rewards/judge_quality/std': 0.35443115234375, 'rewards/total_composite/mean': 0.41041338443756104, 'rewards/total_composite/std': 0.22805547714233398, 'reward': 0.41041338443756104, 'reward_std': 0.22805547714233398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22107474505901337, 'sampling/sampling_logp_difference/max': 1.2895097732543945, 'sampling/importance_sampling_ratio/min': 0.27540576457977295, 'sampling/importance_sampling_ratio/mean': 1.0179635286331177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.865700677037239, 'clip_ratio/low_mean': 0.08136807568371296, 'clip_ratio/low_min': 0.08136807568371296, 'clip_ratio/high_mean': 0.09546291641891003, 'clip_ratio/high_max': 0.09546291641891003, 'clip_ratio/region_mean': 0.17683099210262299, 'reward_total_mean': 0.41041338443756104, 'reward_meter_mean': 0.19163626432418823, 'reward_meter_std': 0.26619383692741394, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9973886609077454, 'reward_repeat_soft_std': 0.0026149640325456858, 'reward_judge_quality_mean': 0.6025000214576721, 'reward_judge_quality_std': 0.35443115234375, 'reward_total_composite_mean': 0.41041338443756104, 'reward_total_composite_std': 0.22805547714233398, 'epoch': 0.02}

  7%|▋         | 241/3300 [33:57<8:21:38,  9.84s/it]INFO 04-13 07:54:55 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 242/3300 [34:03<7:27:49,  8.79s/it]2026-04-13 07:55:01,288 | INFO | train_grpo_train | metrics_logged mode=train step=242

                                                    
{'loss': 0.0012, 'grad_norm': 24.344709396362305, 'learning_rate': 9.26969696969697e-06, 'num_tokens': 447409.0, 'completions/mean_length': 22.5, 'completions/min_length': 20.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7496224641799927, 'rewards/meter/std': 0.4358169138431549, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5515627264976501, 'rewards/total_composite/std': 0.12039294838905334, 'reward': 0.5515627264976501, 'reward_std': 0.12039293348789215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07915819436311722, 'sampling/sampling_logp_difference/max': 0.7866744995117188, 'sampling/importance_sampling_ratio/min': 0.45535656809806824, 'sampling/importance_sampling_ratio/mean': 0.9973716139793396, 'sampling/importance_sampling_ratio/max': 1.668638825416565, 'entropy': 0.6392558962106705, 'clip_ratio/low_mean': 0.07604166679084301, 'clip_ratio/low_min': 0.07604166679084301, 'clip_ratio/high_mean': 0.02766798483207822, 'clip_ratio/high_max': 0.02766798483207822, 'clip_ratio/region_mean': 0.10370965162292123, 'reward_total_mean': 0.5515627264976501, 'reward_meter_mean': 0.7496224641799927, 'reward_meter_std': 0.4358169138431549, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5515627264976501, 'reward_total_composite_std': 0.12039294838905334, 'epoch': 0.02}

  7%|▋         | 242/3300 [34:03<7:27:49,  8.79s/it]INFO 04-13 07:55:01 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 243/3300 [34:10<7:07:41,  8.39s/it]2026-04-13 07:55:08,768 | INFO | train_grpo_train | metrics_logged mode=train step=243

                                                    
{'loss': 0.1421, 'grad_norm': 12.978425025939941, 'learning_rate': 9.266666666666667e-06, 'num_tokens': 449253.0, 'completions/mean_length': 67.5, 'completions/min_length': 44.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.4874016046524048, 'rewards/meter/std': 0.4216694235801697, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921243190765381, 'rewards/repeat_soft/std': 0.012025803327560425, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4441221356391907, 'rewards/total_composite/std': 0.13617344200611115, 'reward': 0.4441221356391907, 'reward_std': 0.13617344200611115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2546895444393158, 'sampling/sampling_logp_difference/max': 1.8767824172973633, 'sampling/importance_sampling_ratio/min': 0.15308187901973724, 'sampling/importance_sampling_ratio/mean': 1.0500438213348389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.821613907814026, 'clip_ratio/low_mean': 0.1031344085931778, 'clip_ratio/low_min': 0.1031344085931778, 'clip_ratio/high_mean': 0.10280172526836395, 'clip_ratio/high_max': 0.10280172526836395, 'clip_ratio/region_mean': 0.20593613386154175, 'reward_total_mean': 0.4441221356391907, 'reward_meter_mean': 0.4874016046524048, 'reward_meter_std': 0.4216694235801697, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921243190765381, 'reward_repeat_soft_std': 0.012025803327560425, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4441221356391907, 'reward_total_composite_std': 0.13617344200611115, 'epoch': 0.02}

  7%|▋         | 243/3300 [34:10<7:07:41,  8.39s/it]INFO 04-13 07:55:08 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 244/3300 [34:18<6:55:33,  8.16s/it]2026-04-13 07:55:16,376 | INFO | train_grpo_train | metrics_logged mode=train step=244

                                                    
{'loss': -0.0164, 'grad_norm': 10.075081825256348, 'learning_rate': 9.263636363636364e-06, 'num_tokens': 450906.0, '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': 0.27877378463745117, 'rewards/meter/std': 0.3891329765319824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9942314624786377, 'rewards/repeat_soft/std': 0.011857480742037296, 'rewards/judge_quality/mean': 0.4337500333786011, 'rewards/judge_quality/std': 0.15528200566768646, 'rewards/total_composite/mean': 0.4079081416130066, 'rewards/total_composite/std': 0.23522789776325226, 'reward': 0.4079081416130066, 'reward_std': 0.23522788286209106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22137722373008728, 'sampling/sampling_logp_difference/max': 1.596456527709961, 'sampling/importance_sampling_ratio/min': 0.2026132047176361, 'sampling/importance_sampling_ratio/mean': 1.0437744855880737, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6175760626792908, 'clip_ratio/low_mean': 0.09715724922716618, 'clip_ratio/low_min': 0.09715724922716618, 'clip_ratio/high_mean': 0.05981331318616867, 'clip_ratio/high_max': 0.05981331318616867, 'clip_ratio/region_mean': 0.15697056241333485, 'reward_total_mean': 0.4079081416130066, 'reward_meter_mean': 0.27877378463745117, 'reward_meter_std': 0.3891329765319824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9942314624786377, 'reward_repeat_soft_std': 0.011857480742037296, 'reward_judge_quality_mean': 0.4337500333786011, 'reward_judge_quality_std': 0.15528200566768646, 'reward_total_composite_mean': 0.4079081416130066, 'reward_total_composite_std': 0.23522789776325226, 'epoch': 0.02}

  7%|▋         | 244/3300 [34:18<6:55:33,  8.16s/it]INFO 04-13 07:55:16 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 245/3300 [34:26<6:51:35,  8.08s/it]2026-04-13 07:55:24,283 | INFO | train_grpo_train | metrics_logged mode=train step=245

                                                    
{'loss': 0.0556, 'grad_norm': 13.549051284790039, 'learning_rate': 9.260606060606062e-06, 'num_tokens': 452696.0, 'completions/mean_length': 49.75, 'completions/min_length': 36.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7952672243118286, 'rewards/meter/std': 0.2641472816467285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9996022582054138, 'rewards/repeat_soft/std': 0.001124941511079669, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.10336308926343918, 'rewards/total_composite/mean': 0.4710841178894043, 'rewards/total_composite/std': 0.20893138647079468, 'reward': 0.4710841178894043, 'reward_std': 0.20893137156963348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22622895240783691, 'sampling/sampling_logp_difference/max': 1.295872688293457, 'sampling/importance_sampling_ratio/min': 0.2736589312553406, 'sampling/importance_sampling_ratio/mean': 1.0102380514144897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.124099910259247, 'clip_ratio/low_mean': 0.043450853787362576, 'clip_ratio/low_min': 0.043450853787362576, 'clip_ratio/high_mean': 0.15173144079744816, 'clip_ratio/high_max': 0.15173144079744816, 'clip_ratio/region_mean': 0.19518229458481073, 'reward_total_mean': 0.4710841178894043, 'reward_meter_mean': 0.7952672243118286, 'reward_meter_std': 0.2641472816467285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9996022582054138, 'reward_repeat_soft_std': 0.001124941511079669, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.10336308926343918, 'reward_total_composite_mean': 0.4710841178894043, 'reward_total_composite_std': 0.20893138647079468, 'epoch': 0.02}

  7%|▋         | 245/3300 [34:26<6:51:35,  8.08s/it]INFO 04-13 07:55:24 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 246/3300 [34:37<7:28:33,  8.81s/it]2026-04-13 07:55:34,800 | INFO | train_grpo_train | metrics_logged mode=train step=246

                                                    
{'loss': -0.0181, 'grad_norm': 5.99717378616333, 'learning_rate': 9.257575757575759e-06, 'num_tokens': 455847.0, 'completions/mean_length': 197.875, 'completions/min_length': 136.0, 'completions/max_length': 296.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 197.875, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 296.0, 'rewards/meter/mean': 0.19821536540985107, 'rewards/meter/std': 0.17813539505004883, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9967764616012573, 'rewards/repeat_soft/std': 0.00359341804869473, 'rewards/judge_quality/mean': 0.2137499898672104, 'rewards/judge_quality/std': 0.10155048221349716, 'rewards/total_composite/mean': 0.3095589280128479, 'rewards/total_composite/std': 0.13021212816238403, 'reward': 0.3095589280128479, 'reward_std': 0.13021212816238403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23955827951431274, 'sampling/sampling_logp_difference/max': 1.8856773376464844, 'sampling/importance_sampling_ratio/min': 0.15172624588012695, 'sampling/importance_sampling_ratio/mean': 1.078747272491455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.056632041931152, 'clip_ratio/low_mean': 0.01977401040494442, 'clip_ratio/low_min': 0.01977401040494442, 'clip_ratio/high_mean': 0.15362998005002737, 'clip_ratio/high_max': 0.15362998005002737, 'clip_ratio/region_mean': 0.1734039904549718, 'reward_total_mean': 0.3095589280128479, 'reward_meter_mean': 0.19821536540985107, 'reward_meter_std': 0.17813539505004883, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9967764616012573, 'reward_repeat_soft_std': 0.00359341804869473, 'reward_judge_quality_mean': 0.2137499898672104, 'reward_judge_quality_std': 0.10155048221349716, 'reward_total_composite_mean': 0.3095589280128479, 'reward_total_composite_std': 0.13021212816238403, 'epoch': 0.02}

  7%|▋         | 246/3300 [34:37<7:28:33,  8.81s/it]INFO 04-13 07:55:35 [block_pool.py:378] Successfully reset prefix cache

  7%|▋         | 247/3300 [34:45<7:28:06,  8.81s/it]2026-04-13 07:55:43,592 | INFO | train_grpo_train | metrics_logged mode=train step=247

                                                    
{'loss': -0.0468, 'grad_norm': 12.217455863952637, 'learning_rate': 9.254545454545454e-06, 'num_tokens': 457542.0, 'completions/mean_length': 51.875, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.40306398272514343, 'rewards/meter/std': 0.3892579972743988, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.997988760471344, 'rewards/repeat_soft/std': 0.0030363963451236486, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.2583388388156891, 'rewards/total_composite/std': 0.2296733856201172, 'reward': 0.2583388388156891, 'reward_std': 0.22967340052127838, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24870580434799194, 'sampling/sampling_logp_difference/max': 1.519944190979004, 'sampling/importance_sampling_ratio/min': 0.2187241166830063, 'sampling/importance_sampling_ratio/mean': 1.0517011880874634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1832044422626495, 'clip_ratio/low_mean': 0.07400570623576641, 'clip_ratio/low_min': 0.07400570623576641, 'clip_ratio/high_mean': 0.1226517204195261, 'clip_ratio/high_max': 0.1226517204195261, 'clip_ratio/region_mean': 0.1966574266552925, 'reward_total_mean': 0.2583388388156891, 'reward_meter_mean': 0.40306398272514343, 'reward_meter_std': 0.3892579972743988, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.997988760471344, 'reward_repeat_soft_std': 0.0030363963451236486, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.2583388388156891, 'reward_total_composite_std': 0.2296733856201172, 'epoch': 0.02}

  7%|▋         | 247/3300 [34:45<7:28:06,  8.81s/it]INFO 04-13 07:55:43 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 248/3300 [34:57<8:15:57,  9.75s/it]2026-04-13 07:55:55,543 | INFO | train_grpo_train | metrics_logged mode=train step=248

                                                    
{'loss': -0.0577, 'grad_norm': 5.521716117858887, 'learning_rate': 9.251515151515152e-06, 'num_tokens': 459522.0, 'completions/mean_length': 151.5, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 100.00000762939453, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.10642732679843903, 'rewards/meter/std': 0.13759547472000122, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9973168969154358, 'rewards/repeat_soft/std': 0.00279581593349576, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.21022097766399384, 'rewards/total_composite/mean': 0.27132105827331543, 'rewards/total_composite/std': 0.17161008715629578, 'reward': 0.27132105827331543, 'reward_std': 0.17161008715629578, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23343904316425323, 'sampling/sampling_logp_difference/max': 1.4936189651489258, 'sampling/importance_sampling_ratio/min': 0.2245585173368454, 'sampling/importance_sampling_ratio/mean': 1.0569976568222046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9061760902404785, 'clip_ratio/low_mean': 0.01772388070821762, 'clip_ratio/low_min': 0.01772388070821762, 'clip_ratio/high_mean': 0.15801917761564255, 'clip_ratio/high_max': 0.15801917761564255, 'clip_ratio/region_mean': 0.17574305832386017, 'reward_total_mean': 0.27132105827331543, 'reward_meter_mean': 0.10642732679843903, 'reward_meter_std': 0.13759547472000122, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9973168969154358, 'reward_repeat_soft_std': 0.00279581593349576, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.21022097766399384, 'reward_total_composite_mean': 0.27132105827331543, 'reward_total_composite_std': 0.17161008715629578, 'epoch': 0.02}

  8%|▊         | 248/3300 [34:57<8:15:57,  9.75s/it]INFO 04-13 07:55:55 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 249/3300 [35:03<7:21:53,  8.69s/it]2026-04-13 07:56:01,761 | INFO | train_grpo_train | metrics_logged mode=train step=249

                                                    
{'loss': 0.0666, 'grad_norm': 23.50229263305664, 'learning_rate': 9.248484848484849e-06, 'num_tokens': 460936.0, 'completions/mean_length': 31.75, 'completions/min_length': 24.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6344245076179504, 'rewards/meter/std': 0.3527393937110901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9875197410583496, 'rewards/repeat_soft/std': 0.01588606648147106, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5660109519958496, 'rewards/total_composite/std': 0.13045436143875122, 'reward': 0.5660109519958496, 'reward_std': 0.13045434653759003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18733884394168854, 'sampling/sampling_logp_difference/max': 1.8012290000915527, 'sampling/importance_sampling_ratio/min': 0.16509586572647095, 'sampling/importance_sampling_ratio/mean': 1.0040596723556519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1648800373077393, 'clip_ratio/low_mean': 0.06251262687146664, 'clip_ratio/low_min': 0.06251262687146664, 'clip_ratio/high_mean': 0.11953509598970413, 'clip_ratio/high_max': 0.11953509598970413, 'clip_ratio/region_mean': 0.18204772286117077, 'reward_total_mean': 0.5660109519958496, 'reward_meter_mean': 0.6344245076179504, 'reward_meter_std': 0.3527393937110901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9875197410583496, 'reward_repeat_soft_std': 0.01588606648147106, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5660109519958496, 'reward_total_composite_std': 0.13045436143875122, 'epoch': 0.03}

  8%|▊         | 249/3300 [35:03<7:21:53,  8.69s/it]INFO 04-13 07:56:01 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 250/3300 [35:30<11:50:02, 13.97s/it]2026-04-13 07:56:28,043 | INFO | train_grpo_train | metrics_logged mode=train step=250

                                                     
{'loss': -0.2321, 'grad_norm': 2.7613444328308105, 'learning_rate': 9.245454545454546e-06, 'num_tokens': 463523.0, 'completions/mean_length': 309.375, 'completions/min_length': 160.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 187.8000030517578, 'completions/min_terminated_length': 160.0, 'completions/max_terminated_length': 234.0, 'rewards/meter/mean': 0.1741102635860443, 'rewards/meter/std': 0.27033716440200806, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.1511857807636261, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.999889612197876, 'rewards/repeat_soft/std': 0.00021023683075327426, 'rewards/judge_quality/mean': 0.11749999970197678, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.16109412908554077, 'rewards/total_composite/std': 0.17332017421722412, 'reward': 0.16109412908554077, 'reward_std': 0.17332017421722412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24947352707386017, 'sampling/sampling_logp_difference/max': 1.2111396789550781, 'sampling/importance_sampling_ratio/min': 0.2978576123714447, 'sampling/importance_sampling_ratio/mean': 1.0666618347167969, 'sampling/importance_sampling_ratio/max': 1.9700634479522705, 'entropy': 2.7018612027168274, 'clip_ratio/low_mean': 0.01875000074505806, 'clip_ratio/low_min': 0.01875000074505806, 'clip_ratio/high_mean': 0.08803261630237103, 'clip_ratio/high_max': 0.08803261630237103, 'clip_ratio/region_mean': 0.10678261704742908, 'reward_total_mean': 0.16109412908554077, 'reward_meter_mean': 0.1741102635860443, 'reward_meter_std': 0.27033716440200806, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.1511857807636261, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.999889612197876, 'reward_repeat_soft_std': 0.00021023683075327426, 'reward_judge_quality_mean': 0.11749999970197678, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.16109412908554077, 'reward_total_composite_std': 0.17332017421722412, 'epoch': 0.03}

  8%|▊         | 250/3300 [35:30<11:50:02, 13.97s/it]INFO 04-13 07:56:28 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:19,  2.42s/it]

 30%|███       | 3/10 [00:14<00:38,  5.52s/it]

 40%|████      | 4/10 [00:23<00:41,  6.84s/it]

 50%|█████     | 5/10 [00:28<00:31,  6.27s/it]

 60%|██████    | 6/10 [00:43<00:36,  9.16s/it]

 70%|███████   | 7/10 [00:50<00:24,  8.26s/it]

 80%|████████  | 8/10 [00:54<00:14,  7.13s/it]

 90%|█████████ | 9/10 [00:59<00:06,  6.40s/it]

100%|██████████| 10/10 [01:03<00:00,  5.61s/it]2026-04-13 07:57:42,149 | INFO | train_grpo_train | metrics_logged mode=eval step=250

                                                     

                                               
{'eval_loss': nan, 'eval_runtime': 74.0879, 'eval_samples_per_second': 1.08, 'eval_steps_per_second': 0.135, 'eval_num_tokens': 463523.0, 'eval_completions/mean_length': 105.1875, 'eval_completions/min_length': 33.0, 'eval_completions/max_length': 279.8, 'eval_completions/clipped_ratio': 0.0375, 'eval_completions/mean_terminated_length': 89.50357284545899, 'eval_completions/min_terminated_length': 33.0, 'eval_completions/max_terminated_length': 187.5, 'eval_rewards/meter/mean': 0.4927772879600525, 'eval_rewards/meter/std': 0.36216977536678313, 'eval_rewards/count_adherence/mean': 0.9197916686534882, 'eval_rewards/count_adherence/std': 0.15851535201072692, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.2230676978826523, 'eval_rewards/repeat_soft/mean': 0.9888128161430358, 'eval_rewards/repeat_soft/std': 0.02121427790261805, 'eval_rewards/judge_quality/mean': 0.3487500011920929, 'eval_rewards/judge_quality/std': 0.1130803014151752, 'eval_rewards/total_composite/mean': 0.41938657462596896, 'eval_rewards/total_composite/std': 0.14896008148789405, 'eval_reward': 0.41938657462596896, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.18254771083593369, 'eval_sampling/sampling_logp_difference/max': 1.249609899520874, 'eval_sampling/importance_sampling_ratio/min': 0.29058967232704164, 'eval_sampling/importance_sampling_ratio/mean': 1.0555893182754517, 'eval_sampling/importance_sampling_ratio/max': 1.7040584444999696, 'eval_entropy': 3.2519503593444825, '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': 0.41938657462596896, 'eval_reward_meter_mean': 0.4927772879600525, 'eval_reward_meter_std': 0.36216977536678313, 'eval_reward_count_adherence_mean': 0.9197916686534882, 'eval_reward_count_adherence_std': 0.15851535201072692, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.2230676978826523, 'eval_reward_repeat_soft_mean': 0.9888128161430358, 'eval_reward_repeat_soft_std': 0.02121427790261805, 'eval_reward_judge_quality_mean': 0.3487500011920929, 'eval_reward_judge_quality_std': 0.1130803014151752, 'eval_reward_total_composite_mean': 0.41938657462596896, 'eval_reward_total_composite_std': 0.14896008148789405, 'epoch': 0.03}

  8%|▊         | 250/3300 [36:44<11:50:02, 13.97s/it]

100%|██████████| 10/10 [01:03<00:00,  5.61s/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 [36:55<29:50:36, 35.24s/it]2026-04-13 07:57:52,908 | INFO | train_grpo_train | metrics_logged mode=train step=251

                                                     
{'loss': 0.0505, 'grad_norm': 15.902443885803223, 'learning_rate': 9.242424242424244e-06, 'num_tokens': 465093.0, 'completions/mean_length': 40.25, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.18963100016117096, 'rewards/meter/std': 0.2852773070335388, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9908022880554199, 'rewards/repeat_soft/std': 0.007465601898729801, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.40558940172195435, 'rewards/total_composite/std': 0.0778096541762352, 'reward': 0.40558940172195435, 'reward_std': 0.0778096541762352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22490194439888, 'sampling/sampling_logp_difference/max': 1.2477350234985352, 'sampling/importance_sampling_ratio/min': 0.2871544659137726, 'sampling/importance_sampling_ratio/mean': 1.0595271587371826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8220799267292023, 'clip_ratio/low_mean': 0.12860641907900572, 'clip_ratio/low_min': 0.12860641907900572, 'clip_ratio/high_mean': 0.031746032647788525, 'clip_ratio/high_max': 0.031746032647788525, 'clip_ratio/region_mean': 0.16035245172679424, 'reward_total_mean': 0.40558940172195435, 'reward_meter_mean': 0.18963100016117096, 'reward_meter_std': 0.2852773070335388, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9908022880554199, 'reward_repeat_soft_std': 0.007465601898729801, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.40558940172195435, 'reward_total_composite_std': 0.0778096541762352, 'epoch': 0.03}

  8%|▊         | 251/3300 [36:55<29:50:36, 35.24s/it]INFO 04-13 07:57:53 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 252/3300 [37:03<22:56:43, 27.10s/it]2026-04-13 07:58:01,025 | INFO | train_grpo_train | metrics_logged mode=train step=252

                                                     
{'loss': 0.0901, 'grad_norm': 11.631461143493652, 'learning_rate': 9.23939393939394e-06, 'num_tokens': 466985.0, 'completions/mean_length': 68.5, 'completions/min_length': 47.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.23365449905395508, 'rewards/meter/std': 0.22981607913970947, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9940780401229858, 'rewards/repeat_soft/std': 0.012517235241830349, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40456604957580566, 'rewards/total_composite/std': 0.0727265328168869, 'reward': 0.40456604957580566, 'reward_std': 0.0727265328168869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2311936616897583, 'sampling/sampling_logp_difference/max': 1.5864734649658203, 'sampling/importance_sampling_ratio/min': 0.2046460211277008, 'sampling/importance_sampling_ratio/mean': 1.0510509014129639, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.064756900072098, 'clip_ratio/low_mean': 0.11715154815465212, 'clip_ratio/low_min': 0.11715154815465212, 'clip_ratio/high_mean': 0.06316489353775978, 'clip_ratio/high_max': 0.06316489353775978, 'clip_ratio/region_mean': 0.1803164416924119, 'reward_total_mean': 0.40456604957580566, 'reward_meter_mean': 0.23365449905395508, 'reward_meter_std': 0.22981607913970947, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9940780401229858, 'reward_repeat_soft_std': 0.012517235241830349, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40456604957580566, 'reward_total_composite_std': 0.0727265328168869, 'epoch': 0.03}

  8%|▊         | 252/3300 [37:03<22:56:43, 27.10s/it]INFO 04-13 07:58:01 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 253/3300 [37:10<17:51:19, 21.10s/it]2026-04-13 07:58:08,108 | INFO | train_grpo_train | metrics_logged mode=train step=253

                                                     
{'loss': 0.2207, 'grad_norm': 14.660782814025879, 'learning_rate': 9.236363636363636e-06, 'num_tokens': 468575.0, 'completions/mean_length': 46.75, 'completions/min_length': 26.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.6230303049087524, 'rewards/meter/std': 0.45745065808296204, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9883996844291687, 'rewards/repeat_soft/std': 0.020203450694680214, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.2499571591615677, 'rewards/total_composite/mean': 0.5688772201538086, 'rewards/total_composite/std': 0.24092671275138855, 'reward': 0.5688772201538086, 'reward_std': 0.24092671275138855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21159371733665466, 'sampling/sampling_logp_difference/max': 1.2312355041503906, 'sampling/importance_sampling_ratio/min': 0.2919316589832306, 'sampling/importance_sampling_ratio/mean': 1.074669599533081, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.669960081577301, 'clip_ratio/low_mean': 0.08137993142008781, 'clip_ratio/low_min': 0.08137993142008781, 'clip_ratio/high_mean': 0.0769733702763915, 'clip_ratio/high_max': 0.0769733702763915, 'clip_ratio/region_mean': 0.15835330169647932, 'reward_total_mean': 0.5688772201538086, 'reward_meter_mean': 0.6230303049087524, 'reward_meter_std': 0.45745065808296204, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9883996844291687, 'reward_repeat_soft_std': 0.020203450694680214, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.2499571591615677, 'reward_total_composite_mean': 0.5688772201538086, 'reward_total_composite_std': 0.24092671275138855, 'epoch': 0.03}

  8%|▊         | 253/3300 [37:10<17:51:19, 21.10s/it]INFO 04-13 07:58:08 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 254/3300 [37:30<17:44:34, 20.97s/it]2026-04-13 07:58:28,784 | INFO | train_grpo_train | metrics_logged mode=train step=254

                                                     
{'loss': 0.0185, 'grad_norm': 16.987140655517578, 'learning_rate': 9.233333333333334e-06, 'num_tokens': 470489.0, 'completions/mean_length': 71.25, 'completions/min_length': 57.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.5752930641174316, 'rewards/meter/std': 0.3816414475440979, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9927878975868225, 'rewards/repeat_soft/std': 0.007032928988337517, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.46241429448127747, 'rewards/total_composite/std': 0.10644890367984772, 'reward': 0.46241429448127747, 'reward_std': 0.10644891113042831, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22552765905857086, 'sampling/sampling_logp_difference/max': 2.155546188354492, 'sampling/importance_sampling_ratio/min': 0.11583990603685379, 'sampling/importance_sampling_ratio/mean': 1.0416035652160645, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.875844269990921, 'clip_ratio/low_mean': 0.11048714444041252, 'clip_ratio/low_min': 0.11048714444041252, 'clip_ratio/high_mean': 0.10103101842105389, 'clip_ratio/high_max': 0.10103101842105389, 'clip_ratio/region_mean': 0.2115181628614664, 'reward_total_mean': 0.46241429448127747, 'reward_meter_mean': 0.5752930641174316, 'reward_meter_std': 0.3816414475440979, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9927878975868225, 'reward_repeat_soft_std': 0.007032928988337517, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.46241429448127747, 'reward_total_composite_std': 0.10644890367984772, 'epoch': 0.03}

  8%|▊         | 254/3300 [37:31<17:44:34, 20.97s/it]INFO 04-13 07:58:29 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 255/3300 [37:38<14:20:59, 16.97s/it]2026-04-13 07:58:36,404 | INFO | train_grpo_train | metrics_logged mode=train step=255

                                                     
{'loss': -0.0058, 'grad_norm': 9.851515769958496, 'learning_rate': 9.23030303030303e-06, 'num_tokens': 472370.0, 'completions/mean_length': 61.125, 'completions/min_length': 34.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.6299946308135986, 'rewards/meter/std': 0.4327425956726074, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9937536120414734, 'rewards/repeat_soft/std': 0.012771286070346832, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.09538455307483673, 'rewards/total_composite/mean': 0.4724371135234833, 'rewards/total_composite/std': 0.14866700768470764, 'reward': 0.4724371135234833, 'reward_std': 0.14866699278354645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23003731667995453, 'sampling/sampling_logp_difference/max': 1.5137968063354492, 'sampling/importance_sampling_ratio/min': 0.22007280588150024, 'sampling/importance_sampling_ratio/mean': 1.0665873289108276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4186623990535736, 'clip_ratio/low_mean': 0.12104959040880203, 'clip_ratio/low_min': 0.12104959040880203, 'clip_ratio/high_mean': 0.1058238660916686, 'clip_ratio/high_max': 0.1058238660916686, 'clip_ratio/region_mean': 0.22687345650047064, 'reward_total_mean': 0.4724371135234833, 'reward_meter_mean': 0.6299946308135986, 'reward_meter_std': 0.4327425956726074, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9937536120414734, 'reward_repeat_soft_std': 0.012771286070346832, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.09538455307483673, 'reward_total_composite_mean': 0.4724371135234833, 'reward_total_composite_std': 0.14866700768470764, 'epoch': 0.03}

  8%|▊         | 255/3300 [37:38<14:20:59, 16.97s/it]INFO 04-13 07:58:36 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 256/3300 [37:45<11:45:09, 13.90s/it]2026-04-13 07:58:43,151 | INFO | train_grpo_train | metrics_logged mode=train step=256

                                                     
{'loss': 0.0585, 'grad_norm': 13.359979629516602, 'learning_rate': 9.227272727272728e-06, 'num_tokens': 474129.0, 'completions/mean_length': 46.875, 'completions/min_length': 40.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7647510170936584, 'rewards/meter/std': 0.263922780752182, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.991895854473114, 'rewards/repeat_soft/std': 0.01966020278632641, 'rewards/judge_quality/mean': 0.48499998450279236, 'rewards/judge_quality/std': 0.15937379002571106, 'rewards/total_composite/mean': 0.5823091864585876, 'rewards/total_composite/std': 0.11638780683279037, 'reward': 0.5823091864585876, 'reward_std': 0.11638780683279037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2153816670179367, 'sampling/sampling_logp_difference/max': 1.2624378204345703, 'sampling/importance_sampling_ratio/min': 0.282963365316391, 'sampling/importance_sampling_ratio/mean': 1.0510693788528442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.654299169778824, 'clip_ratio/low_mean': 0.08597616292536259, 'clip_ratio/low_min': 0.08597616292536259, 'clip_ratio/high_mean': 0.10753452405333519, 'clip_ratio/high_max': 0.10753452405333519, 'clip_ratio/region_mean': 0.19351068697869778, 'reward_total_mean': 0.5823091864585876, 'reward_meter_mean': 0.7647510170936584, 'reward_meter_std': 0.263922780752182, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.991895854473114, 'reward_repeat_soft_std': 0.01966020278632641, 'reward_judge_quality_mean': 0.48499998450279236, 'reward_judge_quality_std': 0.15937379002571106, 'reward_total_composite_mean': 0.5823091864585876, 'reward_total_composite_std': 0.11638780683279037, 'epoch': 0.03}

  8%|▊         | 256/3300 [37:45<11:45:09, 13.90s/it]INFO 04-13 07:58:43 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 257/3300 [37:56<10:57:30, 12.96s/it]2026-04-13 07:58:53,932 | INFO | train_grpo_train | metrics_logged mode=train step=257

                                                     
{'loss': 0.0172, 'grad_norm': 5.30784273147583, 'learning_rate': 9.224242424242424e-06, 'num_tokens': 477513.0, 'completions/mean_length': 210.0, 'completions/min_length': 158.0, 'completions/max_length': 399.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 210.0, 'completions/min_terminated_length': 158.0, 'completions/max_terminated_length': 399.0, 'rewards/meter/mean': 0.6121722459793091, 'rewards/meter/std': 0.3405526876449585, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9984015226364136, 'rewards/repeat_soft/std': 0.0018976051360368729, 'rewards/judge_quality/mean': 0.17249999940395355, 'rewards/judge_quality/std': 0.12623673677444458, 'rewards/total_composite/mean': 0.35219359397888184, 'rewards/total_composite/std': 0.1643473207950592, 'reward': 0.35219359397888184, 'reward_std': 0.164347305893898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23884940147399902, 'sampling/sampling_logp_difference/max': 1.5050020217895508, 'sampling/importance_sampling_ratio/min': 0.2220168560743332, 'sampling/importance_sampling_ratio/mean': 1.0711714029312134, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 4.338009357452393, 'clip_ratio/low_mean': 0.0566375432536006, 'clip_ratio/low_min': 0.0566375432536006, 'clip_ratio/high_mean': 0.09318997152149677, 'clip_ratio/high_max': 0.09318997152149677, 'clip_ratio/region_mean': 0.14982751477509737, 'reward_total_mean': 0.35219359397888184, 'reward_meter_mean': 0.6121722459793091, 'reward_meter_std': 0.3405526876449585, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9984015226364136, 'reward_repeat_soft_std': 0.0018976051360368729, 'reward_judge_quality_mean': 0.17249999940395355, 'reward_judge_quality_std': 0.12623673677444458, 'reward_total_composite_mean': 0.35219359397888184, 'reward_total_composite_std': 0.1643473207950592, 'epoch': 0.03}

  8%|▊         | 257/3300 [37:56<10:57:30, 12.96s/it]INFO 04-13 07:58:54 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 258/3300 [38:03<9:31:20, 11.27s/it] 2026-04-13 07:59:01,248 | INFO | train_grpo_train | metrics_logged mode=train step=258

                                                    
{'loss': 0.0871, 'grad_norm': 11.726977348327637, 'learning_rate': 9.221212121212123e-06, 'num_tokens': 479486.0, 'completions/mean_length': 87.625, 'completions/min_length': 50.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.625, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.5028528571128845, 'rewards/meter/std': 0.2695654034614563, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9962344169616699, 'rewards/repeat_soft/std': 0.004130613524466753, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.49441421031951904, 'rewards/total_composite/std': 0.08856210857629776, 'reward': 0.49441421031951904, 'reward_std': 0.08856210857629776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22967101633548737, 'sampling/sampling_logp_difference/max': 1.4539060592651367, 'sampling/importance_sampling_ratio/min': 0.23365582525730133, 'sampling/importance_sampling_ratio/mean': 1.0647671222686768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.430586889386177, 'clip_ratio/low_mean': 0.10441025160253048, 'clip_ratio/low_min': 0.10441025160253048, 'clip_ratio/high_mean': 0.09899860806763172, 'clip_ratio/high_max': 0.09899860806763172, 'clip_ratio/region_mean': 0.2034088596701622, 'reward_total_mean': 0.49441421031951904, 'reward_meter_mean': 0.5028528571128845, 'reward_meter_std': 0.2695654034614563, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9962344169616699, 'reward_repeat_soft_std': 0.004130613524466753, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.49441421031951904, 'reward_total_composite_std': 0.08856210857629776, 'epoch': 0.03}

  8%|▊         | 258/3300 [38:03<9:31:20, 11.27s/it]INFO 04-13 07:59:01 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 259/3300 [38:15<9:40:05, 11.45s/it]2026-04-13 07:59:13,106 | INFO | train_grpo_train | metrics_logged mode=train step=259

                                                    
{'loss': -0.0702, 'grad_norm': 1.9631386995315552, 'learning_rate': 9.21818181818182e-06, 'num_tokens': 480908.0, 'completions/mean_length': 85.75, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.85714340209961, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6726659536361694, 'rewards/meter/std': 0.3912229835987091, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9622172117233276, 'rewards/repeat_soft/std': 0.0007998839719220996, 'rewards/judge_quality/mean': 0.32625001668930054, 'rewards/judge_quality/std': 0.15592464804649353, 'rewards/total_composite/mean': 0.44956642389297485, 'rewards/total_composite/std': 0.20386633276939392, 'reward': 0.44956642389297485, 'reward_std': 0.20386633276939392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1834326833486557, 'sampling/sampling_logp_difference/max': 0.8966245651245117, 'sampling/importance_sampling_ratio/min': 0.40794432163238525, 'sampling/importance_sampling_ratio/mean': 1.0076311826705933, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7231449484825134, 'clip_ratio/low_mean': 0.04589160904288292, 'clip_ratio/low_min': 0.04589160904288292, 'clip_ratio/high_mean': 0.13169217109680176, 'clip_ratio/high_max': 0.13169217109680176, 'clip_ratio/region_mean': 0.17758378013968468, 'reward_total_mean': 0.44956642389297485, 'reward_meter_mean': 0.6726659536361694, 'reward_meter_std': 0.3912229835987091, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9622172117233276, 'reward_repeat_soft_std': 0.0007998839719220996, 'reward_judge_quality_mean': 0.32625001668930054, 'reward_judge_quality_std': 0.15592464804649353, 'reward_total_composite_mean': 0.44956642389297485, 'reward_total_composite_std': 0.20386633276939392, 'epoch': 0.03}

  8%|▊         | 259/3300 [38:15<9:40:05, 11.45s/it]INFO 04-13 07:59:13 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 260/3300 [38:22<8:39:20, 10.25s/it]2026-04-13 07:59:20,565 | INFO | train_grpo_train | metrics_logged mode=train step=260

                                                    
{'loss': -0.0215, 'grad_norm': 12.083062171936035, 'learning_rate': 9.215151515151515e-06, 'num_tokens': 482528.0, 'completions/mean_length': 46.5, 'completions/min_length': 37.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.44499480724334717, 'rewards/meter/std': 0.3269158899784088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9971332550048828, 'rewards/repeat_soft/std': 0.0036495414096862078, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.47453099489212036, 'rewards/total_composite/std': 0.08867525309324265, 'reward': 0.47453099489212036, 'reward_std': 0.08867524564266205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21196623146533966, 'sampling/sampling_logp_difference/max': 1.0982379913330078, 'sampling/importance_sampling_ratio/min': 0.3334581255912781, 'sampling/importance_sampling_ratio/mean': 1.0493136644363403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3388479202985764, 'clip_ratio/low_mean': 0.08928120322525501, 'clip_ratio/low_min': 0.08928120322525501, 'clip_ratio/high_mean': 0.08731626439839602, 'clip_ratio/high_max': 0.08731626439839602, 'clip_ratio/region_mean': 0.17659746762365103, 'reward_total_mean': 0.47453099489212036, 'reward_meter_mean': 0.44499480724334717, 'reward_meter_std': 0.3269158899784088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9971332550048828, 'reward_repeat_soft_std': 0.0036495414096862078, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.47453099489212036, 'reward_total_composite_std': 0.08867525309324265, 'epoch': 0.03}

  8%|▊         | 260/3300 [38:22<8:39:20, 10.25s/it]INFO 04-13 07:59:20 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 261/3300 [38:30<7:57:26,  9.43s/it]2026-04-13 07:59:28,073 | INFO | train_grpo_train | metrics_logged mode=train step=261

                                                    
{'loss': 0.066, 'grad_norm': 13.19766616821289, 'learning_rate': 9.212121212121213e-06, 'num_tokens': 484183.0, 'completions/mean_length': 48.875, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.3864549398422241, 'rewards/meter/std': 0.3447204530239105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9985442757606506, 'rewards/repeat_soft/std': 0.00270977895706892, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.1529705971479416, 'rewards/total_composite/mean': 0.32365351915359497, 'rewards/total_composite/std': 0.21631422638893127, 'reward': 0.32365351915359497, 'reward_std': 0.21631422638893127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23518016934394836, 'sampling/sampling_logp_difference/max': 1.751021385192871, 'sampling/importance_sampling_ratio/min': 0.17359653115272522, 'sampling/importance_sampling_ratio/mean': 1.0690120458602905, 'sampling/importance_sampling_ratio/max': 1.978293776512146, 'entropy': 3.245684951543808, 'clip_ratio/low_mean': 0.0569117646664381, 'clip_ratio/low_min': 0.0569117646664381, 'clip_ratio/high_mean': 0.14734195545315742, 'clip_ratio/high_max': 0.14734195545315742, 'clip_ratio/region_mean': 0.20425372011959553, 'reward_total_mean': 0.32365351915359497, 'reward_meter_mean': 0.3864549398422241, 'reward_meter_std': 0.3447204530239105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9985442757606506, 'reward_repeat_soft_std': 0.00270977895706892, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.1529705971479416, 'reward_total_composite_mean': 0.32365351915359497, 'reward_total_composite_std': 0.21631422638893127, 'epoch': 0.03}

  8%|▊         | 261/3300 [38:30<7:57:26,  9.43s/it]INFO 04-13 07:59:28 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 262/3300 [38:36<7:13:57,  8.57s/it]2026-04-13 07:59:34,643 | INFO | train_grpo_train | metrics_logged mode=train step=262

                                                    
{'loss': 0.0519, 'grad_norm': 16.314903259277344, 'learning_rate': 9.20909090909091e-06, 'num_tokens': 485761.0, 'completions/mean_length': 42.25, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7184869647026062, 'rewards/meter/std': 0.397292822599411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9871054887771606, 'rewards/repeat_soft/std': 0.02460484392940998, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.5435694456100464, 'rewards/total_composite/std': 0.10916414856910706, 'reward': 0.5435694456100464, 'reward_std': 0.10916414856910706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2332073450088501, 'sampling/sampling_logp_difference/max': 1.291060447692871, 'sampling/importance_sampling_ratio/min': 0.27497902512550354, 'sampling/importance_sampling_ratio/mean': 1.053661584854126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5997531712055206, 'clip_ratio/low_mean': 0.037946428172290325, 'clip_ratio/low_min': 0.037946428172290325, 'clip_ratio/high_mean': 0.14962166547775269, 'clip_ratio/high_max': 0.14962166547775269, 'clip_ratio/region_mean': 0.187568093650043, 'reward_total_mean': 0.5435694456100464, 'reward_meter_mean': 0.7184869647026062, 'reward_meter_std': 0.397292822599411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9871054887771606, 'reward_repeat_soft_std': 0.02460484392940998, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.5435694456100464, 'reward_total_composite_std': 0.10916414856910706, 'epoch': 0.03}

  8%|▊         | 262/3300 [38:36<7:13:57,  8.57s/it]INFO 04-13 07:59:34 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 263/3300 [38:44<6:53:11,  8.16s/it]2026-04-13 07:59:41,856 | INFO | train_grpo_train | metrics_logged mode=train step=263

                                                    
{'loss': -0.032, 'grad_norm': 10.155993461608887, 'learning_rate': 9.206060606060607e-06, 'num_tokens': 487787.0, 'completions/mean_length': 81.25, 'completions/min_length': 66.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.5438829660415649, 'rewards/meter/std': 0.40771350264549255, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.984946608543396, 'rewards/repeat_soft/std': 0.015459898859262466, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.4312058687210083, 'rewards/total_composite/std': 0.19852708280086517, 'reward': 0.4312058687210083, 'reward_std': 0.19852708280086517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2361883819103241, 'sampling/sampling_logp_difference/max': 1.3174123764038086, 'sampling/importance_sampling_ratio/min': 0.26782742142677307, 'sampling/importance_sampling_ratio/mean': 1.0664438009262085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.335661917924881, 'clip_ratio/low_mean': 0.07798272278159857, 'clip_ratio/low_min': 0.07798272278159857, 'clip_ratio/high_mean': 0.0985830519348383, 'clip_ratio/high_max': 0.0985830519348383, 'clip_ratio/region_mean': 0.17656577471643686, 'reward_total_mean': 0.4312058687210083, 'reward_meter_mean': 0.5438829660415649, 'reward_meter_std': 0.40771350264549255, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.984946608543396, 'reward_repeat_soft_std': 0.015459898859262466, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.4312058687210083, 'reward_total_composite_std': 0.19852708280086517, 'epoch': 0.03}

  8%|▊         | 263/3300 [38:44<6:53:11,  8.16s/it]INFO 04-13 07:59:42 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 264/3300 [38:51<6:37:32,  7.86s/it]2026-04-13 07:59:48,996 | INFO | train_grpo_train | metrics_logged mode=train step=264

                                                    
{'loss': 0.0807, 'grad_norm': 9.939325332641602, 'learning_rate': 9.203030303030304e-06, 'num_tokens': 489954.0, 'completions/mean_length': 95.875, 'completions/min_length': 68.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.875, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.39289143681526184, 'rewards/meter/std': 0.3128518760204315, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9843380451202393, 'rewards/repeat_soft/std': 0.029751332476735115, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.44051191210746765, 'rewards/total_composite/std': 0.0717383399605751, 'reward': 0.44051191210746765, 'reward_std': 0.0717383399605751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21802948415279388, 'sampling/sampling_logp_difference/max': 1.4610528945922852, 'sampling/importance_sampling_ratio/min': 0.23199188709259033, 'sampling/importance_sampling_ratio/mean': 1.0672013759613037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0699095129966736, 'clip_ratio/low_mean': 0.07578411884605885, 'clip_ratio/low_min': 0.07578411884605885, 'clip_ratio/high_mean': 0.13615627773106098, 'clip_ratio/high_max': 0.13615627773106098, 'clip_ratio/region_mean': 0.21194039657711983, 'reward_total_mean': 0.44051191210746765, 'reward_meter_mean': 0.39289143681526184, 'reward_meter_std': 0.3128518760204315, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9843380451202393, 'reward_repeat_soft_std': 0.029751332476735115, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.44051191210746765, 'reward_total_composite_std': 0.0717383399605751, 'epoch': 0.03}

  8%|▊         | 264/3300 [38:51<6:37:32,  7.86s/it]INFO 04-13 07:59:49 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 265/3300 [38:58<6:33:32,  7.78s/it]2026-04-13 07:59:56,598 | INFO | train_grpo_train | metrics_logged mode=train step=265

                                                    
{'loss': 0.0817, 'grad_norm': 13.237419128417969, 'learning_rate': 9.200000000000002e-06, 'num_tokens': 492168.0, 'completions/mean_length': 99.75, 'completions/min_length': 83.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.75, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8052846193313599, 'rewards/meter/std': 0.24584336578845978, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9940801858901978, 'rewards/repeat_soft/std': 0.0032723494805395603, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.1642080694437027, 'rewards/total_composite/mean': 0.6095613837242126, 'rewards/total_composite/std': 0.12963679432868958, 'reward': 0.6095613837242126, 'reward_std': 0.12963679432868958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20225085318088531, 'sampling/sampling_logp_difference/max': 2.628049373626709, 'sampling/importance_sampling_ratio/min': 0.07221920043230057, 'sampling/importance_sampling_ratio/mean': 1.0143296718597412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3046093508601189, 'clip_ratio/low_mean': 0.0953658726066351, 'clip_ratio/low_min': 0.0953658726066351, 'clip_ratio/high_mean': 0.07775945216417313, 'clip_ratio/high_max': 0.07775945216417313, 'clip_ratio/region_mean': 0.17312532477080822, 'reward_total_mean': 0.6095613837242126, 'reward_meter_mean': 0.8052846193313599, 'reward_meter_std': 0.24584336578845978, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9940801858901978, 'reward_repeat_soft_std': 0.0032723494805395603, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.1642080694437027, 'reward_total_composite_mean': 0.6095613837242126, 'reward_total_composite_std': 0.12963679432868958, 'epoch': 0.03}

  8%|▊         | 265/3300 [38:58<6:33:32,  7.78s/it]INFO 04-13 07:59:56 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 266/3300 [39:05<6:23:32,  7.58s/it]2026-04-13 08:00:03,727 | INFO | train_grpo_train | metrics_logged mode=train step=266

                                                    
{'loss': 0.1828, 'grad_norm': 18.29632568359375, 'learning_rate': 9.196969696969697e-06, 'num_tokens': 493579.0, 'completions/mean_length': 26.375, 'completions/min_length': 20.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8836090564727783, 'rewards/meter/std': 0.26969626545906067, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.09258200973272324, 'rewards/total_composite/mean': 0.583101749420166, 'rewards/total_composite/std': 0.09299547225236893, 'reward': 0.583101749420166, 'reward_std': 0.09299547970294952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23211590945720673, 'sampling/sampling_logp_difference/max': 1.0317764282226562, 'sampling/importance_sampling_ratio/min': 0.3563733398914337, 'sampling/importance_sampling_ratio/mean': 1.0721259117126465, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2353347837924957, 'clip_ratio/low_mean': 0.050245098769664764, 'clip_ratio/low_min': 0.050245098769664764, 'clip_ratio/high_mean': 0.17944775335490704, 'clip_ratio/high_max': 0.17944775335490704, 'clip_ratio/region_mean': 0.2296928521245718, 'reward_total_mean': 0.583101749420166, 'reward_meter_mean': 0.8836090564727783, 'reward_meter_std': 0.26969626545906067, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.09258200973272324, 'reward_total_composite_mean': 0.583101749420166, 'reward_total_composite_std': 0.09299547225236893, 'epoch': 0.03}

  8%|▊         | 266/3300 [39:05<6:23:32,  7.58s/it]INFO 04-13 08:00:03 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 267/3300 [39:13<6:25:30,  7.63s/it]2026-04-13 08:00:11,451 | INFO | train_grpo_train | metrics_logged mode=train step=267

                                                    
{'loss': 0.1007, 'grad_norm': 11.504515647888184, 'learning_rate': 9.193939393939395e-06, 'num_tokens': 495297.0, 'completions/mean_length': 54.75, 'completions/min_length': 43.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.24899473786354065, 'rewards/meter/std': 0.3889537453651428, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957927465438843, 'rewards/repeat_soft/std': 0.00977754034101963, 'rewards/judge_quality/mean': 0.3675000071525574, 'rewards/judge_quality/std': 0.18100908398628235, 'rewards/total_composite/mean': 0.4044429063796997, 'rewards/total_composite/std': 0.09395794570446014, 'reward': 0.4044429063796997, 'reward_std': 0.09395796060562134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23803894221782684, 'sampling/sampling_logp_difference/max': 1.2172846794128418, 'sampling/importance_sampling_ratio/min': 0.31349968910217285, 'sampling/importance_sampling_ratio/mean': 1.0595011711120605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.626153290271759, 'clip_ratio/low_mean': 0.11126524023711681, 'clip_ratio/low_min': 0.11126524023711681, 'clip_ratio/high_mean': 0.044706499204039574, 'clip_ratio/high_max': 0.044706499204039574, 'clip_ratio/region_mean': 0.1559717394411564, 'reward_total_mean': 0.4044429063796997, 'reward_meter_mean': 0.24899473786354065, 'reward_meter_std': 0.3889537453651428, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957927465438843, 'reward_repeat_soft_std': 0.00977754034101963, 'reward_judge_quality_mean': 0.3675000071525574, 'reward_judge_quality_std': 0.18100908398628235, 'reward_total_composite_mean': 0.4044429063796997, 'reward_total_composite_std': 0.09395794570446014, 'epoch': 0.03}

  8%|▊         | 267/3300 [39:13<6:25:30,  7.63s/it]INFO 04-13 08:00:11 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 268/3300 [39:21<6:25:13,  7.62s/it]2026-04-13 08:00:19,067 | INFO | train_grpo_train | metrics_logged mode=train step=268

                                                    
{'loss': 0.1384, 'grad_norm': 18.148683547973633, 'learning_rate': 9.190909090909092e-06, 'num_tokens': 496902.0, 'completions/mean_length': 33.625, 'completions/min_length': 28.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7711975574493408, 'rewards/meter/std': 0.2813066840171814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9876835346221924, 'rewards/repeat_soft/std': 0.009118892252445221, 'rewards/judge_quality/mean': 0.7737500071525574, 'rewards/judge_quality/std': 0.22032040357589722, 'rewards/total_composite/mean': 0.7391894459724426, 'rewards/total_composite/std': 0.189705029129982, 'reward': 0.7391894459724426, 'reward_std': 0.189705029129982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11177510023117065, 'sampling/sampling_logp_difference/max': 1.8327703475952148, 'sampling/importance_sampling_ratio/min': 0.15996979176998138, 'sampling/importance_sampling_ratio/mean': 0.9933525919914246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7010860815644264, 'clip_ratio/low_mean': 0.055822014808654785, 'clip_ratio/low_min': 0.055822014808654785, 'clip_ratio/high_mean': 0.0606780918315053, 'clip_ratio/high_max': 0.0606780918315053, 'clip_ratio/region_mean': 0.11650010664016008, 'reward_total_mean': 0.7391894459724426, 'reward_meter_mean': 0.7711975574493408, 'reward_meter_std': 0.2813066840171814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9876835346221924, 'reward_repeat_soft_std': 0.009118892252445221, 'reward_judge_quality_mean': 0.7737500071525574, 'reward_judge_quality_std': 0.22032040357589722, 'reward_total_composite_mean': 0.7391894459724426, 'reward_total_composite_std': 0.189705029129982, 'epoch': 0.03}

  8%|▊         | 268/3300 [39:21<6:25:13,  7.62s/it]INFO 04-13 08:00:19 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 269/3300 [39:27<6:10:33,  7.34s/it]2026-04-13 08:00:25,731 | INFO | train_grpo_train | metrics_logged mode=train step=269

                                                    
{'loss': 0.0359, 'grad_norm': 19.865617752075195, 'learning_rate': 9.187878787878789e-06, 'num_tokens': 498302.0, 'completions/mean_length': 25.0, 'completions/min_length': 17.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.903567910194397, 'rewards/meter/std': 0.24369066953659058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9612975120544434, 'rewards/repeat_soft/std': 0.0034009767696261406, 'rewards/judge_quality/mean': 0.48374998569488525, 'rewards/judge_quality/std': 0.18965664505958557, 'rewards/total_composite/mean': 0.6303813457489014, 'rewards/total_composite/std': 0.14658081531524658, 'reward': 0.6303813457489014, 'reward_std': 0.1465807855129242, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20733590424060822, 'sampling/sampling_logp_difference/max': 1.2930784225463867, 'sampling/importance_sampling_ratio/min': 0.27442467212677, 'sampling/importance_sampling_ratio/mean': 1.0204358100891113, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.115874543786049, 'clip_ratio/low_mean': 0.023291926365345716, 'clip_ratio/low_min': 0.023291926365345716, 'clip_ratio/high_mean': 0.1650573741644621, 'clip_ratio/high_max': 0.1650573741644621, 'clip_ratio/region_mean': 0.1883493005298078, 'reward_total_mean': 0.6303813457489014, 'reward_meter_mean': 0.903567910194397, 'reward_meter_std': 0.24369066953659058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9612975120544434, 'reward_repeat_soft_std': 0.0034009767696261406, 'reward_judge_quality_mean': 0.48374998569488525, 'reward_judge_quality_std': 0.18965664505958557, 'reward_total_composite_mean': 0.6303813457489014, 'reward_total_composite_std': 0.14658081531524658, 'epoch': 0.03}

  8%|▊         | 269/3300 [39:27<6:10:33,  7.34s/it]INFO 04-13 08:00:25 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 270/3300 [39:34<6:05:09,  7.23s/it]2026-04-13 08:00:32,718 | INFO | train_grpo_train | metrics_logged mode=train step=270

                                                    
{'loss': -0.0944, 'grad_norm': 13.092336654663086, 'learning_rate': 9.184848484848485e-06, 'num_tokens': 499890.0, 'completions/mean_length': 43.5, 'completions/min_length': 27.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.45151951909065247, 'rewards/meter/std': 0.3828032910823822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 1.0, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.4131723642349243, 'rewards/total_composite/std': 0.18584245443344116, 'reward': 0.4131723642349243, 'reward_std': 0.18584245443344116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2269778698682785, 'sampling/sampling_logp_difference/max': 1.8083457946777344, 'sampling/importance_sampling_ratio/min': 0.16392506659030914, 'sampling/importance_sampling_ratio/mean': 1.041564702987671, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1637730598449707, 'clip_ratio/low_mean': 0.07881944440305233, 'clip_ratio/low_min': 0.07881944440305233, 'clip_ratio/high_mean': 0.13781122490763664, 'clip_ratio/high_max': 0.13781122490763664, 'clip_ratio/region_mean': 0.21663066931068897, 'reward_total_mean': 0.4131723642349243, 'reward_meter_mean': 0.45151951909065247, 'reward_meter_std': 0.3828032910823822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 1.0, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.4131723642349243, 'reward_total_composite_std': 0.18584245443344116, 'epoch': 0.03}

  8%|▊         | 270/3300 [39:34<6:05:09,  7.23s/it]INFO 04-13 08:00:32 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 271/3300 [39:41<5:54:56,  7.03s/it]2026-04-13 08:00:39,281 | INFO | train_grpo_train | metrics_logged mode=train step=271

                                                    
{'loss': 0.0151, 'grad_norm': 14.901123046875, 'learning_rate': 9.181818181818184e-06, 'num_tokens': 501475.0, 'completions/mean_length': 38.125, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5740398168563843, 'rewards/meter/std': 0.3888934552669525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9911396503448486, 'rewards/repeat_soft/std': 0.006866309326142073, 'rewards/judge_quality/mean': 0.3712500035762787, 'rewards/judge_quality/std': 0.10091544687747955, 'rewards/total_composite/mean': 0.5012853741645813, 'rewards/total_composite/std': 0.11350028216838837, 'reward': 0.5012853741645813, 'reward_std': 0.11350028216838837, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2077735960483551, 'sampling/sampling_logp_difference/max': 1.6415634155273438, 'sampling/importance_sampling_ratio/min': 0.19367700815200806, 'sampling/importance_sampling_ratio/mean': 1.0471086502075195, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.64426988363266, 'clip_ratio/low_mean': 0.07613284746184945, 'clip_ratio/low_min': 0.07613284746184945, 'clip_ratio/high_mean': 0.09764429181814194, 'clip_ratio/high_max': 0.09764429181814194, 'clip_ratio/region_mean': 0.1737771392799914, 'reward_total_mean': 0.5012853741645813, 'reward_meter_mean': 0.5740398168563843, 'reward_meter_std': 0.3888934552669525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9911396503448486, 'reward_repeat_soft_std': 0.006866309326142073, 'reward_judge_quality_mean': 0.3712500035762787, 'reward_judge_quality_std': 0.10091544687747955, 'reward_total_composite_mean': 0.5012853741645813, 'reward_total_composite_std': 0.11350028216838837, 'epoch': 0.03}

  8%|▊         | 271/3300 [39:41<5:54:56,  7.03s/it]INFO 04-13 08:00:39 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 272/3300 [39:48<5:52:11,  6.98s/it]2026-04-13 08:00:46,139 | INFO | train_grpo_train | metrics_logged mode=train step=272

                                                    
{'loss': 0.1787, 'grad_norm': 12.871830940246582, 'learning_rate': 9.178787878787879e-06, 'num_tokens': 503081.0, 'completions/mean_length': 43.75, 'completions/min_length': 38.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9809567928314209, 'rewards/meter/std': 0.020499568432569504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9782898426055908, 'rewards/repeat_soft/std': 0.02181963250041008, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1860443502664566, 'rewards/total_composite/mean': 0.609815239906311, 'rewards/total_composite/std': 0.12109959870576859, 'reward': 0.609815239906311, 'reward_std': 0.12109959870576859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25148677825927734, 'sampling/sampling_logp_difference/max': 1.4036140441894531, 'sampling/importance_sampling_ratio/min': 0.24570736289024353, 'sampling/importance_sampling_ratio/mean': 1.060373306274414, 'sampling/importance_sampling_ratio/max': 1.7493022680282593, 'entropy': 3.5093444883823395, 'clip_ratio/low_mean': 0.05145998392254114, 'clip_ratio/low_min': 0.05145998392254114, 'clip_ratio/high_mean': 0.16260988265275955, 'clip_ratio/high_max': 0.16260988265275955, 'clip_ratio/region_mean': 0.2140698665753007, 'reward_total_mean': 0.609815239906311, 'reward_meter_mean': 0.9809567928314209, 'reward_meter_std': 0.020499568432569504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9782898426055908, 'reward_repeat_soft_std': 0.02181963250041008, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1860443502664566, 'reward_total_composite_mean': 0.609815239906311, 'reward_total_composite_std': 0.12109959870576859, 'epoch': 0.03}

  8%|▊         | 272/3300 [39:48<5:52:11,  6.98s/it]INFO 04-13 08:00:46 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 273/3300 [39:55<5:55:36,  7.05s/it]2026-04-13 08:00:53,350 | INFO | train_grpo_train | metrics_logged mode=train step=273

                                                    
{'loss': -0.0372, 'grad_norm': 11.124080657958984, 'learning_rate': 9.175757575757576e-06, 'num_tokens': 504972.0, 'completions/mean_length': 70.375, 'completions/min_length': 61.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5856324434280396, 'rewards/meter/std': 0.3497793972492218, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9933459758758545, 'rewards/repeat_soft/std': 0.006139182019978762, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.515045702457428, 'rewards/total_composite/std': 0.09783965349197388, 'reward': 0.515045702457428, 'reward_std': 0.09783965349197388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1920226514339447, 'sampling/sampling_logp_difference/max': 1.841231346130371, 'sampling/importance_sampling_ratio/min': 0.15862199664115906, 'sampling/importance_sampling_ratio/mean': 1.0478473901748657, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1079849153757095, 'clip_ratio/low_mean': 0.0917767696082592, 'clip_ratio/low_min': 0.0917767696082592, 'clip_ratio/high_mean': 0.08189899660646915, 'clip_ratio/high_max': 0.08189899660646915, 'clip_ratio/region_mean': 0.17367576621472836, 'reward_total_mean': 0.515045702457428, 'reward_meter_mean': 0.5856324434280396, 'reward_meter_std': 0.3497793972492218, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9933459758758545, 'reward_repeat_soft_std': 0.006139182019978762, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.515045702457428, 'reward_total_composite_std': 0.09783965349197388, 'epoch': 0.03}

  8%|▊         | 273/3300 [39:55<5:55:36,  7.05s/it]INFO 04-13 08:00:53 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 274/3300 [40:02<5:49:02,  6.92s/it]2026-04-13 08:00:59,973 | INFO | train_grpo_train | metrics_logged mode=train step=274

                                                    
{'loss': 0.1249, 'grad_norm': 18.43695068359375, 'learning_rate': 9.172727272727274e-06, 'num_tokens': 506379.0, 'completions/mean_length': 22.875, 'completions/min_length': 18.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6266977190971375, 'rewards/meter/std': 0.4282899796962738, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9495941400527954, 'rewards/repeat_soft/std': 0.03650323674082756, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.5095803737640381, 'rewards/total_composite/std': 0.12357732653617859, 'reward': 0.5095803737640381, 'reward_std': 0.12357732653617859, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23109155893325806, 'sampling/sampling_logp_difference/max': 1.5454392433166504, 'sampling/importance_sampling_ratio/min': 0.21321819722652435, 'sampling/importance_sampling_ratio/mean': 1.0291531085968018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1662951186299324, 'clip_ratio/low_mean': 0.10648148227483034, 'clip_ratio/low_min': 0.10648148227483034, 'clip_ratio/high_mean': 0.11796537041664124, 'clip_ratio/high_max': 0.11796537041664124, 'clip_ratio/region_mean': 0.22444685269147158, 'reward_total_mean': 0.5095803737640381, 'reward_meter_mean': 0.6266977190971375, 'reward_meter_std': 0.4282899796962738, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9495941400527954, 'reward_repeat_soft_std': 0.03650323674082756, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.5095803737640381, 'reward_total_composite_std': 0.12357732653617859, 'epoch': 0.03}

  8%|▊         | 274/3300 [40:02<5:49:02,  6.92s/it]INFO 04-13 08:01:00 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 275/3300 [40:10<6:15:50,  7.45s/it]2026-04-13 08:01:08,673 | INFO | train_grpo_train | metrics_logged mode=train step=275

                                                    
{'loss': 0.1211, 'grad_norm': 10.326603889465332, 'learning_rate': 9.169696969696971e-06, 'num_tokens': 508478.0, 'completions/mean_length': 91.375, 'completions/min_length': 69.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.4572254419326782, 'rewards/meter/std': 0.34324416518211365, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9886662364006042, 'rewards/repeat_soft/std': 0.013447579927742481, 'rewards/judge_quality/mean': 0.36124998331069946, 'rewards/judge_quality/std': 0.1141975075006485, 'rewards/total_composite/mean': 0.3660357594490051, 'rewards/total_composite/std': 0.24225392937660217, 'reward': 0.3660357594490051, 'reward_std': 0.24225391447544098, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21642006933689117, 'sampling/sampling_logp_difference/max': 1.1390104293823242, 'sampling/importance_sampling_ratio/min': 0.3201356828212738, 'sampling/importance_sampling_ratio/mean': 1.062514066696167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1374784111976624, 'clip_ratio/low_mean': 0.05686148162931204, 'clip_ratio/low_min': 0.05686148162931204, 'clip_ratio/high_mean': 0.12252083979547024, 'clip_ratio/high_max': 0.12252083979547024, 'clip_ratio/region_mean': 0.17938232142478228, 'reward_total_mean': 0.3660357594490051, 'reward_meter_mean': 0.4572254419326782, 'reward_meter_std': 0.34324416518211365, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9886662364006042, 'reward_repeat_soft_std': 0.013447579927742481, 'reward_judge_quality_mean': 0.36124998331069946, 'reward_judge_quality_std': 0.1141975075006485, 'reward_total_composite_mean': 0.3660357594490051, 'reward_total_composite_std': 0.24225392937660217, 'epoch': 0.03}

  8%|▊         | 275/3300 [40:10<6:15:50,  7.45s/it]INFO 04-13 08:01:08 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 276/3300 [40:18<6:10:50,  7.36s/it]2026-04-13 08:01:15,807 | INFO | train_grpo_train | metrics_logged mode=train step=276

                                                    
{'loss': 0.3342, 'grad_norm': 20.758947372436523, 'learning_rate': 9.166666666666666e-06, 'num_tokens': 510069.0, 'completions/mean_length': 38.875, 'completions/min_length': 30.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7617350816726685, 'rewards/meter/std': 0.314482718706131, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9918779134750366, 'rewards/repeat_soft/std': 0.01899106800556183, 'rewards/judge_quality/mean': 0.44749999046325684, 'rewards/judge_quality/std': 0.12848013639450073, 'rewards/total_composite/mean': 0.5732759237289429, 'rewards/total_composite/std': 0.149704709649086, 'reward': 0.5732759237289429, 'reward_std': 0.1497046947479248, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.241552472114563, 'sampling/sampling_logp_difference/max': 1.3904485702514648, 'sampling/importance_sampling_ratio/min': 0.2489636093378067, 'sampling/importance_sampling_ratio/mean': 1.0301902294158936, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.613171875476837, 'clip_ratio/low_mean': 0.06453399732708931, 'clip_ratio/low_min': 0.06453399732708931, 'clip_ratio/high_mean': 0.13725490681827068, 'clip_ratio/high_max': 0.13725490681827068, 'clip_ratio/region_mean': 0.20178890414536, 'reward_total_mean': 0.5732759237289429, 'reward_meter_mean': 0.7617350816726685, 'reward_meter_std': 0.314482718706131, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9918779134750366, 'reward_repeat_soft_std': 0.01899106800556183, 'reward_judge_quality_mean': 0.44749999046325684, 'reward_judge_quality_std': 0.12848013639450073, 'reward_total_composite_mean': 0.5732759237289429, 'reward_total_composite_std': 0.149704709649086, 'epoch': 0.03}

  8%|▊         | 276/3300 [40:18<6:10:50,  7.36s/it]INFO 04-13 08:01:16 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 277/3300 [40:25<6:07:07,  7.29s/it]2026-04-13 08:01:22,927 | INFO | train_grpo_train | metrics_logged mode=train step=277

                                                    
{'loss': -0.0043, 'grad_norm': 19.438020706176758, 'learning_rate': 9.163636363636365e-06, 'num_tokens': 511499.0, 'completions/mean_length': 32.75, 'completions/min_length': 26.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7640655040740967, 'rewards/meter/std': 0.31606751680374146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9978388547897339, 'rewards/repeat_soft/std': 0.0030812383629381657, 'rewards/judge_quality/mean': 0.7075000405311584, 'rewards/judge_quality/std': 0.21001702547073364, 'rewards/total_composite/mean': 0.7123679518699646, 'rewards/total_composite/std': 0.16456754505634308, 'reward': 0.7123679518699646, 'reward_std': 0.16456754505634308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20448407530784607, 'sampling/sampling_logp_difference/max': 1.4917001724243164, 'sampling/importance_sampling_ratio/min': 0.22498980164527893, 'sampling/importance_sampling_ratio/mean': 1.0104999542236328, 'sampling/importance_sampling_ratio/max': 1.8533438444137573, 'entropy': 1.5110199749469757, 'clip_ratio/low_mean': 0.035912297666072845, 'clip_ratio/low_min': 0.035912297666072845, 'clip_ratio/high_mean': 0.15360071696341038, 'clip_ratio/high_max': 0.15360071696341038, 'clip_ratio/region_mean': 0.18951301462948322, 'reward_total_mean': 0.7123679518699646, 'reward_meter_mean': 0.7640655040740967, 'reward_meter_std': 0.31606751680374146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9978388547897339, 'reward_repeat_soft_std': 0.0030812383629381657, 'reward_judge_quality_mean': 0.7075000405311584, 'reward_judge_quality_std': 0.21001702547073364, 'reward_total_composite_mean': 0.7123679518699646, 'reward_total_composite_std': 0.16456754505634308, 'epoch': 0.03}

  8%|▊         | 277/3300 [40:25<6:07:07,  7.29s/it]INFO 04-13 08:01:23 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 278/3300 [40:31<5:56:03,  7.07s/it]2026-04-13 08:01:29,489 | INFO | train_grpo_train | metrics_logged mode=train step=278

                                                    
{'loss': 0.0259, 'grad_norm': 34.60313415527344, 'learning_rate': 9.160606060606061e-06, 'num_tokens': 512918.0, 'completions/mean_length': 24.375, 'completions/min_length': 21.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9747086763381958, 'rewards/meter/std': 0.05066248029470444, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9778183102607727, 'rewards/repeat_soft/std': 0.021749334409832954, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.9295485019683838, 'rewards/total_composite/std': 0.03065398521721363, 'reward': 0.9295485019683838, 'reward_std': 0.03065398521721363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2493032068014145, 'sampling/sampling_logp_difference/max': 2.6578667163848877, 'sampling/importance_sampling_ratio/min': 0.07009760290384293, 'sampling/importance_sampling_ratio/mean': 0.9814552068710327, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.998925194144249, 'clip_ratio/low_mean': 0.02604166604578495, 'clip_ratio/low_min': 0.02604166604578495, 'clip_ratio/high_mean': 0.16956652328372002, 'clip_ratio/high_max': 0.16956652328372002, 'clip_ratio/region_mean': 0.19560818932950497, 'reward_total_mean': 0.9295485019683838, 'reward_meter_mean': 0.9747086763381958, 'reward_meter_std': 0.05066248029470444, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9778183102607727, 'reward_repeat_soft_std': 0.021749334409832954, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.9295485019683838, 'reward_total_composite_std': 0.03065398521721363, 'epoch': 0.03}

  8%|▊         | 278/3300 [40:31<5:56:03,  7.07s/it]INFO 04-13 08:01:29 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 279/3300 [40:39<6:03:56,  7.23s/it]2026-04-13 08:01:37,087 | INFO | train_grpo_train | metrics_logged mode=train step=279

                                                    
{'loss': 0.069, 'grad_norm': 12.388896942138672, 'learning_rate': 9.157575757575758e-06, 'num_tokens': 514522.0, 'completions/mean_length': 45.5, 'completions/min_length': 34.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8167272806167603, 'rewards/meter/std': 0.3358098864555359, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9955008029937744, 'rewards/repeat_soft/std': 0.006966626737266779, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.25999659299850464, 'rewards/total_composite/mean': 0.6498563885688782, 'rewards/total_composite/std': 0.20248964428901672, 'reward': 0.6498563885688782, 'reward_std': 0.20248964428901672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20259158313274384, 'sampling/sampling_logp_difference/max': 1.0752041339874268, 'sampling/importance_sampling_ratio/min': 0.36106279492378235, 'sampling/importance_sampling_ratio/mean': 1.0507522821426392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3180495500564575, 'clip_ratio/low_mean': 0.1202126294374466, 'clip_ratio/low_min': 0.1202126294374466, 'clip_ratio/high_mean': 0.08005169313400984, 'clip_ratio/high_max': 0.08005169313400984, 'clip_ratio/region_mean': 0.20026432257145643, 'reward_total_mean': 0.6498563885688782, 'reward_meter_mean': 0.8167272806167603, 'reward_meter_std': 0.3358098864555359, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9955008029937744, 'reward_repeat_soft_std': 0.006966626737266779, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.25999659299850464, 'reward_total_composite_mean': 0.6498563885688782, 'reward_total_composite_std': 0.20248964428901672, 'epoch': 0.03}

  8%|▊         | 279/3300 [40:39<6:03:56,  7.23s/it]INFO 04-13 08:01:37 [block_pool.py:378] Successfully reset prefix cache

  8%|▊         | 280/3300 [40:48<6:27:27,  7.70s/it]2026-04-13 08:01:45,881 | INFO | train_grpo_train | metrics_logged mode=train step=280

                                                    
{'loss': 0.102, 'grad_norm': 7.115893363952637, 'learning_rate': 9.154545454545455e-06, 'num_tokens': 516764.0, 'completions/mean_length': 114.25, 'completions/min_length': 93.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.25, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.8475615978240967, 'rewards/meter/std': 0.24203528463840485, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9811174869537354, 'rewards/repeat_soft/std': 0.020043261349201202, 'rewards/judge_quality/mean': 0.26874998211860657, 'rewards/judge_quality/std': 0.13505950570106506, 'rewards/total_composite/mean': 0.46794646978378296, 'rewards/total_composite/std': 0.09037885069847107, 'reward': 0.46794646978378296, 'reward_std': 0.09037884324789047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23452864587306976, 'sampling/sampling_logp_difference/max': 1.3950977325439453, 'sampling/importance_sampling_ratio/min': 0.24780882894992828, 'sampling/importance_sampling_ratio/mean': 1.0599229335784912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.9277411103248596, 'clip_ratio/low_mean': 0.07516323309391737, 'clip_ratio/low_min': 0.07516323309391737, 'clip_ratio/high_mean': 0.07142539694905281, 'clip_ratio/high_max': 0.07142539694905281, 'clip_ratio/region_mean': 0.14658863004297018, 'reward_total_mean': 0.46794646978378296, 'reward_meter_mean': 0.8475615978240967, 'reward_meter_std': 0.24203528463840485, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9811174869537354, 'reward_repeat_soft_std': 0.020043261349201202, 'reward_judge_quality_mean': 0.26874998211860657, 'reward_judge_quality_std': 0.13505950570106506, 'reward_total_composite_mean': 0.46794646978378296, 'reward_total_composite_std': 0.09037885069847107, 'epoch': 0.03}

  8%|▊         | 280/3300 [40:48<6:27:27,  7.70s/it]INFO 04-13 08:01:46 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 281/3300 [40:54<6:13:25,  7.42s/it]2026-04-13 08:01:52,658 | INFO | train_grpo_train | metrics_logged mode=train step=281

                                                    
{'loss': 0.0678, 'grad_norm': 13.96617317199707, 'learning_rate': 9.151515151515153e-06, 'num_tokens': 518319.0, 'completions/mean_length': 40.375, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9172787666320801, 'rewards/meter/std': 0.10134030878543854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.989300549030304, 'rewards/repeat_soft/std': 0.013068420812487602, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.6098494529724121, 'rewards/total_composite/std': 0.02921118587255478, 'reward': 0.6098494529724121, 'reward_std': 0.02921118400990963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20314764976501465, 'sampling/sampling_logp_difference/max': 1.6208076477050781, 'sampling/importance_sampling_ratio/min': 0.19773893058300018, 'sampling/importance_sampling_ratio/mean': 1.0205670595169067, 'sampling/importance_sampling_ratio/max': 1.9314621686935425, 'entropy': 2.173624560236931, 'clip_ratio/low_mean': 0.032751716673374176, 'clip_ratio/low_min': 0.032751716673374176, 'clip_ratio/high_mean': 0.14773373864591122, 'clip_ratio/high_max': 0.14773373864591122, 'clip_ratio/region_mean': 0.1804854553192854, 'reward_total_mean': 0.6098494529724121, 'reward_meter_mean': 0.9172787666320801, 'reward_meter_std': 0.10134030878543854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.989300549030304, 'reward_repeat_soft_std': 0.013068420812487602, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.6098494529724121, 'reward_total_composite_std': 0.02921118587255478, 'epoch': 0.03}

  9%|▊         | 281/3300 [40:54<6:13:25,  7.42s/it]INFO 04-13 08:01:52 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 282/3300 [41:01<6:08:09,  7.32s/it]2026-04-13 08:01:59,738 | INFO | train_grpo_train | metrics_logged mode=train step=282

                                                    
{'loss': 0.2188, 'grad_norm': 14.918357849121094, 'learning_rate': 9.148484848484848e-06, 'num_tokens': 519838.0, 'completions/mean_length': 39.875, 'completions/min_length': 29.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.35328733921051025, 'rewards/meter/std': 0.3507688641548157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9957976937294006, 'rewards/repeat_soft/std': 0.009099088609218597, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.215406596660614, 'rewards/total_composite/mean': 0.38957157731056213, 'rewards/total_composite/std': 0.18033328652381897, 'reward': 0.38957157731056213, 'reward_std': 0.18033325672149658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23948487639427185, 'sampling/sampling_logp_difference/max': 1.5781440734863281, 'sampling/importance_sampling_ratio/min': 0.20635771751403809, 'sampling/importance_sampling_ratio/mean': 1.0777320861816406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4532696157693863, 'clip_ratio/low_mean': 0.09098005574196577, 'clip_ratio/low_min': 0.09098005574196577, 'clip_ratio/high_mean': 0.09778170101344585, 'clip_ratio/high_max': 0.09778170101344585, 'clip_ratio/region_mean': 0.18876175675541162, 'reward_total_mean': 0.38957157731056213, 'reward_meter_mean': 0.35328733921051025, 'reward_meter_std': 0.3507688641548157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9957976937294006, 'reward_repeat_soft_std': 0.009099088609218597, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.215406596660614, 'reward_total_composite_mean': 0.38957157731056213, 'reward_total_composite_std': 0.18033328652381897, 'epoch': 0.03}

  9%|▊         | 282/3300 [41:01<6:08:09,  7.32s/it]INFO 04-13 08:01:59 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 283/3300 [41:10<6:20:58,  7.58s/it]2026-04-13 08:02:07,916 | INFO | train_grpo_train | metrics_logged mode=train step=283

                                                    
{'loss': 0.099, 'grad_norm': 9.182372093200684, 'learning_rate': 9.145454545454546e-06, 'num_tokens': 522270.0, 'completions/mean_length': 124.0, 'completions/min_length': 74.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.0, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.5346735715866089, 'rewards/meter/std': 0.31673410534858704, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9935542345046997, 'rewards/repeat_soft/std': 0.008262968622148037, 'rewards/judge_quality/mean': 0.26374998688697815, 'rewards/judge_quality/std': 0.10901343077421188, 'rewards/total_composite/mean': 0.37938377261161804, 'rewards/total_composite/std': 0.17373378574848175, 'reward': 0.37938377261161804, 'reward_std': 0.17373378574848175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23359964787960052, 'sampling/sampling_logp_difference/max': 1.3233938217163086, 'sampling/importance_sampling_ratio/min': 0.2662302553653717, 'sampling/importance_sampling_ratio/mean': 1.052581548690796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.300060272216797, 'clip_ratio/low_mean': 0.09318131022155285, 'clip_ratio/low_min': 0.09318131022155285, 'clip_ratio/high_mean': 0.1137061920017004, 'clip_ratio/high_max': 0.1137061920017004, 'clip_ratio/region_mean': 0.20688750222325325, 'reward_total_mean': 0.37938377261161804, 'reward_meter_mean': 0.5346735715866089, 'reward_meter_std': 0.31673410534858704, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9935542345046997, 'reward_repeat_soft_std': 0.008262968622148037, 'reward_judge_quality_mean': 0.26374998688697815, 'reward_judge_quality_std': 0.10901343077421188, 'reward_total_composite_mean': 0.37938377261161804, 'reward_total_composite_std': 0.17373378574848175, 'epoch': 0.03}

  9%|▊         | 283/3300 [41:10<6:20:58,  7.58s/it]INFO 04-13 08:02:08 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 284/3300 [41:18<6:26:25,  7.69s/it]2026-04-13 08:02:15,862 | INFO | train_grpo_train | metrics_logged mode=train step=284

                                                    
{'loss': 0.1768, 'grad_norm': 10.7915620803833, 'learning_rate': 9.142424242424243e-06, 'num_tokens': 524321.0, 'completions/mean_length': 81.375, 'completions/min_length': 56.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.3346508741378784, 'rewards/meter/std': 0.24749335646629333, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9945473670959473, 'rewards/repeat_soft/std': 0.004762158263474703, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.426018089056015, 'rewards/total_composite/std': 0.09572703391313553, 'reward': 0.426018089056015, 'reward_std': 0.09572703391313553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23324817419052124, 'sampling/sampling_logp_difference/max': 1.6429443359375, 'sampling/importance_sampling_ratio/min': 0.193409726023674, 'sampling/importance_sampling_ratio/mean': 1.0439000129699707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6393192410469055, 'clip_ratio/low_mean': 0.08876687660813332, 'clip_ratio/low_min': 0.08876687660813332, 'clip_ratio/high_mean': 0.11488107591867447, 'clip_ratio/high_max': 0.11488107591867447, 'clip_ratio/region_mean': 0.20364795252680779, 'reward_total_mean': 0.426018089056015, 'reward_meter_mean': 0.3346508741378784, 'reward_meter_std': 0.24749335646629333, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.18600596487522125, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9945473670959473, 'reward_repeat_soft_std': 0.004762158263474703, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.426018089056015, 'reward_total_composite_std': 0.09572703391313553, 'epoch': 0.03}

  9%|▊         | 284/3300 [41:18<6:26:25,  7.69s/it]INFO 04-13 08:02:16 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 285/3300 [41:24<6:12:47,  7.42s/it]2026-04-13 08:02:22,653 | INFO | train_grpo_train | metrics_logged mode=train step=285

                                                    
{'loss': 0.0998, 'grad_norm': 18.474149703979492, 'learning_rate': 9.13939393939394e-06, 'num_tokens': 525809.0, 'completions/mean_length': 24.0, 'completions/min_length': 19.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.5993695855140686, 'rewards/meter/std': 0.4427065849304199, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9621384143829346, 'rewards/repeat_soft/std': 0.0010226722806692123, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5071156620979309, 'rewards/total_composite/std': 0.12619905173778534, 'reward': 0.5071156620979309, 'reward_std': 0.12619903683662415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19200460612773895, 'sampling/sampling_logp_difference/max': 1.185643196105957, 'sampling/importance_sampling_ratio/min': 0.30554959177970886, 'sampling/importance_sampling_ratio/mean': 1.0759772062301636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0072427690029144, 'clip_ratio/low_mean': 0.07683908101171255, 'clip_ratio/low_min': 0.07683908101171255, 'clip_ratio/high_mean': 0.08359994180500507, 'clip_ratio/high_max': 0.08359994180500507, 'clip_ratio/region_mean': 0.16043902281671762, 'reward_total_mean': 0.5071156620979309, 'reward_meter_mean': 0.5993695855140686, 'reward_meter_std': 0.4427065849304199, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9621384143829346, 'reward_repeat_soft_std': 0.0010226722806692123, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5071156620979309, 'reward_total_composite_std': 0.12619905173778534, 'epoch': 0.03}

  9%|▊         | 285/3300 [41:24<6:12:47,  7.42s/it]INFO 04-13 08:02:22 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 286/3300 [41:36<7:23:31,  8.83s/it]2026-04-13 08:02:34,775 | INFO | train_grpo_train | metrics_logged mode=train step=286

                                                    
{'loss': -0.1503, 'grad_norm': 2.696510076522827, 'learning_rate': 9.136363636363637e-06, 'num_tokens': 527855.0, 'completions/mean_length': 191.75, 'completions/min_length': 67.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 85.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.22511011362075806, 'rewards/meter/std': 0.22926172614097595, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9987255334854126, 'rewards/repeat_soft/std': 0.001067682751454413, 'rewards/judge_quality/mean': 0.2224999964237213, 'rewards/judge_quality/std': 0.16799233853816986, 'rewards/total_composite/mean': 0.283405065536499, 'rewards/total_composite/std': 0.19441770017147064, 'reward': 0.283405065536499, 'reward_std': 0.19441771507263184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24778270721435547, 'sampling/sampling_logp_difference/max': 1.1723203659057617, 'sampling/importance_sampling_ratio/min': 0.3096476197242737, 'sampling/importance_sampling_ratio/mean': 1.1020667552947998, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.5178698301315308, 'clip_ratio/low_mean': 0.023989899083971977, 'clip_ratio/low_min': 0.023989899083971977, 'clip_ratio/high_mean': 0.11538269557058811, 'clip_ratio/high_max': 0.11538269557058811, 'clip_ratio/region_mean': 0.1393725946545601, 'reward_total_mean': 0.283405065536499, 'reward_meter_mean': 0.22511011362075806, 'reward_meter_std': 0.22926172614097595, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9987255334854126, 'reward_repeat_soft_std': 0.001067682751454413, 'reward_judge_quality_mean': 0.2224999964237213, 'reward_judge_quality_std': 0.16799233853816986, 'reward_total_composite_mean': 0.283405065536499, 'reward_total_composite_std': 0.19441770017147064, 'epoch': 0.03}

  9%|▊         | 286/3300 [41:37<7:23:31,  8.83s/it]INFO 04-13 08:02:35 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 287/3300 [41:49<8:11:47,  9.79s/it]2026-04-13 08:02:46,816 | INFO | train_grpo_train | metrics_logged mode=train step=287

                                                    
{'loss': -0.191, 'grad_norm': 2.264235496520996, 'learning_rate': 9.133333333333335e-06, 'num_tokens': 530125.0, 'completions/mean_length': 237.75, 'completions/min_length': 121.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 146.33334350585938, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.4126812219619751, 'rewards/meter/std': 0.27017462253570557, 'rewards/count_adherence/mean': 0.7750000357627869, 'rewards/count_adherence/std': 0.16690459847450256, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.999116063117981, 'rewards/repeat_soft/std': 0.0009569707908667624, 'rewards/judge_quality/mean': 0.2512499988079071, 'rewards/judge_quality/std': 0.15887439250946045, 'rewards/total_composite/mean': 0.33408790826797485, 'rewards/total_composite/std': 0.16096235811710358, 'reward': 0.33408790826797485, 'reward_std': 0.16096235811710358, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23185819387435913, 'sampling/sampling_logp_difference/max': 1.263564109802246, 'sampling/importance_sampling_ratio/min': 0.2826448380947113, 'sampling/importance_sampling_ratio/mean': 1.054463505744934, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1233180463314056, 'clip_ratio/low_mean': 0.006369426846504211, 'clip_ratio/low_min': 0.006369426846504211, 'clip_ratio/high_mean': 0.13328539207577705, 'clip_ratio/high_max': 0.13328539207577705, 'clip_ratio/region_mean': 0.13965481892228127, 'reward_total_mean': 0.33408790826797485, 'reward_meter_mean': 0.4126812219619751, 'reward_meter_std': 0.27017462253570557, 'reward_count_adherence_mean': 0.7750000357627869, 'reward_count_adherence_std': 0.16690459847450256, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.999116063117981, 'reward_repeat_soft_std': 0.0009569707908667624, 'reward_judge_quality_mean': 0.2512499988079071, 'reward_judge_quality_std': 0.15887439250946045, 'reward_total_composite_mean': 0.33408790826797485, 'reward_total_composite_std': 0.16096235811710358, 'epoch': 0.03}

  9%|▊         | 287/3300 [41:49<8:11:47,  9.79s/it]INFO 04-13 08:02:47 [block_pool.py:378] Successfully reset prefix cache

  9%|▊         | 288/3300 [41:55<7:23:24,  8.83s/it]2026-04-13 08:02:53,409 | INFO | train_grpo_train | metrics_logged mode=train step=288

                                                    
{'loss': 0.0965, 'grad_norm': 18.33196258544922, 'learning_rate': 9.130303030303032e-06, 'num_tokens': 531671.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5347582101821899, 'rewards/meter/std': 0.4431576132774353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9937200546264648, 'rewards/repeat_soft/std': 0.005315437447279692, 'rewards/judge_quality/mean': 0.5225000381469727, 'rewards/judge_quality/std': 0.2406984120607376, 'rewards/total_composite/mean': 0.5168812870979309, 'rewards/total_composite/std': 0.26617226004600525, 'reward': 0.5168812870979309, 'reward_std': 0.26617226004600525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22125428915023804, 'sampling/sampling_logp_difference/max': 1.2057805061340332, 'sampling/importance_sampling_ratio/min': 0.29945817589759827, 'sampling/importance_sampling_ratio/mean': 1.0432169437408447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7232500463724136, 'clip_ratio/low_mean': 0.08229638077318668, 'clip_ratio/low_min': 0.08229638077318668, 'clip_ratio/high_mean': 0.11801720969378948, 'clip_ratio/high_max': 0.11801720969378948, 'clip_ratio/region_mean': 0.20031359046697617, 'reward_total_mean': 0.5168812870979309, 'reward_meter_mean': 0.5347582101821899, 'reward_meter_std': 0.4431576132774353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9937200546264648, 'reward_repeat_soft_std': 0.005315437447279692, 'reward_judge_quality_mean': 0.5225000381469727, 'reward_judge_quality_std': 0.2406984120607376, 'reward_total_composite_mean': 0.5168812870979309, 'reward_total_composite_std': 0.26617226004600525, 'epoch': 0.03}

  9%|▊         | 288/3300 [41:55<7:23:24,  8.83s/it]INFO 04-13 08:02:53 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 289/3300 [42:03<7:09:45,  8.56s/it]2026-04-13 08:03:01,344 | INFO | train_grpo_train | metrics_logged mode=train step=289

                                                    
{'loss': 0.0646, 'grad_norm': 16.360633850097656, 'learning_rate': 9.127272727272727e-06, 'num_tokens': 533543.0, 'completions/mean_length': 70.0, 'completions/min_length': 45.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.5416287183761597, 'rewards/meter/std': 0.33586210012435913, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9970144033432007, 'rewards/repeat_soft/std': 0.0027460602577775717, 'rewards/judge_quality/mean': 0.6200000047683716, 'rewards/judge_quality/std': 0.21380899846553802, 'rewards/total_composite/mean': 0.5797946453094482, 'rewards/total_composite/std': 0.17210178077220917, 'reward': 0.5797946453094482, 'reward_std': 0.17210179567337036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2408500611782074, 'sampling/sampling_logp_difference/max': 1.9898395538330078, 'sampling/importance_sampling_ratio/min': 0.13671736419200897, 'sampling/importance_sampling_ratio/mean': 1.0045815706253052, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4601524621248245, 'clip_ratio/low_mean': 0.09364362061023712, 'clip_ratio/low_min': 0.09364362061023712, 'clip_ratio/high_mean': 0.10282890498638153, 'clip_ratio/high_max': 0.10282890498638153, 'clip_ratio/region_mean': 0.19647252559661865, 'reward_total_mean': 0.5797946453094482, 'reward_meter_mean': 0.5416287183761597, 'reward_meter_std': 0.33586210012435913, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9970144033432007, 'reward_repeat_soft_std': 0.0027460602577775717, 'reward_judge_quality_mean': 0.6200000047683716, 'reward_judge_quality_std': 0.21380899846553802, 'reward_total_composite_mean': 0.5797946453094482, 'reward_total_composite_std': 0.17210178077220917, 'epoch': 0.03}

  9%|▉         | 289/3300 [42:03<7:09:45,  8.56s/it]INFO 04-13 08:03:01 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 290/3300 [42:10<6:46:11,  8.10s/it]2026-04-13 08:03:08,354 | INFO | train_grpo_train | metrics_logged mode=train step=290

                                                    
{'loss': 0.0531, 'grad_norm': 13.182306289672852, 'learning_rate': 9.124242424242425e-06, 'num_tokens': 535226.0, 'completions/mean_length': 44.375, 'completions/min_length': 40.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9862470626831055, 'rewards/meter/std': 0.010380273684859276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9979305863380432, 'rewards/repeat_soft/std': 0.003956858534365892, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.26762181520462036, 'rewards/total_composite/mean': 0.7850335836410522, 'rewards/total_composite/std': 0.17486260831356049, 'reward': 0.7850335836410522, 'reward_std': 0.17486260831356049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1653902381658554, 'sampling/sampling_logp_difference/max': 1.4844894409179688, 'sampling/importance_sampling_ratio/min': 0.22661800682544708, 'sampling/importance_sampling_ratio/mean': 1.0317223072052002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5290031135082245, 'clip_ratio/low_mean': 0.06294930493459105, 'clip_ratio/low_min': 0.06294930493459105, 'clip_ratio/high_mean': 0.06362293194979429, 'clip_ratio/high_max': 0.06362293194979429, 'clip_ratio/region_mean': 0.12657223688438535, 'reward_total_mean': 0.7850335836410522, 'reward_meter_mean': 0.9862470626831055, 'reward_meter_std': 0.010380273684859276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9979305863380432, 'reward_repeat_soft_std': 0.003956858534365892, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.26762181520462036, 'reward_total_composite_mean': 0.7850335836410522, 'reward_total_composite_std': 0.17486260831356049, 'epoch': 0.03}

  9%|▉         | 290/3300 [42:10<6:46:11,  8.10s/it]INFO 04-13 08:03:08 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 291/3300 [42:17<6:28:39,  7.75s/it]2026-04-13 08:03:15,294 | INFO | train_grpo_train | metrics_logged mode=train step=291

                                                    
{'loss': -0.0157, 'grad_norm': 13.291495323181152, 'learning_rate': 9.121212121212122e-06, 'num_tokens': 536954.0, 'completions/mean_length': 39.0, 'completions/min_length': 30.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7474240064620972, 'rewards/meter/std': 0.3927259147167206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9828974008560181, 'rewards/repeat_soft/std': 0.016468608751893044, 'rewards/judge_quality/mean': 0.3725000023841858, 'rewards/judge_quality/std': 0.11055056750774384, 'rewards/total_composite/mean': 0.5376838445663452, 'rewards/total_composite/std': 0.11802530288696289, 'reward': 0.5376838445663452, 'reward_std': 0.1180252879858017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20696787536144257, 'sampling/sampling_logp_difference/max': 0.9229650497436523, 'sampling/importance_sampling_ratio/min': 0.4048669636249542, 'sampling/importance_sampling_ratio/mean': 1.0405880212783813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7855406999588013, 'clip_ratio/low_mean': 0.09184157848358154, 'clip_ratio/low_min': 0.09184157848358154, 'clip_ratio/high_mean': 0.0927211381494999, 'clip_ratio/high_max': 0.0927211381494999, 'clip_ratio/region_mean': 0.18456271663308144, 'reward_total_mean': 0.5376838445663452, 'reward_meter_mean': 0.7474240064620972, 'reward_meter_std': 0.3927259147167206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9828974008560181, 'reward_repeat_soft_std': 0.016468608751893044, 'reward_judge_quality_mean': 0.3725000023841858, 'reward_judge_quality_std': 0.11055056750774384, 'reward_total_composite_mean': 0.5376838445663452, 'reward_total_composite_std': 0.11802530288696289, 'epoch': 0.03}

  9%|▉         | 291/3300 [42:17<6:28:39,  7.75s/it]INFO 04-13 08:03:15 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 292/3300 [42:25<6:33:51,  7.86s/it]2026-04-13 08:03:23,397 | INFO | train_grpo_train | metrics_logged mode=train step=292

                                                    
{'loss': 0.0883, 'grad_norm': 10.179062843322754, 'learning_rate': 9.118181818181819e-06, 'num_tokens': 539060.0, 'completions/mean_length': 82.25, 'completions/min_length': 56.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.25, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.4219987988471985, 'rewards/meter/std': 0.2983505427837372, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.99605393409729, 'rewards/repeat_soft/std': 0.002786876168102026, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.4893326759338379, 'rewards/total_composite/std': 0.10986044257879257, 'reward': 0.4893326759338379, 'reward_std': 0.10986043512821198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23603616654872894, 'sampling/sampling_logp_difference/max': 2.265481948852539, 'sampling/importance_sampling_ratio/min': 0.10378000885248184, 'sampling/importance_sampling_ratio/mean': 1.0513489246368408, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.693507432937622, 'clip_ratio/low_mean': 0.1513797529041767, 'clip_ratio/low_min': 0.1513797529041767, 'clip_ratio/high_mean': 0.06801920384168625, 'clip_ratio/high_max': 0.06801920384168625, 'clip_ratio/region_mean': 0.21939895674586296, 'reward_total_mean': 0.4893326759338379, 'reward_meter_mean': 0.4219987988471985, 'reward_meter_std': 0.2983505427837372, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.99605393409729, 'reward_repeat_soft_std': 0.002786876168102026, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.4893326759338379, 'reward_total_composite_std': 0.10986044257879257, 'epoch': 0.03}

  9%|▉         | 292/3300 [42:25<6:33:51,  7.86s/it]INFO 04-13 08:03:23 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 293/3300 [42:33<6:28:52,  7.76s/it]2026-04-13 08:03:30,930 | INFO | train_grpo_train | metrics_logged mode=train step=293

                                                    
{'loss': -0.0211, 'grad_norm': 15.809340476989746, 'learning_rate': 9.115151515151516e-06, 'num_tokens': 540735.0, 'completions/mean_length': 34.375, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5477538108825684, 'rewards/meter/std': 0.3368853032588959, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996740460395813, 'rewards/repeat_soft/std': 0.005701808724552393, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5014572739601135, 'rewards/total_composite/std': 0.09065254777669907, 'reward': 0.5014572739601135, 'reward_std': 0.09065254032611847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20220136642456055, 'sampling/sampling_logp_difference/max': 1.4310646057128906, 'sampling/importance_sampling_ratio/min': 0.23905429244041443, 'sampling/importance_sampling_ratio/mean': 1.0401586294174194, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.230154663324356, 'clip_ratio/low_mean': 0.0651041679084301, 'clip_ratio/low_min': 0.0651041679084301, 'clip_ratio/high_mean': 0.12616414576768875, 'clip_ratio/high_max': 0.12616414576768875, 'clip_ratio/region_mean': 0.19126831367611885, 'reward_total_mean': 0.5014572739601135, 'reward_meter_mean': 0.5477538108825684, 'reward_meter_std': 0.3368853032588959, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996740460395813, 'reward_repeat_soft_std': 0.005701808724552393, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5014572739601135, 'reward_total_composite_std': 0.09065254777669907, 'epoch': 0.03}

  9%|▉         | 293/3300 [42:33<6:28:52,  7.76s/it]INFO 04-13 08:03:31 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 294/3300 [42:42<6:52:48,  8.24s/it]2026-04-13 08:03:40,291 | INFO | train_grpo_train | metrics_logged mode=train step=294

                                                    
{'loss': 0.063, 'grad_norm': 11.58062744140625, 'learning_rate': 9.112121212121214e-06, 'num_tokens': 542607.0, 'completions/mean_length': 56.0, 'completions/min_length': 45.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.6475329399108887, 'rewards/meter/std': 0.3424645960330963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984492063522339, 'rewards/repeat_soft/std': 0.0020309877581894398, 'rewards/judge_quality/mean': 0.4049999713897705, 'rewards/judge_quality/std': 0.16716118156909943, 'rewards/total_composite/mean': 0.5189603567123413, 'rewards/total_composite/std': 0.1285088062286377, 'reward': 0.5189603567123413, 'reward_std': 0.1285088211297989, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22326365113258362, 'sampling/sampling_logp_difference/max': 1.262068748474121, 'sampling/importance_sampling_ratio/min': 0.28306782245635986, 'sampling/importance_sampling_ratio/mean': 1.0421050786972046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6189921498298645, 'clip_ratio/low_mean': 0.12590462155640125, 'clip_ratio/low_min': 0.12590462155640125, 'clip_ratio/high_mean': 0.07572508230805397, 'clip_ratio/high_max': 0.07572508230805397, 'clip_ratio/region_mean': 0.20162970386445522, 'reward_total_mean': 0.5189603567123413, 'reward_meter_mean': 0.6475329399108887, 'reward_meter_std': 0.3424645960330963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984492063522339, 'reward_repeat_soft_std': 0.0020309877581894398, 'reward_judge_quality_mean': 0.4049999713897705, 'reward_judge_quality_std': 0.16716118156909943, 'reward_total_composite_mean': 0.5189603567123413, 'reward_total_composite_std': 0.1285088062286377, 'epoch': 0.03}

  9%|▉         | 294/3300 [42:42<6:52:48,  8.24s/it]INFO 04-13 08:03:40 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 295/3300 [42:49<6:27:56,  7.75s/it]2026-04-13 08:03:46,885 | INFO | train_grpo_train | metrics_logged mode=train step=295

                                                    
{'loss': 0.0501, 'grad_norm': 14.44871997833252, 'learning_rate': 9.10909090909091e-06, 'num_tokens': 543998.0, 'completions/mean_length': 19.875, 'completions/min_length': 15.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.7429407238960266, 'rewards/meter/std': 0.37553855776786804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9585226774215698, 'rewards/repeat_soft/std': 0.011249415576457977, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5907266139984131, 'rewards/total_composite/std': 0.1698986142873764, 'reward': 0.5907266139984131, 'reward_std': 0.1698986291885376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18262533843517303, 'sampling/sampling_logp_difference/max': 1.3003034591674805, 'sampling/importance_sampling_ratio/min': 0.2724491059780121, 'sampling/importance_sampling_ratio/mean': 1.0100187063217163, 'sampling/importance_sampling_ratio/max': 1.863359808921814, 'entropy': 1.568077489733696, 'clip_ratio/low_mean': 0.053125000558793545, 'clip_ratio/low_min': 0.053125000558793545, 'clip_ratio/high_mean': 0.14995477348566055, 'clip_ratio/high_max': 0.14995477348566055, 'clip_ratio/region_mean': 0.2030797740444541, 'reward_total_mean': 0.5907266139984131, 'reward_meter_mean': 0.7429407238960266, 'reward_meter_std': 0.37553855776786804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9585226774215698, 'reward_repeat_soft_std': 0.011249415576457977, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5907266139984131, 'reward_total_composite_std': 0.1698986142873764, 'epoch': 0.03}

  9%|▉         | 295/3300 [42:49<6:27:56,  7.75s/it]INFO 04-13 08:03:47 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 296/3300 [42:56<6:18:22,  7.56s/it]2026-04-13 08:03:54,001 | INFO | train_grpo_train | metrics_logged mode=train step=296

                                                    
{'loss': 0.1309, 'grad_norm': 22.907466888427734, 'learning_rate': 9.106060606060606e-06, 'num_tokens': 545273.0, 'completions/mean_length': 19.375, 'completions/min_length': 14.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.4516580104827881, 'rewards/meter/std': 0.4649323523044586, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8865073919296265, 'rewards/repeat_soft/std': 0.18254618346691132, 'rewards/judge_quality/mean': 0.38749998807907104, 'rewards/judge_quality/std': 0.11877349019050598, 'rewards/total_composite/mean': 0.4390692710876465, 'rewards/total_composite/std': 0.12529818713665009, 'reward': 0.4390692710876465, 'reward_std': 0.12529820203781128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21530936658382416, 'sampling/sampling_logp_difference/max': 1.4255495071411133, 'sampling/importance_sampling_ratio/min': 0.2403763383626938, 'sampling/importance_sampling_ratio/mean': 1.0328723192214966, 'sampling/importance_sampling_ratio/max': 1.8340866565704346, 'entropy': 2.569779545068741, 'clip_ratio/low_mean': 0.0794306411407888, 'clip_ratio/low_min': 0.0794306411407888, 'clip_ratio/high_mean': 0.05716765904799104, 'clip_ratio/high_max': 0.05716765904799104, 'clip_ratio/region_mean': 0.13659830018877983, 'reward_total_mean': 0.4390692710876465, 'reward_meter_mean': 0.4516580104827881, 'reward_meter_std': 0.4649323523044586, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8865073919296265, 'reward_repeat_soft_std': 0.18254618346691132, 'reward_judge_quality_mean': 0.38749998807907104, 'reward_judge_quality_std': 0.11877349019050598, 'reward_total_composite_mean': 0.4390692710876465, 'reward_total_composite_std': 0.12529818713665009, 'epoch': 0.03}

  9%|▉         | 296/3300 [42:56<6:18:22,  7.56s/it]INFO 04-13 08:03:54 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 297/3300 [43:03<6:16:14,  7.52s/it]2026-04-13 08:04:01,426 | INFO | train_grpo_train | metrics_logged mode=train step=297

                                                    
{'loss': 0.0847, 'grad_norm': 11.290495872497559, 'learning_rate': 9.103030303030304e-06, 'num_tokens': 546975.0, 'completions/mean_length': 57.75, 'completions/min_length': 51.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8845210075378418, 'rewards/meter/std': 0.14227625727653503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.997162938117981, 'rewards/repeat_soft/std': 0.002737692091614008, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.16446885466575623, 'rewards/total_composite/mean': 0.5384292602539062, 'rewards/total_composite/std': 0.2386161834001541, 'reward': 0.5384292602539062, 'reward_std': 0.2386161834001541, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23331233859062195, 'sampling/sampling_logp_difference/max': 2.008185863494873, 'sampling/importance_sampling_ratio/min': 0.13423196971416473, 'sampling/importance_sampling_ratio/mean': 1.0346351861953735, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.3425499498844147, 'clip_ratio/low_mean': 0.04545697756111622, 'clip_ratio/low_min': 0.04545697756111622, 'clip_ratio/high_mean': 0.1748992707580328, 'clip_ratio/high_max': 0.1748992707580328, 'clip_ratio/region_mean': 0.22035624831914902, 'reward_total_mean': 0.5384292602539062, 'reward_meter_mean': 0.8845210075378418, 'reward_meter_std': 0.14227625727653503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.997162938117981, 'reward_repeat_soft_std': 0.002737692091614008, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.16446885466575623, 'reward_total_composite_mean': 0.5384292602539062, 'reward_total_composite_std': 0.2386161834001541, 'epoch': 0.03}

  9%|▉         | 297/3300 [43:03<6:16:14,  7.52s/it]INFO 04-13 08:04:01 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 298/3300 [43:11<6:19:54,  7.59s/it]2026-04-13 08:04:09,197 | INFO | train_grpo_train | metrics_logged mode=train step=298

                                                    
{'loss': 0.0433, 'grad_norm': 19.058700561523438, 'learning_rate': 9.100000000000001e-06, 'num_tokens': 548888.0, 'completions/mean_length': 50.125, 'completions/min_length': 39.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.3783789873123169, 'rewards/meter/std': 0.25831398367881775, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9826259613037109, 'rewards/repeat_soft/std': 0.015653958544135094, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.43459272384643555, 'rewards/total_composite/std': 0.08175422251224518, 'reward': 0.43459272384643555, 'reward_std': 0.08175422251224518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22767361998558044, 'sampling/sampling_logp_difference/max': 1.933873176574707, 'sampling/importance_sampling_ratio/min': 0.14458711445331573, 'sampling/importance_sampling_ratio/mean': 1.0235506296157837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6796212941408157, 'clip_ratio/low_mean': 0.10850738547742367, 'clip_ratio/low_min': 0.10850738547742367, 'clip_ratio/high_mean': 0.07100156880915165, 'clip_ratio/high_max': 0.07100156880915165, 'clip_ratio/region_mean': 0.17950895428657532, 'reward_total_mean': 0.43459272384643555, 'reward_meter_mean': 0.3783789873123169, 'reward_meter_std': 0.25831398367881775, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9826259613037109, 'reward_repeat_soft_std': 0.015653958544135094, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.43459272384643555, 'reward_total_composite_std': 0.08175422251224518, 'epoch': 0.03}

  9%|▉         | 298/3300 [43:11<6:19:54,  7.59s/it]INFO 04-13 08:04:09 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 299/3300 [43:18<6:08:04,  7.36s/it]2026-04-13 08:04:16,009 | INFO | train_grpo_train | metrics_logged mode=train step=299

                                                    
{'loss': 0.1628, 'grad_norm': 12.718059539794922, 'learning_rate': 9.096969696969698e-06, 'num_tokens': 550751.0, 'completions/mean_length': 60.875, 'completions/min_length': 47.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6708670258522034, 'rewards/meter/std': 0.3672305643558502, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9817839860916138, 'rewards/repeat_soft/std': 0.02569853514432907, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.5338301658630371, 'rewards/total_composite/std': 0.10163712501525879, 'reward': 0.5338301658630371, 'reward_std': 0.1016371101140976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19892556965351105, 'sampling/sampling_logp_difference/max': 1.3361644744873047, 'sampling/importance_sampling_ratio/min': 0.26285192370414734, 'sampling/importance_sampling_ratio/mean': 1.06418776512146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8512686491012573, 'clip_ratio/low_mean': 0.061188691295683384, 'clip_ratio/low_min': 0.061188691295683384, 'clip_ratio/high_mean': 0.09996799007058144, 'clip_ratio/high_max': 0.09996799007058144, 'clip_ratio/region_mean': 0.16115668136626482, 'reward_total_mean': 0.5338301658630371, 'reward_meter_mean': 0.6708670258522034, 'reward_meter_std': 0.3672305643558502, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9817839860916138, 'reward_repeat_soft_std': 0.02569853514432907, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.5338301658630371, 'reward_total_composite_std': 0.10163712501525879, 'epoch': 0.03}

  9%|▉         | 299/3300 [43:18<6:08:04,  7.36s/it]INFO 04-13 08:04:16 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 300/3300 [43:25<6:01:55,  7.24s/it]2026-04-13 08:04:22,970 | INFO | train_grpo_train | metrics_logged mode=train step=300

                                                    
{'loss': 0.1161, 'grad_norm': 18.334640502929688, 'learning_rate': 9.093939393939395e-06, 'num_tokens': 552255.0, 'completions/mean_length': 33.0, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7040756940841675, 'rewards/meter/std': 0.40580132603645325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9876308441162109, 'rewards/repeat_soft/std': 0.019401581957936287, 'rewards/judge_quality/mean': 0.3962499797344208, 'rewards/judge_quality/std': 0.09085899591445923, 'rewards/total_composite/mean': 0.5045775175094604, 'rewards/total_composite/std': 0.22371245920658112, 'reward': 0.5045775175094604, 'reward_std': 0.22371245920658112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2362433224916458, 'sampling/sampling_logp_difference/max': 1.2512593269348145, 'sampling/importance_sampling_ratio/min': 0.2861442267894745, 'sampling/importance_sampling_ratio/mean': 1.0532236099243164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6142437160015106, 'clip_ratio/low_mean': 0.06884342804551125, 'clip_ratio/low_min': 0.06884342804551125, 'clip_ratio/high_mean': 0.17780407890677452, 'clip_ratio/high_max': 0.17780407890677452, 'clip_ratio/region_mean': 0.24664750695228577, 'reward_total_mean': 0.5045775175094604, 'reward_meter_mean': 0.7040756940841675, 'reward_meter_std': 0.40580132603645325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9876308441162109, 'reward_repeat_soft_std': 0.019401581957936287, 'reward_judge_quality_mean': 0.3962499797344208, 'reward_judge_quality_std': 0.09085899591445923, 'reward_total_composite_mean': 0.5045775175094604, 'reward_total_composite_std': 0.22371245920658112, 'epoch': 0.03}

  9%|▉         | 300/3300 [43:25<6:01:55,  7.24s/it]INFO 04-13 08:04:23 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:19,  2.39s/it]

 30%|███       | 3/10 [00:14<00:37,  5.38s/it]

 40%|████      | 4/10 [00:18<00:29,  5.00s/it]

 50%|█████     | 5/10 [00:24<00:25,  5.17s/it]

 60%|██████    | 6/10 [00:28<00:19,  4.99s/it]

 70%|███████   | 7/10 [00:33<00:14,  4.80s/it]

 80%|████████  | 8/10 [00:37<00:09,  4.60s/it]

 90%|█████████ | 9/10 [00:42<00:04,  4.66s/it]

100%|██████████| 10/10 [00:46<00:00,  4.44s/it]2026-04-13 08:05:13,469 | INFO | train_grpo_train | metrics_logged mode=eval step=300

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 50.4786, 'eval_samples_per_second': 1.585, 'eval_steps_per_second': 0.198, 'eval_num_tokens': 552255.0, 'eval_completions/mean_length': 60.25, 'eval_completions/min_length': 26.1, 'eval_completions/max_length': 135.1, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 54.56607170104981, 'eval_completions/min_terminated_length': 26.1, 'eval_completions/max_terminated_length': 92.7, 'eval_rewards/meter/mean': 0.4739754617214203, 'eval_rewards/meter/std': 0.35049658715724946, 'eval_rewards/count_adherence/mean': 0.962291669845581, 'eval_rewards/count_adherence/std': 0.08423538282513618, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.9883093476295471, 'eval_rewards/repeat_soft/std': 0.02110612459946424, 'eval_rewards/judge_quality/mean': 0.46175000071525574, 'eval_rewards/judge_quality/std': 0.16253908574581147, 'eval_rewards/total_composite/mean': 0.4752084851264954, 'eval_rewards/total_composite/std': 0.13369336500763893, 'eval_reward': 0.4752084851264954, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15603995621204375, 'eval_sampling/sampling_logp_difference/max': 1.1039558410644532, 'eval_sampling/importance_sampling_ratio/min': 0.3371889188885689, 'eval_sampling/importance_sampling_ratio/mean': 1.05187269449234, 'eval_sampling/importance_sampling_ratio/max': 1.5952545404434204, 'eval_entropy': 2.5126993536949156, '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': 0.4752084851264954, 'eval_reward_meter_mean': 0.4739754617214203, 'eval_reward_meter_std': 0.35049658715724946, 'eval_reward_count_adherence_mean': 0.962291669845581, 'eval_reward_count_adherence_std': 0.08423538282513618, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.9883093476295471, 'eval_reward_repeat_soft_std': 0.02110612459946424, 'eval_reward_judge_quality_mean': 0.46175000071525574, 'eval_reward_judge_quality_std': 0.16253908574581147, 'eval_reward_total_composite_mean': 0.4752084851264954, 'eval_reward_total_composite_std': 0.13369336500763893, 'epoch': 0.03}

  9%|▉         | 300/3300 [44:15<6:01:55,  7.24s/it]

100%|██████████| 10/10 [00:46<00:00,  4.44s/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 [44:25<19:16:45, 23.14s/it]2026-04-13 08:05:23,219 | INFO | train_grpo_train | metrics_logged mode=train step=301

                                                     
{'loss': 0.0372, 'grad_norm': 13.64366626739502, 'learning_rate': 9.090909090909091e-06, 'num_tokens': 554193.0, 'completions/mean_length': 64.25, 'completions/min_length': 57.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.2980521023273468, 'rewards/meter/std': 0.3184024691581726, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939161539077759, 'rewards/repeat_soft/std': 0.009360043331980705, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.42934858798980713, 'rewards/total_composite/std': 0.08828973025083542, 'reward': 0.42934858798980713, 'reward_std': 0.08828973025083542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2365809828042984, 'sampling/sampling_logp_difference/max': 1.6053876876831055, 'sampling/importance_sampling_ratio/min': 0.2008116990327835, 'sampling/importance_sampling_ratio/mean': 1.0825746059417725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.1526523530483246, 'clip_ratio/low_mean': 0.10929449833929539, 'clip_ratio/low_min': 0.10929449833929539, 'clip_ratio/high_mean': 0.06331565603613853, 'clip_ratio/high_max': 0.06331565603613853, 'clip_ratio/region_mean': 0.17261015437543392, 'reward_total_mean': 0.42934858798980713, 'reward_meter_mean': 0.2980521023273468, 'reward_meter_std': 0.3184024691581726, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939161539077759, 'reward_repeat_soft_std': 0.009360043331980705, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.42934858798980713, 'reward_total_composite_std': 0.08828973025083542, 'epoch': 0.03}

  9%|▉         | 301/3300 [44:25<19:16:45, 23.14s/it]INFO 04-13 08:05:23 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 302/3300 [44:33<15:30:17, 18.62s/it]2026-04-13 08:05:31,282 | INFO | train_grpo_train | metrics_logged mode=train step=302

                                                     
{'loss': 0.0041, 'grad_norm': 11.884839057922363, 'learning_rate': 9.087878787878788e-06, 'num_tokens': 556028.0, 'completions/mean_length': 58.375, 'completions/min_length': 50.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7205952405929565, 'rewards/meter/std': 0.37002670764923096, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9917113780975342, 'rewards/repeat_soft/std': 0.008996184915304184, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5415973663330078, 'rewards/total_composite/std': 0.10051526874303818, 'reward': 0.5415973663330078, 'reward_std': 0.10051526874303818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22755461931228638, 'sampling/sampling_logp_difference/max': 1.2427740097045898, 'sampling/importance_sampling_ratio/min': 0.28858259320259094, 'sampling/importance_sampling_ratio/mean': 1.03282630443573, 'sampling/importance_sampling_ratio/max': 1.9246867895126343, 'entropy': 2.900486648082733, 'clip_ratio/low_mean': 0.03722166828811169, 'clip_ratio/low_min': 0.03722166828811169, 'clip_ratio/high_mean': 0.14813154190778732, 'clip_ratio/high_max': 0.14813154190778732, 'clip_ratio/region_mean': 0.185353210195899, 'reward_total_mean': 0.5415973663330078, 'reward_meter_mean': 0.7205952405929565, 'reward_meter_std': 0.37002670764923096, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9917113780975342, 'reward_repeat_soft_std': 0.008996184915304184, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5415973663330078, 'reward_total_composite_std': 0.10051526874303818, 'epoch': 0.03}

  9%|▉         | 302/3300 [44:33<15:30:17, 18.62s/it]INFO 04-13 08:05:31 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 303/3300 [44:40<12:41:08, 15.24s/it]2026-04-13 08:05:38,631 | INFO | train_grpo_train | metrics_logged mode=train step=303

                                                     
{'loss': 0.0515, 'grad_norm': 10.325447082519531, 'learning_rate': 9.084848484848486e-06, 'num_tokens': 558213.0, 'completions/mean_length': 87.125, 'completions/min_length': 78.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.125, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.33736515045166016, 'rewards/meter/std': 0.2612042725086212, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9919346570968628, 'rewards/repeat_soft/std': 0.0070076147094368935, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.4575711488723755, 'rewards/total_composite/std': 0.07796068489551544, 'reward': 0.4575711488723755, 'reward_std': 0.07796068489551544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20860318839550018, 'sampling/sampling_logp_difference/max': 1.5033361911773682, 'sampling/importance_sampling_ratio/min': 0.22238700091838837, 'sampling/importance_sampling_ratio/mean': 1.048304796218872, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4992952048778534, 'clip_ratio/low_mean': 0.09519331157207489, 'clip_ratio/low_min': 0.09519331157207489, 'clip_ratio/high_mean': 0.09584271814674139, 'clip_ratio/high_max': 0.09584271814674139, 'clip_ratio/region_mean': 0.19103602971881628, 'reward_total_mean': 0.4575711488723755, 'reward_meter_mean': 0.33736515045166016, 'reward_meter_std': 0.2612042725086212, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9919346570968628, 'reward_repeat_soft_std': 0.0070076147094368935, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.4575711488723755, 'reward_total_composite_std': 0.07796068489551544, 'epoch': 0.03}

  9%|▉         | 303/3300 [44:40<12:41:08, 15.24s/it]INFO 04-13 08:05:38 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 304/3300 [44:48<10:50:00, 13.02s/it]2026-04-13 08:05:46,469 | INFO | train_grpo_train | metrics_logged mode=train step=304

                                                     
{'loss': 0.047, 'grad_norm': 11.980401992797852, 'learning_rate': 9.081818181818183e-06, 'num_tokens': 560301.0, 'completions/mean_length': 83.0, 'completions/min_length': 73.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.0, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.3610544800758362, 'rewards/meter/std': 0.272571325302124, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9927531480789185, 'rewards/repeat_soft/std': 0.009791922755539417, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.4092235267162323, 'rewards/total_composite/std': 0.20892708003520966, 'reward': 0.4092235267162323, 'reward_std': 0.20892708003520966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22895309329032898, 'sampling/sampling_logp_difference/max': 2.0566253662109375, 'sampling/importance_sampling_ratio/min': 0.12788480520248413, 'sampling/importance_sampling_ratio/mean': 1.0300662517547607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2296884953975677, 'clip_ratio/low_mean': 0.10625382140278816, 'clip_ratio/low_min': 0.10625382140278816, 'clip_ratio/high_mean': 0.07721620984375477, 'clip_ratio/high_max': 0.07721620984375477, 'clip_ratio/region_mean': 0.18347003124654293, 'reward_total_mean': 0.4092235267162323, 'reward_meter_mean': 0.3610544800758362, 'reward_meter_std': 0.272571325302124, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9927531480789185, 'reward_repeat_soft_std': 0.009791922755539417, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.4092235267162323, 'reward_total_composite_std': 0.20892708003520966, 'epoch': 0.03}

  9%|▉         | 304/3300 [44:48<10:50:00, 13.02s/it]INFO 04-13 08:05:46 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 305/3300 [44:56<9:30:53, 11.44s/it] 2026-04-13 08:05:54,216 | INFO | train_grpo_train | metrics_logged mode=train step=305

                                                    
{'loss': 0.0564, 'grad_norm': 19.926694869995117, 'learning_rate': 9.078787878787878e-06, 'num_tokens': 561755.0, 'completions/mean_length': 36.75, 'completions/min_length': 25.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.1998346596956253, 'rewards/meter/std': 0.3175995647907257, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9959381818771362, 'rewards/repeat_soft/std': 0.011488578282296658, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.36549967527389526, 'rewards/total_composite/std': 0.1727069914340973, 'reward': 0.36549967527389526, 'reward_std': 0.1727069914340973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1940957009792328, 'sampling/sampling_logp_difference/max': 1.5790209770202637, 'sampling/importance_sampling_ratio/min': 0.20617684721946716, 'sampling/importance_sampling_ratio/mean': 1.0141253471374512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7443688213825226, 'clip_ratio/low_mean': 0.05058035673573613, 'clip_ratio/low_min': 0.05058035673573613, 'clip_ratio/high_mean': 0.11338341608643532, 'clip_ratio/high_max': 0.11338341608643532, 'clip_ratio/region_mean': 0.16396377282217145, 'reward_total_mean': 0.36549967527389526, 'reward_meter_mean': 0.1998346596956253, 'reward_meter_std': 0.3175995647907257, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9959381818771362, 'reward_repeat_soft_std': 0.011488578282296658, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.36549967527389526, 'reward_total_composite_std': 0.1727069914340973, 'epoch': 0.03}

  9%|▉         | 305/3300 [44:56<9:30:53, 11.44s/it]INFO 04-13 08:05:54 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 306/3300 [45:04<8:39:17, 10.41s/it]2026-04-13 08:06:02,218 | INFO | train_grpo_train | metrics_logged mode=train step=306

                                                    
{'loss': -0.025, 'grad_norm': 17.35762596130371, 'learning_rate': 9.075757575757577e-06, 'num_tokens': 563522.0, 'completions/mean_length': 49.875, 'completions/min_length': 41.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5379660129547119, 'rewards/meter/std': 0.389585018157959, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9965270757675171, 'rewards/repeat_soft/std': 0.0026339867617934942, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.12291111052036285, 'rewards/total_composite/mean': 0.4408089816570282, 'rewards/total_composite/std': 0.11375391483306885, 'reward': 0.4408089816570282, 'reward_std': 0.11375392228364944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2487066686153412, 'sampling/sampling_logp_difference/max': 1.7364892959594727, 'sampling/importance_sampling_ratio/min': 0.17613768577575684, 'sampling/importance_sampling_ratio/mean': 1.030597448348999, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.687733069062233, 'clip_ratio/low_mean': 0.1220129169523716, 'clip_ratio/low_min': 0.1220129169523716, 'clip_ratio/high_mean': 0.09736341051757336, 'clip_ratio/high_max': 0.09736341051757336, 'clip_ratio/region_mean': 0.21937632746994495, 'reward_total_mean': 0.4408089816570282, 'reward_meter_mean': 0.5379660129547119, 'reward_meter_std': 0.389585018157959, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9965270757675171, 'reward_repeat_soft_std': 0.0026339867617934942, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.12291111052036285, 'reward_total_composite_mean': 0.4408089816570282, 'reward_total_composite_std': 0.11375391483306885, 'epoch': 0.03}

  9%|▉         | 306/3300 [45:04<8:39:17, 10.41s/it]INFO 04-13 08:06:02 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 307/3300 [45:11<7:48:45,  9.40s/it]2026-04-13 08:06:09,262 | INFO | train_grpo_train | metrics_logged mode=train step=307

                                                    
{'loss': 0.0418, 'grad_norm': 19.14022445678711, 'learning_rate': 9.072727272727273e-06, 'num_tokens': 564926.0, 'completions/mean_length': 33.5, 'completions/min_length': 27.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.45821669697761536, 'rewards/meter/std': 0.345441997051239, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9934337139129639, 'rewards/repeat_soft/std': 0.008110122755169868, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2353682667016983, 'rewards/total_composite/mean': 0.5238087773323059, 'rewards/total_composite/std': 0.20000897347927094, 'reward': 0.5238087773323059, 'reward_std': 0.20000897347927094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1961425095796585, 'sampling/sampling_logp_difference/max': 1.7668094635009766, 'sampling/importance_sampling_ratio/min': 0.17087732255458832, 'sampling/importance_sampling_ratio/mean': 1.030051589012146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5763475075364113, 'clip_ratio/low_mean': 0.07996872556395829, 'clip_ratio/low_min': 0.07996872556395829, 'clip_ratio/high_mean': 0.062245048582553864, 'clip_ratio/high_max': 0.062245048582553864, 'clip_ratio/region_mean': 0.14221377414651215, 'reward_total_mean': 0.5238087773323059, 'reward_meter_mean': 0.45821669697761536, 'reward_meter_std': 0.345441997051239, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9934337139129639, 'reward_repeat_soft_std': 0.008110122755169868, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2353682667016983, 'reward_total_composite_mean': 0.5238087773323059, 'reward_total_composite_std': 0.20000897347927094, 'epoch': 0.03}

  9%|▉         | 307/3300 [45:11<7:48:45,  9.40s/it]INFO 04-13 08:06:09 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 308/3300 [45:19<7:21:37,  8.86s/it]2026-04-13 08:06:16,858 | INFO | train_grpo_train | metrics_logged mode=train step=308

                                                    
{'loss': -0.0583, 'grad_norm': 16.181547164916992, 'learning_rate': 9.06969696969697e-06, 'num_tokens': 566707.0, 'completions/mean_length': 37.625, 'completions/min_length': 30.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8855193853378296, 'rewards/meter/std': 0.16866350173950195, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974081516265869, 'rewards/repeat_soft/std': 0.0032312325201928616, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5825735330581665, 'rewards/total_composite/std': 0.05710341036319733, 'reward': 0.5825735330581665, 'reward_std': 0.057103417813777924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21633878350257874, 'sampling/sampling_logp_difference/max': 1.501877784729004, 'sampling/importance_sampling_ratio/min': 0.22271157801151276, 'sampling/importance_sampling_ratio/mean': 1.0522710084915161, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1090298891067505, 'clip_ratio/low_mean': 0.04930555634200573, 'clip_ratio/low_min': 0.04930555634200573, 'clip_ratio/high_mean': 0.15730290859937668, 'clip_ratio/high_max': 0.15730290859937668, 'clip_ratio/region_mean': 0.2066084649413824, 'reward_total_mean': 0.5825735330581665, 'reward_meter_mean': 0.8855193853378296, 'reward_meter_std': 0.16866350173950195, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974081516265869, 'reward_repeat_soft_std': 0.0032312325201928616, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5825735330581665, 'reward_total_composite_std': 0.05710341036319733, 'epoch': 0.03}

  9%|▉         | 308/3300 [45:19<7:21:37,  8.86s/it]INFO 04-13 08:06:17 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 309/3300 [45:26<7:02:45,  8.48s/it]2026-04-13 08:06:24,460 | INFO | train_grpo_train | metrics_logged mode=train step=309

                                                    
{'loss': 0.0176, 'grad_norm': 11.218291282653809, 'learning_rate': 9.066666666666667e-06, 'num_tokens': 568710.0, 'completions/mean_length': 62.375, 'completions/min_length': 50.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.3145693242549896, 'rewards/meter/std': 0.29673534631729126, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9977501630783081, 'rewards/repeat_soft/std': 0.0027243790682405233, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.19949938356876373, 'rewards/total_composite/mean': 0.44601163268089294, 'rewards/total_composite/std': 0.1473219394683838, 'reward': 0.44601163268089294, 'reward_std': 0.1473219245672226, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23919273912906647, 'sampling/sampling_logp_difference/max': 1.101022720336914, 'sampling/importance_sampling_ratio/min': 0.33253082633018494, 'sampling/importance_sampling_ratio/mean': 1.0720115900039673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.621856689453125, 'clip_ratio/low_mean': 0.1271504582837224, 'clip_ratio/low_min': 0.1271504582837224, 'clip_ratio/high_mean': 0.04959016479551792, 'clip_ratio/high_max': 0.04959016479551792, 'clip_ratio/region_mean': 0.17674062307924032, 'reward_total_mean': 0.44601163268089294, 'reward_meter_mean': 0.3145693242549896, 'reward_meter_std': 0.29673534631729126, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9977501630783081, 'reward_repeat_soft_std': 0.0027243790682405233, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.19949938356876373, 'reward_total_composite_mean': 0.44601163268089294, 'reward_total_composite_std': 0.1473219394683838, 'epoch': 0.03}

  9%|▉         | 309/3300 [45:26<7:02:45,  8.48s/it]INFO 04-13 08:06:24 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 310/3300 [45:35<7:07:05,  8.57s/it]2026-04-13 08:06:33,240 | INFO | train_grpo_train | metrics_logged mode=train step=310

                                                    
{'loss': 0.0655, 'grad_norm': 8.005025863647461, 'learning_rate': 9.063636363636365e-06, 'num_tokens': 570990.0, 'completions/mean_length': 107.0, 'completions/min_length': 92.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.0, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.8105195760726929, 'rewards/meter/std': 0.17381274700164795, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.12400396168231964, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.948270320892334, 'rewards/repeat_soft/std': 0.10612338781356812, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.5282342433929443, 'rewards/total_composite/std': 0.09169287234544754, 'reward': 0.5282342433929443, 'reward_std': 0.09169287234544754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22060352563858032, 'sampling/sampling_logp_difference/max': 1.4738521575927734, 'sampling/importance_sampling_ratio/min': 0.22904148697853088, 'sampling/importance_sampling_ratio/mean': 1.0602737665176392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.581006407737732, 'clip_ratio/low_mean': 0.04999029543250799, 'clip_ratio/low_min': 0.04999029543250799, 'clip_ratio/high_mean': 0.1028110533952713, 'clip_ratio/high_max': 0.1028110533952713, 'clip_ratio/region_mean': 0.1528013488277793, 'reward_total_mean': 0.5282342433929443, 'reward_meter_mean': 0.8105195760726929, 'reward_meter_std': 0.17381274700164795, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.12400396168231964, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.948270320892334, 'reward_repeat_soft_std': 0.10612338781356812, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.5282342433929443, 'reward_total_composite_std': 0.09169287234544754, 'epoch': 0.03}

  9%|▉         | 310/3300 [45:35<7:07:05,  8.57s/it]INFO 04-13 08:06:33 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 311/3300 [45:43<6:58:19,  8.40s/it]2026-04-13 08:06:41,238 | INFO | train_grpo_train | metrics_logged mode=train step=311

                                                    
{'loss': -0.0212, 'grad_norm': 13.041028022766113, 'learning_rate': 9.06060606060606e-06, 'num_tokens': 572897.0, 'completions/mean_length': 66.375, 'completions/min_length': 56.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.375, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.30376312136650085, 'rewards/meter/std': 0.3042370080947876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9922004342079163, 'rewards/repeat_soft/std': 0.011386986821889877, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.3983754515647888, 'rewards/total_composite/std': 0.18190890550613403, 'reward': 0.3983754515647888, 'reward_std': 0.18190890550613403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25558942556381226, 'sampling/sampling_logp_difference/max': 1.3604774475097656, 'sampling/importance_sampling_ratio/min': 0.2565382719039917, 'sampling/importance_sampling_ratio/mean': 1.054573893547058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.934918701648712, 'clip_ratio/low_mean': 0.08660714328289032, 'clip_ratio/low_min': 0.08660714328289032, 'clip_ratio/high_mean': 0.1264073895290494, 'clip_ratio/high_max': 0.1264073895290494, 'clip_ratio/region_mean': 0.21301453281193972, 'reward_total_mean': 0.3983754515647888, 'reward_meter_mean': 0.30376312136650085, 'reward_meter_std': 0.3042370080947876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9922004342079163, 'reward_repeat_soft_std': 0.011386986821889877, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.3983754515647888, 'reward_total_composite_std': 0.18190890550613403, 'epoch': 0.03}

  9%|▉         | 311/3300 [45:43<6:58:19,  8.40s/it]INFO 04-13 08:06:41 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 312/3300 [45:50<6:42:16,  8.08s/it]2026-04-13 08:06:48,571 | INFO | train_grpo_train | metrics_logged mode=train step=312

                                                    
{'loss': 0.0742, 'grad_norm': 29.928604125976562, 'learning_rate': 9.057575757575759e-06, 'num_tokens': 574291.0, 'completions/mean_length': 20.25, 'completions/min_length': 17.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6075140833854675, 'rewards/meter/std': 0.39379948377609253, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.10336308926343918, 'rewards/total_composite/mean': 0.46354204416275024, 'rewards/total_composite/std': 0.07323381304740906, 'reward': 0.46354204416275024, 'reward_std': 0.07323379814624786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23011985421180725, 'sampling/sampling_logp_difference/max': 1.4686622619628906, 'sampling/importance_sampling_ratio/min': 0.23023326694965363, 'sampling/importance_sampling_ratio/mean': 1.016001582145691, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.196314051747322, 'clip_ratio/low_mean': 0.0753822373226285, 'clip_ratio/low_min': 0.0753822373226285, 'clip_ratio/high_mean': 0.1280034352093935, 'clip_ratio/high_max': 0.1280034352093935, 'clip_ratio/region_mean': 0.203385672532022, 'reward_total_mean': 0.46354204416275024, 'reward_meter_mean': 0.6075140833854675, 'reward_meter_std': 0.39379948377609253, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.10336308926343918, 'reward_total_composite_mean': 0.46354204416275024, 'reward_total_composite_std': 0.07323381304740906, 'epoch': 0.03}

  9%|▉         | 312/3300 [45:50<6:42:16,  8.08s/it]INFO 04-13 08:06:48 [block_pool.py:378] Successfully reset prefix cache

  9%|▉         | 313/3300 [45:59<6:49:49,  8.23s/it]2026-04-13 08:06:57,158 | INFO | train_grpo_train | metrics_logged mode=train step=313

                                                    
{'loss': -0.052, 'grad_norm': 12.364789962768555, 'learning_rate': 9.054545454545455e-06, 'num_tokens': 576479.0, 'completions/mean_length': 77.5, 'completions/min_length': 58.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.4774291515350342, 'rewards/meter/std': 0.23937858641147614, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9925657510757446, 'rewards/repeat_soft/std': 0.0036443264689296484, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4078633189201355, 'rewards/total_composite/std': 0.17348188161849976, 'reward': 0.4078633189201355, 'reward_std': 0.17348188161849976, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21338839828968048, 'sampling/sampling_logp_difference/max': 2.0070290565490723, 'sampling/importance_sampling_ratio/min': 0.13438734412193298, 'sampling/importance_sampling_ratio/mean': 1.0475268363952637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.310043603181839, 'clip_ratio/low_mean': 0.017241379246115685, 'clip_ratio/low_min': 0.017241379246115685, 'clip_ratio/high_mean': 0.19226273521780968, 'clip_ratio/high_max': 0.19226273521780968, 'clip_ratio/region_mean': 0.20950411446392536, 'reward_total_mean': 0.4078633189201355, 'reward_meter_mean': 0.4774291515350342, 'reward_meter_std': 0.23937858641147614, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9925657510757446, 'reward_repeat_soft_std': 0.0036443264689296484, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4078633189201355, 'reward_total_composite_std': 0.17348188161849976, 'epoch': 0.03}

  9%|▉         | 313/3300 [45:59<6:49:49,  8.23s/it]INFO 04-13 08:06:57 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 314/3300 [46:06<6:26:49,  7.77s/it]2026-04-13 08:07:03,863 | INFO | train_grpo_train | metrics_logged mode=train step=314

                                                    
{'loss': 0.0678, 'grad_norm': 21.277273178100586, 'learning_rate': 9.051515151515152e-06, 'num_tokens': 577769.0, 'completions/mean_length': 25.25, 'completions/min_length': 17.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9609196186065674, 'rewards/meter/std': 0.09015358984470367, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2921227812767029, 'rewards/total_composite/mean': 0.7461719512939453, 'rewards/total_composite/std': 0.17513616383075714, 'reward': 0.7461719512939453, 'reward_std': 0.17513614892959595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17423312366008759, 'sampling/sampling_logp_difference/max': 1.3621187210083008, 'sampling/importance_sampling_ratio/min': 0.2561175525188446, 'sampling/importance_sampling_ratio/mean': 1.039286494255066, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6310700699687004, 'clip_ratio/low_mean': 0.05719997314736247, 'clip_ratio/low_min': 0.05719997314736247, 'clip_ratio/high_mean': 0.07325160969048738, 'clip_ratio/high_max': 0.07325160969048738, 'clip_ratio/region_mean': 0.13045158283784986, 'reward_total_mean': 0.7461719512939453, 'reward_meter_mean': 0.9609196186065674, 'reward_meter_std': 0.09015358984470367, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2921227812767029, 'reward_total_composite_mean': 0.7461719512939453, 'reward_total_composite_std': 0.17513616383075714, 'epoch': 0.03}

 10%|▉         | 314/3300 [46:06<6:26:49,  7.77s/it]INFO 04-13 08:07:04 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 315/3300 [46:13<6:25:43,  7.75s/it]2026-04-13 08:07:11,572 | INFO | train_grpo_train | metrics_logged mode=train step=315

                                                    
{'loss': 0.0864, 'grad_norm': 20.46292495727539, 'learning_rate': 9.04848484848485e-06, 'num_tokens': 579251.0, 'completions/mean_length': 31.25, 'completions/min_length': 29.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.5624465942382812, 'rewards/meter/std': 0.43712732195854187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9856735467910767, 'rewards/repeat_soft/std': 0.027919858694076538, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5045914053916931, 'rewards/total_composite/std': 0.12228681147098541, 'reward': 0.5045914053916931, 'reward_std': 0.12228679656982422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2290685772895813, 'sampling/sampling_logp_difference/max': 1.2177257537841797, 'sampling/importance_sampling_ratio/min': 0.2959023416042328, 'sampling/importance_sampling_ratio/mean': 1.0391243696212769, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4925708770751953, 'clip_ratio/low_mean': 0.08946078643202782, 'clip_ratio/low_min': 0.08946078643202782, 'clip_ratio/high_mean': 0.0903023574501276, 'clip_ratio/high_max': 0.0903023574501276, 'clip_ratio/region_mean': 0.17976314388215542, 'reward_total_mean': 0.5045914053916931, 'reward_meter_mean': 0.5624465942382812, 'reward_meter_std': 0.43712732195854187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9856735467910767, 'reward_repeat_soft_std': 0.027919858694076538, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5045914053916931, 'reward_total_composite_std': 0.12228681147098541, 'epoch': 0.03}

 10%|▉         | 315/3300 [46:13<6:25:43,  7.75s/it]INFO 04-13 08:07:11 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 316/3300 [46:20<6:12:34,  7.49s/it]2026-04-13 08:07:18,448 | INFO | train_grpo_train | metrics_logged mode=train step=316

                                                    
{'loss': 0.0748, 'grad_norm': 13.913999557495117, 'learning_rate': 9.045454545454546e-06, 'num_tokens': 580891.0, 'completions/mean_length': 48.0, 'completions/min_length': 37.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.3270801603794098, 'rewards/meter/std': 0.34035104513168335, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9914262294769287, 'rewards/repeat_soft/std': 0.011263171210885048, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.2218751311302185, 'rewards/total_composite/mean': 0.4752548336982727, 'rewards/total_composite/std': 0.16100160777568817, 'reward': 0.4752548336982727, 'reward_std': 0.16100160777568817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2388077676296234, 'sampling/sampling_logp_difference/max': 1.3835439682006836, 'sampling/importance_sampling_ratio/min': 0.2506885528564453, 'sampling/importance_sampling_ratio/mean': 1.0462722778320312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.769525781273842, 'clip_ratio/low_mean': 0.1359204314649105, 'clip_ratio/low_min': 0.1359204314649105, 'clip_ratio/high_mean': 0.0889878123998642, 'clip_ratio/high_max': 0.0889878123998642, 'clip_ratio/region_mean': 0.2249082438647747, 'reward_total_mean': 0.4752548336982727, 'reward_meter_mean': 0.3270801603794098, 'reward_meter_std': 0.34035104513168335, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9914262294769287, 'reward_repeat_soft_std': 0.011263171210885048, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.2218751311302185, 'reward_total_composite_mean': 0.4752548336982727, 'reward_total_composite_std': 0.16100160777568817, 'epoch': 0.03}

 10%|▉         | 316/3300 [46:20<6:12:34,  7.49s/it]INFO 04-13 08:07:18 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 317/3300 [46:27<6:06:46,  7.38s/it]2026-04-13 08:07:25,581 | INFO | train_grpo_train | metrics_logged mode=train step=317

                                                    
{'loss': 0.0683, 'grad_norm': 18.654348373413086, 'learning_rate': 9.042424242424244e-06, 'num_tokens': 582591.0, 'completions/mean_length': 35.5, 'completions/min_length': 29.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6177049875259399, 'rewards/meter/std': 0.4753933548927307, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9828039407730103, 'rewards/repeat_soft/std': 0.04082963243126869, 'rewards/judge_quality/mean': 0.4925000071525574, 'rewards/judge_quality/std': 0.16446885466575623, 'rewards/total_composite/mean': 0.5686860680580139, 'rewards/total_composite/std': 0.18503132462501526, 'reward': 0.5686860680580139, 'reward_std': 0.18503132462501526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2160700559616089, 'sampling/sampling_logp_difference/max': 1.0445046424865723, 'sampling/importance_sampling_ratio/min': 0.35186606645584106, 'sampling/importance_sampling_ratio/mean': 1.014116883277893, 'sampling/importance_sampling_ratio/max': 1.9738162755966187, 'entropy': 2.671849325299263, 'clip_ratio/low_mean': 0.04806985380128026, 'clip_ratio/low_min': 0.04806985380128026, 'clip_ratio/high_mean': 0.1266771275550127, 'clip_ratio/high_max': 0.1266771275550127, 'clip_ratio/region_mean': 0.17474698135629296, 'reward_total_mean': 0.5686860680580139, 'reward_meter_mean': 0.6177049875259399, 'reward_meter_std': 0.4753933548927307, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9828039407730103, 'reward_repeat_soft_std': 0.04082963243126869, 'reward_judge_quality_mean': 0.4925000071525574, 'reward_judge_quality_std': 0.16446885466575623, 'reward_total_composite_mean': 0.5686860680580139, 'reward_total_composite_std': 0.18503132462501526, 'epoch': 0.03}

 10%|▉         | 317/3300 [46:27<6:06:46,  7.38s/it]INFO 04-13 08:07:25 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 318/3300 [46:36<6:23:41,  7.72s/it]2026-04-13 08:07:34,083 | INFO | train_grpo_train | metrics_logged mode=train step=318

                                                    
{'loss': 0.1333, 'grad_norm': 16.658180236816406, 'learning_rate': 9.03939393939394e-06, 'num_tokens': 584484.0, 'completions/mean_length': 68.625, 'completions/min_length': 57.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5067485570907593, 'rewards/meter/std': 0.3414824604988098, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948186874389648, 'rewards/repeat_soft/std': 0.003807717002928257, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.478193461894989, 'rewards/total_composite/std': 0.09166455268859863, 'reward': 0.478193461894989, 'reward_std': 0.09166455268859863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25587788224220276, 'sampling/sampling_logp_difference/max': 2.545060157775879, 'sampling/importance_sampling_ratio/min': 0.07846833020448685, 'sampling/importance_sampling_ratio/mean': 1.0415856838226318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1368319243192673, 'clip_ratio/low_mean': 0.1427593883126974, 'clip_ratio/low_min': 0.1427593883126974, 'clip_ratio/high_mean': 0.09945355169475079, 'clip_ratio/high_max': 0.09945355169475079, 'clip_ratio/region_mean': 0.2422129400074482, 'reward_total_mean': 0.478193461894989, 'reward_meter_mean': 0.5067485570907593, 'reward_meter_std': 0.3414824604988098, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948186874389648, 'reward_repeat_soft_std': 0.003807717002928257, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.478193461894989, 'reward_total_composite_std': 0.09166455268859863, 'epoch': 0.03}

 10%|▉         | 318/3300 [46:36<6:23:41,  7.72s/it]INFO 04-13 08:07:34 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 319/3300 [46:43<6:12:35,  7.50s/it]2026-04-13 08:07:41,100 | INFO | train_grpo_train | metrics_logged mode=train step=319

                                                    
{'loss': -0.0281, 'grad_norm': 18.42127799987793, 'learning_rate': 9.036363636363638e-06, 'num_tokens': 585887.0, 'completions/mean_length': 25.375, 'completions/min_length': 19.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7930670976638794, 'rewards/meter/std': 0.3408731520175934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.11501552164554596, 'rewards/total_composite/mean': 0.532849133014679, 'rewards/total_composite/std': 0.1017293781042099, 'reward': 0.532849133014679, 'reward_std': 0.1017293632030487, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20987407863140106, 'sampling/sampling_logp_difference/max': 1.3737754821777344, 'sampling/importance_sampling_ratio/min': 0.2531493902206421, 'sampling/importance_sampling_ratio/mean': 1.0605309009552002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9036002308130264, 'clip_ratio/low_mean': 0.060653116554021835, 'clip_ratio/low_min': 0.060653116554021835, 'clip_ratio/high_mean': 0.06566179590299726, 'clip_ratio/high_max': 0.06566179590299726, 'clip_ratio/region_mean': 0.1263149124570191, 'reward_total_mean': 0.532849133014679, 'reward_meter_mean': 0.7930670976638794, 'reward_meter_std': 0.3408731520175934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.11501552164554596, 'reward_total_composite_mean': 0.532849133014679, 'reward_total_composite_std': 0.1017293781042099, 'epoch': 0.03}

 10%|▉         | 319/3300 [46:43<6:12:35,  7.50s/it]INFO 04-13 08:07:41 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 320/3300 [46:50<6:12:57,  7.51s/it]2026-04-13 08:07:48,604 | INFO | train_grpo_train | metrics_logged mode=train step=320

                                                    
{'loss': 0.045, 'grad_norm': 23.47793960571289, 'learning_rate': 9.033333333333334e-06, 'num_tokens': 587425.0, 'completions/mean_length': 30.25, 'completions/min_length': 24.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7791892886161804, 'rewards/meter/std': 0.29988420009613037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9933795928955078, 'rewards/repeat_soft/std': 0.008616972714662552, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5498079657554626, 'rewards/total_composite/std': 0.08107728511095047, 'reward': 0.5498079657554626, 'reward_std': 0.08107728511095047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1847909539937973, 'sampling/sampling_logp_difference/max': 2.1464381217956543, 'sampling/importance_sampling_ratio/min': 0.3022194504737854, 'sampling/importance_sampling_ratio/mean': 1.0040720701217651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9304276630282402, 'clip_ratio/low_mean': 0.09081439627334476, 'clip_ratio/low_min': 0.09081439627334476, 'clip_ratio/high_mean': 0.07184555940330029, 'clip_ratio/high_max': 0.07184555940330029, 'clip_ratio/region_mean': 0.16265995567664504, 'reward_total_mean': 0.5498079657554626, 'reward_meter_mean': 0.7791892886161804, 'reward_meter_std': 0.29988420009613037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9933795928955078, 'reward_repeat_soft_std': 0.008616972714662552, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5498079657554626, 'reward_total_composite_std': 0.08107728511095047, 'epoch': 0.03}

 10%|▉         | 320/3300 [46:50<6:12:57,  7.51s/it]INFO 04-13 08:07:48 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 321/3300 [46:58<6:20:12,  7.66s/it]2026-04-13 08:07:56,603 | INFO | train_grpo_train | metrics_logged mode=train step=321

                                                    
{'loss': 0.0122, 'grad_norm': 11.308182716369629, 'learning_rate': 9.030303030303031e-06, 'num_tokens': 589232.0, 'completions/mean_length': 63.875, 'completions/min_length': 56.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.6346712112426758, 'rewards/meter/std': 0.3175996243953705, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9911579489707947, 'rewards/repeat_soft/std': 0.0027391016483306885, 'rewards/judge_quality/mean': 0.6825000047683716, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.6405850648880005, 'rewards/total_composite/std': 0.1904749721288681, 'reward': 0.6405850648880005, 'reward_std': 0.1904749721288681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13196319341659546, 'sampling/sampling_logp_difference/max': 2.581531047821045, 'sampling/importance_sampling_ratio/min': 0.07565807551145554, 'sampling/importance_sampling_ratio/mean': 1.0030826330184937, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5553000271320343, 'clip_ratio/low_mean': 0.07165405922569335, 'clip_ratio/low_min': 0.07165405922569335, 'clip_ratio/high_mean': 0.0394513588398695, 'clip_ratio/high_max': 0.0394513588398695, 'clip_ratio/region_mean': 0.11110541806556284, 'reward_total_mean': 0.6405850648880005, 'reward_meter_mean': 0.6346712112426758, 'reward_meter_std': 0.3175996243953705, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9911579489707947, 'reward_repeat_soft_std': 0.0027391016483306885, 'reward_judge_quality_mean': 0.6825000047683716, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.6405850648880005, 'reward_total_composite_std': 0.1904749721288681, 'epoch': 0.03}

 10%|▉         | 321/3300 [46:58<6:20:12,  7.66s/it]INFO 04-13 08:07:56 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 322/3300 [47:06<6:20:56,  7.68s/it]2026-04-13 08:08:04,351 | INFO | train_grpo_train | metrics_logged mode=train step=322

                                                    
{'loss': 0.1136, 'grad_norm': 10.704634666442871, 'learning_rate': 9.027272727272728e-06, 'num_tokens': 591174.0, 'completions/mean_length': 79.75, 'completions/min_length': 67.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.75, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.24984392523765564, 'rewards/meter/std': 0.15510807931423187, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9884793758392334, 'rewards/repeat_soft/std': 0.009767917916178703, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.41147929430007935, 'rewards/total_composite/std': 0.047434739768505096, 'reward': 0.41147929430007935, 'reward_std': 0.047434743493795395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24083074927330017, 'sampling/sampling_logp_difference/max': 1.3724861145019531, 'sampling/importance_sampling_ratio/min': 0.25347599387168884, 'sampling/importance_sampling_ratio/mean': 1.0627409219741821, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.4511457979679108, 'clip_ratio/low_mean': 0.11510831117630005, 'clip_ratio/low_min': 0.11510831117630005, 'clip_ratio/high_mean': 0.10612666979432106, 'clip_ratio/high_max': 0.10612666979432106, 'clip_ratio/region_mean': 0.2212349809706211, 'reward_total_mean': 0.41147929430007935, 'reward_meter_mean': 0.24984392523765564, 'reward_meter_std': 0.15510807931423187, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9884793758392334, 'reward_repeat_soft_std': 0.009767917916178703, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.41147929430007935, 'reward_total_composite_std': 0.047434739768505096, 'epoch': 0.03}

 10%|▉         | 322/3300 [47:06<6:20:56,  7.68s/it]INFO 04-13 08:08:04 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 323/3300 [47:19<7:34:52,  9.17s/it]2026-04-13 08:08:16,980 | INFO | train_grpo_train | metrics_logged mode=train step=323

                                                    
{'loss': -0.0703, 'grad_norm': 3.8716347217559814, 'learning_rate': 9.024242424242426e-06, 'num_tokens': 592750.0, 'completions/mean_length': 94.0, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.28571701049805, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5991547703742981, 'rewards/meter/std': 0.3987487852573395, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9954737424850464, 'rewards/repeat_soft/std': 0.006524989847093821, 'rewards/judge_quality/mean': 0.44749999046325684, 'rewards/judge_quality/std': 0.23407875001430511, 'rewards/total_composite/mean': 0.44736433029174805, 'rewards/total_composite/std': 0.22696553170681, 'reward': 0.44736433029174805, 'reward_std': 0.22696553170681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22250120341777802, 'sampling/sampling_logp_difference/max': 1.0450553894042969, 'sampling/importance_sampling_ratio/min': 0.35167235136032104, 'sampling/importance_sampling_ratio/mean': 1.0371218919754028, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9466101080179214, 'clip_ratio/low_mean': 0.04711751639842987, 'clip_ratio/low_min': 0.04711751639842987, 'clip_ratio/high_mean': 0.14875785633921623, 'clip_ratio/high_max': 0.14875785633921623, 'clip_ratio/region_mean': 0.1958753727376461, 'reward_total_mean': 0.44736433029174805, 'reward_meter_mean': 0.5991547703742981, 'reward_meter_std': 0.3987487852573395, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9954737424850464, 'reward_repeat_soft_std': 0.006524989847093821, 'reward_judge_quality_mean': 0.44749999046325684, 'reward_judge_quality_std': 0.23407875001430511, 'reward_total_composite_mean': 0.44736433029174805, 'reward_total_composite_std': 0.22696553170681, 'epoch': 0.03}

 10%|▉         | 323/3300 [47:19<7:34:52,  9.17s/it]INFO 04-13 08:08:17 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 324/3300 [47:26<7:12:56,  8.73s/it]2026-04-13 08:08:24,674 | INFO | train_grpo_train | metrics_logged mode=train step=324

                                                    
{'loss': 0.075, 'grad_norm': 13.876107215881348, 'learning_rate': 9.021212121212121e-06, 'num_tokens': 594402.0, 'completions/mean_length': 46.5, 'completions/min_length': 41.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5066701173782349, 'rewards/meter/std': 0.32948172092437744, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9983240365982056, 'rewards/repeat_soft/std': 0.0026988540776073933, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4880695343017578, 'rewards/total_composite/std': 0.08996383100748062, 'reward': 0.4880695343017578, 'reward_std': 0.08996384590864182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2152794450521469, 'sampling/sampling_logp_difference/max': 1.6851024627685547, 'sampling/importance_sampling_ratio/min': 0.18542543053627014, 'sampling/importance_sampling_ratio/mean': 1.0551681518554688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.030674993991852, 'clip_ratio/low_mean': 0.10947934351861477, 'clip_ratio/low_min': 0.10947934351861477, 'clip_ratio/high_mean': 0.08234127052128315, 'clip_ratio/high_max': 0.08234127052128315, 'clip_ratio/region_mean': 0.19182061403989792, 'reward_total_mean': 0.4880695343017578, 'reward_meter_mean': 0.5066701173782349, 'reward_meter_std': 0.32948172092437744, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9983240365982056, 'reward_repeat_soft_std': 0.0026988540776073933, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4880695343017578, 'reward_total_composite_std': 0.08996383100748062, 'epoch': 0.03}

 10%|▉         | 324/3300 [47:26<7:12:56,  8.73s/it]INFO 04-13 08:08:24 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 325/3300 [47:34<6:58:41,  8.44s/it]2026-04-13 08:08:32,499 | INFO | train_grpo_train | metrics_logged mode=train step=325

                                                    
{'loss': 0.0071, 'grad_norm': 16.47154426574707, 'learning_rate': 9.01818181818182e-06, 'num_tokens': 596013.0, 'completions/mean_length': 38.375, 'completions/min_length': 33.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7837358713150024, 'rewards/meter/std': 0.24594815075397491, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950377941131592, 'rewards/repeat_soft/std': 0.006041733082383871, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.220891073346138, 'rewards/total_composite/mean': 0.6733725666999817, 'rewards/total_composite/std': 0.19483089447021484, 'reward': 0.6733725666999817, 'reward_std': 0.19483089447021484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20909790694713593, 'sampling/sampling_logp_difference/max': 1.372457504272461, 'sampling/importance_sampling_ratio/min': 0.25348323583602905, 'sampling/importance_sampling_ratio/mean': 1.057938575744629, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6385374814271927, 'clip_ratio/low_mean': 0.11629138700664043, 'clip_ratio/low_min': 0.11629138700664043, 'clip_ratio/high_mean': 0.05888960137963295, 'clip_ratio/high_max': 0.05888960137963295, 'clip_ratio/region_mean': 0.17518098838627338, 'reward_total_mean': 0.6733725666999817, 'reward_meter_mean': 0.7837358713150024, 'reward_meter_std': 0.24594815075397491, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950377941131592, 'reward_repeat_soft_std': 0.006041733082383871, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.220891073346138, 'reward_total_composite_mean': 0.6733725666999817, 'reward_total_composite_std': 0.19483089447021484, 'epoch': 0.03}

 10%|▉         | 325/3300 [47:34<6:58:41,  8.44s/it]INFO 04-13 08:08:32 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 326/3300 [47:41<6:41:49,  8.11s/it]2026-04-13 08:08:39,866 | INFO | train_grpo_train | metrics_logged mode=train step=326

                                                    
{'loss': 0.0723, 'grad_norm': 15.22860336303711, 'learning_rate': 9.015151515151516e-06, 'num_tokens': 597468.0, 'completions/mean_length': 37.875, 'completions/min_length': 28.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7003387212753296, 'rewards/meter/std': 0.3372787833213806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9941596388816833, 'rewards/repeat_soft/std': 0.008668807335197926, 'rewards/judge_quality/mean': 0.6487500071525574, 'rewards/judge_quality/std': 0.2507951855659485, 'rewards/total_composite/mean': 0.6733263731002808, 'rewards/total_composite/std': 0.22199413180351257, 'reward': 0.6733263731002808, 'reward_std': 0.22199410200119019, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1748155802488327, 'sampling/sampling_logp_difference/max': 1.2661004066467285, 'sampling/importance_sampling_ratio/min': 0.2819288969039917, 'sampling/importance_sampling_ratio/mean': 1.0314687490463257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3604123145341873, 'clip_ratio/low_mean': 0.11356301605701447, 'clip_ratio/low_min': 0.11356301605701447, 'clip_ratio/high_mean': 0.0538545292802155, 'clip_ratio/high_max': 0.0538545292802155, 'clip_ratio/region_mean': 0.16741754533722997, 'reward_total_mean': 0.6733263731002808, 'reward_meter_mean': 0.7003387212753296, 'reward_meter_std': 0.3372787833213806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9941596388816833, 'reward_repeat_soft_std': 0.008668807335197926, 'reward_judge_quality_mean': 0.6487500071525574, 'reward_judge_quality_std': 0.2507951855659485, 'reward_total_composite_mean': 0.6733263731002808, 'reward_total_composite_std': 0.22199413180351257, 'epoch': 0.03}

 10%|▉         | 326/3300 [47:42<6:41:49,  8.11s/it]INFO 04-13 08:08:40 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 327/3300 [47:49<6:31:46,  7.91s/it]2026-04-13 08:08:47,208 | INFO | train_grpo_train | metrics_logged mode=train step=327

                                                    
{'loss': 0.1348, 'grad_norm': 21.305774688720703, 'learning_rate': 9.012121212121213e-06, 'num_tokens': 598802.0, 'completions/mean_length': 30.75, 'completions/min_length': 22.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8316705226898193, 'rewards/meter/std': 0.2706790566444397, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9990543127059937, 'rewards/repeat_soft/std': 0.001751757226884365, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6967610120773315, 'rewards/total_composite/std': 0.17174653708934784, 'reward': 0.6967610120773315, 'reward_std': 0.17174653708934784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22484582662582397, 'sampling/sampling_logp_difference/max': 1.5765159130096436, 'sampling/importance_sampling_ratio/min': 0.20669399201869965, 'sampling/importance_sampling_ratio/mean': 1.0373828411102295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4714982882142067, 'clip_ratio/low_mean': 0.10420633293688297, 'clip_ratio/low_min': 0.10420633293688297, 'clip_ratio/high_mean': 0.07759773265570402, 'clip_ratio/high_max': 0.07759773265570402, 'clip_ratio/region_mean': 0.181804065592587, 'reward_total_mean': 0.6967610120773315, 'reward_meter_mean': 0.8316705226898193, 'reward_meter_std': 0.2706790566444397, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9990543127059937, 'reward_repeat_soft_std': 0.001751757226884365, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6967610120773315, 'reward_total_composite_std': 0.17174653708934784, 'epoch': 0.03}

 10%|▉         | 327/3300 [47:49<6:31:46,  7.91s/it]INFO 04-13 08:08:47 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 328/3300 [47:57<6:31:03,  7.89s/it]2026-04-13 08:08:55,079 | INFO | train_grpo_train | metrics_logged mode=train step=328

                                                    
{'loss': -0.1995, 'grad_norm': 25.675718307495117, 'learning_rate': 9.00909090909091e-06, 'num_tokens': 600250.0, 'completions/mean_length': 22.0, 'completions/min_length': 11.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 11.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.5685776472091675, 'rewards/meter/std': 0.465649276971817, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9718749523162842, 'rewards/repeat_soft/std': 0.017359137535095215, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.5208595395088196, 'rewards/total_composite/std': 0.3706021010875702, 'reward': 0.5208595395088196, 'reward_std': 0.3706021010875702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1883666068315506, 'sampling/sampling_logp_difference/max': 1.1565308570861816, 'sampling/importance_sampling_ratio/min': 0.31457558274269104, 'sampling/importance_sampling_ratio/mean': 1.034841537475586, 'sampling/importance_sampling_ratio/max': 1.9202556610107422, 'entropy': 2.0162010490894318, 'clip_ratio/low_mean': 0.02864583395421505, 'clip_ratio/low_min': 0.02864583395421505, 'clip_ratio/high_mean': 0.1330465618520975, 'clip_ratio/high_max': 0.1330465618520975, 'clip_ratio/region_mean': 0.16169239580631256, 'reward_total_mean': 0.5208595395088196, 'reward_meter_mean': 0.5685776472091675, 'reward_meter_std': 0.465649276971817, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9718749523162842, 'reward_repeat_soft_std': 0.017359137535095215, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.5208595395088196, 'reward_total_composite_std': 0.3706021010875702, 'epoch': 0.03}

 10%|▉         | 328/3300 [47:57<6:31:03,  7.89s/it]INFO 04-13 08:08:55 [block_pool.py:378] Successfully reset prefix cache

 10%|▉         | 329/3300 [48:04<6:21:56,  7.71s/it]2026-04-13 08:09:02,426 | INFO | train_grpo_train | metrics_logged mode=train step=329

                                                    
{'loss': 0.0671, 'grad_norm': 23.811079025268555, 'learning_rate': 9.006060606060607e-06, 'num_tokens': 601833.0, 'completions/mean_length': 22.875, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.3704826235771179, 'rewards/meter/std': 0.4448740482330322, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.3687499761581421, 'rewards/judge_quality/std': 0.10802611708641052, 'rewards/total_composite/mean': 0.4032096266746521, 'rewards/total_composite/std': 0.19858667254447937, 'reward': 0.4032096266746521, 'reward_std': 0.19858667254447937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2445867955684662, 'sampling/sampling_logp_difference/max': 1.0363855361938477, 'sampling/importance_sampling_ratio/min': 0.3547345697879791, 'sampling/importance_sampling_ratio/mean': 1.0885217189788818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.910661995410919, 'clip_ratio/low_mean': 0.11830841470509768, 'clip_ratio/low_min': 0.11830841470509768, 'clip_ratio/high_mean': 0.04718535486608744, 'clip_ratio/high_max': 0.04718535486608744, 'clip_ratio/region_mean': 0.1654937695711851, 'reward_total_mean': 0.4032096266746521, 'reward_meter_mean': 0.3704826235771179, 'reward_meter_std': 0.4448740482330322, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.3687499761581421, 'reward_judge_quality_std': 0.10802611708641052, 'reward_total_composite_mean': 0.4032096266746521, 'reward_total_composite_std': 0.19858667254447937, 'epoch': 0.03}

 10%|▉         | 329/3300 [48:04<6:21:56,  7.71s/it]INFO 04-13 08:09:02 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 330/3300 [48:12<6:22:38,  7.73s/it]2026-04-13 08:09:10,136 | INFO | train_grpo_train | metrics_logged mode=train step=330

                                                    
{'loss': 0.0256, 'grad_norm': 17.825525283813477, 'learning_rate': 9.003030303030303e-06, 'num_tokens': 603503.0, 'completions/mean_length': 38.75, 'completions/min_length': 35.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.33915719389915466, 'rewards/meter/std': 0.30123084783554077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9892999529838562, 'rewards/repeat_soft/std': 0.008622348308563232, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.4413272738456726, 'rewards/total_composite/std': 0.08239845186471939, 'reward': 0.4413272738456726, 'reward_std': 0.0823984369635582, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1875453144311905, 'sampling/sampling_logp_difference/max': 1.3236665725708008, 'sampling/importance_sampling_ratio/min': 0.2661576271057129, 'sampling/importance_sampling_ratio/mean': 1.0321362018585205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4364940524101257, 'clip_ratio/low_mean': 0.09573934972286224, 'clip_ratio/low_min': 0.09573934972286224, 'clip_ratio/high_mean': 0.08293084055185318, 'clip_ratio/high_max': 0.08293084055185318, 'clip_ratio/region_mean': 0.17867019027471542, 'reward_total_mean': 0.4413272738456726, 'reward_meter_mean': 0.33915719389915466, 'reward_meter_std': 0.30123084783554077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9892999529838562, 'reward_repeat_soft_std': 0.008622348308563232, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.4413272738456726, 'reward_total_composite_std': 0.08239845186471939, 'epoch': 0.03}

 10%|█         | 330/3300 [48:12<6:22:38,  7.73s/it]INFO 04-13 08:09:10 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 331/3300 [48:19<6:16:40,  7.61s/it]2026-04-13 08:09:17,472 | INFO | train_grpo_train | metrics_logged mode=train step=331

                                                    
{'loss': 0.0365, 'grad_norm': 19.763568878173828, 'learning_rate': 9e-06, 'num_tokens': 604977.0, 'completions/mean_length': 28.25, 'completions/min_length': 23.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.30294397473335266, 'rewards/meter/std': 0.34484946727752686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9890438318252563, 'rewards/repeat_soft/std': 0.022879350930452347, 'rewards/judge_quality/mean': 0.4675000011920929, 'rewards/judge_quality/std': 0.21022097766399384, 'rewards/total_composite/mean': 0.4555625915527344, 'rewards/total_composite/std': 0.13149835169315338, 'reward': 0.4555625915527344, 'reward_std': 0.1314983367919922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24701382219791412, 'sampling/sampling_logp_difference/max': 2.6028099060058594, 'sampling/importance_sampling_ratio/min': 0.07406517118215561, 'sampling/importance_sampling_ratio/mean': 1.0688800811767578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4246515184640884, 'clip_ratio/low_mean': 0.11589226126670837, 'clip_ratio/low_min': 0.11589226126670837, 'clip_ratio/high_mean': 0.078125, 'clip_ratio/high_max': 0.078125, 'clip_ratio/region_mean': 0.19401726126670837, 'reward_total_mean': 0.4555625915527344, 'reward_meter_mean': 0.30294397473335266, 'reward_meter_std': 0.34484946727752686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9890438318252563, 'reward_repeat_soft_std': 0.022879350930452347, 'reward_judge_quality_mean': 0.4675000011920929, 'reward_judge_quality_std': 0.21022097766399384, 'reward_total_composite_mean': 0.4555625915527344, 'reward_total_composite_std': 0.13149835169315338, 'epoch': 0.03}

 10%|█         | 331/3300 [48:19<6:16:40,  7.61s/it]INFO 04-13 08:09:17 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 332/3300 [48:26<6:05:58,  7.40s/it]2026-04-13 08:09:24,382 | INFO | train_grpo_train | metrics_logged mode=train step=332

                                                    
{'loss': 0.054, 'grad_norm': 14.886153221130371, 'learning_rate': 8.996969696969697e-06, 'num_tokens': 606595.0, 'completions/mean_length': 34.25, 'completions/min_length': 28.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.4881669580936432, 'rewards/meter/std': 0.41594117879867554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9882906079292297, 'rewards/repeat_soft/std': 0.011169468984007835, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5594117641448975, 'rewards/total_composite/std': 0.23872269690036774, 'reward': 0.5594117641448975, 'reward_std': 0.23872269690036774, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18812359869480133, 'sampling/sampling_logp_difference/max': 1.2091703414916992, 'sampling/importance_sampling_ratio/min': 0.2984447777271271, 'sampling/importance_sampling_ratio/mean': 1.0254172086715698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7859932333230972, 'clip_ratio/low_mean': 0.08856897242367268, 'clip_ratio/low_min': 0.08856897242367268, 'clip_ratio/high_mean': 0.05078725144267082, 'clip_ratio/high_max': 0.05078725144267082, 'clip_ratio/region_mean': 0.1393562238663435, 'reward_total_mean': 0.5594117641448975, 'reward_meter_mean': 0.4881669580936432, 'reward_meter_std': 0.41594117879867554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9882906079292297, 'reward_repeat_soft_std': 0.011169468984007835, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5594117641448975, 'reward_total_composite_std': 0.23872269690036774, 'epoch': 0.03}

 10%|█         | 332/3300 [48:26<6:05:58,  7.40s/it]INFO 04-13 08:09:24 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 333/3300 [48:40<7:37:11,  9.25s/it]2026-04-13 08:09:38,081 | INFO | train_grpo_train | metrics_logged mode=train step=333

                                                    
{'loss': -0.064, 'grad_norm': 2.7316510677337646, 'learning_rate': 8.993939393939395e-06, 'num_tokens': 608741.0, 'completions/mean_length': 210.25, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 109.66667175292969, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.799657940864563, 'rewards/meter/std': 0.20600946247577667, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9958186149597168, 'rewards/repeat_soft/std': 0.002802198752760887, 'rewards/judge_quality/mean': 0.21875, 'rewards/judge_quality/std': 0.17016273736953735, 'rewards/total_composite/mean': 0.333798885345459, 'rewards/total_composite/std': 0.2898932695388794, 'reward': 0.333798885345459, 'reward_std': 0.2898932695388794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22127531468868256, 'sampling/sampling_logp_difference/max': 1.5819416046142578, 'sampling/importance_sampling_ratio/min': 0.20557557046413422, 'sampling/importance_sampling_ratio/mean': 1.0461658239364624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.653821587562561, 'clip_ratio/low_mean': 0.013698630034923553, 'clip_ratio/low_min': 0.013698630034923553, 'clip_ratio/high_mean': 0.11938321124762297, 'clip_ratio/high_max': 0.11938321124762297, 'clip_ratio/region_mean': 0.13308184128254652, 'reward_total_mean': 0.333798885345459, 'reward_meter_mean': 0.799657940864563, 'reward_meter_std': 0.20600946247577667, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9958186149597168, 'reward_repeat_soft_std': 0.002802198752760887, 'reward_judge_quality_mean': 0.21875, 'reward_judge_quality_std': 0.17016273736953735, 'reward_total_composite_mean': 0.333798885345459, 'reward_total_composite_std': 0.2898932695388794, 'epoch': 0.03}

 10%|█         | 333/3300 [48:40<7:37:11,  9.25s/it]INFO 04-13 08:09:38 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 334/3300 [48:47<7:13:54,  8.78s/it]2026-04-13 08:09:45,618 | INFO | train_grpo_train | metrics_logged mode=train step=334

                                                    
{'loss': 0.0524, 'grad_norm': 13.678813934326172, 'learning_rate': 8.990909090909092e-06, 'num_tokens': 610686.0, 'completions/mean_length': 74.125, 'completions/min_length': 58.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.125, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.3780759572982788, 'rewards/meter/std': 0.31160059571266174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9950718879699707, 'rewards/repeat_soft/std': 0.006842370145022869, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.3830219507217407, 'rewards/total_composite/std': 0.17284251749515533, 'reward': 0.3830219507217407, 'reward_std': 0.17284250259399414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23534409701824188, 'sampling/sampling_logp_difference/max': 1.1867132186889648, 'sampling/importance_sampling_ratio/min': 0.3052228093147278, 'sampling/importance_sampling_ratio/mean': 1.060614824295044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.6963296830654144, 'clip_ratio/low_mean': 0.061289879493415356, 'clip_ratio/low_min': 0.061289879493415356, 'clip_ratio/high_mean': 0.11982089839875698, 'clip_ratio/high_max': 0.11982089839875698, 'clip_ratio/region_mean': 0.18111077789217234, 'reward_total_mean': 0.3830219507217407, 'reward_meter_mean': 0.3780759572982788, 'reward_meter_std': 0.31160059571266174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9950718879699707, 'reward_repeat_soft_std': 0.006842370145022869, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.3830219507217407, 'reward_total_composite_std': 0.17284251749515533, 'epoch': 0.03}

 10%|█         | 334/3300 [48:47<7:13:54,  8.78s/it]INFO 04-13 08:09:45 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 335/3300 [48:54<6:48:05,  8.26s/it]2026-04-13 08:09:52,665 | INFO | train_grpo_train | metrics_logged mode=train step=335

                                                    
{'loss': 0.0727, 'grad_norm': 15.958833694458008, 'learning_rate': 8.98787878787879e-06, 'num_tokens': 612672.0, 'completions/mean_length': 63.25, 'completions/min_length': 57.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.25, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.35783955454826355, 'rewards/meter/std': 0.2124491184949875, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9926632642745972, 'rewards/repeat_soft/std': 0.005155050661414862, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.46032220125198364, 'rewards/total_composite/std': 0.06749068945646286, 'reward': 0.46032220125198364, 'reward_std': 0.06749069690704346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19007080793380737, 'sampling/sampling_logp_difference/max': 1.824625849723816, 'sampling/importance_sampling_ratio/min': 0.16127797961235046, 'sampling/importance_sampling_ratio/mean': 1.0112042427062988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5455197468400002, 'clip_ratio/low_mean': 0.044313523918390274, 'clip_ratio/low_min': 0.044313523918390274, 'clip_ratio/high_mean': 0.13582287728786469, 'clip_ratio/high_max': 0.13582287728786469, 'clip_ratio/region_mean': 0.18013640120625496, 'reward_total_mean': 0.46032220125198364, 'reward_meter_mean': 0.35783955454826355, 'reward_meter_std': 0.2124491184949875, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9926632642745972, 'reward_repeat_soft_std': 0.005155050661414862, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.46032220125198364, 'reward_total_composite_std': 0.06749068945646286, 'epoch': 0.03}

 10%|█         | 335/3300 [48:54<6:48:05,  8.26s/it]INFO 04-13 08:09:52 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 336/3300 [49:02<6:33:18,  7.96s/it]2026-04-13 08:09:59,950 | INFO | train_grpo_train | metrics_logged mode=train step=336

                                                    
{'loss': 0.0209, 'grad_norm': 19.562091827392578, 'learning_rate': 8.984848484848485e-06, 'num_tokens': 614193.0, 'completions/mean_length': 34.125, 'completions/min_length': 25.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5130771398544312, 'rewards/meter/std': 0.39404621720314026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9882612228393555, 'rewards/repeat_soft/std': 0.009143383242189884, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.6394940614700317, 'rewards/total_composite/std': 0.2444366216659546, 'reward': 0.6394940614700317, 'reward_std': 0.2444366216659546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1709323227405548, 'sampling/sampling_logp_difference/max': 3.006063938140869, 'sampling/importance_sampling_ratio/min': 0.04948607459664345, 'sampling/importance_sampling_ratio/mean': 1.0001240968704224, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0319463834166527, 'clip_ratio/low_mean': 0.09706691186875105, 'clip_ratio/low_min': 0.09706691186875105, 'clip_ratio/high_mean': 0.08584709116257727, 'clip_ratio/high_max': 0.08584709116257727, 'clip_ratio/region_mean': 0.18291400303132832, 'reward_total_mean': 0.6394940614700317, 'reward_meter_mean': 0.5130771398544312, 'reward_meter_std': 0.39404621720314026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9882612228393555, 'reward_repeat_soft_std': 0.009143383242189884, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.6394940614700317, 'reward_total_composite_std': 0.2444366216659546, 'epoch': 0.03}

 10%|█         | 336/3300 [49:02<6:33:18,  7.96s/it]INFO 04-13 08:10:00 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 337/3300 [49:09<6:30:50,  7.91s/it]2026-04-13 08:10:07,757 | INFO | train_grpo_train | metrics_logged mode=train step=337

                                                    
{'loss': 0.018, 'grad_norm': 11.302577018737793, 'learning_rate': 8.981818181818182e-06, 'num_tokens': 616565.0, 'completions/mean_length': 96.5, 'completions/min_length': 83.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.22483281791210175, 'rewards/meter/std': 0.31735825538635254, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9898207187652588, 'rewards/repeat_soft/std': 0.006588577758520842, 'rewards/judge_quality/mean': 0.3725000023841858, 'rewards/judge_quality/std': 0.16368524730205536, 'rewards/total_composite/mean': 0.3557148277759552, 'rewards/total_composite/std': 0.2066221833229065, 'reward': 0.3557148277759552, 'reward_std': 0.2066221833229065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23147423565387726, 'sampling/sampling_logp_difference/max': 1.458292007446289, 'sampling/importance_sampling_ratio/min': 0.23263327777385712, 'sampling/importance_sampling_ratio/mean': 1.0710952281951904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.368072062730789, 'clip_ratio/low_mean': 0.10687562357634306, 'clip_ratio/low_min': 0.10687562357634306, 'clip_ratio/high_mean': 0.07749764993786812, 'clip_ratio/high_max': 0.07749764993786812, 'clip_ratio/region_mean': 0.18437327351421118, 'reward_total_mean': 0.3557148277759552, 'reward_meter_mean': 0.22483281791210175, 'reward_meter_std': 0.31735825538635254, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9898207187652588, 'reward_repeat_soft_std': 0.006588577758520842, 'reward_judge_quality_mean': 0.3725000023841858, 'reward_judge_quality_std': 0.16368524730205536, 'reward_total_composite_mean': 0.3557148277759552, 'reward_total_composite_std': 0.2066221833229065, 'epoch': 0.03}

 10%|█         | 337/3300 [49:09<6:30:50,  7.91s/it]INFO 04-13 08:10:08 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 338/3300 [49:22<7:41:35,  9.35s/it]2026-04-13 08:10:20,444 | INFO | train_grpo_train | metrics_logged mode=train step=338

                                                    
{'loss': -0.1492, 'grad_norm': 2.613356113433838, 'learning_rate': 8.97878787878788e-06, 'num_tokens': 618287.0, 'completions/mean_length': 108.25, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.57143020629883, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8069460391998291, 'rewards/meter/std': 0.3264923691749573, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9952407479286194, 'rewards/repeat_soft/std': 0.004081881605088711, 'rewards/judge_quality/mean': 0.33124998211860657, 'rewards/judge_quality/std': 0.13715866208076477, 'rewards/total_composite/mean': 0.498996376991272, 'rewards/total_composite/std': 0.2086392194032669, 'reward': 0.498996376991272, 'reward_std': 0.2086392045021057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23468974232673645, 'sampling/sampling_logp_difference/max': 1.2172784805297852, 'sampling/importance_sampling_ratio/min': 0.29603472352027893, 'sampling/importance_sampling_ratio/mean': 1.0711236000061035, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9825278222560883, 'clip_ratio/low_mean': 0.014705882407724857, 'clip_ratio/low_min': 0.014705882407724857, 'clip_ratio/high_mean': 0.14897500537335873, 'clip_ratio/high_max': 0.14897500537335873, 'clip_ratio/region_mean': 0.16368088778108358, 'reward_total_mean': 0.498996376991272, 'reward_meter_mean': 0.8069460391998291, 'reward_meter_std': 0.3264923691749573, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9952407479286194, 'reward_repeat_soft_std': 0.004081881605088711, 'reward_judge_quality_mean': 0.33124998211860657, 'reward_judge_quality_std': 0.13715866208076477, 'reward_total_composite_mean': 0.498996376991272, 'reward_total_composite_std': 0.2086392194032669, 'epoch': 0.03}

 10%|█         | 338/3300 [49:22<7:41:35,  9.35s/it]INFO 04-13 08:10:20 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 339/3300 [49:31<7:39:28,  9.31s/it]2026-04-13 08:10:29,683 | INFO | train_grpo_train | metrics_logged mode=train step=339

                                                    
{'loss': 0.4037, 'grad_norm': 11.293525695800781, 'learning_rate': 8.975757575757577e-06, 'num_tokens': 620037.0, 'completions/mean_length': 58.75, 'completions/min_length': 32.0, 'completions/max_length': 233.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 233.0, 'rewards/meter/mean': 0.47527629137039185, 'rewards/meter/std': 0.39625003933906555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9971331357955933, 'rewards/repeat_soft/std': 0.004550741985440254, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.460955947637558, 'rewards/total_composite/std': 0.1209416389465332, 'reward': 0.460955947637558, 'reward_std': 0.1209416389465332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20768223702907562, 'sampling/sampling_logp_difference/max': 1.2325115203857422, 'sampling/importance_sampling_ratio/min': 0.2915593981742859, 'sampling/importance_sampling_ratio/mean': 1.0328776836395264, 'sampling/importance_sampling_ratio/max': 1.9464133977890015, 'entropy': 3.656894564628601, 'clip_ratio/low_mean': 0.09924108954146504, 'clip_ratio/low_min': 0.09924108954146504, 'clip_ratio/high_mean': 0.0744554940611124, 'clip_ratio/high_max': 0.0744554940611124, 'clip_ratio/region_mean': 0.17369658360257745, 'reward_total_mean': 0.460955947637558, 'reward_meter_mean': 0.47527629137039185, 'reward_meter_std': 0.39625003933906555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9971331357955933, 'reward_repeat_soft_std': 0.004550741985440254, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.460955947637558, 'reward_total_composite_std': 0.1209416389465332, 'epoch': 0.03}

 10%|█         | 339/3300 [49:31<7:39:28,  9.31s/it]INFO 04-13 08:10:29 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 340/3300 [49:39<7:13:47,  8.79s/it]2026-04-13 08:10:37,251 | INFO | train_grpo_train | metrics_logged mode=train step=340

                                                    
{'loss': 0.0406, 'grad_norm': 17.970109939575195, 'learning_rate': 8.972727272727272e-06, 'num_tokens': 621493.0, 'completions/mean_length': 33.0, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7923741340637207, 'rewards/meter/std': 0.34158453345298767, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9971886873245239, 'rewards/repeat_soft/std': 0.003208628622815013, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.2655452787876129, 'rewards/total_composite/mean': 0.6965111494064331, 'rewards/total_composite/std': 0.20136895775794983, 'reward': 0.6965111494064331, 'reward_std': 0.20136895775794983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16407504677772522, 'sampling/sampling_logp_difference/max': 1.1980657577514648, 'sampling/importance_sampling_ratio/min': 0.30177733302116394, 'sampling/importance_sampling_ratio/mean': 1.0395454168319702, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.998213142156601, 'clip_ratio/low_mean': 0.07138069625943899, 'clip_ratio/low_min': 0.07138069625943899, 'clip_ratio/high_mean': 0.0566600551828742, 'clip_ratio/high_max': 0.0566600551828742, 'clip_ratio/region_mean': 0.1280407514423132, 'reward_total_mean': 0.6965111494064331, 'reward_meter_mean': 0.7923741340637207, 'reward_meter_std': 0.34158453345298767, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9971886873245239, 'reward_repeat_soft_std': 0.003208628622815013, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.2655452787876129, 'reward_total_composite_mean': 0.6965111494064331, 'reward_total_composite_std': 0.20136895775794983, 'epoch': 0.03}

 10%|█         | 340/3300 [49:39<7:13:47,  8.79s/it]INFO 04-13 08:10:37 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 341/3300 [49:47<6:56:36,  8.45s/it]2026-04-13 08:10:44,895 | INFO | train_grpo_train | metrics_logged mode=train step=341

                                                    
{'loss': -0.2348, 'grad_norm': 13.813849449157715, 'learning_rate': 8.969696969696971e-06, 'num_tokens': 623318.0, 'completions/mean_length': 54.125, 'completions/min_length': 36.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.6467094421386719, 'rewards/meter/std': 0.24001756310462952, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9904347658157349, 'rewards/repeat_soft/std': 0.006750911008566618, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.516783595085144, 'rewards/total_composite/std': 0.06690096855163574, 'reward': 0.516783595085144, 'reward_std': 0.06690097600221634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2130843698978424, 'sampling/sampling_logp_difference/max': 0.9128122329711914, 'sampling/importance_sampling_ratio/min': 0.4013938307762146, 'sampling/importance_sampling_ratio/mean': 1.0654175281524658, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.926282674074173, 'clip_ratio/low_mean': 0.061325354501605034, 'clip_ratio/low_min': 0.061325354501605034, 'clip_ratio/high_mean': 0.1286101434379816, 'clip_ratio/high_max': 0.1286101434379816, 'clip_ratio/region_mean': 0.18993549793958664, 'reward_total_mean': 0.516783595085144, 'reward_meter_mean': 0.6467094421386719, 'reward_meter_std': 0.24001756310462952, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9904347658157349, 'reward_repeat_soft_std': 0.006750911008566618, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.516783595085144, 'reward_total_composite_std': 0.06690096855163574, 'epoch': 0.03}

 10%|█         | 341/3300 [49:47<6:56:36,  8.45s/it]INFO 04-13 08:10:45 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 342/3300 [49:58<7:46:18,  9.46s/it]2026-04-13 08:10:56,747 | INFO | train_grpo_train | metrics_logged mode=train step=342

                                                    
{'loss': 0.3122, 'grad_norm': 4.728224754333496, 'learning_rate': 8.966666666666667e-06, 'num_tokens': 626172.0, 'completions/mean_length': 176.75, 'completions/min_length': 66.0, 'completions/max_length': 488.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 176.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 488.0, 'rewards/meter/mean': 0.2066187709569931, 'rewards/meter/std': 0.3268250823020935, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973129034042358, 'rewards/repeat_soft/std': 0.002697502262890339, 'rewards/judge_quality/mean': 0.26375001668930054, 'rewards/judge_quality/std': 0.2464570552110672, 'rewards/total_composite/mean': 0.39993906021118164, 'rewards/total_composite/std': 0.17000128328800201, 'reward': 0.39993906021118164, 'reward_std': 0.1700012981891632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21022287011146545, 'sampling/sampling_logp_difference/max': 2.8367786407470703, 'sampling/importance_sampling_ratio/min': 0.058614179491996765, 'sampling/importance_sampling_ratio/mean': 1.0411263704299927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 5.513430565595627, 'clip_ratio/low_mean': 0.07977114617824554, 'clip_ratio/low_min': 0.07977114617824554, 'clip_ratio/high_mean': 0.04984848573803902, 'clip_ratio/high_max': 0.04984848573803902, 'clip_ratio/region_mean': 0.12961963191628456, 'reward_total_mean': 0.39993906021118164, 'reward_meter_mean': 0.2066187709569931, 'reward_meter_std': 0.3268250823020935, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973129034042358, 'reward_repeat_soft_std': 0.002697502262890339, 'reward_judge_quality_mean': 0.26375001668930054, 'reward_judge_quality_std': 0.2464570552110672, 'reward_total_composite_mean': 0.39993906021118164, 'reward_total_composite_std': 0.17000128328800201, 'epoch': 0.03}

 10%|█         | 342/3300 [49:58<7:46:18,  9.46s/it]INFO 04-13 08:10:56 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 343/3300 [50:07<7:27:56,  9.09s/it]2026-04-13 08:11:04,926 | INFO | train_grpo_train | metrics_logged mode=train step=343

                                                    
{'loss': 0.027, 'grad_norm': 24.89185905456543, 'learning_rate': 8.963636363636364e-06, 'num_tokens': 627570.0, 'completions/mean_length': 20.75, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.694798469543457, 'rewards/meter/std': 0.35779568552970886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9567934274673462, 'rewards/repeat_soft/std': 0.016140474006533623, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404788017273, 'rewards/total_composite/mean': 0.5756093263626099, 'rewards/total_composite/std': 0.15500064194202423, 'reward': 0.5756093263626099, 'reward_std': 0.15500065684318542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20464152097702026, 'sampling/sampling_logp_difference/max': 1.7080192565917969, 'sampling/importance_sampling_ratio/min': 0.18122439086437225, 'sampling/importance_sampling_ratio/mean': 1.0239838361740112, 'sampling/importance_sampling_ratio/max': 1.7914179563522339, 'entropy': 2.4688137769699097, 'clip_ratio/low_mean': 0.04615705972537398, 'clip_ratio/low_min': 0.04615705972537398, 'clip_ratio/high_mean': 0.1430090507492423, 'clip_ratio/high_max': 0.1430090507492423, 'clip_ratio/region_mean': 0.1891661104746163, 'reward_total_mean': 0.5756093263626099, 'reward_meter_mean': 0.694798469543457, 'reward_meter_std': 0.35779568552970886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9567934274673462, 'reward_repeat_soft_std': 0.016140474006533623, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404788017273, 'reward_total_composite_mean': 0.5756093263626099, 'reward_total_composite_std': 0.15500064194202423, 'epoch': 0.03}

 10%|█         | 343/3300 [50:07<7:27:56,  9.09s/it]INFO 04-13 08:11:05 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 344/3300 [50:18<7:56:52,  9.68s/it]2026-04-13 08:11:15,981 | INFO | train_grpo_train | metrics_logged mode=train step=344

                                                    
{'loss': 0.0324, 'grad_norm': 15.108419418334961, 'learning_rate': 8.960606060606061e-06, 'num_tokens': 629196.0, 'completions/mean_length': 38.25, 'completions/min_length': 29.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6917316317558289, 'rewards/meter/std': 0.3612944185733795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9986319541931152, 'rewards/repeat_soft/std': 0.0038694811519235373, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.22398583590984344, 'rewards/total_composite/mean': 0.6348282098770142, 'rewards/total_composite/std': 0.19019661843776703, 'reward': 0.6348282098770142, 'reward_std': 0.19019660353660583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21258258819580078, 'sampling/sampling_logp_difference/max': 1.3830804824829102, 'sampling/importance_sampling_ratio/min': 0.2508047819137573, 'sampling/importance_sampling_ratio/mean': 1.0417386293411255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1024929881095886, 'clip_ratio/low_mean': 0.10580164100974798, 'clip_ratio/low_min': 0.10580164100974798, 'clip_ratio/high_mean': 0.07892816513776779, 'clip_ratio/high_max': 0.07892816513776779, 'clip_ratio/region_mean': 0.18472980614751577, 'reward_total_mean': 0.6348282098770142, 'reward_meter_mean': 0.6917316317558289, 'reward_meter_std': 0.3612944185733795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9986319541931152, 'reward_repeat_soft_std': 0.0038694811519235373, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.22398583590984344, 'reward_total_composite_mean': 0.6348282098770142, 'reward_total_composite_std': 0.19019661843776703, 'epoch': 0.03}

 10%|█         | 344/3300 [50:18<7:56:52,  9.68s/it]INFO 04-13 08:11:16 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 345/3300 [50:25<7:23:42,  9.01s/it]2026-04-13 08:11:23,428 | INFO | train_grpo_train | metrics_logged mode=train step=345

                                                    
{'loss': 0.1219, 'grad_norm': 16.092153549194336, 'learning_rate': 8.957575757575758e-06, 'num_tokens': 630735.0, 'completions/mean_length': 35.375, 'completions/min_length': 24.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.402677983045578, 'rewards/meter/std': 0.3711997866630554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928983449935913, 'rewards/repeat_soft/std': 0.00860940758138895, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.1524970829486847, 'rewards/total_composite/mean': 0.44169631600379944, 'rewards/total_composite/std': 0.08157047629356384, 'reward': 0.44169631600379944, 'reward_std': 0.08157046884298325, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23123355209827423, 'sampling/sampling_logp_difference/max': 1.5154838562011719, 'sampling/importance_sampling_ratio/min': 0.2197018563747406, 'sampling/importance_sampling_ratio/mean': 1.0942494869232178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8523354530334473, 'clip_ratio/low_mean': 0.11519742850214243, 'clip_ratio/low_min': 0.11519742850214243, 'clip_ratio/high_mean': 0.08987254276871681, 'clip_ratio/high_max': 0.08987254276871681, 'clip_ratio/region_mean': 0.20506997127085924, 'reward_total_mean': 0.44169631600379944, 'reward_meter_mean': 0.402677983045578, 'reward_meter_std': 0.3711997866630554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928983449935913, 'reward_repeat_soft_std': 0.00860940758138895, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.1524970829486847, 'reward_total_composite_mean': 0.44169631600379944, 'reward_total_composite_std': 0.08157047629356384, 'epoch': 0.03}

 10%|█         | 345/3300 [50:25<7:23:42,  9.01s/it]INFO 04-13 08:11:23 [block_pool.py:378] Successfully reset prefix cache

 10%|█         | 346/3300 [50:36<7:51:30,  9.58s/it]2026-04-13 08:11:34,361 | INFO | train_grpo_train | metrics_logged mode=train step=346

                                                    
{'loss': 0.1177, 'grad_norm': 17.083423614501953, 'learning_rate': 8.954545454545456e-06, 'num_tokens': 632355.0, 'completions/mean_length': 48.5, 'completions/min_length': 40.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7290892601013184, 'rewards/meter/std': 0.23382572829723358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975104331970215, 'rewards/repeat_soft/std': 0.0042014517821371555, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5872535705566406, 'rewards/total_composite/std': 0.1103053092956543, 'reward': 0.5872535705566406, 'reward_std': 0.1103053092956543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24400284886360168, 'sampling/sampling_logp_difference/max': 1.9545016288757324, 'sampling/importance_sampling_ratio/min': 0.14163504540920258, 'sampling/importance_sampling_ratio/mean': 1.0451273918151855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.0295895636081696, 'clip_ratio/low_mean': 0.11783511005342007, 'clip_ratio/low_min': 0.11783511005342007, 'clip_ratio/high_mean': 0.08385869488120079, 'clip_ratio/high_max': 0.08385869488120079, 'clip_ratio/region_mean': 0.20169380493462086, 'reward_total_mean': 0.5872535705566406, 'reward_meter_mean': 0.7290892601013184, 'reward_meter_std': 0.23382572829723358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975104331970215, 'reward_repeat_soft_std': 0.0042014517821371555, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5872535705566406, 'reward_total_composite_std': 0.1103053092956543, 'epoch': 0.03}

 10%|█         | 346/3300 [50:36<7:51:30,  9.58s/it]INFO 04-13 08:11:34 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 347/3300 [50:51<9:04:28, 11.06s/it]2026-04-13 08:11:48,907 | INFO | train_grpo_train | metrics_logged mode=train step=347

                                                    
{'loss': -0.0959, 'grad_norm': 2.079007863998413, 'learning_rate': 8.951515151515153e-06, 'num_tokens': 634374.0, 'completions/mean_length': 406.375, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 89.5, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.40235579013824463, 'rewards/meter/std': 0.2786467671394348, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/hard_gate/mean': 0.375, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9947296380996704, 'rewards/repeat_soft/std': 0.013047036714851856, 'rewards/judge_quality/mean': 0.07500000298023224, 'rewards/judge_quality/std': 0.0707106813788414, 'rewards/total_composite/mean': 0.1072046309709549, 'rewards/total_composite/std': 0.16362407803535461, 'reward': 0.1072046309709549, 'reward_std': 0.16362407803535461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2555040121078491, 'sampling/sampling_logp_difference/max': 1.3137226104736328, 'sampling/importance_sampling_ratio/min': 0.2688175141811371, 'sampling/importance_sampling_ratio/mean': 1.0996003150939941, 'sampling/importance_sampling_ratio/max': 1.9200503826141357, 'entropy': 1.4671365022659302, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.04957912489771843, 'clip_ratio/high_max': 0.04957912489771843, 'clip_ratio/region_mean': 0.04957912489771843, 'reward_total_mean': 0.1072046309709549, 'reward_meter_mean': 0.40235579013824463, 'reward_meter_std': 0.2786467671394348, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.2519763112068176, 'reward_hard_gate_mean': 0.375, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9947296380996704, 'reward_repeat_soft_std': 0.013047036714851856, 'reward_judge_quality_mean': 0.07500000298023224, 'reward_judge_quality_std': 0.0707106813788414, 'reward_total_composite_mean': 0.1072046309709549, 'reward_total_composite_std': 0.16362407803535461, 'epoch': 0.03}

 11%|█         | 347/3300 [50:51<9:04:28, 11.06s/it]INFO 04-13 08:11:49 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 348/3300 [50:58<8:11:57, 10.00s/it]2026-04-13 08:11:56,378 | INFO | train_grpo_train | metrics_logged mode=train step=348

                                                    
{'loss': -0.0053, 'grad_norm': 15.67885684967041, 'learning_rate': 8.94848484848485e-06, 'num_tokens': 635801.0, 'completions/mean_length': 35.375, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.22712120413780212, 'rewards/meter/std': 0.3341773748397827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9984456896781921, 'rewards/repeat_soft/std': 0.00291926390491426, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.4229438900947571, 'rewards/total_composite/std': 0.09702915698289871, 'reward': 0.4229438900947571, 'reward_std': 0.09702915698289871, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2014777958393097, 'sampling/sampling_logp_difference/max': 1.14703369140625, 'sampling/importance_sampling_ratio/min': 0.31757739186286926, 'sampling/importance_sampling_ratio/mean': 1.0919189453125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.308030426502228, 'clip_ratio/low_mean': 0.10102496016770601, 'clip_ratio/low_min': 0.10102496016770601, 'clip_ratio/high_mean': 0.056480538100004196, 'clip_ratio/high_max': 0.056480538100004196, 'clip_ratio/region_mean': 0.1575054982677102, 'reward_total_mean': 0.4229438900947571, 'reward_meter_mean': 0.22712120413780212, 'reward_meter_std': 0.3341773748397827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9984456896781921, 'reward_repeat_soft_std': 0.00291926390491426, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.4229438900947571, 'reward_total_composite_std': 0.09702915698289871, 'epoch': 0.03}

 11%|█         | 348/3300 [50:58<8:11:57, 10.00s/it]INFO 04-13 08:11:56 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 349/3300 [51:06<7:39:37,  9.35s/it]2026-04-13 08:12:04,232 | INFO | train_grpo_train | metrics_logged mode=train step=349

                                                    
{'loss': -0.0018, 'grad_norm': 22.244646072387695, 'learning_rate': 8.945454545454546e-06, 'num_tokens': 637409.0, 'completions/mean_length': 38.0, 'completions/min_length': 35.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9537421464920044, 'rewards/meter/std': 0.07749338448047638, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9891766309738159, 'rewards/repeat_soft/std': 0.016941606998443604, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.6537644863128662, 'rewards/total_composite/std': 0.11940761655569077, 'reward': 0.6537644863128662, 'reward_std': 0.11940759420394897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21780113875865936, 'sampling/sampling_logp_difference/max': 1.3470020294189453, 'sampling/importance_sampling_ratio/min': 0.26001861691474915, 'sampling/importance_sampling_ratio/mean': 1.0623605251312256, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.431287869811058, 'clip_ratio/low_mean': 0.12143311556428671, 'clip_ratio/low_min': 0.12143311556428671, 'clip_ratio/high_mean': 0.02631578966975212, 'clip_ratio/high_max': 0.02631578966975212, 'clip_ratio/region_mean': 0.14774890523403883, 'reward_total_mean': 0.6537644863128662, 'reward_meter_mean': 0.9537421464920044, 'reward_meter_std': 0.07749338448047638, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9891766309738159, 'reward_repeat_soft_std': 0.016941606998443604, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.6537644863128662, 'reward_total_composite_std': 0.11940761655569077, 'epoch': 0.04}

 11%|█         | 349/3300 [51:06<7:39:37,  9.35s/it]INFO 04-13 08:12:04 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 350/3300 [51:14<7:17:11,  8.89s/it]2026-04-13 08:12:12,030 | INFO | train_grpo_train | metrics_logged mode=train step=350

                                                    
{'loss': 0.1051, 'grad_norm': 17.096710205078125, 'learning_rate': 8.942424242424243e-06, 'num_tokens': 639036.0, 'completions/mean_length': 35.375, 'completions/min_length': 29.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.4493858218193054, 'rewards/meter/std': 0.3458157181739807, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9930947422981262, 'rewards/repeat_soft/std': 0.012191902846097946, 'rewards/judge_quality/mean': 0.45875000953674316, 'rewards/judge_quality/std': 0.20711885392665863, 'rewards/total_composite/mean': 0.4564903974533081, 'rewards/total_composite/std': 0.07396651059389114, 'reward': 0.4564903974533081, 'reward_std': 0.07396651059389114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25389793515205383, 'sampling/sampling_logp_difference/max': 2.7773337364196777, 'sampling/importance_sampling_ratio/min': 0.06220414116978645, 'sampling/importance_sampling_ratio/mean': 1.0388528108596802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.771635800600052, 'clip_ratio/low_mean': 0.12376289814710617, 'clip_ratio/low_min': 0.12376289814710617, 'clip_ratio/high_mean': 0.10631196945905685, 'clip_ratio/high_max': 0.10631196945905685, 'clip_ratio/region_mean': 0.23007486760616302, 'reward_total_mean': 0.4564903974533081, 'reward_meter_mean': 0.4493858218193054, 'reward_meter_std': 0.3458157181739807, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9930947422981262, 'reward_repeat_soft_std': 0.012191902846097946, 'reward_judge_quality_mean': 0.45875000953674316, 'reward_judge_quality_std': 0.20711885392665863, 'reward_total_composite_mean': 0.4564903974533081, 'reward_total_composite_std': 0.07396651059389114, 'epoch': 0.04}

 11%|█         | 350/3300 [51:14<7:17:11,  8.89s/it]INFO 04-13 08:12:12 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:07<00:28,  3.51s/it]

 30%|███       | 3/10 [00:17<00:44,  6.32s/it]

 40%|████      | 4/10 [00:21<00:33,  5.67s/it]

 50%|█████     | 5/10 [00:35<00:42,  8.50s/it]

 60%|██████    | 6/10 [00:43<00:33,  8.37s/it]

 70%|███████   | 7/10 [00:54<00:27,  9.24s/it]

 80%|████████  | 8/10 [01:03<00:18,  9.10s/it]

 90%|█████████ | 9/10 [01:12<00:08,  8.92s/it]

100%|██████████| 10/10 [01:16<00:00,  7.46s/it]2026-04-13 08:13:35,414 | INFO | train_grpo_train | metrics_logged mode=eval step=350

                                                    

                                               
{'eval_loss': nan, 'eval_runtime': 83.3587, 'eval_samples_per_second': 0.96, 'eval_steps_per_second': 0.12, 'eval_num_tokens': 639036.0, 'eval_completions/mean_length': 84.8875, 'eval_completions/min_length': 25.6, 'eval_completions/max_length': 246.1, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 62.49464378356934, 'eval_completions/min_terminated_length': 25.6, 'eval_completions/max_terminated_length': 133.1, 'eval_rewards/meter/mean': 0.534323850274086, 'eval_rewards/meter/std': 0.3636632800102234, 'eval_rewards/count_adherence/mean': 0.9616666615009308, 'eval_rewards/count_adherence/std': 0.08326617777347564, 'eval_rewards/hard_gate/mean': 0.875, 'eval_rewards/hard_gate/std': 0.25587469935417173, 'eval_rewards/repeat_soft/mean': 0.9886297285556793, 'eval_rewards/repeat_soft/std': 0.01991980553139001, 'eval_rewards/judge_quality/mean': 0.4411250025033951, 'eval_rewards/judge_quality/std': 0.19366998383775352, 'eval_rewards/total_composite/mean': 0.46733091175556185, 'eval_rewards/total_composite/std': 0.22581544667482376, 'eval_reward': 0.46733091175556185, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.17469898462295533, 'eval_sampling/sampling_logp_difference/max': 1.1589813232421875, 'eval_sampling/importance_sampling_ratio/min': 0.32134189903736116, 'eval_sampling/importance_sampling_ratio/mean': 1.05313880443573, 'eval_sampling/importance_sampling_ratio/max': 1.7286567568778992, 'eval_entropy': 3.2651810169219972, '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': 0.46733091175556185, 'eval_reward_meter_mean': 0.534323850274086, 'eval_reward_meter_std': 0.3636632800102234, 'eval_reward_count_adherence_mean': 0.9616666615009308, 'eval_reward_count_adherence_std': 0.08326617777347564, 'eval_reward_hard_gate_mean': 0.875, 'eval_reward_hard_gate_std': 0.25587469935417173, 'eval_reward_repeat_soft_mean': 0.9886297285556793, 'eval_reward_repeat_soft_std': 0.01991980553139001, 'eval_reward_judge_quality_mean': 0.4411250025033951, 'eval_reward_judge_quality_std': 0.19366998383775352, 'eval_reward_total_composite_mean': 0.46733091175556185, 'eval_reward_total_composite_std': 0.22581544667482376, 'epoch': 0.04}

 11%|█         | 350/3300 [52:37<7:17:11,  8.89s/it]

100%|██████████| 10/10 [01:16<00:00,  7.46s/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 [52:56<30:17:35, 36.98s/it]2026-04-13 08:13:54,554 | INFO | train_grpo_train | metrics_logged mode=train step=351

                                                     
{'loss': -0.0436, 'grad_norm': 3.864333152770996, 'learning_rate': 8.93939393939394e-06, 'num_tokens': 640362.0, 'completions/mean_length': 79.75, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 18.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.35554254055023193, 'rewards/meter/std': 0.4149716794490814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.33417007327079773, 'rewards/total_composite/mean': 0.4977003335952759, 'rewards/total_composite/std': 0.3186092674732208, 'reward': 0.4977003335952759, 'reward_std': 0.3186092674732208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24403917789459229, 'sampling/sampling_logp_difference/max': 1.3631038665771484, 'sampling/importance_sampling_ratio/min': 0.2558653950691223, 'sampling/importance_sampling_ratio/mean': 1.0408650636672974, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.172814294695854, 'clip_ratio/low_mean': 0.07537839887663722, 'clip_ratio/low_min': 0.07537839887663722, 'clip_ratio/high_mean': 0.09027778077870607, 'clip_ratio/high_max': 0.09027778077870607, 'clip_ratio/region_mean': 0.1656561796553433, 'reward_total_mean': 0.4977003335952759, 'reward_meter_mean': 0.35554254055023193, 'reward_meter_std': 0.4149716794490814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.33417007327079773, 'reward_total_composite_mean': 0.4977003335952759, 'reward_total_composite_std': 0.3186092674732208, 'epoch': 0.04}

 11%|█         | 351/3300 [52:56<30:17:35, 36.98s/it]INFO 04-13 08:13:54 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 352/3300 [53:06<23:29:11, 28.68s/it]2026-04-13 08:14:03,869 | INFO | train_grpo_train | metrics_logged mode=train step=352

                                                     
{'loss': 0.0587, 'grad_norm': 17.360363006591797, 'learning_rate': 8.936363636363638e-06, 'num_tokens': 641851.0, 'completions/mean_length': 33.125, 'completions/min_length': 28.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7030478715896606, 'rewards/meter/std': 0.40410366654396057, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.989193320274353, 'rewards/repeat_soft/std': 0.01741654798388481, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.22965426743030548, 'rewards/total_composite/mean': 0.6477556228637695, 'rewards/total_composite/std': 0.22216816246509552, 'reward': 0.6477556228637695, 'reward_std': 0.22216814756393433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22128766775131226, 'sampling/sampling_logp_difference/max': 1.7048578262329102, 'sampling/importance_sampling_ratio/min': 0.18179823458194733, 'sampling/importance_sampling_ratio/mean': 1.0677930116653442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.238470569252968, 'clip_ratio/low_mean': 0.11027043638750911, 'clip_ratio/low_min': 0.11027043638750911, 'clip_ratio/high_mean': 0.051785715855658054, 'clip_ratio/high_max': 0.051785715855658054, 'clip_ratio/region_mean': 0.16205615224316716, 'reward_total_mean': 0.6477556228637695, 'reward_meter_mean': 0.7030478715896606, 'reward_meter_std': 0.40410366654396057, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.989193320274353, 'reward_repeat_soft_std': 0.01741654798388481, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.22965426743030548, 'reward_total_composite_mean': 0.6477556228637695, 'reward_total_composite_std': 0.22216816246509552, 'epoch': 0.04}

 11%|█         | 352/3300 [53:06<23:29:11, 28.68s/it]INFO 04-13 08:14:04 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 353/3300 [53:17<19:18:42, 23.59s/it]2026-04-13 08:14:15,581 | INFO | train_grpo_train | metrics_logged mode=train step=353

                                                     
{'loss': 0.161, 'grad_norm': 14.564729690551758, 'learning_rate': 8.933333333333333e-06, 'num_tokens': 643586.0, 'completions/mean_length': 37.875, 'completions/min_length': 28.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5424282550811768, 'rewards/meter/std': 0.37805813550949097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9600545763969421, 'rewards/repeat_soft/std': 0.09529281407594681, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.23790381848812103, 'rewards/total_composite/mean': 0.5965635776519775, 'rewards/total_composite/std': 0.22408899664878845, 'reward': 0.5965635776519775, 'reward_std': 0.22408899664878845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18450625240802765, 'sampling/sampling_logp_difference/max': 3.021036148071289, 'sampling/importance_sampling_ratio/min': 0.048750679939985275, 'sampling/importance_sampling_ratio/mean': 1.0115100145339966, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6726677417755127, 'clip_ratio/low_mean': 0.07034111954271793, 'clip_ratio/low_min': 0.07034111954271793, 'clip_ratio/high_mean': 0.07265911251306534, 'clip_ratio/high_max': 0.07265911251306534, 'clip_ratio/region_mean': 0.14300023205578327, 'reward_total_mean': 0.5965635776519775, 'reward_meter_mean': 0.5424282550811768, 'reward_meter_std': 0.37805813550949097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9600545763969421, 'reward_repeat_soft_std': 0.09529281407594681, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.23790381848812103, 'reward_total_composite_mean': 0.5965635776519775, 'reward_total_composite_std': 0.22408899664878845, 'epoch': 0.04}

 11%|█         | 353/3300 [53:17<19:18:42, 23.59s/it]INFO 04-13 08:14:15 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 354/3300 [53:25<15:19:34, 18.73s/it]2026-04-13 08:14:22,968 | INFO | train_grpo_train | metrics_logged mode=train step=354

                                                     
{'loss': 0.0566, 'grad_norm': 19.15772819519043, 'learning_rate': 8.930303030303032e-06, 'num_tokens': 645169.0, 'completions/mean_length': 33.875, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.48136985301971436, 'rewards/meter/std': 0.3475463092327118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976884126663208, 'rewards/repeat_soft/std': 0.002912714844569564, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.582542359828949, 'rewards/total_composite/std': 0.1957775503396988, 'reward': 0.582542359828949, 'reward_std': 0.1957775354385376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17875388264656067, 'sampling/sampling_logp_difference/max': 1.5533866882324219, 'sampling/importance_sampling_ratio/min': 0.21153037250041962, 'sampling/importance_sampling_ratio/mean': 1.0230615139007568, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3181058838963509, 'clip_ratio/low_mean': 0.09575557895004749, 'clip_ratio/low_min': 0.09575557895004749, 'clip_ratio/high_mean': 0.05081847682595253, 'clip_ratio/high_max': 0.05081847682595253, 'clip_ratio/region_mean': 0.14657405577600002, 'reward_total_mean': 0.582542359828949, 'reward_meter_mean': 0.48136985301971436, 'reward_meter_std': 0.3475463092327118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976884126663208, 'reward_repeat_soft_std': 0.002912714844569564, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.582542359828949, 'reward_total_composite_std': 0.1957775503396988, 'epoch': 0.04}

 11%|█         | 354/3300 [53:25<15:19:34, 18.73s/it]INFO 04-13 08:14:23 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 355/3300 [53:38<13:52:48, 16.97s/it]2026-04-13 08:14:35,822 | INFO | train_grpo_train | metrics_logged mode=train step=355

                                                     
{'loss': 0.0768, 'grad_norm': 12.950587272644043, 'learning_rate': 8.927272727272728e-06, 'num_tokens': 646596.0, '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': 0.27759212255477905, 'rewards/meter/std': 0.1785835325717926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9925885200500488, 'rewards/repeat_soft/std': 0.007694505620747805, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.44296160340309143, 'rewards/total_composite/std': 0.08418701589107513, 'reward': 0.44296160340309143, 'reward_std': 0.08418700098991394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22627484798431396, 'sampling/sampling_logp_difference/max': 2.520127296447754, 'sampling/importance_sampling_ratio/min': 0.08044936507940292, 'sampling/importance_sampling_ratio/mean': 1.015830159187317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.899744674563408, 'clip_ratio/low_mean': 0.08606674242764711, 'clip_ratio/low_min': 0.08606674242764711, 'clip_ratio/high_mean': 0.14343082159757614, 'clip_ratio/high_max': 0.14343082159757614, 'clip_ratio/region_mean': 0.22949756402522326, 'reward_total_mean': 0.44296160340309143, 'reward_meter_mean': 0.27759212255477905, 'reward_meter_std': 0.1785835325717926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9925885200500488, 'reward_repeat_soft_std': 0.007694505620747805, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.44296160340309143, 'reward_total_composite_std': 0.08418701589107513, 'epoch': 0.04}

 11%|█         | 355/3300 [53:38<13:52:48, 16.97s/it]INFO 04-13 08:14:36 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 356/3300 [53:56<14:19:08, 17.51s/it]2026-04-13 08:14:54,597 | INFO | train_grpo_train | metrics_logged mode=train step=356

                                                     
{'loss': -0.1054, 'grad_norm': 3.625981330871582, 'learning_rate': 8.924242424242425e-06, 'num_tokens': 648292.0, 'completions/mean_length': 182.0, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.43331071734428406, 'rewards/meter/std': 0.39519160985946655, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9938517808914185, 'rewards/repeat_soft/std': 0.008353594690561295, 'rewards/judge_quality/mean': 0.22999998927116394, 'rewards/judge_quality/std': 0.1437259316444397, 'rewards/total_composite/mean': 0.2760443389415741, 'rewards/total_composite/std': 0.23782400786876678, 'reward': 0.2760443389415741, 'reward_std': 0.23782400786876678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24254828691482544, 'sampling/sampling_logp_difference/max': 1.1382880210876465, 'sampling/importance_sampling_ratio/min': 0.3203670382499695, 'sampling/importance_sampling_ratio/mean': 1.0883963108062744, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.182831972837448, 'clip_ratio/low_mean': 0.024671053513884544, 'clip_ratio/low_min': 0.024671053513884544, 'clip_ratio/high_mean': 0.12995248660445213, 'clip_ratio/high_max': 0.12995248660445213, 'clip_ratio/region_mean': 0.15462354011833668, 'reward_total_mean': 0.2760443389415741, 'reward_meter_mean': 0.43331071734428406, 'reward_meter_std': 0.39519160985946655, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9938517808914185, 'reward_repeat_soft_std': 0.008353594690561295, 'reward_judge_quality_mean': 0.22999998927116394, 'reward_judge_quality_std': 0.1437259316444397, 'reward_total_composite_mean': 0.2760443389415741, 'reward_total_composite_std': 0.23782400786876678, 'epoch': 0.04}

 11%|█         | 356/3300 [53:56<14:19:08, 17.51s/it]INFO 04-13 08:14:54 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 357/3300 [54:12<13:59:15, 17.11s/it]2026-04-13 08:15:10,852 | INFO | train_grpo_train | metrics_logged mode=train step=357

                                                     
{'loss': 0.0534, 'grad_norm': 4.924477577209473, 'learning_rate': 8.921212121212122e-06, 'num_tokens': 650441.0, 'completions/mean_length': 151.625, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 100.14286041259766, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 214.0, 'rewards/meter/mean': 0.8353273868560791, 'rewards/meter/std': 0.1646488457918167, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2121320366859436, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8883692622184753, 'rewards/repeat_soft/std': 0.17075568437576294, 'rewards/judge_quality/mean': 0.2724999785423279, 'rewards/judge_quality/std': 0.14119388163089752, 'rewards/total_composite/mean': 0.353649377822876, 'rewards/total_composite/std': 0.23280341923236847, 'reward': 0.353649377822876, 'reward_std': 0.23280341923236847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22231870889663696, 'sampling/sampling_logp_difference/max': 1.5295047760009766, 'sampling/importance_sampling_ratio/min': 0.21664293110370636, 'sampling/importance_sampling_ratio/mean': 1.046348214149475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9584960341453552, 'clip_ratio/low_mean': 0.03665999136865139, 'clip_ratio/low_min': 0.03665999136865139, 'clip_ratio/high_mean': 0.11384666338562965, 'clip_ratio/high_max': 0.11384666338562965, 'clip_ratio/region_mean': 0.15050665475428104, 'reward_total_mean': 0.353649377822876, 'reward_meter_mean': 0.8353273868560791, 'reward_meter_std': 0.1646488457918167, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2121320366859436, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8883692622184753, 'reward_repeat_soft_std': 0.17075568437576294, 'reward_judge_quality_mean': 0.2724999785423279, 'reward_judge_quality_std': 0.14119388163089752, 'reward_total_composite_mean': 0.353649377822876, 'reward_total_composite_std': 0.23280341923236847, 'epoch': 0.04}

 11%|█         | 357/3300 [54:13<13:59:15, 17.11s/it]INFO 04-13 08:15:11 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 358/3300 [54:20<11:38:36, 14.25s/it]2026-04-13 08:15:18,343 | INFO | train_grpo_train | metrics_logged mode=train step=358

                                                     
{'loss': 0.1867, 'grad_norm': 12.765982627868652, 'learning_rate': 8.91818181818182e-06, 'num_tokens': 652183.0, 'completions/mean_length': 59.75, 'completions/min_length': 41.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.5446306467056274, 'rewards/meter/std': 0.4207291305065155, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9915409088134766, 'rewards/repeat_soft/std': 0.010725677944719791, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.47466665506362915, 'rewards/total_composite/std': 0.1031181737780571, 'reward': 0.47466665506362915, 'reward_std': 0.1031181663274765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2030535191297531, 'sampling/sampling_logp_difference/max': 1.5430059432983398, 'sampling/importance_sampling_ratio/min': 0.21373766660690308, 'sampling/importance_sampling_ratio/mean': 1.0571295022964478, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6188998222351074, 'clip_ratio/low_mean': 0.0892687551677227, 'clip_ratio/low_min': 0.0892687551677227, 'clip_ratio/high_mean': 0.07147535216063261, 'clip_ratio/high_max': 0.07147535216063261, 'clip_ratio/region_mean': 0.1607441073283553, 'reward_total_mean': 0.47466665506362915, 'reward_meter_mean': 0.5446306467056274, 'reward_meter_std': 0.4207291305065155, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9915409088134766, 'reward_repeat_soft_std': 0.010725677944719791, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.47466665506362915, 'reward_total_composite_std': 0.1031181737780571, 'epoch': 0.04}

 11%|█         | 358/3300 [54:20<11:38:36, 14.25s/it]INFO 04-13 08:15:18 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 359/3300 [54:30<10:35:17, 12.96s/it]2026-04-13 08:15:28,367 | INFO | train_grpo_train | metrics_logged mode=train step=359

                                                     
{'loss': 0.5258, 'grad_norm': 9.785633087158203, 'learning_rate': 8.915151515151515e-06, 'num_tokens': 654457.0, 'completions/mean_length': 101.25, 'completions/min_length': 66.0, 'completions/max_length': 275.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 275.0, 'rewards/meter/mean': 0.6811692714691162, 'rewards/meter/std': 0.37058714032173157, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9920492172241211, 'rewards/repeat_soft/std': 0.01068015955388546, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22211645543575287, 'rewards/total_composite/mean': 0.5265733003616333, 'rewards/total_composite/std': 0.2566831409931183, 'reward': 0.5265733003616333, 'reward_std': 0.2566831111907959, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22395886480808258, 'sampling/sampling_logp_difference/max': 2.0679333209991455, 'sampling/importance_sampling_ratio/min': 0.12644682824611664, 'sampling/importance_sampling_ratio/mean': 1.0653128623962402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.974060446023941, 'clip_ratio/low_mean': 0.05053030326962471, 'clip_ratio/low_min': 0.05053030326962471, 'clip_ratio/high_mean': 0.13463346101343632, 'clip_ratio/high_max': 0.13463346101343632, 'clip_ratio/region_mean': 0.18516376428306103, 'reward_total_mean': 0.5265733003616333, 'reward_meter_mean': 0.6811692714691162, 'reward_meter_std': 0.37058714032173157, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9920492172241211, 'reward_repeat_soft_std': 0.01068015955388546, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22211645543575287, 'reward_total_composite_mean': 0.5265733003616333, 'reward_total_composite_std': 0.2566831409931183, 'epoch': 0.04}

 11%|█         | 359/3300 [54:30<10:35:17, 12.96s/it]INFO 04-13 08:15:28 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 360/3300 [54:40<9:52:09, 12.08s/it] 2026-04-13 08:15:38,342 | INFO | train_grpo_train | metrics_logged mode=train step=360

                                                    
{'loss': 0.5822, 'grad_norm': 9.903790473937988, 'learning_rate': 8.912121212121214e-06, 'num_tokens': 656499.0, 'completions/mean_length': 91.25, 'completions/min_length': 46.0, 'completions/max_length': 340.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 340.0, 'rewards/meter/mean': 0.3474387526512146, 'rewards/meter/std': 0.32828089594841003, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9982553720474243, 'rewards/repeat_soft/std': 0.0026103267446160316, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.13845446705818176, 'rewards/total_composite/mean': 0.32856714725494385, 'rewards/total_composite/std': 0.2217308133840561, 'reward': 0.32856714725494385, 'reward_std': 0.2217308133840561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20932020246982574, 'sampling/sampling_logp_difference/max': 1.7218084335327148, 'sampling/importance_sampling_ratio/min': 0.17874261736869812, 'sampling/importance_sampling_ratio/mean': 1.0371640920639038, 'sampling/importance_sampling_ratio/max': 1.958345651626587, 'entropy': 4.028614193201065, 'clip_ratio/low_mean': 0.022341629955917597, 'clip_ratio/low_min': 0.022341629955917597, 'clip_ratio/high_mean': 0.14197558723390102, 'clip_ratio/high_max': 0.14197558723390102, 'clip_ratio/region_mean': 0.16431721718981862, 'reward_total_mean': 0.32856714725494385, 'reward_meter_mean': 0.3474387526512146, 'reward_meter_std': 0.32828089594841003, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9982553720474243, 'reward_repeat_soft_std': 0.0026103267446160316, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.13845446705818176, 'reward_total_composite_mean': 0.32856714725494385, 'reward_total_composite_std': 0.2217308133840561, 'epoch': 0.04}

 11%|█         | 360/3300 [54:40<9:52:09, 12.08s/it]INFO 04-13 08:15:38 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 361/3300 [54:52<9:47:17, 11.99s/it]2026-04-13 08:15:50,110 | INFO | train_grpo_train | metrics_logged mode=train step=361

                                                    
{'loss': -0.1497, 'grad_norm': 3.4068005084991455, 'learning_rate': 8.90909090909091e-06, 'num_tokens': 658196.0, 'completions/mean_length': 113.125, 'completions/min_length': 51.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.142860412597656, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7484950423240662, 'rewards/meter/std': 0.31213483214378357, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9944742918014526, 'rewards/repeat_soft/std': 0.007948680780827999, 'rewards/judge_quality/mean': 0.29750001430511475, 'rewards/judge_quality/std': 0.14518460631370544, 'rewards/total_composite/mean': 0.44896334409713745, 'rewards/total_composite/std': 0.20162814855575562, 'reward': 0.44896334409713745, 'reward_std': 0.2016281634569168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25105538964271545, 'sampling/sampling_logp_difference/max': 2.721737861633301, 'sampling/importance_sampling_ratio/min': 0.06576037406921387, 'sampling/importance_sampling_ratio/mean': 1.0676491260528564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2537022531032562, 'clip_ratio/low_mean': 0.04337772913277149, 'clip_ratio/low_min': 0.04337772913277149, 'clip_ratio/high_mean': 0.12420917861163616, 'clip_ratio/high_max': 0.12420917861163616, 'clip_ratio/region_mean': 0.16758690774440765, 'reward_total_mean': 0.44896334409713745, 'reward_meter_mean': 0.7484950423240662, 'reward_meter_std': 0.31213483214378357, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9944742918014526, 'reward_repeat_soft_std': 0.007948680780827999, 'reward_judge_quality_mean': 0.29750001430511475, 'reward_judge_quality_std': 0.14518460631370544, 'reward_total_composite_mean': 0.44896334409713745, 'reward_total_composite_std': 0.20162814855575562, 'epoch': 0.04}

 11%|█         | 361/3300 [54:52<9:47:17, 11.99s/it]INFO 04-13 08:15:50 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 362/3300 [54:58<8:27:44, 10.37s/it]2026-04-13 08:15:56,698 | INFO | train_grpo_train | metrics_logged mode=train step=362

                                                    
{'loss': -0.0354, 'grad_norm': 19.476285934448242, 'learning_rate': 8.906060606060607e-06, 'num_tokens': 659666.0, 'completions/mean_length': 22.75, 'completions/min_length': 19.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6831570863723755, 'rewards/meter/std': 0.31254181265830994, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9976022839546204, 'rewards/repeat_soft/std': 0.003977146465331316, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.539945125579834, 'rewards/total_composite/std': 0.0878436267375946, 'reward': 0.539945125579834, 'reward_std': 0.08784361928701401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2205607146024704, 'sampling/sampling_logp_difference/max': 1.830787181854248, 'sampling/importance_sampling_ratio/min': 0.16028733551502228, 'sampling/importance_sampling_ratio/mean': 1.0382581949234009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4231555983424187, 'clip_ratio/low_mean': 0.06328320689499378, 'clip_ratio/low_min': 0.06328320689499378, 'clip_ratio/high_mean': 0.1684735119342804, 'clip_ratio/high_max': 0.1684735119342804, 'clip_ratio/region_mean': 0.23175671882927418, 'reward_total_mean': 0.539945125579834, 'reward_meter_mean': 0.6831570863723755, 'reward_meter_std': 0.31254181265830994, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9976022839546204, 'reward_repeat_soft_std': 0.003977146465331316, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.539945125579834, 'reward_total_composite_std': 0.0878436267375946, 'epoch': 0.04}

 11%|█         | 362/3300 [54:58<8:27:44, 10.37s/it]INFO 04-13 08:15:56 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 363/3300 [55:10<8:52:33, 10.88s/it]2026-04-13 08:16:08,773 | INFO | train_grpo_train | metrics_logged mode=train step=363

                                                    
{'loss': -0.1141, 'grad_norm': 4.718713760375977, 'learning_rate': 8.903030303030304e-06, 'num_tokens': 661668.0, 'completions/mean_length': 132.25, 'completions/min_length': 55.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 78.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.3226236402988434, 'rewards/meter/std': 0.3042055368423462, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.987562894821167, 'rewards/repeat_soft/std': 0.01699833758175373, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.25427767634391785, 'rewards/total_composite/mean': 0.4441455006599426, 'rewards/total_composite/std': 0.2318834662437439, 'reward': 0.4441455006599426, 'reward_std': 0.2318834513425827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21444430947303772, 'sampling/sampling_logp_difference/max': 1.1028151512145996, 'sampling/importance_sampling_ratio/min': 0.331935316324234, 'sampling/importance_sampling_ratio/mean': 1.0466419458389282, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.461810067296028, 'clip_ratio/low_mean': 0.085496561601758, 'clip_ratio/low_min': 0.085496561601758, 'clip_ratio/high_mean': 0.06669638678431511, 'clip_ratio/high_max': 0.06669638678431511, 'clip_ratio/region_mean': 0.1521929483860731, 'reward_total_mean': 0.4441455006599426, 'reward_meter_mean': 0.3226236402988434, 'reward_meter_std': 0.3042055368423462, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.987562894821167, 'reward_repeat_soft_std': 0.01699833758175373, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.25427767634391785, 'reward_total_composite_mean': 0.4441455006599426, 'reward_total_composite_std': 0.2318834662437439, 'epoch': 0.04}

 11%|█         | 363/3300 [55:11<8:52:33, 10.88s/it]INFO 04-13 08:16:09 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 364/3300 [55:17<7:49:44,  9.60s/it]2026-04-13 08:16:15,384 | INFO | train_grpo_train | metrics_logged mode=train step=364

                                                    
{'loss': 0.0201, 'grad_norm': 17.85727882385254, 'learning_rate': 8.900000000000001e-06, 'num_tokens': 663105.0, 'completions/mean_length': 33.625, 'completions/min_length': 30.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8145196437835693, 'rewards/meter/std': 0.3323499262332916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9631528854370117, 'rewards/repeat_soft/std': 0.031031915917992592, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.2188117355108261, 'rewards/total_composite/mean': 0.7170701026916504, 'rewards/total_composite/std': 0.19379006326198578, 'reward': 0.7170701026916504, 'reward_std': 0.19379007816314697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15746693313121796, 'sampling/sampling_logp_difference/max': 1.4978742599487305, 'sampling/importance_sampling_ratio/min': 0.2236049920320511, 'sampling/importance_sampling_ratio/mean': 1.0556191205978394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5243082791566849, 'clip_ratio/low_mean': 0.04288856312632561, 'clip_ratio/low_min': 0.04288856312632561, 'clip_ratio/high_mean': 0.09164384286850691, 'clip_ratio/high_max': 0.09164384286850691, 'clip_ratio/region_mean': 0.13453240599483252, 'reward_total_mean': 0.7170701026916504, 'reward_meter_mean': 0.8145196437835693, 'reward_meter_std': 0.3323499262332916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9631528854370117, 'reward_repeat_soft_std': 0.031031915917992592, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.2188117355108261, 'reward_total_composite_mean': 0.7170701026916504, 'reward_total_composite_std': 0.19379006326198578, 'epoch': 0.04}

 11%|█         | 364/3300 [55:17<7:49:44,  9.60s/it]INFO 04-13 08:16:15 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 365/3300 [55:25<7:24:01,  9.08s/it]2026-04-13 08:16:23,239 | INFO | train_grpo_train | metrics_logged mode=train step=365

                                                    
{'loss': 0.0139, 'grad_norm': 11.384407043457031, 'learning_rate': 8.896969696969697e-06, 'num_tokens': 665024.0, 'completions/mean_length': 63.875, 'completions/min_length': 60.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.5080638527870178, 'rewards/meter/std': 0.3114967942237854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9727797508239746, 'rewards/repeat_soft/std': 0.061003249138593674, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.47228801250457764, 'rewards/total_composite/std': 0.07512867450714111, 'reward': 0.47228801250457764, 'reward_std': 0.07512866705656052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.207632914185524, 'sampling/sampling_logp_difference/max': 1.5510435104370117, 'sampling/importance_sampling_ratio/min': 0.21202661097049713, 'sampling/importance_sampling_ratio/mean': 1.0509876012802124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7623439729213715, 'clip_ratio/low_mean': 0.09796436131000519, 'clip_ratio/low_min': 0.09796436131000519, 'clip_ratio/high_mean': 0.06974686309695244, 'clip_ratio/high_max': 0.06974686309695244, 'clip_ratio/region_mean': 0.16771122440695763, 'reward_total_mean': 0.47228801250457764, 'reward_meter_mean': 0.5080638527870178, 'reward_meter_std': 0.3114967942237854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9727797508239746, 'reward_repeat_soft_std': 0.061003249138593674, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.47228801250457764, 'reward_total_composite_std': 0.07512867450714111, 'epoch': 0.04}

 11%|█         | 365/3300 [55:25<7:24:01,  9.08s/it]INFO 04-13 08:16:23 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 366/3300 [55:31<6:45:02,  8.28s/it]2026-04-13 08:16:29,671 | INFO | train_grpo_train | metrics_logged mode=train step=366

                                                    
{'loss': 0.0729, 'grad_norm': 22.787700653076172, 'learning_rate': 8.893939393939394e-06, 'num_tokens': 666534.0, 'completions/mean_length': 28.75, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7012413740158081, 'rewards/meter/std': 0.42376846075057983, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9953725934028625, 'rewards/repeat_soft/std': 0.004975995514541864, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.6596815586090088, 'rewards/total_composite/std': 0.2462015450000763, 'reward': 0.6596815586090088, 'reward_std': 0.2462015450000763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.25858786702156067, 'sampling/sampling_logp_difference/max': 2.060631275177002, 'sampling/importance_sampling_ratio/min': 0.12737353146076202, 'sampling/importance_sampling_ratio/mean': 1.077356219291687, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6123074293136597, 'clip_ratio/low_mean': 0.13816550932824612, 'clip_ratio/low_min': 0.13816550932824612, 'clip_ratio/high_mean': 0.053494623862206936, 'clip_ratio/high_max': 0.053494623862206936, 'clip_ratio/region_mean': 0.19166013319045305, 'reward_total_mean': 0.6596815586090088, 'reward_meter_mean': 0.7012413740158081, 'reward_meter_std': 0.42376846075057983, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9953725934028625, 'reward_repeat_soft_std': 0.004975995514541864, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.6596815586090088, 'reward_total_composite_std': 0.2462015450000763, 'epoch': 0.04}

 11%|█         | 366/3300 [55:31<6:45:02,  8.28s/it]INFO 04-13 08:16:29 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 367/3300 [55:43<7:39:41,  9.40s/it]2026-04-13 08:16:41,690 | INFO | train_grpo_train | metrics_logged mode=train step=367

                                                    
{'loss': -0.1425, 'grad_norm': 2.678206205368042, 'learning_rate': 8.890909090909091e-06, 'num_tokens': 668583.0, 'completions/mean_length': 206.125, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 104.16667175292969, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.27708929777145386, 'rewards/meter/std': 0.27479180693626404, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9985931515693665, 'rewards/repeat_soft/std': 0.0012916075065732002, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.22238561511039734, 'rewards/total_composite/mean': 0.3461523652076721, 'rewards/total_composite/std': 0.24402405321598053, 'reward': 0.3461523652076721, 'reward_std': 0.24402403831481934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22072133421897888, 'sampling/sampling_logp_difference/max': 1.3632798194885254, 'sampling/importance_sampling_ratio/min': 0.2558203637599945, 'sampling/importance_sampling_ratio/mean': 1.029079556465149, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9468597173690796, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.15607979893684387, 'clip_ratio/high_max': 0.15607979893684387, 'clip_ratio/region_mean': 0.15607979893684387, 'reward_total_mean': 0.3461523652076721, 'reward_meter_mean': 0.27708929777145386, 'reward_meter_std': 0.27479180693626404, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9985931515693665, 'reward_repeat_soft_std': 0.0012916075065732002, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.22238561511039734, 'reward_total_composite_mean': 0.3461523652076721, 'reward_total_composite_std': 0.24402405321598053, 'epoch': 0.04}

 11%|█         | 367/3300 [55:43<7:39:41,  9.40s/it]INFO 04-13 08:16:41 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 368/3300 [55:59<9:07:50, 11.21s/it]2026-04-13 08:16:57,119 | INFO | train_grpo_train | metrics_logged mode=train step=368

                                                    
{'loss': -0.0744, 'grad_norm': 4.494978427886963, 'learning_rate': 8.887878787878789e-06, 'num_tokens': 670159.0, 'completions/mean_length': 96.0, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.57143020629883, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5928663015365601, 'rewards/meter/std': 0.37849077582359314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9926130175590515, 'rewards/repeat_soft/std': 0.012723951600492, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.2958281636238098, 'rewards/total_composite/mean': 0.49554625153541565, 'rewards/total_composite/std': 0.2634750008583069, 'reward': 0.49554625153541565, 'reward_std': 0.2634750008583069, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21571017801761627, 'sampling/sampling_logp_difference/max': 1.3126220703125, 'sampling/importance_sampling_ratio/min': 0.2691135108470917, 'sampling/importance_sampling_ratio/mean': 1.060481071472168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.405510425567627, 'clip_ratio/low_mean': 0.057502121198922396, 'clip_ratio/low_min': 0.057502121198922396, 'clip_ratio/high_mean': 0.08323585148900747, 'clip_ratio/high_max': 0.08323585148900747, 'clip_ratio/region_mean': 0.14073797268792987, 'reward_total_mean': 0.49554625153541565, 'reward_meter_mean': 0.5928663015365601, 'reward_meter_std': 0.37849077582359314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9926130175590515, 'reward_repeat_soft_std': 0.012723951600492, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.2958281636238098, 'reward_total_composite_mean': 0.49554625153541565, 'reward_total_composite_std': 0.2634750008583069, 'epoch': 0.04}

 11%|█         | 368/3300 [55:59<9:07:50, 11.21s/it]INFO 04-13 08:16:57 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 369/3300 [56:06<8:14:29, 10.12s/it]2026-04-13 08:17:04,703 | INFO | train_grpo_train | metrics_logged mode=train step=369

                                                    
{'loss': 0.0322, 'grad_norm': 12.926513671875, 'learning_rate': 8.884848484848486e-06, 'num_tokens': 671859.0, 'completions/mean_length': 55.5, 'completions/min_length': 43.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.3802972435951233, 'rewards/meter/std': 0.39531365036964417, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928926229476929, 'rewards/repeat_soft/std': 0.010803640820086002, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.45972201228141785, 'rewards/total_composite/std': 0.17263850569725037, 'reward': 0.45972201228141785, 'reward_std': 0.17263852059841156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22510534524917603, 'sampling/sampling_logp_difference/max': 1.3513402938842773, 'sampling/importance_sampling_ratio/min': 0.25889304280281067, 'sampling/importance_sampling_ratio/mean': 1.0685721635818481, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.410489469766617, 'clip_ratio/low_mean': 0.08347007166594267, 'clip_ratio/low_min': 0.08347007166594267, 'clip_ratio/high_mean': 0.05272738356143236, 'clip_ratio/high_max': 0.05272738356143236, 'clip_ratio/region_mean': 0.13619745522737503, 'reward_total_mean': 0.45972201228141785, 'reward_meter_mean': 0.3802972435951233, 'reward_meter_std': 0.39531365036964417, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928926229476929, 'reward_repeat_soft_std': 0.010803640820086002, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.45972201228141785, 'reward_total_composite_std': 0.17263850569725037, 'epoch': 0.04}

 11%|█         | 369/3300 [56:06<8:14:29, 10.12s/it]INFO 04-13 08:17:04 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 370/3300 [56:13<7:23:22,  9.08s/it]2026-04-13 08:17:11,347 | INFO | train_grpo_train | metrics_logged mode=train step=370

                                                    
{'loss': 0.0327, 'grad_norm': 38.586856842041016, 'learning_rate': 8.881818181818183e-06, 'num_tokens': 673312.0, 'completions/mean_length': 23.625, 'completions/min_length': 22.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.2130904495716095, 'rewards/meter/std': 0.22513271868228912, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9993371367454529, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.7649999856948853, 'rewards/judge_quality/std': 0.2979933023452759, 'rewards/total_composite/mean': 0.4061920940876007, 'rewards/total_composite/std': 0.2044784426689148, 'reward': 0.4061920940876007, 'reward_std': 0.2044784426689148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19987602531909943, 'sampling/sampling_logp_difference/max': 1.7916243076324463, 'sampling/importance_sampling_ratio/min': 0.1666891872882843, 'sampling/importance_sampling_ratio/mean': 1.0216882228851318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1096134707331657, 'clip_ratio/low_mean': 0.1515974961221218, 'clip_ratio/low_min': 0.1515974961221218, 'clip_ratio/high_mean': 0.04985007457435131, 'clip_ratio/high_max': 0.04985007457435131, 'clip_ratio/region_mean': 0.20144757069647312, 'reward_total_mean': 0.4061920940876007, 'reward_meter_mean': 0.2130904495716095, 'reward_meter_std': 0.22513271868228912, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9993371367454529, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.7649999856948853, 'reward_judge_quality_std': 0.2979933023452759, 'reward_total_composite_mean': 0.4061920940876007, 'reward_total_composite_std': 0.2044784426689148, 'epoch': 0.04}

 11%|█         | 370/3300 [56:13<7:23:22,  9.08s/it]INFO 04-13 08:17:11 [block_pool.py:378] Successfully reset prefix cache

 11%|█         | 371/3300 [56:20<6:57:43,  8.56s/it]2026-04-13 08:17:18,685 | INFO | train_grpo_train | metrics_logged mode=train step=371

                                                    
{'loss': 0.0258, 'grad_norm': 17.992992401123047, 'learning_rate': 8.87878787878788e-06, 'num_tokens': 674861.0, 'completions/mean_length': 31.625, 'completions/min_length': 28.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.815354585647583, 'rewards/meter/std': 0.30784371495246887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9822838306427002, 'rewards/repeat_soft/std': 0.011374717578291893, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.22158117592334747, 'rewards/total_composite/mean': 0.6992222666740417, 'rewards/total_composite/std': 0.1985442191362381, 'reward': 0.6992222666740417, 'reward_std': 0.1985442191362381, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18011300265789032, 'sampling/sampling_logp_difference/max': 1.3447229862213135, 'sampling/importance_sampling_ratio/min': 0.260611891746521, 'sampling/importance_sampling_ratio/mean': 1.0152212381362915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.919819250702858, 'clip_ratio/low_mean': 0.06414490658789873, 'clip_ratio/low_min': 0.06414490658789873, 'clip_ratio/high_mean': 0.06656280532479286, 'clip_ratio/high_max': 0.06656280532479286, 'clip_ratio/region_mean': 0.1307077119126916, 'reward_total_mean': 0.6992222666740417, 'reward_meter_mean': 0.815354585647583, 'reward_meter_std': 0.30784371495246887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9822838306427002, 'reward_repeat_soft_std': 0.011374717578291893, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.22158117592334747, 'reward_total_composite_mean': 0.6992222666740417, 'reward_total_composite_std': 0.1985442191362381, 'epoch': 0.04}

 11%|█         | 371/3300 [56:20<6:57:43,  8.56s/it]INFO 04-13 08:17:18 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 372/3300 [56:28<6:36:43,  8.13s/it]2026-04-13 08:17:25,820 | INFO | train_grpo_train | metrics_logged mode=train step=372

                                                    
{'loss': 0.0014, 'grad_norm': 16.33405303955078, 'learning_rate': 8.875757575757576e-06, 'num_tokens': 676335.0, 'completions/mean_length': 21.25, 'completions/min_length': 16.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.16977465152740479, 'rewards/meter/std': 0.28688836097717285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9377456903457642, 'rewards/repeat_soft/std': 0.032807815819978714, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3870103359222412, 'rewards/total_composite/std': 0.07520600408315659, 'reward': 0.3870103359222412, 'reward_std': 0.07520599663257599, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17178981006145477, 'sampling/sampling_logp_difference/max': 1.4719525575637817, 'sampling/importance_sampling_ratio/min': 0.22947698831558228, 'sampling/importance_sampling_ratio/mean': 1.0394285917282104, 'sampling/importance_sampling_ratio/max': 1.8622790575027466, 'entropy': 1.3809120282530785, 'clip_ratio/low_mean': 0.09722041804343462, 'clip_ratio/low_min': 0.09722041804343462, 'clip_ratio/high_mean': 0.01733954483643174, 'clip_ratio/high_max': 0.01733954483643174, 'clip_ratio/region_mean': 0.11455996287986636, 'reward_total_mean': 0.3870103359222412, 'reward_meter_mean': 0.16977465152740479, 'reward_meter_std': 0.28688836097717285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9377456903457642, 'reward_repeat_soft_std': 0.032807815819978714, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3870103359222412, 'reward_total_composite_std': 0.07520600408315659, 'epoch': 0.04}

 11%|█▏        | 372/3300 [56:28<6:36:43,  8.13s/it]INFO 04-13 08:17:26 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 373/3300 [56:36<6:37:59,  8.16s/it]2026-04-13 08:17:34,043 | INFO | train_grpo_train | metrics_logged mode=train step=373

                                                    
{'loss': 0.5053, 'grad_norm': 17.594009399414062, 'learning_rate': 8.872727272727275e-06, 'num_tokens': 678000.0, 'completions/mean_length': 46.125, 'completions/min_length': 29.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.49487289786338806, 'rewards/meter/std': 0.493295282125473, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9968066215515137, 'rewards/repeat_soft/std': 0.004193334840238094, 'rewards/judge_quality/mean': 0.2749999761581421, 'rewards/judge_quality/std': 0.155838742852211, 'rewards/total_composite/mean': 0.41905075311660767, 'rewards/total_composite/std': 0.2089303880929947, 'reward': 0.41905075311660767, 'reward_std': 0.2089303880929947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2209448367357254, 'sampling/sampling_logp_difference/max': 1.3029823303222656, 'sampling/importance_sampling_ratio/min': 0.2717202305793762, 'sampling/importance_sampling_ratio/mean': 1.0277589559555054, 'sampling/importance_sampling_ratio/max': 1.9530292749404907, 'entropy': 4.145709306001663, 'clip_ratio/low_mean': 0.0727820647880435, 'clip_ratio/low_min': 0.0727820647880435, 'clip_ratio/high_mean': 0.1434063334017992, 'clip_ratio/high_max': 0.1434063334017992, 'clip_ratio/region_mean': 0.2161883981898427, 'reward_total_mean': 0.41905075311660767, 'reward_meter_mean': 0.49487289786338806, 'reward_meter_std': 0.493295282125473, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9968066215515137, 'reward_repeat_soft_std': 0.004193334840238094, 'reward_judge_quality_mean': 0.2749999761581421, 'reward_judge_quality_std': 0.155838742852211, 'reward_total_composite_mean': 0.41905075311660767, 'reward_total_composite_std': 0.2089303880929947, 'epoch': 0.04}

 11%|█▏        | 373/3300 [56:36<6:37:59,  8.16s/it]INFO 04-13 08:17:34 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 374/3300 [56:42<6:12:10,  7.63s/it]2026-04-13 08:17:40,469 | INFO | train_grpo_train | metrics_logged mode=train step=374

                                                    
{'loss': 0.0485, 'grad_norm': 15.2440767288208, 'learning_rate': 8.86969696969697e-06, 'num_tokens': 679437.0, 'completions/mean_length': 31.625, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7328658699989319, 'rewards/meter/std': 0.37381434440612793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985873699188232, 'rewards/repeat_soft/std': 0.0030330861918628216, 'rewards/judge_quality/mean': 0.4649999737739563, 'rewards/judge_quality/std': 0.18423588573932648, 'rewards/total_composite/mean': 0.599024772644043, 'rewards/total_composite/std': 0.1625562608242035, 'reward': 0.599024772644043, 'reward_std': 0.1625562459230423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21124495565891266, 'sampling/sampling_logp_difference/max': 1.1573410034179688, 'sampling/importance_sampling_ratio/min': 0.3143208622932434, 'sampling/importance_sampling_ratio/mean': 1.0387402772903442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.707137167453766, 'clip_ratio/low_mean': 0.0766433198004961, 'clip_ratio/low_min': 0.0766433198004961, 'clip_ratio/high_mean': 0.10901551321148872, 'clip_ratio/high_max': 0.10901551321148872, 'clip_ratio/region_mean': 0.18565883301198483, 'reward_total_mean': 0.599024772644043, 'reward_meter_mean': 0.7328658699989319, 'reward_meter_std': 0.37381434440612793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985873699188232, 'reward_repeat_soft_std': 0.0030330861918628216, 'reward_judge_quality_mean': 0.4649999737739563, 'reward_judge_quality_std': 0.18423588573932648, 'reward_total_composite_mean': 0.599024772644043, 'reward_total_composite_std': 0.1625562608242035, 'epoch': 0.04}

 11%|█▏        | 374/3300 [56:42<6:12:10,  7.63s/it]INFO 04-13 08:17:40 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 375/3300 [56:49<6:01:51,  7.42s/it]2026-04-13 08:17:47,381 | INFO | train_grpo_train | metrics_logged mode=train step=375

                                                    
{'loss': 0.0524, 'grad_norm': 24.896814346313477, 'learning_rate': 8.866666666666668e-06, 'num_tokens': 681064.0, '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': 0.7659133672714233, 'rewards/meter/std': 0.3337141275405884, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985313415527344, 'rewards/repeat_soft/std': 0.002027380047366023, 'rewards/judge_quality/mean': 0.8324999809265137, 'rewards/judge_quality/std': 0.18077217042446136, 'rewards/total_composite/mean': 0.7856926321983337, 'rewards/total_composite/std': 0.20556603372097015, 'reward': 0.7856926321983337, 'reward_std': 0.20556604862213135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1270153820514679, 'sampling/sampling_logp_difference/max': 1.3475301265716553, 'sampling/importance_sampling_ratio/min': 0.3013465404510498, 'sampling/importance_sampling_ratio/mean': 1.014768123626709, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8150487747043371, 'clip_ratio/low_mean': 0.08566176518797874, 'clip_ratio/low_min': 0.08566176518797874, 'clip_ratio/high_mean': 0.06432462367229164, 'clip_ratio/high_max': 0.06432462367229164, 'clip_ratio/region_mean': 0.14998638886027038, 'reward_total_mean': 0.7856926321983337, 'reward_meter_mean': 0.7659133672714233, 'reward_meter_std': 0.3337141275405884, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985313415527344, 'reward_repeat_soft_std': 0.002027380047366023, 'reward_judge_quality_mean': 0.8324999809265137, 'reward_judge_quality_std': 0.18077217042446136, 'reward_total_composite_mean': 0.7856926321983337, 'reward_total_composite_std': 0.20556603372097015, 'epoch': 0.04}

 11%|█▏        | 375/3300 [56:49<6:01:51,  7.42s/it]INFO 04-13 08:17:47 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 376/3300 [56:55<5:43:34,  7.05s/it]2026-04-13 08:17:53,560 | INFO | train_grpo_train | metrics_logged mode=train step=376

                                                    
{'loss': 0.0104, 'grad_norm': 21.54682731628418, 'learning_rate': 8.863636363636365e-06, 'num_tokens': 682490.0, 'completions/mean_length': 20.25, 'completions/min_length': 18.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8666355013847351, 'rewards/meter/std': 0.23355844616889954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3687500059604645, 'rewards/judge_quality/std': 0.10802611708641052, 'rewards/total_composite/mean': 0.5619431734085083, 'rewards/total_composite/std': 0.08701276779174805, 'reward': 0.5619431734085083, 'reward_std': 0.08701277524232864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2312190681695938, 'sampling/sampling_logp_difference/max': 1.1098909378051758, 'sampling/importance_sampling_ratio/min': 0.3295949101448059, 'sampling/importance_sampling_ratio/mean': 1.0532430410385132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.901041753590107, 'clip_ratio/low_mean': 0.0575396828353405, 'clip_ratio/low_min': 0.0575396828353405, 'clip_ratio/high_mean': 0.11208952125161886, 'clip_ratio/high_max': 0.11208952125161886, 'clip_ratio/region_mean': 0.16962920408695936, 'reward_total_mean': 0.5619431734085083, 'reward_meter_mean': 0.8666355013847351, 'reward_meter_std': 0.23355844616889954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3687500059604645, 'reward_judge_quality_std': 0.10802611708641052, 'reward_total_composite_mean': 0.5619431734085083, 'reward_total_composite_std': 0.08701276779174805, 'epoch': 0.04}

 11%|█▏        | 376/3300 [56:55<5:43:34,  7.05s/it]INFO 04-13 08:17:53 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 377/3300 [57:02<5:40:42,  6.99s/it]2026-04-13 08:18:00,424 | INFO | train_grpo_train | metrics_logged mode=train step=377

                                                    
{'loss': 0.0295, 'grad_norm': 13.482222557067871, 'learning_rate': 8.860606060606062e-06, 'num_tokens': 684398.0, 'completions/mean_length': 63.5, 'completions/min_length': 61.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8272577524185181, 'rewards/meter/std': 0.24018819630146027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.994218647480011, 'rewards/repeat_soft/std': 0.005557127296924591, 'rewards/judge_quality/mean': 0.5149999856948853, 'rewards/judge_quality/std': 0.18031719326972961, 'rewards/total_composite/mean': 0.6300148963928223, 'rewards/total_composite/std': 0.13305138051509857, 'reward': 0.6300148963928223, 'reward_std': 0.133051335811615, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22164733707904816, 'sampling/sampling_logp_difference/max': 1.8124122619628906, 'sampling/importance_sampling_ratio/min': 0.1632598340511322, 'sampling/importance_sampling_ratio/mean': 1.0644114017486572, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9146456122398376, 'clip_ratio/low_mean': 0.09237921983003616, 'clip_ratio/low_min': 0.09237921983003616, 'clip_ratio/high_mean': 0.09039351716637611, 'clip_ratio/high_max': 0.09039351716637611, 'clip_ratio/region_mean': 0.18277273699641228, 'reward_total_mean': 0.6300148963928223, 'reward_meter_mean': 0.8272577524185181, 'reward_meter_std': 0.24018819630146027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.994218647480011, 'reward_repeat_soft_std': 0.005557127296924591, 'reward_judge_quality_mean': 0.5149999856948853, 'reward_judge_quality_std': 0.18031719326972961, 'reward_total_composite_mean': 0.6300148963928223, 'reward_total_composite_std': 0.13305138051509857, 'epoch': 0.04}

 11%|█▏        | 377/3300 [57:02<5:40:42,  6.99s/it]INFO 04-13 08:18:00 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 378/3300 [57:09<5:36:57,  6.92s/it]2026-04-13 08:18:07,177 | INFO | train_grpo_train | metrics_logged mode=train step=378

                                                    
{'loss': 0.0172, 'grad_norm': 11.616564750671387, 'learning_rate': 8.857575757575758e-06, 'num_tokens': 686446.0, 'completions/mean_length': 67.0, 'completions/min_length': 61.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.827986478805542, 'rewards/meter/std': 0.24381479620933533, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9585182070732117, 'rewards/repeat_soft/std': 0.05319053307175636, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.5980088114738464, 'rewards/total_composite/std': 0.11041803658008575, 'reward': 0.5980088114738464, 'reward_std': 0.11041803658008575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21592731773853302, 'sampling/sampling_logp_difference/max': 1.395995020866394, 'sampling/importance_sampling_ratio/min': 0.2612544894218445, 'sampling/importance_sampling_ratio/mean': 1.0480549335479736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.346274048089981, 'clip_ratio/low_mean': 0.07993471156805754, 'clip_ratio/low_min': 0.07993471156805754, 'clip_ratio/high_mean': 0.07965071126818657, 'clip_ratio/high_max': 0.07965071126818657, 'clip_ratio/region_mean': 0.1595854228362441, 'reward_total_mean': 0.5980088114738464, 'reward_meter_mean': 0.827986478805542, 'reward_meter_std': 0.24381479620933533, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9585182070732117, 'reward_repeat_soft_std': 0.05319053307175636, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.5980088114738464, 'reward_total_composite_std': 0.11041803658008575, 'epoch': 0.04}

 11%|█▏        | 378/3300 [57:09<5:36:57,  6.92s/it]INFO 04-13 08:18:07 [block_pool.py:378] Successfully reset prefix cache

 11%|█▏        | 379/3300 [57:16<5:42:20,  7.03s/it]2026-04-13 08:18:14,462 | INFO | train_grpo_train | metrics_logged mode=train step=379

                                                    
{'loss': 0.0941, 'grad_norm': 20.9310359954834, 'learning_rate': 8.854545454545455e-06, 'num_tokens': 688064.0, 'completions/mean_length': 32.25, 'completions/min_length': 28.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8344127535820007, 'rewards/meter/std': 0.31773367524147034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9943695068359375, 'rewards/repeat_soft/std': 0.0056973411701619625, 'rewards/judge_quality/mean': 0.518750011920929, 'rewards/judge_quality/std': 0.17908000946044922, 'rewards/total_composite/mean': 0.5999100804328918, 'rewards/total_composite/std': 0.08593261241912842, 'reward': 0.5999100804328918, 'reward_std': 0.08593261241912842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20274046063423157, 'sampling/sampling_logp_difference/max': 1.5597805976867676, 'sampling/importance_sampling_ratio/min': 0.21018217504024506, 'sampling/importance_sampling_ratio/mean': 1.037693738937378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8272703140974045, 'clip_ratio/low_mean': 0.027027027681469917, 'clip_ratio/low_min': 0.027027027681469917, 'clip_ratio/high_mean': 0.17235427629202604, 'clip_ratio/high_max': 0.17235427629202604, 'clip_ratio/region_mean': 0.19938130397349596, 'reward_total_mean': 0.5999100804328918, 'reward_meter_mean': 0.8344127535820007, 'reward_meter_std': 0.31773367524147034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9943695068359375, 'reward_repeat_soft_std': 0.0056973411701619625, 'reward_judge_quality_mean': 0.518750011920929, 'reward_judge_quality_std': 0.17908000946044922, 'reward_total_composite_mean': 0.5999100804328918, 'reward_total_composite_std': 0.08593261241912842, 'epoch': 0.04}

 11%|█▏        | 379/3300 [57:16<5:42:20,  7.03s/it]INFO 04-13 08:18:14 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 380/3300 [57:24<5:49:41,  7.19s/it]2026-04-13 08:18:22,021 | INFO | train_grpo_train | metrics_logged mode=train step=380

                                                    
{'loss': -0.0215, 'grad_norm': 14.866976737976074, 'learning_rate': 8.851515151515152e-06, 'num_tokens': 689762.0, 'completions/mean_length': 49.25, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.2547784149646759, 'rewards/meter/std': 0.32982468605041504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9860686659812927, 'rewards/repeat_soft/std': 0.012328761629760265, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.22937417030334473, 'rewards/total_composite/mean': 0.4310709834098816, 'rewards/total_composite/std': 0.09128675609827042, 'reward': 0.4310709834098816, 'reward_std': 0.09128675609827042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2316330373287201, 'sampling/sampling_logp_difference/max': 3.0058579444885254, 'sampling/importance_sampling_ratio/min': 0.04949627071619034, 'sampling/importance_sampling_ratio/mean': 1.0419652462005615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.449590340256691, 'clip_ratio/low_mean': 0.11093355249613523, 'clip_ratio/low_min': 0.11093355249613523, 'clip_ratio/high_mean': 0.07752490043640137, 'clip_ratio/high_max': 0.07752490043640137, 'clip_ratio/region_mean': 0.1884584529325366, 'reward_total_mean': 0.4310709834098816, 'reward_meter_mean': 0.2547784149646759, 'reward_meter_std': 0.32982468605041504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9860686659812927, 'reward_repeat_soft_std': 0.012328761629760265, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.22937417030334473, 'reward_total_composite_mean': 0.4310709834098816, 'reward_total_composite_std': 0.09128675609827042, 'epoch': 0.04}

 12%|█▏        | 380/3300 [57:24<5:49:41,  7.19s/it]INFO 04-13 08:18:22 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 381/3300 [57:31<5:44:33,  7.08s/it]2026-04-13 08:18:28,848 | INFO | train_grpo_train | metrics_logged mode=train step=381

                                                    
{'loss': 0.0262, 'grad_norm': 22.278461456298828, 'learning_rate': 8.84848484848485e-06, 'num_tokens': 691310.0, 'completions/mean_length': 33.5, 'completions/min_length': 26.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6267951130867004, 'rewards/meter/std': 0.3689000904560089, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9895439147949219, 'rewards/repeat_soft/std': 0.009762046858668327, 'rewards/judge_quality/mean': 0.6487500071525574, 'rewards/judge_quality/std': 0.2456151396036148, 'rewards/total_composite/mean': 0.5943039059638977, 'rewards/total_composite/std': 0.31039348244667053, 'reward': 0.5943039059638977, 'reward_std': 0.31039348244667053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21351630985736847, 'sampling/sampling_logp_difference/max': 2.1638095378875732, 'sampling/importance_sampling_ratio/min': 0.11488661915063858, 'sampling/importance_sampling_ratio/mean': 1.0162560939788818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.262278027832508, 'clip_ratio/low_mean': 0.09266827069222927, 'clip_ratio/low_min': 0.09266827069222927, 'clip_ratio/high_mean': 0.08275891002267599, 'clip_ratio/high_max': 0.08275891002267599, 'clip_ratio/region_mean': 0.17542718071490526, 'reward_total_mean': 0.5943039059638977, 'reward_meter_mean': 0.6267951130867004, 'reward_meter_std': 0.3689000904560089, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9895439147949219, 'reward_repeat_soft_std': 0.009762046858668327, 'reward_judge_quality_mean': 0.6487500071525574, 'reward_judge_quality_std': 0.2456151396036148, 'reward_total_composite_mean': 0.5943039059638977, 'reward_total_composite_std': 0.31039348244667053, 'epoch': 0.04}

 12%|█▏        | 381/3300 [57:31<5:44:33,  7.08s/it]INFO 04-13 08:18:29 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 382/3300 [57:37<5:37:11,  6.93s/it]2026-04-13 08:18:35,537 | INFO | train_grpo_train | metrics_logged mode=train step=382

                                                    
{'loss': 0.0568, 'grad_norm': 13.969280242919922, 'learning_rate': 8.845454545454547e-06, 'num_tokens': 693068.0, 'completions/mean_length': 58.75, 'completions/min_length': 53.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.4683726131916046, 'rewards/meter/std': 0.27449503540992737, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.994033694267273, 'rewards/repeat_soft/std': 0.00757418991997838, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.4857119619846344, 'rewards/total_composite/std': 0.09801746904850006, 'reward': 0.4857119619846344, 'reward_std': 0.09801746904850006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21199508011341095, 'sampling/sampling_logp_difference/max': 1.337010383605957, 'sampling/importance_sampling_ratio/min': 0.2626296579837799, 'sampling/importance_sampling_ratio/mean': 1.0761182308197021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5622952729463577, 'clip_ratio/low_mean': 0.12251085415482521, 'clip_ratio/low_min': 0.12251085415482521, 'clip_ratio/high_mean': 0.0687596295028925, 'clip_ratio/high_max': 0.0687596295028925, 'clip_ratio/region_mean': 0.1912704836577177, 'reward_total_mean': 0.4857119619846344, 'reward_meter_mean': 0.4683726131916046, 'reward_meter_std': 0.27449503540992737, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.994033694267273, 'reward_repeat_soft_std': 0.00757418991997838, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.4857119619846344, 'reward_total_composite_std': 0.09801746904850006, 'epoch': 0.04}

 12%|█▏        | 382/3300 [57:37<5:37:11,  6.93s/it]INFO 04-13 08:18:35 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 383/3300 [57:45<5:45:20,  7.10s/it]2026-04-13 08:18:42,946 | INFO | train_grpo_train | metrics_logged mode=train step=383

                                                    
{'loss': 0.0829, 'grad_norm': 14.486379623413086, 'learning_rate': 8.842424242424244e-06, 'num_tokens': 695375.0, 'completions/mean_length': 85.375, 'completions/min_length': 78.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.41754233837127686, 'rewards/meter/std': 0.2040727734565735, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9971754550933838, 'rewards/repeat_soft/std': 0.0023173748049885035, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.39841270446777344, 'rewards/total_composite/std': 0.17113541066646576, 'reward': 0.39841270446777344, 'reward_std': 0.17113539576530457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24437576532363892, 'sampling/sampling_logp_difference/max': 1.599085807800293, 'sampling/importance_sampling_ratio/min': 0.20208117365837097, 'sampling/importance_sampling_ratio/mean': 1.0334339141845703, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.323674261569977, 'clip_ratio/low_mean': 0.051868095993995667, 'clip_ratio/low_min': 0.051868095993995667, 'clip_ratio/high_mean': 0.17507386952638626, 'clip_ratio/high_max': 0.17507386952638626, 'clip_ratio/region_mean': 0.22694196552038193, 'reward_total_mean': 0.39841270446777344, 'reward_meter_mean': 0.41754233837127686, 'reward_meter_std': 0.2040727734565735, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9971754550933838, 'reward_repeat_soft_std': 0.0023173748049885035, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.39841270446777344, 'reward_total_composite_std': 0.17113541066646576, 'epoch': 0.04}

 12%|█▏        | 383/3300 [57:45<5:45:20,  7.10s/it]INFO 04-13 08:18:43 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 384/3300 [57:52<5:45:03,  7.10s/it]2026-04-13 08:18:50,026 | INFO | train_grpo_train | metrics_logged mode=train step=384

                                                    
{'loss': -0.0105, 'grad_norm': 17.235023498535156, 'learning_rate': 8.83939393939394e-06, 'num_tokens': 696932.0, 'completions/mean_length': 35.625, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8582241535186768, 'rewards/meter/std': 0.31072792410850525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995353102684021, 'rewards/repeat_soft/std': 0.0044983429834246635, 'rewards/judge_quality/mean': 0.5399999618530273, 'rewards/judge_quality/std': 0.2218751311302185, 'rewards/total_composite/mean': 0.658287763595581, 'rewards/total_composite/std': 0.1771986037492752, 'reward': 0.658287763595581, 'reward_std': 0.1771986037492752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22254599630832672, 'sampling/sampling_logp_difference/max': 1.4940357208251953, 'sampling/importance_sampling_ratio/min': 0.22446495294570923, 'sampling/importance_sampling_ratio/mean': 1.022464632987976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8612878322601318, 'clip_ratio/low_mean': 0.10957883857190609, 'clip_ratio/low_min': 0.10957883857190609, 'clip_ratio/high_mean': 0.06023901887238026, 'clip_ratio/high_max': 0.06023901887238026, 'clip_ratio/region_mean': 0.16981785744428635, 'reward_total_mean': 0.658287763595581, 'reward_meter_mean': 0.8582241535186768, 'reward_meter_std': 0.31072792410850525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995353102684021, 'reward_repeat_soft_std': 0.0044983429834246635, 'reward_judge_quality_mean': 0.5399999618530273, 'reward_judge_quality_std': 0.2218751311302185, 'reward_total_composite_mean': 0.658287763595581, 'reward_total_composite_std': 0.1771986037492752, 'epoch': 0.04}

 12%|█▏        | 384/3300 [57:52<5:45:03,  7.10s/it]INFO 04-13 08:18:50 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 385/3300 [57:58<5:34:06,  6.88s/it]2026-04-13 08:18:56,383 | INFO | train_grpo_train | metrics_logged mode=train step=385

                                                    
{'loss': -0.0119, 'grad_norm': 19.83598518371582, 'learning_rate': 8.836363636363637e-06, 'num_tokens': 698256.0, 'completions/mean_length': 18.5, 'completions/min_length': 15.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.5, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.5626965761184692, 'rewards/meter/std': 0.42163652181625366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5029791593551636, 'rewards/total_composite/std': 0.12039375305175781, 'reward': 0.5029791593551636, 'reward_std': 0.12039375305175781, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13141751289367676, 'sampling/sampling_logp_difference/max': 1.0829339027404785, 'sampling/importance_sampling_ratio/min': 0.33860063552856445, 'sampling/importance_sampling_ratio/mean': 1.0016568899154663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6041314825415611, 'clip_ratio/low_mean': 0.050520834513008595, 'clip_ratio/low_min': 0.050520834513008595, 'clip_ratio/high_mean': 0.05680868960916996, 'clip_ratio/high_max': 0.05680868960916996, 'clip_ratio/region_mean': 0.10732952412217855, 'reward_total_mean': 0.5029791593551636, 'reward_meter_mean': 0.5626965761184692, 'reward_meter_std': 0.42163652181625366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5029791593551636, 'reward_total_composite_std': 0.12039375305175781, 'epoch': 0.04}

 12%|█▏        | 385/3300 [57:58<5:34:06,  6.88s/it]INFO 04-13 08:18:56 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 386/3300 [58:05<5:34:33,  6.89s/it]2026-04-13 08:19:03,299 | INFO | train_grpo_train | metrics_logged mode=train step=386

                                                    
{'loss': -0.0405, 'grad_norm': 16.2897891998291, 'learning_rate': 8.833333333333334e-06, 'num_tokens': 700003.0, 'completions/mean_length': 46.375, 'completions/min_length': 42.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6400235891342163, 'rewards/meter/std': 0.370656281709671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928402900695801, 'rewards/repeat_soft/std': 0.011342632584273815, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.1538088172674179, 'rewards/total_composite/mean': 0.5819841027259827, 'rewards/total_composite/std': 0.1263429969549179, 'reward': 0.5819841027259827, 'reward_std': 0.1263429820537567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2091495394706726, 'sampling/sampling_logp_difference/max': 2.2717998027801514, 'sampling/importance_sampling_ratio/min': 0.1031264066696167, 'sampling/importance_sampling_ratio/mean': 1.0420584678649902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9817766696214676, 'clip_ratio/low_mean': 0.06466450355947018, 'clip_ratio/low_min': 0.06466450355947018, 'clip_ratio/high_mean': 0.1348732877522707, 'clip_ratio/high_max': 0.1348732877522707, 'clip_ratio/region_mean': 0.19953779131174088, 'reward_total_mean': 0.5819841027259827, 'reward_meter_mean': 0.6400235891342163, 'reward_meter_std': 0.370656281709671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928402900695801, 'reward_repeat_soft_std': 0.011342632584273815, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.1538088172674179, 'reward_total_composite_mean': 0.5819841027259827, 'reward_total_composite_std': 0.1263429969549179, 'epoch': 0.04}

 12%|█▏        | 386/3300 [58:05<5:34:33,  6.89s/it]INFO 04-13 08:19:03 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 387/3300 [58:12<5:31:07,  6.82s/it]2026-04-13 08:19:09,960 | INFO | train_grpo_train | metrics_logged mode=train step=387

                                                    
{'loss': 0.0497, 'grad_norm': 12.300729751586914, 'learning_rate': 8.830303030303031e-06, 'num_tokens': 701732.0, 'completions/mean_length': 50.125, 'completions/min_length': 44.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4203408360481262, 'rewards/meter/std': 0.4558182656764984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9930034875869751, 'rewards/repeat_soft/std': 0.014712625183165073, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.1302744597196579, 'rewards/total_composite/mean': 0.4715820252895355, 'rewards/total_composite/std': 0.1289324164390564, 'reward': 0.4715820252895355, 'reward_std': 0.1289324015378952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16170267760753632, 'sampling/sampling_logp_difference/max': 2.8092567920684814, 'sampling/importance_sampling_ratio/min': 0.060249753296375275, 'sampling/importance_sampling_ratio/mean': 1.0188674926757812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0982524752616882, 'clip_ratio/low_mean': 0.0752681796438992, 'clip_ratio/low_min': 0.0752681796438992, 'clip_ratio/high_mean': 0.06160464696586132, 'clip_ratio/high_max': 0.06160464696586132, 'clip_ratio/region_mean': 0.13687282660976052, 'reward_total_mean': 0.4715820252895355, 'reward_meter_mean': 0.4203408360481262, 'reward_meter_std': 0.4558182656764984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9930034875869751, 'reward_repeat_soft_std': 0.014712625183165073, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.1302744597196579, 'reward_total_composite_mean': 0.4715820252895355, 'reward_total_composite_std': 0.1289324164390564, 'epoch': 0.04}

 12%|█▏        | 387/3300 [58:12<5:31:07,  6.82s/it]INFO 04-13 08:19:10 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 388/3300 [58:23<6:41:07,  8.26s/it]2026-04-13 08:19:21,595 | INFO | train_grpo_train | metrics_logged mode=train step=388

                                                    
{'loss': -0.1057, 'grad_norm': 2.612093448638916, 'learning_rate': 8.827272727272727e-06, 'num_tokens': 703180.0, 'completions/mean_length': 95.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.42857360839844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9530330896377563, 'rewards/meter/std': 0.08797551691532135, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9803513288497925, 'rewards/repeat_soft/std': 0.03263489902019501, 'rewards/judge_quality/mean': 0.6299999952316284, 'rewards/judge_quality/std': 0.2958281338214874, 'rewards/total_composite/mean': 0.7036803364753723, 'rewards/total_composite/std': 0.3078506886959076, 'reward': 0.7036803364753723, 'reward_std': 0.3078506886959076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21610069274902344, 'sampling/sampling_logp_difference/max': 1.2582550048828125, 'sampling/importance_sampling_ratio/min': 0.2841494381427765, 'sampling/importance_sampling_ratio/mean': 1.0802019834518433, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3301824927330017, 'clip_ratio/low_mean': 0.04922780115157366, 'clip_ratio/low_min': 0.04922780115157366, 'clip_ratio/high_mean': 0.10317971557378769, 'clip_ratio/high_max': 0.10317971557378769, 'clip_ratio/region_mean': 0.15240751672536135, 'reward_total_mean': 0.7036803364753723, 'reward_meter_mean': 0.9530330896377563, 'reward_meter_std': 0.08797551691532135, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9803513288497925, 'reward_repeat_soft_std': 0.03263489902019501, 'reward_judge_quality_mean': 0.6299999952316284, 'reward_judge_quality_std': 0.2958281338214874, 'reward_total_composite_mean': 0.7036803364753723, 'reward_total_composite_std': 0.3078506886959076, 'epoch': 0.04}

 12%|█▏        | 388/3300 [58:23<6:41:07,  8.26s/it]INFO 04-13 08:19:21 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 389/3300 [58:35<7:33:18,  9.34s/it]2026-04-13 08:19:33,456 | INFO | train_grpo_train | metrics_logged mode=train step=389

                                                    
{'loss': -0.0934, 'grad_norm': 5.493145942687988, 'learning_rate': 8.824242424242426e-06, 'num_tokens': 705116.0, 'completions/mean_length': 129.0, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.28572082519531, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.8792860507965088, 'rewards/meter/std': 0.23752202093601227, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9941399097442627, 'rewards/repeat_soft/std': 0.0030691882129758596, 'rewards/judge_quality/mean': 0.3687500059604645, 'rewards/judge_quality/std': 0.1940867006778717, 'rewards/total_composite/mean': 0.43528708815574646, 'rewards/total_composite/std': 0.2920048236846924, 'reward': 0.43528708815574646, 'reward_std': 0.2920048236846924, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23220102488994598, 'sampling/sampling_logp_difference/max': 1.0492897033691406, 'sampling/importance_sampling_ratio/min': 0.35018637776374817, 'sampling/importance_sampling_ratio/mean': 1.0767548084259033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9076579213142395, 'clip_ratio/low_mean': 0.04516545124351978, 'clip_ratio/low_min': 0.04516545124351978, 'clip_ratio/high_mean': 0.09360717888921499, 'clip_ratio/high_max': 0.09360717888921499, 'clip_ratio/region_mean': 0.13877263013273478, 'reward_total_mean': 0.43528708815574646, 'reward_meter_mean': 0.8792860507965088, 'reward_meter_std': 0.23752202093601227, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9941399097442627, 'reward_repeat_soft_std': 0.0030691882129758596, 'reward_judge_quality_mean': 0.3687500059604645, 'reward_judge_quality_std': 0.1940867006778717, 'reward_total_composite_mean': 0.43528708815574646, 'reward_total_composite_std': 0.2920048236846924, 'epoch': 0.04}

 12%|█▏        | 389/3300 [58:35<7:33:18,  9.34s/it]INFO 04-13 08:19:33 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 390/3300 [58:42<6:59:10,  8.64s/it]2026-04-13 08:19:40,465 | INFO | train_grpo_train | metrics_logged mode=train step=390

                                                    
{'loss': 0.0821, 'grad_norm': 15.537460327148438, 'learning_rate': 8.821212121212121e-06, 'num_tokens': 707297.0, 'completions/mean_length': 82.625, 'completions/min_length': 73.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.625, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.6493761539459229, 'rewards/meter/std': 0.27126702666282654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9950517416000366, 'rewards/repeat_soft/std': 0.0026858015917241573, 'rewards/judge_quality/mean': 0.6574999690055847, 'rewards/judge_quality/std': 0.1505940854549408, 'rewards/total_composite/mean': 0.6419288516044617, 'rewards/total_composite/std': 0.1497366577386856, 'reward': 0.6419288516044617, 'reward_std': 0.1497366577386856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23744165897369385, 'sampling/sampling_logp_difference/max': 3.6390271186828613, 'sampling/importance_sampling_ratio/min': 0.026277896016836166, 'sampling/importance_sampling_ratio/mean': 0.9899482727050781, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4409097135066986, 'clip_ratio/low_mean': 0.1076983530074358, 'clip_ratio/low_min': 0.1076983530074358, 'clip_ratio/high_mean': 0.10639841854572296, 'clip_ratio/high_max': 0.10639841854572296, 'clip_ratio/region_mean': 0.21409677155315876, 'reward_total_mean': 0.6419288516044617, 'reward_meter_mean': 0.6493761539459229, 'reward_meter_std': 0.27126702666282654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9950517416000366, 'reward_repeat_soft_std': 0.0026858015917241573, 'reward_judge_quality_mean': 0.6574999690055847, 'reward_judge_quality_std': 0.1505940854549408, 'reward_total_composite_mean': 0.6419288516044617, 'reward_total_composite_std': 0.1497366577386856, 'epoch': 0.04}

 12%|█▏        | 390/3300 [58:42<6:59:10,  8.64s/it]INFO 04-13 08:19:40 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 391/3300 [58:48<6:23:10,  7.90s/it]2026-04-13 08:19:46,640 | INFO | train_grpo_train | metrics_logged mode=train step=391

                                                    
{'loss': 0.0345, 'grad_norm': 19.484699249267578, 'learning_rate': 8.818181818181819e-06, 'num_tokens': 708686.0, 'completions/mean_length': 33.625, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6881487965583801, 'rewards/meter/std': 0.4074837863445282, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9906495213508606, 'rewards/repeat_soft/std': 0.014289024285972118, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5396751165390015, 'rewards/total_composite/std': 0.1060907319188118, 'reward': 0.5396751165390015, 'reward_std': 0.1060907244682312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19514882564544678, 'sampling/sampling_logp_difference/max': 1.788926601409912, 'sampling/importance_sampling_ratio/min': 0.16713947057724, 'sampling/importance_sampling_ratio/mean': 1.0347812175750732, 'sampling/importance_sampling_ratio/max': 1.9738614559173584, 'entropy': 1.9678066968917847, 'clip_ratio/low_mean': 0.055244164541363716, 'clip_ratio/low_min': 0.055244164541363716, 'clip_ratio/high_mean': 0.14656863175332546, 'clip_ratio/high_max': 0.14656863175332546, 'clip_ratio/region_mean': 0.20181279629468918, 'reward_total_mean': 0.5396751165390015, 'reward_meter_mean': 0.6881487965583801, 'reward_meter_std': 0.4074837863445282, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9906495213508606, 'reward_repeat_soft_std': 0.014289024285972118, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5396751165390015, 'reward_total_composite_std': 0.1060907319188118, 'epoch': 0.04}

 12%|█▏        | 391/3300 [58:48<6:23:10,  7.90s/it]INFO 04-13 08:19:46 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 392/3300 [58:55<6:08:00,  7.59s/it]2026-04-13 08:19:53,513 | INFO | train_grpo_train | metrics_logged mode=train step=392

                                                    
{'loss': -0.0108, 'grad_norm': 13.432182312011719, 'learning_rate': 8.815151515151516e-06, 'num_tokens': 710620.0, 'completions/mean_length': 65.75, 'completions/min_length': 46.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.17572955787181854, 'rewards/meter/std': 0.13979798555374146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9765012264251709, 'rewards/repeat_soft/std': 0.026776380836963654, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.39454472064971924, 'rewards/total_composite/std': 0.03480697423219681, 'reward': 0.39454472064971924, 'reward_std': 0.03480697423219681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21580170094966888, 'sampling/sampling_logp_difference/max': 1.3005542755126953, 'sampling/importance_sampling_ratio/min': 0.2723807692527771, 'sampling/importance_sampling_ratio/mean': 1.0618659257888794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9697498977184296, 'clip_ratio/low_mean': 0.05944700539112091, 'clip_ratio/low_min': 0.05944700539112091, 'clip_ratio/high_mean': 0.1420013289898634, 'clip_ratio/high_max': 0.1420013289898634, 'clip_ratio/region_mean': 0.2014483343809843, 'reward_total_mean': 0.39454472064971924, 'reward_meter_mean': 0.17572955787181854, 'reward_meter_std': 0.13979798555374146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9765012264251709, 'reward_repeat_soft_std': 0.026776380836963654, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.39454472064971924, 'reward_total_composite_std': 0.03480697423219681, 'epoch': 0.04}

 12%|█▏        | 392/3300 [58:55<6:08:00,  7.59s/it]INFO 04-13 08:19:53 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 393/3300 [59:02<6:00:25,  7.44s/it]2026-04-13 08:20:00,590 | INFO | train_grpo_train | metrics_logged mode=train step=393

                                                    
{'loss': 0.1136, 'grad_norm': 13.407106399536133, 'learning_rate': 8.812121212121213e-06, 'num_tokens': 712500.0, 'completions/mean_length': 68.0, 'completions/min_length': 56.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.56876540184021, 'rewards/meter/std': 0.38598161935806274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9751752614974976, 'rewards/repeat_soft/std': 0.026960745453834534, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.5403320789337158, 'rewards/total_composite/std': 0.17929324507713318, 'reward': 0.5403320789337158, 'reward_std': 0.17929324507713318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22484685480594635, 'sampling/sampling_logp_difference/max': 1.744100570678711, 'sampling/importance_sampling_ratio/min': 0.17480213940143585, 'sampling/importance_sampling_ratio/mean': 1.0354881286621094, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3533735424280167, 'clip_ratio/low_mean': 0.1128325741738081, 'clip_ratio/low_min': 0.1128325741738081, 'clip_ratio/high_mean': 0.05654893722385168, 'clip_ratio/high_max': 0.05654893722385168, 'clip_ratio/region_mean': 0.16938151139765978, 'reward_total_mean': 0.5403320789337158, 'reward_meter_mean': 0.56876540184021, 'reward_meter_std': 0.38598161935806274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9751752614974976, 'reward_repeat_soft_std': 0.026960745453834534, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.5403320789337158, 'reward_total_composite_std': 0.17929324507713318, 'epoch': 0.04}

 12%|█▏        | 393/3300 [59:02<6:00:25,  7.44s/it]INFO 04-13 08:20:00 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 394/3300 [59:10<5:57:14,  7.38s/it]2026-04-13 08:20:07,819 | INFO | train_grpo_train | metrics_logged mode=train step=394

                                                    
{'loss': 0.0969, 'grad_norm': 14.463814735412598, 'learning_rate': 8.809090909090909e-06, 'num_tokens': 714150.0, 'completions/mean_length': 39.25, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.49386802315711975, 'rewards/meter/std': 0.3387821912765503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.992102861404419, 'rewards/repeat_soft/std': 0.008454841561615467, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.15191635489463806, 'rewards/total_composite/mean': 0.5453914403915405, 'rewards/total_composite/std': 0.1739407628774643, 'reward': 0.5453914403915405, 'reward_std': 0.1739407628774643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18999524414539337, 'sampling/sampling_logp_difference/max': 2.017080307006836, 'sampling/importance_sampling_ratio/min': 0.1330433487892151, 'sampling/importance_sampling_ratio/mean': 1.034308910369873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5738208591938019, 'clip_ratio/low_mean': 0.12540349457412958, 'clip_ratio/low_min': 0.12540349457412958, 'clip_ratio/high_mean': 0.04475806374102831, 'clip_ratio/high_max': 0.04475806374102831, 'clip_ratio/region_mean': 0.1701615583151579, 'reward_total_mean': 0.5453914403915405, 'reward_meter_mean': 0.49386802315711975, 'reward_meter_std': 0.3387821912765503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.992102861404419, 'reward_repeat_soft_std': 0.008454841561615467, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.15191635489463806, 'reward_total_composite_mean': 0.5453914403915405, 'reward_total_composite_std': 0.1739407628774643, 'epoch': 0.04}

 12%|█▏        | 394/3300 [59:10<5:57:14,  7.38s/it]INFO 04-13 08:20:08 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 395/3300 [59:21<7:02:43,  8.73s/it]2026-04-13 08:20:19,711 | INFO | train_grpo_train | metrics_logged mode=train step=395

                                                    
{'loss': -0.0986, 'grad_norm': 5.845329284667969, 'learning_rate': 8.806060606060608e-06, 'num_tokens': 716294.0, 'completions/mean_length': 158.0, 'completions/min_length': 86.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 107.42857360839844, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8389735221862793, 'rewards/meter/std': 0.2774248421192169, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.976121723651886, 'rewards/repeat_soft/std': 0.017020486295223236, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.14961259067058563, 'rewards/total_composite/mean': 0.45580732822418213, 'rewards/total_composite/std': 0.20585764944553375, 'reward': 0.45580732822418213, 'reward_std': 0.20585763454437256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2123926728963852, 'sampling/sampling_logp_difference/max': 1.4099535942077637, 'sampling/importance_sampling_ratio/min': 0.24415461719036102, 'sampling/importance_sampling_ratio/mean': 1.0625028610229492, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.605967551469803, 'clip_ratio/low_mean': 0.044079553335905075, 'clip_ratio/low_min': 0.044079553335905075, 'clip_ratio/high_mean': 0.09661378152668476, 'clip_ratio/high_max': 0.09661378152668476, 'clip_ratio/region_mean': 0.14069333486258984, 'reward_total_mean': 0.45580732822418213, 'reward_meter_mean': 0.8389735221862793, 'reward_meter_std': 0.2774248421192169, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.976121723651886, 'reward_repeat_soft_std': 0.017020486295223236, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.14961259067058563, 'reward_total_composite_mean': 0.45580732822418213, 'reward_total_composite_std': 0.20585764944553375, 'epoch': 0.04}

 12%|█▏        | 395/3300 [59:21<7:02:43,  8.73s/it]INFO 04-13 08:20:19 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 396/3300 [59:30<7:06:31,  8.81s/it]2026-04-13 08:20:28,714 | INFO | train_grpo_train | metrics_logged mode=train step=396

                                                    
{'loss': 0.3118, 'grad_norm': 10.84103012084961, 'learning_rate': 8.803030303030303e-06, 'num_tokens': 718666.0, 'completions/mean_length': 107.5, 'completions/min_length': 70.0, 'completions/max_length': 221.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 221.0, 'rewards/meter/mean': 0.6435260772705078, 'rewards/meter/std': 0.40282678604125977, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.21380899846553802, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9662259817123413, 'rewards/repeat_soft/std': 0.02086596004664898, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.2121320515871048, 'rewards/total_composite/mean': 0.447258323431015, 'rewards/total_composite/std': 0.31164485216140747, 'reward': 0.447258323431015, 'reward_std': 0.31164485216140747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21521738171577454, 'sampling/sampling_logp_difference/max': 1.9176673889160156, 'sampling/importance_sampling_ratio/min': 0.14694933593273163, 'sampling/importance_sampling_ratio/mean': 1.041980504989624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2222574949264526, 'clip_ratio/low_mean': 0.05741437803953886, 'clip_ratio/low_min': 0.05741437803953886, 'clip_ratio/high_mean': 0.11796465516090393, 'clip_ratio/high_max': 0.11796465516090393, 'clip_ratio/region_mean': 0.1753790332004428, 'reward_total_mean': 0.447258323431015, 'reward_meter_mean': 0.6435260772705078, 'reward_meter_std': 0.40282678604125977, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.21380899846553802, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9662259817123413, 'reward_repeat_soft_std': 0.02086596004664898, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.2121320515871048, 'reward_total_composite_mean': 0.447258323431015, 'reward_total_composite_std': 0.31164485216140747, 'epoch': 0.04}

 12%|█▏        | 396/3300 [59:30<7:06:31,  8.81s/it]INFO 04-13 08:20:28 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 397/3300 [59:37<6:32:37,  8.11s/it]2026-04-13 08:20:35,202 | INFO | train_grpo_train | metrics_logged mode=train step=397

                                                    
{'loss': -0.029, 'grad_norm': 13.39407730102539, 'learning_rate': 8.8e-06, 'num_tokens': 720361.0, 'completions/mean_length': 40.875, 'completions/min_length': 36.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7563626766204834, 'rewards/meter/std': 0.34741678833961487, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957313537597656, 'rewards/repeat_soft/std': 0.004118890967220068, 'rewards/judge_quality/mean': 0.6474999785423279, 'rewards/judge_quality/std': 0.2365073412656784, 'rewards/total_composite/mean': 0.6738553047180176, 'rewards/total_composite/std': 0.18831169605255127, 'reward': 0.6738553047180176, 'reward_std': 0.18831169605255127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1969454139471054, 'sampling/sampling_logp_difference/max': 1.505446434020996, 'sampling/importance_sampling_ratio/min': 0.22191819548606873, 'sampling/importance_sampling_ratio/mean': 1.0099977254867554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6303172707557678, 'clip_ratio/low_mean': 0.09531616419553757, 'clip_ratio/low_min': 0.09531616419553757, 'clip_ratio/high_mean': 0.06741962395608425, 'clip_ratio/high_max': 0.06741962395608425, 'clip_ratio/region_mean': 0.16273578815162182, 'reward_total_mean': 0.6738553047180176, 'reward_meter_mean': 0.7563626766204834, 'reward_meter_std': 0.34741678833961487, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957313537597656, 'reward_repeat_soft_std': 0.004118890967220068, 'reward_judge_quality_mean': 0.6474999785423279, 'reward_judge_quality_std': 0.2365073412656784, 'reward_total_composite_mean': 0.6738553047180176, 'reward_total_composite_std': 0.18831169605255127, 'epoch': 0.04}

 12%|█▏        | 397/3300 [59:37<6:32:37,  8.11s/it]INFO 04-13 08:20:35 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 398/3300 [59:44<6:16:32,  7.78s/it]2026-04-13 08:20:42,217 | INFO | train_grpo_train | metrics_logged mode=train step=398

                                                    
{'loss': 0.1119, 'grad_norm': 18.618391036987305, 'learning_rate': 8.796969696969698e-06, 'num_tokens': 722027.0, 'completions/mean_length': 39.25, 'completions/min_length': 30.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.692610502243042, 'rewards/meter/std': 0.39874544739723206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921975135803223, 'rewards/repeat_soft/std': 0.006021677982062101, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.6342549324035645, 'rewards/total_composite/std': 0.20082521438598633, 'reward': 0.6342549324035645, 'reward_std': 0.20082521438598633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2037847340106964, 'sampling/sampling_logp_difference/max': 1.1941423416137695, 'sampling/importance_sampling_ratio/min': 0.3029636740684509, 'sampling/importance_sampling_ratio/mean': 1.0290169715881348, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.982935443520546, 'clip_ratio/low_mean': 0.12808520812541246, 'clip_ratio/low_min': 0.12808520812541246, 'clip_ratio/high_mean': 0.07797619327902794, 'clip_ratio/high_max': 0.07797619327902794, 'clip_ratio/region_mean': 0.2060614014044404, 'reward_total_mean': 0.6342549324035645, 'reward_meter_mean': 0.692610502243042, 'reward_meter_std': 0.39874544739723206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921975135803223, 'reward_repeat_soft_std': 0.006021677982062101, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.6342549324035645, 'reward_total_composite_std': 0.20082521438598633, 'epoch': 0.04}

 12%|█▏        | 398/3300 [59:44<6:16:32,  7.78s/it]INFO 04-13 08:20:42 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 399/3300 [59:56<7:25:33,  9.22s/it]2026-04-13 08:20:54,770 | INFO | train_grpo_train | metrics_logged mode=train step=399

                                                    
{'loss': 0.1911, 'grad_norm': 4.706899166107178, 'learning_rate': 8.793939393939395e-06, 'num_tokens': 724384.0, 'completions/mean_length': 162.625, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.71429443359375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 309.0, 'rewards/meter/mean': 0.6536551713943481, 'rewards/meter/std': 0.36277490854263306, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9848779439926147, 'rewards/repeat_soft/std': 0.028398148715496063, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.4322788715362549, 'rewards/total_composite/std': 0.2684842050075531, 'reward': 0.4322788715362549, 'reward_std': 0.2684842050075531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19599947333335876, 'sampling/sampling_logp_difference/max': 1.4805917739868164, 'sampling/importance_sampling_ratio/min': 0.22750303149223328, 'sampling/importance_sampling_ratio/mean': 1.030487298965454, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.985503315925598, 'clip_ratio/low_mean': 0.007281553465873003, 'clip_ratio/low_min': 0.007281553465873003, 'clip_ratio/high_mean': 0.15667317807674408, 'clip_ratio/high_max': 0.15667317807674408, 'clip_ratio/region_mean': 0.16395473154261708, 'reward_total_mean': 0.4322788715362549, 'reward_meter_mean': 0.6536551713943481, 'reward_meter_std': 0.36277490854263306, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9848779439926147, 'reward_repeat_soft_std': 0.028398148715496063, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.4322788715362549, 'reward_total_composite_std': 0.2684842050075531, 'epoch': 0.04}

 12%|█▏        | 399/3300 [59:56<7:25:33,  9.22s/it]INFO 04-13 08:20:54 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 400/3300 [1:00:03<6:50:23,  8.49s/it]2026-04-13 08:21:01,571 | INFO | train_grpo_train | metrics_logged mode=train step=400

                                                      
{'loss': 0.0069, 'grad_norm': 17.07670021057129, 'learning_rate': 8.790909090909092e-06, 'num_tokens': 725874.0, 'completions/mean_length': 34.25, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9133437871932983, 'rewards/meter/std': 0.10935598611831665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9907272458076477, 'rewards/repeat_soft/std': 0.0059598698280751705, 'rewards/judge_quality/mean': 0.7275000214576721, 'rewards/judge_quality/std': 0.20119288563728333, 'rewards/total_composite/mean': 0.7812926769256592, 'rewards/total_composite/std': 0.13477379083633423, 'reward': 0.7812926769256592, 'reward_std': 0.13477377593517303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15207621455192566, 'sampling/sampling_logp_difference/max': 1.8708162307739258, 'sampling/importance_sampling_ratio/min': 0.15399791300296783, 'sampling/importance_sampling_ratio/mean': 1.025518774986267, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.907796710729599, 'clip_ratio/low_mean': 0.044287471333518624, 'clip_ratio/low_min': 0.044287471333518624, 'clip_ratio/high_mean': 0.08598564472049475, 'clip_ratio/high_max': 0.08598564472049475, 'clip_ratio/region_mean': 0.13027311605401337, 'reward_total_mean': 0.7812926769256592, 'reward_meter_mean': 0.9133437871932983, 'reward_meter_std': 0.10935598611831665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9907272458076477, 'reward_repeat_soft_std': 0.0059598698280751705, 'reward_judge_quality_mean': 0.7275000214576721, 'reward_judge_quality_std': 0.20119288563728333, 'reward_total_composite_mean': 0.7812926769256592, 'reward_total_composite_std': 0.13477379083633423, 'epoch': 0.04}

 12%|█▏        | 400/3300 [1:00:03<6:50:23,  8.49s/it]INFO 04-13 08:21:01 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:17,  2.15s/it]

 30%|███       | 3/10 [00:13<00:36,  5.22s/it]

 40%|████      | 4/10 [00:17<00:28,  4.76s/it]

 50%|█████     | 5/10 [00:22<00:23,  4.78s/it]

 60%|██████    | 6/10 [00:27<00:18,  4.72s/it]

 70%|███████   | 7/10 [00:32<00:14,  4.85s/it]

 80%|████████  | 8/10 [00:37<00:09,  4.79s/it]

 90%|█████████ | 9/10 [00:41<00:04,  4.67s/it]

100%|██████████| 10/10 [00:46<00:00,  4.68s/it]2026-04-13 08:21:52,053 | INFO | train_grpo_train | metrics_logged mode=eval step=400

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 50.4575, 'eval_samples_per_second': 1.585, 'eval_steps_per_second': 0.198, 'eval_num_tokens': 725874.0, 'eval_completions/mean_length': 71.0625, 'eval_completions/min_length': 27.7, 'eval_completions/max_length': 147.3, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 59.775, 'eval_completions/min_terminated_length': 27.7, 'eval_completions/max_terminated_length': 104.7, 'eval_rewards/meter/mean': 0.6084526628255844, 'eval_rewards/meter/std': 0.3190243661403656, 'eval_rewards/count_adherence/mean': 0.9627083301544189, 'eval_rewards/count_adherence/std': 0.07116240337491035, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.1767766922712326, 'eval_rewards/repeat_soft/mean': 0.976105797290802, 'eval_rewards/repeat_soft/std': 0.028223066963255404, 'eval_rewards/judge_quality/mean': 0.5106249988079071, 'eval_rewards/judge_quality/std': 0.17371627036482096, 'eval_rewards/total_composite/mean': 0.5150679022073745, 'eval_rewards/total_composite/std': 0.17966578379273415, 'eval_reward': 0.5150679022073745, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.13076039776206017, 'eval_sampling/sampling_logp_difference/max': 1.2090730667114258, 'eval_sampling/importance_sampling_ratio/min': 0.304891861975193, 'eval_sampling/importance_sampling_ratio/mean': 1.0383795976638794, 'eval_sampling/importance_sampling_ratio/max': 1.538509440422058, 'eval_entropy': 1.8213319420814513, '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': 0.5150679022073745, 'eval_reward_meter_mean': 0.6084526628255844, 'eval_reward_meter_std': 0.3190243661403656, 'eval_reward_count_adherence_mean': 0.9627083301544189, 'eval_reward_count_adherence_std': 0.07116240337491035, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.1767766922712326, 'eval_reward_repeat_soft_mean': 0.976105797290802, 'eval_reward_repeat_soft_std': 0.028223066963255404, 'eval_reward_judge_quality_mean': 0.5106249988079071, 'eval_reward_judge_quality_std': 0.17371627036482096, 'eval_reward_total_composite_mean': 0.5150679022073745, 'eval_reward_total_composite_std': 0.17966578379273415, 'epoch': 0.04}

 12%|█▏        | 400/3300 [1:00:54<6:50:23,  8.49s/it]

100%|██████████| 10/10 [00:46<00:00,  4.68s/it]

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

 12%|█▏        | 401/3300 [1:01:05<19:39:37, 24.41s/it]2026-04-13 08:22:03,143 | INFO | train_grpo_train | metrics_logged mode=train step=401

                                                       
{'loss': 0.1463, 'grad_norm': 13.34891128540039, 'learning_rate': 8.787878787878788e-06, 'num_tokens': 727854.0, 'completions/mean_length': 67.5, 'completions/min_length': 58.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.3987351059913635, 'rewards/meter/std': 0.24704766273498535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9943833351135254, 'rewards/repeat_soft/std': 0.003714403137564659, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.5330898761749268, 'rewards/total_composite/std': 0.15060415863990784, 'reward': 0.5330898761749268, 'reward_std': 0.15060415863990784, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19010724127292633, 'sampling/sampling_logp_difference/max': 2.565248489379883, 'sampling/importance_sampling_ratio/min': 0.07690006494522095, 'sampling/importance_sampling_ratio/mean': 1.0112743377685547, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2907120138406754, 'clip_ratio/low_mean': 0.10911228507757187, 'clip_ratio/low_min': 0.10911228507757187, 'clip_ratio/high_mean': 0.08028915338218212, 'clip_ratio/high_max': 0.08028915338218212, 'clip_ratio/region_mean': 0.189401438459754, 'reward_total_mean': 0.5330898761749268, 'reward_meter_mean': 0.3987351059913635, 'reward_meter_std': 0.24704766273498535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9943833351135254, 'reward_repeat_soft_std': 0.003714403137564659, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.5330898761749268, 'reward_total_composite_std': 0.15060415863990784, 'epoch': 0.04}

 12%|█▏        | 401/3300 [1:01:05<19:39:37, 24.41s/it]INFO 04-13 08:22:03 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 402/3300 [1:01:12<15:32:30, 19.31s/it]2026-04-13 08:22:10,661 | INFO | train_grpo_train | metrics_logged mode=train step=402

                                                       
{'loss': 0.0708, 'grad_norm': 25.57284164428711, 'learning_rate': 8.784848484848487e-06, 'num_tokens': 729290.0, 'completions/mean_length': 19.5, 'completions/min_length': 14.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8696723580360413, 'rewards/meter/std': 0.2797050476074219, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9484448432922363, 'rewards/repeat_soft/std': 0.017298834398388863, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.1141975000500679, 'rewards/total_composite/mean': 0.4683290421962738, 'rewards/total_composite/std': 0.2080000638961792, 'reward': 0.4683290421962738, 'reward_std': 0.2080000638961792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20655979216098785, 'sampling/sampling_logp_difference/max': 1.509596347808838, 'sampling/importance_sampling_ratio/min': 0.22099916636943817, 'sampling/importance_sampling_ratio/mean': 1.0491156578063965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.2838480323553085, 'clip_ratio/low_mean': 0.02878289483487606, 'clip_ratio/low_min': 0.02878289483487606, 'clip_ratio/high_mean': 0.14130889158695936, 'clip_ratio/high_max': 0.14130889158695936, 'clip_ratio/region_mean': 0.17009178642183542, 'reward_total_mean': 0.4683290421962738, 'reward_meter_mean': 0.8696723580360413, 'reward_meter_std': 0.2797050476074219, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9484448432922363, 'reward_repeat_soft_std': 0.017298834398388863, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.1141975000500679, 'reward_total_composite_mean': 0.4683290421962738, 'reward_total_composite_std': 0.2080000638961792, 'epoch': 0.04}

 12%|█▏        | 402/3300 [1:01:12<15:32:30, 19.31s/it]INFO 04-13 08:22:10 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 403/3300 [1:01:20<12:43:04, 15.80s/it]2026-04-13 08:22:18,202 | INFO | train_grpo_train | metrics_logged mode=train step=403

                                                       
{'loss': 0.4829, 'grad_norm': 22.063720703125, 'learning_rate': 8.781818181818182e-06, 'num_tokens': 730781.0, 'completions/mean_length': 26.375, 'completions/min_length': 16.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.7853130102157593, 'rewards/meter/std': 0.34553444385528564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9394265413284302, 'rewards/repeat_soft/std': 0.03469981253147125, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.3009004592895508, 'rewards/total_composite/mean': 0.6728497743606567, 'rewards/total_composite/std': 0.22580192983150482, 'reward': 0.6728497743606567, 'reward_std': 0.22580192983150482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21233589947223663, 'sampling/sampling_logp_difference/max': 1.7593927383422852, 'sampling/importance_sampling_ratio/min': 0.17214937508106232, 'sampling/importance_sampling_ratio/mean': 1.0368895530700684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9786519557237625, 'clip_ratio/low_mean': 0.15014297794550657, 'clip_ratio/low_min': 0.15014297794550657, 'clip_ratio/high_mean': 0.0657894741743803, 'clip_ratio/high_max': 0.0657894741743803, 'clip_ratio/region_mean': 0.21593245211988688, 'reward_total_mean': 0.6728497743606567, 'reward_meter_mean': 0.7853130102157593, 'reward_meter_std': 0.34553444385528564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9394265413284302, 'reward_repeat_soft_std': 0.03469981253147125, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.3009004592895508, 'reward_total_composite_mean': 0.6728497743606567, 'reward_total_composite_std': 0.22580192983150482, 'epoch': 0.04}

 12%|█▏        | 403/3300 [1:01:20<12:43:04, 15.80s/it]INFO 04-13 08:22:18 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 404/3300 [1:01:28<10:46:33, 13.40s/it]2026-04-13 08:22:26,037 | INFO | train_grpo_train | metrics_logged mode=train step=404

                                                       
{'loss': 0.0177, 'grad_norm': 9.488207817077637, 'learning_rate': 8.77878787878788e-06, 'num_tokens': 733016.0, 'completions/mean_length': 81.375, 'completions/min_length': 72.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9766528606414795, 'rewards/meter/std': 0.032082848250865936, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9504460096359253, 'rewards/repeat_soft/std': 0.06312248110771179, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.6209337115287781, 'rewards/total_composite/std': 0.0731566920876503, 'reward': 0.6209337115287781, 'reward_std': 0.07315670698881149, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18496696650981903, 'sampling/sampling_logp_difference/max': 1.6487913131713867, 'sampling/importance_sampling_ratio/min': 0.19228217005729675, 'sampling/importance_sampling_ratio/mean': 1.0251444578170776, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8824598789215088, 'clip_ratio/low_mean': 0.15241516008973122, 'clip_ratio/low_min': 0.15241516008973122, 'clip_ratio/high_mean': 0.017241379246115685, 'clip_ratio/high_max': 0.017241379246115685, 'clip_ratio/region_mean': 0.1696565393358469, 'reward_total_mean': 0.6209337115287781, 'reward_meter_mean': 0.9766528606414795, 'reward_meter_std': 0.032082848250865936, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9504460096359253, 'reward_repeat_soft_std': 0.06312248110771179, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.6209337115287781, 'reward_total_composite_std': 0.0731566920876503, 'epoch': 0.04}

 12%|█▏        | 404/3300 [1:01:28<10:46:33, 13.40s/it]INFO 04-13 08:22:26 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 405/3300 [1:01:35<9:24:27, 11.70s/it] 2026-04-13 08:22:33,683 | INFO | train_grpo_train | metrics_logged mode=train step=405

                                                      
{'loss': 0.059, 'grad_norm': 13.412790298461914, 'learning_rate': 8.775757575757577e-06, 'num_tokens': 735115.0, 'completions/mean_length': 78.375, 'completions/min_length': 65.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.589837908744812, 'rewards/meter/std': 0.3526695668697357, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877939224243164, 'rewards/repeat_soft/std': 0.009391658008098602, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5582143664360046, 'rewards/total_composite/std': 0.14221405982971191, 'reward': 0.5582143664360046, 'reward_std': 0.14221405982971191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18767502903938293, 'sampling/sampling_logp_difference/max': 1.452690601348877, 'sampling/importance_sampling_ratio/min': 0.2339400053024292, 'sampling/importance_sampling_ratio/mean': 1.0276070833206177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6444737315177917, 'clip_ratio/low_mean': 0.07759459409862757, 'clip_ratio/low_min': 0.07759459409862757, 'clip_ratio/high_mean': 0.07964119128882885, 'clip_ratio/high_max': 0.07964119128882885, 'clip_ratio/region_mean': 0.15723578538745642, 'reward_total_mean': 0.5582143664360046, 'reward_meter_mean': 0.589837908744812, 'reward_meter_std': 0.3526695668697357, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877939224243164, 'reward_repeat_soft_std': 0.009391658008098602, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5582143664360046, 'reward_total_composite_std': 0.14221405982971191, 'epoch': 0.04}

 12%|█▏        | 405/3300 [1:01:35<9:24:27, 11.70s/it]INFO 04-13 08:22:33 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 406/3300 [1:01:43<8:19:02, 10.35s/it]2026-04-13 08:22:40,868 | INFO | train_grpo_train | metrics_logged mode=train step=406

                                                      
{'loss': 0.1202, 'grad_norm': 18.717052459716797, 'learning_rate': 8.772727272727274e-06, 'num_tokens': 736681.0, 'completions/mean_length': 40.75, 'completions/min_length': 33.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9577003717422485, 'rewards/meter/std': 0.058974817395210266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921804666519165, 'rewards/repeat_soft/std': 0.0125203812494874, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.6150851249694824, 'rewards/total_composite/std': 0.02064206451177597, 'reward': 0.6150851249694824, 'reward_std': 0.020642060786485672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2295956015586853, 'sampling/sampling_logp_difference/max': 1.4451169967651367, 'sampling/importance_sampling_ratio/min': 0.2357185035943985, 'sampling/importance_sampling_ratio/mean': 1.0493429899215698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5007997155189514, 'clip_ratio/low_mean': 0.06051027122884989, 'clip_ratio/low_min': 0.06051027122884989, 'clip_ratio/high_mean': 0.13226395472884178, 'clip_ratio/high_max': 0.13226395472884178, 'clip_ratio/region_mean': 0.19277422595769167, 'reward_total_mean': 0.6150851249694824, 'reward_meter_mean': 0.9577003717422485, 'reward_meter_std': 0.058974817395210266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921804666519165, 'reward_repeat_soft_std': 0.0125203812494874, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.6150851249694824, 'reward_total_composite_std': 0.02064206451177597, 'epoch': 0.04}

 12%|█▏        | 406/3300 [1:01:43<8:19:02, 10.35s/it]INFO 04-13 08:22:41 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 407/3300 [1:01:52<8:01:42,  9.99s/it]2026-04-13 08:22:50,028 | INFO | train_grpo_train | metrics_logged mode=train step=407

                                                      
{'loss': 0.0501, 'grad_norm': 11.843538284301758, 'learning_rate': 8.76969696969697e-06, 'num_tokens': 738337.0, 'completions/mean_length': 37.0, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.17129772901535034, 'rewards/meter/std': 0.25353488326072693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9954594373703003, 'rewards/repeat_soft/std': 0.0074075618758797646, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.4334220290184021, 'rewards/total_composite/std': 0.1568155437707901, 'reward': 0.4334220290184021, 'reward_std': 0.1568155437707901, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18124893307685852, 'sampling/sampling_logp_difference/max': 1.4005489349365234, 'sampling/importance_sampling_ratio/min': 0.2464616298675537, 'sampling/importance_sampling_ratio/mean': 1.0053983926773071, 'sampling/importance_sampling_ratio/max': 1.9805840253829956, 'entropy': 1.377250149846077, 'clip_ratio/low_mean': 0.16454838402569294, 'clip_ratio/low_min': 0.16454838402569294, 'clip_ratio/high_mean': 0.025735294446349144, 'clip_ratio/high_max': 0.025735294446349144, 'clip_ratio/region_mean': 0.19028367847204208, 'reward_total_mean': 0.4334220290184021, 'reward_meter_mean': 0.17129772901535034, 'reward_meter_std': 0.25353488326072693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9954594373703003, 'reward_repeat_soft_std': 0.0074075618758797646, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.4334220290184021, 'reward_total_composite_std': 0.1568155437707901, 'epoch': 0.04}

 12%|█▏        | 407/3300 [1:01:52<8:01:42,  9.99s/it]INFO 04-13 08:22:50 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 408/3300 [1:01:59<7:25:41,  9.25s/it]2026-04-13 08:22:57,765 | INFO | train_grpo_train | metrics_logged mode=train step=408

                                                      
{'loss': 0.0024, 'grad_norm': 19.389144897460938, 'learning_rate': 8.766666666666669e-06, 'num_tokens': 740026.0, 'completions/mean_length': 36.125, 'completions/min_length': 30.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8589898347854614, 'rewards/meter/std': 0.22964926064014435, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9893051981925964, 'rewards/repeat_soft/std': 0.01291816309094429, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5730451345443726, 'rewards/total_composite/std': 0.07296156883239746, 'reward': 0.5730451345443726, 'reward_std': 0.07296158373355865, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24116696417331696, 'sampling/sampling_logp_difference/max': 1.9932687282562256, 'sampling/importance_sampling_ratio/min': 0.1362493336200714, 'sampling/importance_sampling_ratio/mean': 1.0360838174819946, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6887087672948837, 'clip_ratio/low_mean': 0.07655342854559422, 'clip_ratio/low_min': 0.07655342854559422, 'clip_ratio/high_mean': 0.129474725574255, 'clip_ratio/high_max': 0.129474725574255, 'clip_ratio/region_mean': 0.2060281541198492, 'reward_total_mean': 0.5730451345443726, 'reward_meter_mean': 0.8589898347854614, 'reward_meter_std': 0.22964926064014435, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9893051981925964, 'reward_repeat_soft_std': 0.01291816309094429, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5730451345443726, 'reward_total_composite_std': 0.07296156883239746, 'epoch': 0.04}

 12%|█▏        | 408/3300 [1:01:59<7:25:41,  9.25s/it]INFO 04-13 08:22:58 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 409/3300 [1:02:08<7:12:12,  8.97s/it]2026-04-13 08:23:05,870 | INFO | train_grpo_train | metrics_logged mode=train step=409

                                                      
{'loss': 0.0562, 'grad_norm': 11.353716850280762, 'learning_rate': 8.763636363636364e-06, 'num_tokens': 742081.0, 'completions/mean_length': 69.875, 'completions/min_length': 62.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8985353112220764, 'rewards/meter/std': 0.22408749163150787, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9455304145812988, 'rewards/repeat_soft/std': 0.06894509494304657, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6432985067367554, 'rewards/total_composite/std': 0.10527710616588593, 'reward': 0.6432985067367554, 'reward_std': 0.10527710616588593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16822470724582672, 'sampling/sampling_logp_difference/max': 2.2970826625823975, 'sampling/importance_sampling_ratio/min': 0.10055176168680191, 'sampling/importance_sampling_ratio/mean': 1.0406503677368164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5529555454850197, 'clip_ratio/low_mean': 0.132304847240448, 'clip_ratio/low_min': 0.132304847240448, 'clip_ratio/high_mean': 0.0498511902987957, 'clip_ratio/high_max': 0.0498511902987957, 'clip_ratio/region_mean': 0.1821560375392437, 'reward_total_mean': 0.6432985067367554, 'reward_meter_mean': 0.8985353112220764, 'reward_meter_std': 0.22408749163150787, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9455304145812988, 'reward_repeat_soft_std': 0.06894509494304657, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6432985067367554, 'reward_total_composite_std': 0.10527710616588593, 'epoch': 0.04}

 12%|█▏        | 409/3300 [1:02:08<7:12:12,  8.97s/it]INFO 04-13 08:23:06 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 410/3300 [1:02:14<6:37:21,  8.25s/it]2026-04-13 08:23:12,457 | INFO | train_grpo_train | metrics_logged mode=train step=410

                                                      
{'loss': -0.0049, 'grad_norm': 15.925487518310547, 'learning_rate': 8.760606060606061e-06, 'num_tokens': 743731.0, 'completions/mean_length': 41.25, 'completions/min_length': 35.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6764265894889832, 'rewards/meter/std': 0.39743971824645996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9897750616073608, 'rewards/repeat_soft/std': 0.017793547362089157, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.2093527615070343, 'rewards/total_composite/mean': 0.6413960456848145, 'rewards/total_composite/std': 0.21809059381484985, 'reward': 0.6413960456848145, 'reward_std': 0.21809059381484985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17137528955936432, 'sampling/sampling_logp_difference/max': 1.964127540588379, 'sampling/importance_sampling_ratio/min': 0.14027822017669678, 'sampling/importance_sampling_ratio/mean': 1.0164107084274292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1172003895044327, 'clip_ratio/low_mean': 0.0888611413538456, 'clip_ratio/low_min': 0.0888611413538456, 'clip_ratio/high_mean': 0.06415513809770346, 'clip_ratio/high_max': 0.06415513809770346, 'clip_ratio/region_mean': 0.15301627945154905, 'reward_total_mean': 0.6413960456848145, 'reward_meter_mean': 0.6764265894889832, 'reward_meter_std': 0.39743971824645996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9897750616073608, 'reward_repeat_soft_std': 0.017793547362089157, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.2093527615070343, 'reward_total_composite_mean': 0.6413960456848145, 'reward_total_composite_std': 0.21809059381484985, 'epoch': 0.04}

 12%|█▏        | 410/3300 [1:02:14<6:37:21,  8.25s/it]INFO 04-13 08:23:12 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 411/3300 [1:02:22<6:29:04,  8.08s/it]2026-04-13 08:23:20,144 | INFO | train_grpo_train | metrics_logged mode=train step=411

                                                      
{'loss': 0.0128, 'grad_norm': 18.98621368408203, 'learning_rate': 8.757575757575759e-06, 'num_tokens': 745330.0, 'completions/mean_length': 39.875, 'completions/min_length': 31.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.855193018913269, 'rewards/meter/std': 0.27002963423728943, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9820802211761475, 'rewards/repeat_soft/std': 0.020501619204878807, 'rewards/judge_quality/mean': 0.6112499833106995, 'rewards/judge_quality/std': 0.15037456154823303, 'rewards/total_composite/mean': 0.5779845714569092, 'rewards/total_composite/std': 0.261025607585907, 'reward': 0.5779845714569092, 'reward_std': 0.261025607585907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1591351330280304, 'sampling/sampling_logp_difference/max': 1.2452759742736816, 'sampling/importance_sampling_ratio/min': 0.2878614664077759, 'sampling/importance_sampling_ratio/mean': 1.024895191192627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3111989423632622, 'clip_ratio/low_mean': 0.04879032168537378, 'clip_ratio/low_min': 0.04879032168537378, 'clip_ratio/high_mean': 0.12211860902607441, 'clip_ratio/high_max': 0.12211860902607441, 'clip_ratio/region_mean': 0.1709089307114482, 'reward_total_mean': 0.5779845714569092, 'reward_meter_mean': 0.855193018913269, 'reward_meter_std': 0.27002963423728943, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9820802211761475, 'reward_repeat_soft_std': 0.020501619204878807, 'reward_judge_quality_mean': 0.6112499833106995, 'reward_judge_quality_std': 0.15037456154823303, 'reward_total_composite_mean': 0.5779845714569092, 'reward_total_composite_std': 0.261025607585907, 'epoch': 0.04}

 12%|█▏        | 411/3300 [1:02:22<6:29:04,  8.08s/it]INFO 04-13 08:23:20 [block_pool.py:378] Successfully reset prefix cache

 12%|█▏        | 412/3300 [1:02:31<6:46:52,  8.45s/it]2026-04-13 08:23:29,472 | INFO | train_grpo_train | metrics_logged mode=train step=412

                                                      
{'loss': -0.0314, 'grad_norm': 13.070358276367188, 'learning_rate': 8.754545454545456e-06, 'num_tokens': 746713.0, 'completions/mean_length': 32.875, 'completions/min_length': 26.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8472753763198853, 'rewards/meter/std': 0.17025259137153625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9989843964576721, 'rewards/repeat_soft/std': 0.0015448419144377112, 'rewards/judge_quality/mean': 0.6274999976158142, 'rewards/judge_quality/std': 0.21952873468399048, 'rewards/total_composite/mean': 0.6450576782226562, 'rewards/total_composite/std': 0.3009953796863556, 'reward': 0.6450576782226562, 'reward_std': 0.3009953796863556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15432707965373993, 'sampling/sampling_logp_difference/max': 1.2746858596801758, 'sampling/importance_sampling_ratio/min': 0.2795187532901764, 'sampling/importance_sampling_ratio/mean': 1.008492112159729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0406625121831894, 'clip_ratio/low_mean': 0.05918387370184064, 'clip_ratio/low_min': 0.05918387370184064, 'clip_ratio/high_mean': 0.07324465177953243, 'clip_ratio/high_max': 0.07324465177953243, 'clip_ratio/region_mean': 0.13242852548137307, 'reward_total_mean': 0.6450576782226562, 'reward_meter_mean': 0.8472753763198853, 'reward_meter_std': 0.17025259137153625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9989843964576721, 'reward_repeat_soft_std': 0.0015448419144377112, 'reward_judge_quality_mean': 0.6274999976158142, 'reward_judge_quality_std': 0.21952873468399048, 'reward_total_composite_mean': 0.6450576782226562, 'reward_total_composite_std': 0.3009953796863556, 'epoch': 0.04}

 12%|█▏        | 412/3300 [1:02:31<6:46:52,  8.45s/it]INFO 04-13 08:23:29 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 413/3300 [1:02:41<7:06:54,  8.87s/it]2026-04-13 08:23:39,329 | INFO | train_grpo_train | metrics_logged mode=train step=413

                                                      
{'loss': 0.0681, 'grad_norm': 14.127117156982422, 'learning_rate': 8.751515151515151e-06, 'num_tokens': 748539.0, 'completions/mean_length': 52.25, 'completions/min_length': 46.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8255071640014648, 'rewards/meter/std': 0.2958468496799469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9425725936889648, 'rewards/repeat_soft/std': 0.03496149182319641, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6523831486701965, 'rewards/total_composite/std': 0.18113821744918823, 'reward': 0.6523831486701965, 'reward_std': 0.18113821744918823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19584651291370392, 'sampling/sampling_logp_difference/max': 1.7609224319458008, 'sampling/importance_sampling_ratio/min': 0.17188623547554016, 'sampling/importance_sampling_ratio/mean': 1.0101100206375122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2705471515655518, 'clip_ratio/low_mean': 0.0991121381521225, 'clip_ratio/low_min': 0.0991121381521225, 'clip_ratio/high_mean': 0.058673469349741936, 'clip_ratio/high_max': 0.058673469349741936, 'clip_ratio/region_mean': 0.15778560750186443, 'reward_total_mean': 0.6523831486701965, 'reward_meter_mean': 0.8255071640014648, 'reward_meter_std': 0.2958468496799469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9425725936889648, 'reward_repeat_soft_std': 0.03496149182319641, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6523831486701965, 'reward_total_composite_std': 0.18113821744918823, 'epoch': 0.04}

 13%|█▎        | 413/3300 [1:02:41<7:06:54,  8.87s/it]INFO 04-13 08:23:39 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 414/3300 [1:02:48<6:39:06,  8.30s/it]2026-04-13 08:23:46,264 | INFO | train_grpo_train | metrics_logged mode=train step=414

                                                      
{'loss': 0.0386, 'grad_norm': 19.73372459411621, 'learning_rate': 8.748484848484849e-06, 'num_tokens': 750014.0, 'completions/mean_length': 29.375, 'completions/min_length': 16.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.375, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5379759073257446, 'rewards/meter/std': 0.34564530849456787, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9569318294525146, 'rewards/repeat_soft/std': 0.015749182552099228, 'rewards/judge_quality/mean': 0.36375001072883606, 'rewards/judge_quality/std': 0.13265825808048248, 'rewards/total_composite/mean': 0.4616193175315857, 'rewards/total_composite/std': 0.09335094690322876, 'reward': 0.4616193175315857, 'reward_std': 0.09335093945264816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1824444681406021, 'sampling/sampling_logp_difference/max': 1.346390724182129, 'sampling/importance_sampling_ratio/min': 0.2601776123046875, 'sampling/importance_sampling_ratio/mean': 1.032379150390625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6422661319375038, 'clip_ratio/low_mean': 0.08975493349134922, 'clip_ratio/low_min': 0.08975493349134922, 'clip_ratio/high_mean': 0.054734849371016026, 'clip_ratio/high_max': 0.054734849371016026, 'clip_ratio/region_mean': 0.14448978286236525, 'reward_total_mean': 0.4616193175315857, 'reward_meter_mean': 0.5379759073257446, 'reward_meter_std': 0.34564530849456787, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9569318294525146, 'reward_repeat_soft_std': 0.015749182552099228, 'reward_judge_quality_mean': 0.36375001072883606, 'reward_judge_quality_std': 0.13265825808048248, 'reward_total_composite_mean': 0.4616193175315857, 'reward_total_composite_std': 0.09335094690322876, 'epoch': 0.04}

 13%|█▎        | 414/3300 [1:02:48<6:39:06,  8.30s/it]INFO 04-13 08:23:46 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 415/3300 [1:02:55<6:15:02,  7.80s/it]2026-04-13 08:23:52,885 | INFO | train_grpo_train | metrics_logged mode=train step=415

                                                      
{'loss': 0.0945, 'grad_norm': 22.352832794189453, 'learning_rate': 8.745454545454546e-06, 'num_tokens': 751704.0, 'completions/mean_length': 47.25, 'completions/min_length': 44.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7549527883529663, 'rewards/meter/std': 0.26335862278938293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9893888235092163, 'rewards/repeat_soft/std': 0.007539966143667698, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6675245761871338, 'rewards/total_composite/std': 0.15883620083332062, 'reward': 0.6675245761871338, 'reward_std': 0.15883618593215942, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24240705370903015, 'sampling/sampling_logp_difference/max': 1.9810259342193604, 'sampling/importance_sampling_ratio/min': 0.13792765140533447, 'sampling/importance_sampling_ratio/mean': 1.0085594654083252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3290906995534897, 'clip_ratio/low_mean': 0.15271147526800632, 'clip_ratio/low_min': 0.15271147526800632, 'clip_ratio/high_mean': 0.06439394131302834, 'clip_ratio/high_max': 0.06439394131302834, 'clip_ratio/region_mean': 0.21710541658103466, 'reward_total_mean': 0.6675245761871338, 'reward_meter_mean': 0.7549527883529663, 'reward_meter_std': 0.26335862278938293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9893888235092163, 'reward_repeat_soft_std': 0.007539966143667698, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6675245761871338, 'reward_total_composite_std': 0.15883620083332062, 'epoch': 0.04}

 13%|█▎        | 415/3300 [1:02:55<6:15:02,  7.80s/it]INFO 04-13 08:23:53 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 416/3300 [1:03:01<5:53:37,  7.36s/it]2026-04-13 08:23:59,224 | INFO | train_grpo_train | metrics_logged mode=train step=416

                                                      
{'loss': 0.0218, 'grad_norm': 13.643056869506836, 'learning_rate': 8.742424242424243e-06, 'num_tokens': 753220.0, 'completions/mean_length': 40.5, 'completions/min_length': 37.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9194313287734985, 'rewards/meter/std': 0.13804028928279877, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9871430993080139, 'rewards/repeat_soft/std': 0.01945396140217781, 'rewards/judge_quality/mean': 0.6762500405311584, 'rewards/judge_quality/std': 0.2018442004919052, 'rewards/total_composite/mean': 0.7491375207901001, 'rewards/total_composite/std': 0.13408410549163818, 'reward': 0.7491375207901001, 'reward_std': 0.13408410549163818, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17899270355701447, 'sampling/sampling_logp_difference/max': 1.3610343933105469, 'sampling/importance_sampling_ratio/min': 0.38020259141921997, 'sampling/importance_sampling_ratio/mean': 1.0316815376281738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6855877041816711, 'clip_ratio/low_mean': 0.09274809807538986, 'clip_ratio/low_min': 0.09274809807538986, 'clip_ratio/high_mean': 0.08824772294610739, 'clip_ratio/high_max': 0.08824772294610739, 'clip_ratio/region_mean': 0.18099582102149725, 'reward_total_mean': 0.7491375207901001, 'reward_meter_mean': 0.9194313287734985, 'reward_meter_std': 0.13804028928279877, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9871430993080139, 'reward_repeat_soft_std': 0.01945396140217781, 'reward_judge_quality_mean': 0.6762500405311584, 'reward_judge_quality_std': 0.2018442004919052, 'reward_total_composite_mean': 0.7491375207901001, 'reward_total_composite_std': 0.13408410549163818, 'epoch': 0.04}

 13%|█▎        | 416/3300 [1:03:01<5:53:37,  7.36s/it]INFO 04-13 08:23:59 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 417/3300 [1:03:08<5:47:03,  7.22s/it]2026-04-13 08:24:06,125 | INFO | train_grpo_train | metrics_logged mode=train step=417

                                                      
{'loss': 0.0458, 'grad_norm': 9.444697380065918, 'learning_rate': 8.73939393939394e-06, 'num_tokens': 755195.0, 'completions/mean_length': 70.875, 'completions/min_length': 61.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9849028587341309, 'rewards/meter/std': 0.014686250127851963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8227899074554443, 'rewards/repeat_soft/std': 0.1468363106250763, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.6057982444763184, 'rewards/total_composite/std': 0.09362467378377914, 'reward': 0.6057982444763184, 'reward_std': 0.09362467378377914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14614373445510864, 'sampling/sampling_logp_difference/max': 1.246006727218628, 'sampling/importance_sampling_ratio/min': 0.2876511812210083, 'sampling/importance_sampling_ratio/mean': 1.0279991626739502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1420061439275742, 'clip_ratio/low_mean': 0.07545688562095165, 'clip_ratio/low_min': 0.07545688562095165, 'clip_ratio/high_mean': 0.06647481210529804, 'clip_ratio/high_max': 0.06647481210529804, 'clip_ratio/region_mean': 0.1419316977262497, 'reward_total_mean': 0.6057982444763184, 'reward_meter_mean': 0.9849028587341309, 'reward_meter_std': 0.014686250127851963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8227899074554443, 'reward_repeat_soft_std': 0.1468363106250763, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.6057982444763184, 'reward_total_composite_std': 0.09362467378377914, 'epoch': 0.04}

 13%|█▎        | 417/3300 [1:03:08<5:47:03,  7.22s/it]INFO 04-13 08:24:06 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 418/3300 [1:03:15<5:53:12,  7.35s/it]2026-04-13 08:24:13,798 | INFO | train_grpo_train | metrics_logged mode=train step=418

                                                      
{'loss': 0.0944, 'grad_norm': 20.964733123779297, 'learning_rate': 8.736363636363638e-06, 'num_tokens': 756763.0, 'completions/mean_length': 36.0, 'completions/min_length': 26.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7766556739807129, 'rewards/meter/std': 0.33451172709465027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9609910845756531, 'rewards/repeat_soft/std': 0.045096565037965775, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.6375292539596558, 'rewards/total_composite/std': 0.19607028365135193, 'reward': 0.6375292539596558, 'reward_std': 0.19607029855251312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2082994431257248, 'sampling/sampling_logp_difference/max': 4.037148475646973, 'sampling/importance_sampling_ratio/min': 0.017647724598646164, 'sampling/importance_sampling_ratio/mean': 1.0093640089035034, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7629767209291458, 'clip_ratio/low_mean': 0.1277181648183614, 'clip_ratio/low_min': 0.1277181648183614, 'clip_ratio/high_mean': 0.05000000074505806, 'clip_ratio/high_max': 0.05000000074505806, 'clip_ratio/region_mean': 0.17771816556341946, 'reward_total_mean': 0.6375292539596558, 'reward_meter_mean': 0.7766556739807129, 'reward_meter_std': 0.33451172709465027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9609910845756531, 'reward_repeat_soft_std': 0.045096565037965775, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.6375292539596558, 'reward_total_composite_std': 0.19607028365135193, 'epoch': 0.04}

 13%|█▎        | 418/3300 [1:03:16<5:53:12,  7.35s/it]INFO 04-13 08:24:14 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 419/3300 [1:03:23<5:54:44,  7.39s/it]2026-04-13 08:24:21,248 | INFO | train_grpo_train | metrics_logged mode=train step=419

                                                      
{'loss': 0.0766, 'grad_norm': 7.3235578536987305, 'learning_rate': 8.733333333333333e-06, 'num_tokens': 759007.0, 'completions/mean_length': 91.5, 'completions/min_length': 75.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9800565838813782, 'rewards/meter/std': 0.016581416130065918, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9972222447395325, 'rewards/repeat_soft/std': 0.0028130800928920507, 'rewards/judge_quality/mean': 0.5399999618530273, 'rewards/judge_quality/std': 0.22123032808303833, 'rewards/total_composite/mean': 0.6941713094711304, 'rewards/total_composite/std': 0.1425619125366211, 'reward': 0.6941713094711304, 'reward_std': 0.1425618976354599, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1567026972770691, 'sampling/sampling_logp_difference/max': 1.4153804779052734, 'sampling/importance_sampling_ratio/min': 0.242833212018013, 'sampling/importance_sampling_ratio/mean': 1.029744029045105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6311266124248505, 'clip_ratio/low_mean': 0.06705222418531775, 'clip_ratio/low_min': 0.06705222418531775, 'clip_ratio/high_mean': 0.060451144352555275, 'clip_ratio/high_max': 0.060451144352555275, 'clip_ratio/region_mean': 0.12750336853787303, 'reward_total_mean': 0.6941713094711304, 'reward_meter_mean': 0.9800565838813782, 'reward_meter_std': 0.016581416130065918, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9972222447395325, 'reward_repeat_soft_std': 0.0028130800928920507, 'reward_judge_quality_mean': 0.5399999618530273, 'reward_judge_quality_std': 0.22123032808303833, 'reward_total_composite_mean': 0.6941713094711304, 'reward_total_composite_std': 0.1425619125366211, 'epoch': 0.04}

 13%|█▎        | 419/3300 [1:03:23<5:54:44,  7.39s/it]INFO 04-13 08:24:21 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 420/3300 [1:03:30<5:45:00,  7.19s/it]2026-04-13 08:24:27,967 | INFO | train_grpo_train | metrics_logged mode=train step=420

                                                      
{'loss': 0.1067, 'grad_norm': 14.094193458557129, 'learning_rate': 8.73030303030303e-06, 'num_tokens': 760555.0, 'completions/mean_length': 41.5, 'completions/min_length': 30.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.31071364879608154, 'rewards/meter/std': 0.2925502359867096, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957492351531982, 'rewards/repeat_soft/std': 0.005208555143326521, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.43526244163513184, 'rewards/total_composite/std': 0.07953180372714996, 'reward': 0.43526244163513184, 'reward_std': 0.07953178882598877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21063420176506042, 'sampling/sampling_logp_difference/max': 1.7707998752593994, 'sampling/importance_sampling_ratio/min': 0.1701968014240265, 'sampling/importance_sampling_ratio/mean': 1.0322872400283813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9791297540068626, 'clip_ratio/low_mean': 0.13185427524149418, 'clip_ratio/low_min': 0.13185427524149418, 'clip_ratio/high_mean': 0.05131579004228115, 'clip_ratio/high_max': 0.05131579004228115, 'clip_ratio/region_mean': 0.18317006528377533, 'reward_total_mean': 0.43526244163513184, 'reward_meter_mean': 0.31071364879608154, 'reward_meter_std': 0.2925502359867096, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957492351531982, 'reward_repeat_soft_std': 0.005208555143326521, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.43526244163513184, 'reward_total_composite_std': 0.07953180372714996, 'epoch': 0.04}

 13%|█▎        | 420/3300 [1:03:30<5:45:00,  7.19s/it]INFO 04-13 08:24:28 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 421/3300 [1:03:37<5:51:32,  7.33s/it]2026-04-13 08:24:35,666 | INFO | train_grpo_train | metrics_logged mode=train step=421

                                                      
{'loss': 0.031, 'grad_norm': 28.130706787109375, 'learning_rate': 8.727272727272728e-06, 'num_tokens': 762151.0, 'completions/mean_length': 22.5, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8638845086097717, 'rewards/meter/std': 0.3477573096752167, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5874203443527222, 'rewards/total_composite/std': 0.09827587753534317, 'reward': 0.5874203443527222, 'reward_std': 0.09827586263418198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18736852705478668, 'sampling/sampling_logp_difference/max': 2.2155232429504395, 'sampling/importance_sampling_ratio/min': 0.10909641534090042, 'sampling/importance_sampling_ratio/mean': 1.0350109338760376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.808619111776352, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.1561513552442193, 'clip_ratio/high_max': 0.1561513552442193, 'clip_ratio/region_mean': 0.17400849889963865, 'reward_total_mean': 0.5874203443527222, 'reward_meter_mean': 0.8638845086097717, 'reward_meter_std': 0.3477573096752167, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5874203443527222, 'reward_total_composite_std': 0.09827587753534317, 'epoch': 0.04}

 13%|█▎        | 421/3300 [1:03:37<5:51:32,  7.33s/it]INFO 04-13 08:24:35 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 422/3300 [1:03:45<5:50:56,  7.32s/it]2026-04-13 08:24:42,945 | INFO | train_grpo_train | metrics_logged mode=train step=422

                                                      
{'loss': 0.0368, 'grad_norm': 17.89279556274414, 'learning_rate': 8.724242424242425e-06, 'num_tokens': 763698.0, 'completions/mean_length': 31.375, 'completions/min_length': 27.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9931104183197021, 'rewards/meter/std': 0.002878874307498336, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.942622184753418, 'rewards/repeat_soft/std': 0.06343577057123184, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6243409514427185, 'rewards/total_composite/std': 0.4099300503730774, 'reward': 0.6243409514427185, 'reward_std': 0.409930020570755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1619260013103485, 'sampling/sampling_logp_difference/max': 1.5931119918823242, 'sampling/importance_sampling_ratio/min': 0.20329199731349945, 'sampling/importance_sampling_ratio/mean': 1.027131199836731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.269327849149704, 'clip_ratio/low_mean': 0.04277935717254877, 'clip_ratio/low_min': 0.04277935717254877, 'clip_ratio/high_mean': 0.08713672962039709, 'clip_ratio/high_max': 0.08713672962039709, 'clip_ratio/region_mean': 0.12991608679294586, 'reward_total_mean': 0.6243409514427185, 'reward_meter_mean': 0.9931104183197021, 'reward_meter_std': 0.002878874307498336, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.942622184753418, 'reward_repeat_soft_std': 0.06343577057123184, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6243409514427185, 'reward_total_composite_std': 0.4099300503730774, 'epoch': 0.04}

 13%|█▎        | 422/3300 [1:03:45<5:50:56,  7.32s/it]INFO 04-13 08:24:43 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 423/3300 [1:03:53<6:11:19,  7.74s/it]2026-04-13 08:24:51,651 | INFO | train_grpo_train | metrics_logged mode=train step=423

                                                      
{'loss': 0.0938, 'grad_norm': 12.574085235595703, 'learning_rate': 8.72121212121212e-06, 'num_tokens': 765711.0, 'completions/mean_length': 78.625, 'completions/min_length': 51.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.39449363946914673, 'rewards/meter/std': 0.24479669332504272, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9084656834602356, 'rewards/repeat_soft/std': 0.2384941577911377, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18216457962989807, 'rewards/total_composite/mean': 0.4602924585342407, 'rewards/total_composite/std': 0.09776712954044342, 'reward': 0.4602924585342407, 'reward_std': 0.09776712208986282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19461190700531006, 'sampling/sampling_logp_difference/max': 2.4785842895507812, 'sampling/importance_sampling_ratio/min': 0.08386186510324478, 'sampling/importance_sampling_ratio/mean': 1.010611891746521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4617060646414757, 'clip_ratio/low_mean': 0.07937775831669569, 'clip_ratio/low_min': 0.07937775831669569, 'clip_ratio/high_mean': 0.09583958983421326, 'clip_ratio/high_max': 0.09583958983421326, 'clip_ratio/region_mean': 0.17521734815090895, 'reward_total_mean': 0.4602924585342407, 'reward_meter_mean': 0.39449363946914673, 'reward_meter_std': 0.24479669332504272, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9084656834602356, 'reward_repeat_soft_std': 0.2384941577911377, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18216457962989807, 'reward_total_composite_mean': 0.4602924585342407, 'reward_total_composite_std': 0.09776712954044342, 'epoch': 0.04}

 13%|█▎        | 423/3300 [1:03:53<6:11:19,  7.74s/it]INFO 04-13 08:24:51 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 424/3300 [1:04:02<6:26:03,  8.05s/it]2026-04-13 08:25:00,429 | INFO | train_grpo_train | metrics_logged mode=train step=424

                                                      
{'loss': 0.0904, 'grad_norm': 14.377594947814941, 'learning_rate': 8.71818181818182e-06, 'num_tokens': 767528.0, 'completions/mean_length': 56.125, 'completions/min_length': 47.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8052804470062256, 'rewards/meter/std': 0.29650941491127014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9423341751098633, 'rewards/repeat_soft/std': 0.08976280689239502, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.6919419169425964, 'rewards/total_composite/std': 0.19053557515144348, 'reward': 0.6919419169425964, 'reward_std': 0.1905355602502823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20590421557426453, 'sampling/sampling_logp_difference/max': 1.8177146911621094, 'sampling/importance_sampling_ratio/min': 0.16239644587039948, 'sampling/importance_sampling_ratio/mean': 1.0273287296295166, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3588280379772186, 'clip_ratio/low_mean': 0.08734078239649534, 'clip_ratio/low_min': 0.08734078239649534, 'clip_ratio/high_mean': 0.10055535100400448, 'clip_ratio/high_max': 0.10055535100400448, 'clip_ratio/region_mean': 0.18789613340049982, 'reward_total_mean': 0.6919419169425964, 'reward_meter_mean': 0.8052804470062256, 'reward_meter_std': 0.29650941491127014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9423341751098633, 'reward_repeat_soft_std': 0.08976280689239502, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.6919419169425964, 'reward_total_composite_std': 0.19053557515144348, 'epoch': 0.04}

 13%|█▎        | 424/3300 [1:04:02<6:26:03,  8.05s/it]INFO 04-13 08:25:00 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 425/3300 [1:04:10<6:18:49,  7.91s/it]2026-04-13 08:25:07,987 | INFO | train_grpo_train | metrics_logged mode=train step=425

                                                      
{'loss': 0.1283, 'grad_norm': 24.715789794921875, 'learning_rate': 8.715151515151515e-06, 'num_tokens': 769222.0, 'completions/mean_length': 30.75, 'completions/min_length': 28.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8295538425445557, 'rewards/meter/std': 0.3294765055179596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9836940765380859, 'rewards/repeat_soft/std': 0.010605689138174057, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/total_composite/mean': 0.8057328462600708, 'rewards/total_composite/std': 0.20401300489902496, 'reward': 0.8057328462600708, 'reward_std': 0.20401298999786377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12585145235061646, 'sampling/sampling_logp_difference/max': 1.3707711696624756, 'sampling/importance_sampling_ratio/min': 0.2539110481739044, 'sampling/importance_sampling_ratio/mean': 0.9612415432929993, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.530036062002182, 'clip_ratio/low_mean': 0.030766253359615803, 'clip_ratio/low_min': 0.030766253359615803, 'clip_ratio/high_mean': 0.05957115115597844, 'clip_ratio/high_max': 0.05957115115597844, 'clip_ratio/region_mean': 0.09033740451559424, 'reward_total_mean': 0.8057328462600708, 'reward_meter_mean': 0.8295538425445557, 'reward_meter_std': 0.3294765055179596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9836940765380859, 'reward_repeat_soft_std': 0.010605689138174057, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.8057328462600708, 'reward_total_composite_std': 0.20401300489902496, 'epoch': 0.04}

 13%|█▎        | 425/3300 [1:04:10<6:18:49,  7.91s/it]INFO 04-13 08:25:08 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 426/3300 [1:04:17<6:10:28,  7.73s/it]2026-04-13 08:25:15,338 | INFO | train_grpo_train | metrics_logged mode=train step=426

                                                      
{'loss': 0.0414, 'grad_norm': 15.755805015563965, 'learning_rate': 8.712121212121212e-06, 'num_tokens': 770914.0, 'completions/mean_length': 49.5, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7202902436256409, 'rewards/meter/std': 0.3197583556175232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9715853929519653, 'rewards/repeat_soft/std': 0.025107305496931076, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.22038927674293518, 'rewards/total_composite/mean': 0.6496169567108154, 'rewards/total_composite/std': 0.1708659678697586, 'reward': 0.6496169567108154, 'reward_std': 0.1708659678697586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2259758710861206, 'sampling/sampling_logp_difference/max': 1.8252153396606445, 'sampling/importance_sampling_ratio/min': 0.1611829400062561, 'sampling/importance_sampling_ratio/mean': 1.0428838729858398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.802308902144432, 'clip_ratio/low_mean': 0.10848477762192488, 'clip_ratio/low_min': 0.10848477762192488, 'clip_ratio/high_mean': 0.06292749661952257, 'clip_ratio/high_max': 0.06292749661952257, 'clip_ratio/region_mean': 0.17141227424144745, 'reward_total_mean': 0.6496169567108154, 'reward_meter_mean': 0.7202902436256409, 'reward_meter_std': 0.3197583556175232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9715853929519653, 'reward_repeat_soft_std': 0.025107305496931076, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.22038927674293518, 'reward_total_composite_mean': 0.6496169567108154, 'reward_total_composite_std': 0.1708659678697586, 'epoch': 0.04}

 13%|█▎        | 426/3300 [1:04:17<6:10:28,  7.73s/it]INFO 04-13 08:25:15 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 427/3300 [1:04:27<6:36:20,  8.28s/it]2026-04-13 08:25:24,865 | INFO | train_grpo_train | metrics_logged mode=train step=427

                                                      
{'loss': -0.0635, 'grad_norm': 8.554418563842773, 'learning_rate': 8.70909090909091e-06, 'num_tokens': 773425.0, 'completions/mean_length': 132.875, 'completions/min_length': 84.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.875, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.9280627965927124, 'rewards/meter/std': 0.16058552265167236, 'rewards/count_adherence/mean': 0.7916666269302368, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9765769839286804, 'rewards/repeat_soft/std': 0.025364678353071213, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.505779504776001, 'rewards/total_composite/std': 0.3327721655368805, 'reward': 0.505779504776001, 'reward_std': 0.3327721655368805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17246946692466736, 'sampling/sampling_logp_difference/max': 1.9859402179718018, 'sampling/importance_sampling_ratio/min': 0.13725151121616364, 'sampling/importance_sampling_ratio/mean': 1.025324821472168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.558809533715248, 'clip_ratio/low_mean': 0.04445033520460129, 'clip_ratio/low_min': 0.04445033520460129, 'clip_ratio/high_mean': 0.1237593786790967, 'clip_ratio/high_max': 0.1237593786790967, 'clip_ratio/region_mean': 0.168209713883698, 'reward_total_mean': 0.505779504776001, 'reward_meter_mean': 0.9280627965927124, 'reward_meter_std': 0.16058552265167236, 'reward_count_adherence_mean': 0.7916666269302368, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9765769839286804, 'reward_repeat_soft_std': 0.025364678353071213, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.505779504776001, 'reward_total_composite_std': 0.3327721655368805, 'epoch': 0.04}

 13%|█▎        | 427/3300 [1:04:27<6:36:20,  8.28s/it]INFO 04-13 08:25:25 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 428/3300 [1:04:34<6:30:47,  8.16s/it]2026-04-13 08:25:32,767 | INFO | train_grpo_train | metrics_logged mode=train step=428

                                                      
{'loss': 0.0752, 'grad_norm': 17.699542999267578, 'learning_rate': 8.706060606060607e-06, 'num_tokens': 775099.0, 'completions/mean_length': 42.25, 'completions/min_length': 32.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.4861021041870117, 'rewards/meter/std': 0.359035462141037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9769493937492371, 'rewards/repeat_soft/std': 0.029769068583846092, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4864930510520935, 'rewards/total_composite/std': 0.1076832190155983, 'reward': 0.4864930510520935, 'reward_std': 0.1076832041144371, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2017541229724884, 'sampling/sampling_logp_difference/max': 1.6000213623046875, 'sampling/importance_sampling_ratio/min': 0.2018921971321106, 'sampling/importance_sampling_ratio/mean': 1.0370087623596191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.544850543141365, 'clip_ratio/low_mean': 0.12223918549716473, 'clip_ratio/low_min': 0.12223918549716473, 'clip_ratio/high_mean': 0.057329047471284866, 'clip_ratio/high_max': 0.057329047471284866, 'clip_ratio/region_mean': 0.1795682329684496, 'reward_total_mean': 0.4864930510520935, 'reward_meter_mean': 0.4861021041870117, 'reward_meter_std': 0.359035462141037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9769493937492371, 'reward_repeat_soft_std': 0.029769068583846092, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4864930510520935, 'reward_total_composite_std': 0.1076832190155983, 'epoch': 0.04}

 13%|█▎        | 428/3300 [1:04:34<6:30:47,  8.16s/it]INFO 04-13 08:25:33 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 429/3300 [1:04:43<6:33:44,  8.23s/it]2026-04-13 08:25:41,146 | INFO | train_grpo_train | metrics_logged mode=train step=429

                                                      
{'loss': 0.117, 'grad_norm': 11.39891529083252, 'learning_rate': 8.703030303030304e-06, 'num_tokens': 777565.0, 'completions/mean_length': 129.25, 'completions/min_length': 106.0, 'completions/max_length': 177.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.25, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 177.0, 'rewards/meter/mean': 0.9440456628799438, 'rewards/meter/std': 0.10624020546674728, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9840169548988342, 'rewards/repeat_soft/std': 0.019269311800599098, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.5813806653022766, 'rewards/total_composite/std': 0.250386506319046, 'reward': 0.5813806653022766, 'reward_std': 0.250386506319046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20710064470767975, 'sampling/sampling_logp_difference/max': 2.949585437774658, 'sampling/importance_sampling_ratio/min': 0.05236141011118889, 'sampling/importance_sampling_ratio/mean': 1.0383763313293457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0161682814359665, 'clip_ratio/low_mean': 0.09372947178781033, 'clip_ratio/low_min': 0.09372947178781033, 'clip_ratio/high_mean': 0.08874286524951458, 'clip_ratio/high_max': 0.08874286524951458, 'clip_ratio/region_mean': 0.1824723370373249, 'reward_total_mean': 0.5813806653022766, 'reward_meter_mean': 0.9440456628799438, 'reward_meter_std': 0.10624020546674728, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9840169548988342, 'reward_repeat_soft_std': 0.019269311800599098, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.5813806653022766, 'reward_total_composite_std': 0.250386506319046, 'epoch': 0.04}

 13%|█▎        | 429/3300 [1:04:43<6:33:44,  8.23s/it]INFO 04-13 08:25:41 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 430/3300 [1:04:50<6:12:18,  7.78s/it]2026-04-13 08:25:47,891 | INFO | train_grpo_train | metrics_logged mode=train step=430

                                                      
{'loss': 0.0519, 'grad_norm': 15.2701416015625, 'learning_rate': 8.700000000000001e-06, 'num_tokens': 778826.0, 'completions/mean_length': 26.625, 'completions/min_length': 19.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.3605045676231384, 'rewards/meter/std': 0.31919997930526733, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.8237500190734863, 'rewards/judge_quality/std': 0.2722361385822296, 'rewards/total_composite/mean': 0.5100744366645813, 'rewards/total_composite/std': 0.16296890377998352, 'reward': 0.5100744366645813, 'reward_std': 0.16296890377998352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18163342773914337, 'sampling/sampling_logp_difference/max': 2.0839452743530273, 'sampling/importance_sampling_ratio/min': 0.12443829327821732, 'sampling/importance_sampling_ratio/mean': 1.0236661434173584, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6269883215427399, 'clip_ratio/low_mean': 0.07251747930422425, 'clip_ratio/low_min': 0.07251747930422425, 'clip_ratio/high_mean': 0.07679007947444916, 'clip_ratio/high_max': 0.07679007947444916, 'clip_ratio/region_mean': 0.1493075587786734, 'reward_total_mean': 0.5100744366645813, 'reward_meter_mean': 0.3605045676231384, 'reward_meter_std': 0.31919997930526733, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.8237500190734863, 'reward_judge_quality_std': 0.2722361385822296, 'reward_total_composite_mean': 0.5100744366645813, 'reward_total_composite_std': 0.16296890377998352, 'epoch': 0.04}

 13%|█▎        | 430/3300 [1:04:50<6:12:18,  7.78s/it]INFO 04-13 08:25:48 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 431/3300 [1:04:57<6:03:49,  7.61s/it]2026-04-13 08:25:55,092 | INFO | train_grpo_train | metrics_logged mode=train step=431

                                                      
{'loss': -0.0001, 'grad_norm': 17.1655330657959, 'learning_rate': 8.696969696969699e-06, 'num_tokens': 780305.0, 'completions/mean_length': 36.875, 'completions/min_length': 33.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.4635351300239563, 'rewards/meter/std': 0.3749304711818695, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9966304302215576, 'rewards/repeat_soft/std': 0.004131971392780542, 'rewards/judge_quality/mean': 0.6675000190734863, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.48579472303390503, 'rewards/total_composite/std': 0.2634431719779968, 'reward': 0.48579472303390503, 'reward_std': 0.2634431719779968, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2026931643486023, 'sampling/sampling_logp_difference/max': 1.8704214096069336, 'sampling/importance_sampling_ratio/min': 0.15405872464179993, 'sampling/importance_sampling_ratio/mean': 1.00552237033844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.370562307536602, 'clip_ratio/low_mean': 0.09522218257188797, 'clip_ratio/low_min': 0.09522218257188797, 'clip_ratio/high_mean': 0.07766646798700094, 'clip_ratio/high_max': 0.07766646798700094, 'clip_ratio/region_mean': 0.1728886505588889, 'reward_total_mean': 0.48579472303390503, 'reward_meter_mean': 0.4635351300239563, 'reward_meter_std': 0.3749304711818695, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9966304302215576, 'reward_repeat_soft_std': 0.004131971392780542, 'reward_judge_quality_mean': 0.6675000190734863, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.48579472303390503, 'reward_total_composite_std': 0.2634431719779968, 'epoch': 0.04}

 13%|█▎        | 431/3300 [1:04:57<6:03:49,  7.61s/it]INFO 04-13 08:25:55 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 432/3300 [1:05:05<6:13:13,  7.81s/it]2026-04-13 08:26:03,365 | INFO | train_grpo_train | metrics_logged mode=train step=432

                                                      
{'loss': 0.0515, 'grad_norm': 7.66887092590332, 'learning_rate': 8.693939393939394e-06, 'num_tokens': 783010.0, 'completions/mean_length': 123.125, 'completions/min_length': 110.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.125, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.9879873991012573, 'rewards/meter/std': 0.009452181868255138, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8556176424026489, 'rewards/repeat_soft/std': 0.10616997629404068, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5374159216880798, 'rewards/total_composite/std': 0.23432575166225433, 'reward': 0.5374159216880798, 'reward_std': 0.23432573676109314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16589200496673584, 'sampling/sampling_logp_difference/max': 1.4953727722167969, 'sampling/importance_sampling_ratio/min': 0.22416503727436066, 'sampling/importance_sampling_ratio/mean': 1.0411186218261719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5380447953939438, 'clip_ratio/low_mean': 0.028368454426527023, 'clip_ratio/low_min': 0.028368454426527023, 'clip_ratio/high_mean': 0.1173196341842413, 'clip_ratio/high_max': 0.1173196341842413, 'clip_ratio/region_mean': 0.14568808861076832, 'reward_total_mean': 0.5374159216880798, 'reward_meter_mean': 0.9879873991012573, 'reward_meter_std': 0.009452181868255138, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8556176424026489, 'reward_repeat_soft_std': 0.10616997629404068, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5374159216880798, 'reward_total_composite_std': 0.23432575166225433, 'epoch': 0.04}

 13%|█▎        | 432/3300 [1:05:05<6:13:13,  7.81s/it]INFO 04-13 08:26:03 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 433/3300 [1:05:12<5:54:19,  7.42s/it]2026-04-13 08:26:09,863 | INFO | train_grpo_train | metrics_logged mode=train step=433

                                                      
{'loss': 0.0247, 'grad_norm': 17.924205780029297, 'learning_rate': 8.690909090909091e-06, 'num_tokens': 784669.0, 'completions/mean_length': 36.375, 'completions/min_length': 31.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9752165079116821, 'rewards/meter/std': 0.040312111377716064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9120166897773743, 'rewards/repeat_soft/std': 0.10026796162128448, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.6149869561195374, 'rewards/total_composite/std': 0.2904825508594513, 'reward': 0.6149869561195374, 'reward_std': 0.2904825508594513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1944267749786377, 'sampling/sampling_logp_difference/max': 1.400411605834961, 'sampling/importance_sampling_ratio/min': 0.24649548530578613, 'sampling/importance_sampling_ratio/mean': 1.0463149547576904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9324277490377426, 'clip_ratio/low_mean': 0.08154762163758278, 'clip_ratio/low_min': 0.08154762163758278, 'clip_ratio/high_mean': 0.08843881450593472, 'clip_ratio/high_max': 0.08843881450593472, 'clip_ratio/region_mean': 0.1699864361435175, 'reward_total_mean': 0.6149869561195374, 'reward_meter_mean': 0.9752165079116821, 'reward_meter_std': 0.040312111377716064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9120166897773743, 'reward_repeat_soft_std': 0.10026796162128448, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.6149869561195374, 'reward_total_composite_std': 0.2904825508594513, 'epoch': 0.04}

 13%|█▎        | 433/3300 [1:05:12<5:54:19,  7.42s/it]INFO 04-13 08:26:10 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 434/3300 [1:05:18<5:45:59,  7.24s/it]2026-04-13 08:26:16,709 | INFO | train_grpo_train | metrics_logged mode=train step=434

                                                      
{'loss': -0.2025, 'grad_norm': 7.984838008880615, 'learning_rate': 8.687878787878789e-06, 'num_tokens': 786349.0, 'completions/mean_length': 56.0, 'completions/min_length': 17.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7314310073852539, 'rewards/meter/std': 0.2750467360019684, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9881883859634399, 'rewards/repeat_soft/std': 0.013463444076478481, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.2018442004919052, 'rewards/total_composite/mean': 0.5222692489624023, 'rewards/total_composite/std': 0.24057649075984955, 'reward': 0.5222692489624023, 'reward_std': 0.24057647585868835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1608295440673828, 'sampling/sampling_logp_difference/max': 3.9674124717712402, 'sampling/importance_sampling_ratio/min': 0.018922332674264908, 'sampling/importance_sampling_ratio/mean': 1.0018035173416138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7854427397251129, 'clip_ratio/low_mean': 0.05656746681779623, 'clip_ratio/low_min': 0.05656746681779623, 'clip_ratio/high_mean': 0.10639818571507931, 'clip_ratio/high_max': 0.10639818571507931, 'clip_ratio/region_mean': 0.16296565253287554, 'reward_total_mean': 0.5222692489624023, 'reward_meter_mean': 0.7314310073852539, 'reward_meter_std': 0.2750467360019684, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9881883859634399, 'reward_repeat_soft_std': 0.013463444076478481, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.2018442004919052, 'reward_total_composite_mean': 0.5222692489624023, 'reward_total_composite_std': 0.24057649075984955, 'epoch': 0.04}

 13%|█▎        | 434/3300 [1:05:18<5:45:59,  7.24s/it]INFO 04-13 08:26:16 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 435/3300 [1:05:25<5:31:23,  6.94s/it]2026-04-13 08:26:22,940 | INFO | train_grpo_train | metrics_logged mode=train step=435

                                                      
{'loss': 0.0547, 'grad_norm': 30.92169189453125, 'learning_rate': 8.684848484848486e-06, 'num_tokens': 787744.0, 'completions/mean_length': 17.375, 'completions/min_length': 15.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.7917190194129944, 'rewards/meter/std': 0.27918416261672974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9305113554000854, 'rewards/repeat_soft/std': 0.07009010016918182, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5465813279151917, 'rewards/total_composite/std': 0.08320525288581848, 'reward': 0.5465813279151917, 'reward_std': 0.08320525288581848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1961645632982254, 'sampling/sampling_logp_difference/max': 1.2406105995178223, 'sampling/importance_sampling_ratio/min': 0.2892075777053833, 'sampling/importance_sampling_ratio/mean': 1.047686219215393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.589640535414219, 'clip_ratio/low_mean': 0.07426470797508955, 'clip_ratio/low_min': 0.07426470797508955, 'clip_ratio/high_mean': 0.10399305773898959, 'clip_ratio/high_max': 0.10399305773898959, 'clip_ratio/region_mean': 0.17825776571407914, 'reward_total_mean': 0.5465813279151917, 'reward_meter_mean': 0.7917190194129944, 'reward_meter_std': 0.27918416261672974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9305113554000854, 'reward_repeat_soft_std': 0.07009010016918182, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5465813279151917, 'reward_total_composite_std': 0.08320525288581848, 'epoch': 0.04}

 13%|█▎        | 435/3300 [1:05:25<5:31:23,  6.94s/it]INFO 04-13 08:26:23 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 436/3300 [1:05:33<5:47:00,  7.27s/it]2026-04-13 08:26:30,978 | INFO | train_grpo_train | metrics_logged mode=train step=436

                                                      
{'loss': 0.0478, 'grad_norm': 11.516942977905273, 'learning_rate': 8.681818181818182e-06, 'num_tokens': 789814.0, 'completions/mean_length': 76.75, 'completions/min_length': 61.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.75, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.6057450771331787, 'rewards/meter/std': 0.2812247574329376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9964901804924011, 'rewards/repeat_soft/std': 0.003044117009267211, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.55422043800354, 'rewards/total_composite/std': 0.13206170499324799, 'reward': 0.55422043800354, 'reward_std': 0.13206170499324799, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1947910338640213, 'sampling/sampling_logp_difference/max': 1.2177276611328125, 'sampling/importance_sampling_ratio/min': 0.2959018051624298, 'sampling/importance_sampling_ratio/mean': 1.017616629600525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0375936776399612, 'clip_ratio/low_mean': 0.09038877952843904, 'clip_ratio/low_min': 0.09038877952843904, 'clip_ratio/high_mean': 0.07839574478566647, 'clip_ratio/high_max': 0.07839574478566647, 'clip_ratio/region_mean': 0.1687845243141055, 'reward_total_mean': 0.55422043800354, 'reward_meter_mean': 0.6057450771331787, 'reward_meter_std': 0.2812247574329376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9964901804924011, 'reward_repeat_soft_std': 0.003044117009267211, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.55422043800354, 'reward_total_composite_std': 0.13206170499324799, 'epoch': 0.04}

 13%|█▎        | 436/3300 [1:05:33<5:47:00,  7.27s/it]INFO 04-13 08:26:31 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 437/3300 [1:05:40<5:47:10,  7.28s/it]2026-04-13 08:26:38,268 | INFO | train_grpo_train | metrics_logged mode=train step=437

                                                      
{'loss': 0.0099, 'grad_norm': 13.20158863067627, 'learning_rate': 8.67878787878788e-06, 'num_tokens': 791820.0, 'completions/mean_length': 64.75, 'completions/min_length': 59.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.75, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9005851745605469, 'rewards/meter/std': 0.21653418242931366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9773539900779724, 'rewards/repeat_soft/std': 0.027312718331813812, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.6135097742080688, 'rewards/total_composite/std': 0.2840389013290405, 'reward': 0.6135097742080688, 'reward_std': 0.2840389013290405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20088978111743927, 'sampling/sampling_logp_difference/max': 1.7408978939056396, 'sampling/importance_sampling_ratio/min': 0.17536288499832153, 'sampling/importance_sampling_ratio/mean': 1.0183771848678589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6622829139232635, 'clip_ratio/low_mean': 0.1009948905557394, 'clip_ratio/low_min': 0.1009948905557394, 'clip_ratio/high_mean': 0.1130131408572197, 'clip_ratio/high_max': 0.1130131408572197, 'clip_ratio/region_mean': 0.2140080314129591, 'reward_total_mean': 0.6135097742080688, 'reward_meter_mean': 0.9005851745605469, 'reward_meter_std': 0.21653418242931366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9773539900779724, 'reward_repeat_soft_std': 0.027312718331813812, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.6135097742080688, 'reward_total_composite_std': 0.2840389013290405, 'epoch': 0.04}

 13%|█▎        | 437/3300 [1:05:40<5:47:10,  7.28s/it]INFO 04-13 08:26:38 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 438/3300 [1:05:47<5:48:57,  7.32s/it]2026-04-13 08:26:45,677 | INFO | train_grpo_train | metrics_logged mode=train step=438

                                                      
{'loss': -0.0558, 'grad_norm': 15.111183166503906, 'learning_rate': 8.675757575757576e-06, 'num_tokens': 793404.0, 'completions/mean_length': 39.0, 'completions/min_length': 32.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7226242423057556, 'rewards/meter/std': 0.3787357211112976, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969919919967651, 'rewards/repeat_soft/std': 0.006377330049872398, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.5948572158813477, 'rewards/total_composite/std': 0.16125662624835968, 'reward': 0.5948572158813477, 'reward_std': 0.16125662624835968, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19764646887779236, 'sampling/sampling_logp_difference/max': 1.359586238861084, 'sampling/importance_sampling_ratio/min': 0.25676700472831726, 'sampling/importance_sampling_ratio/mean': 1.0119014978408813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7450518012046814, 'clip_ratio/low_mean': 0.05391471879556775, 'clip_ratio/low_min': 0.05391471879556775, 'clip_ratio/high_mean': 0.1409989334642887, 'clip_ratio/high_max': 0.1409989334642887, 'clip_ratio/region_mean': 0.19491365225985646, 'reward_total_mean': 0.5948572158813477, 'reward_meter_mean': 0.7226242423057556, 'reward_meter_std': 0.3787357211112976, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969919919967651, 'reward_repeat_soft_std': 0.006377330049872398, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.5948572158813477, 'reward_total_composite_std': 0.16125662624835968, 'epoch': 0.04}

 13%|█▎        | 438/3300 [1:05:47<5:48:57,  7.32s/it]INFO 04-13 08:26:45 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 439/3300 [1:05:54<5:44:07,  7.22s/it]2026-04-13 08:26:52,665 | INFO | train_grpo_train | metrics_logged mode=train step=439

                                                      
{'loss': -0.0189, 'grad_norm': 16.915756225585938, 'learning_rate': 8.672727272727273e-06, 'num_tokens': 794921.0, 'completions/mean_length': 37.625, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5057054758071899, 'rewards/meter/std': 0.4522855579853058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9985843300819397, 'rewards/repeat_soft/std': 0.0036601137835532427, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.47689583897590637, 'rewards/total_composite/std': 0.26841363310813904, 'reward': 0.47689583897590637, 'reward_std': 0.26841360330581665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21142157912254333, 'sampling/sampling_logp_difference/max': 1.6103746891021729, 'sampling/importance_sampling_ratio/min': 0.36710667610168457, 'sampling/importance_sampling_ratio/mean': 1.0207772254943848, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.390752464532852, 'clip_ratio/low_mean': 0.0848891967907548, 'clip_ratio/low_min': 0.0848891967907548, 'clip_ratio/high_mean': 0.11912698484957218, 'clip_ratio/high_max': 0.11912698484957218, 'clip_ratio/region_mean': 0.20401618164032698, 'reward_total_mean': 0.47689583897590637, 'reward_meter_mean': 0.5057054758071899, 'reward_meter_std': 0.4522855579853058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9985843300819397, 'reward_repeat_soft_std': 0.0036601137835532427, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.47689583897590637, 'reward_total_composite_std': 0.26841363310813904, 'epoch': 0.04}

 13%|█▎        | 439/3300 [1:05:54<5:44:07,  7.22s/it]INFO 04-13 08:26:52 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 440/3300 [1:06:02<5:50:43,  7.36s/it]2026-04-13 08:27:00,350 | INFO | train_grpo_train | metrics_logged mode=train step=440

                                                      
{'loss': 0.0282, 'grad_norm': 10.141862869262695, 'learning_rate': 8.66969696969697e-06, 'num_tokens': 797454.0, 'completions/mean_length': 103.625, 'completions/min_length': 84.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.625, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.6514586806297302, 'rewards/meter/std': 0.2366323471069336, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928970336914062, 'rewards/repeat_soft/std': 0.004234022926539183, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5241789817810059, 'rewards/total_composite/std': 0.11570888757705688, 'reward': 0.5241789817810059, 'reward_std': 0.11570888012647629, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21444502472877502, 'sampling/sampling_logp_difference/max': 1.3234453201293945, 'sampling/importance_sampling_ratio/min': 0.266216516494751, 'sampling/importance_sampling_ratio/mean': 1.0347667932510376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.9164059162139893, 'clip_ratio/low_mean': 0.10031036101281643, 'clip_ratio/low_min': 0.10031036101281643, 'clip_ratio/high_mean': 0.07623598724603653, 'clip_ratio/high_max': 0.07623598724603653, 'clip_ratio/region_mean': 0.17654634825885296, 'reward_total_mean': 0.5241789817810059, 'reward_meter_mean': 0.6514586806297302, 'reward_meter_std': 0.2366323471069336, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928970336914062, 'reward_repeat_soft_std': 0.004234022926539183, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5241789817810059, 'reward_total_composite_std': 0.11570888757705688, 'epoch': 0.04}

 13%|█▎        | 440/3300 [1:06:02<5:50:43,  7.36s/it]INFO 04-13 08:27:00 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 441/3300 [1:06:09<5:44:16,  7.23s/it]2026-04-13 08:27:07,268 | INFO | train_grpo_train | metrics_logged mode=train step=441

                                                      
{'loss': 0.0436, 'grad_norm': 24.817583084106445, 'learning_rate': 8.666666666666668e-06, 'num_tokens': 798797.0, 'completions/mean_length': 28.875, 'completions/min_length': 25.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.5134502649307251, 'rewards/meter/std': 0.39729589223861694, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9864075183868408, 'rewards/repeat_soft/std': 0.013196037150919437, 'rewards/judge_quality/mean': 0.6487500071525574, 'rewards/judge_quality/std': 0.2507951855659485, 'rewards/total_composite/mean': 0.572398841381073, 'rewards/total_composite/std': 0.21716076135635376, 'reward': 0.572398841381073, 'reward_std': 0.21716077625751495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17735087871551514, 'sampling/sampling_logp_difference/max': 1.9800612926483154, 'sampling/importance_sampling_ratio/min': 0.1540224403142929, 'sampling/importance_sampling_ratio/mean': 0.9859204292297363, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0950158163905144, 'clip_ratio/low_mean': 0.08132441435009241, 'clip_ratio/low_min': 0.08132441435009241, 'clip_ratio/high_mean': 0.08135775849223137, 'clip_ratio/high_max': 0.08135775849223137, 'clip_ratio/region_mean': 0.16268217284232378, 'reward_total_mean': 0.572398841381073, 'reward_meter_mean': 0.5134502649307251, 'reward_meter_std': 0.39729589223861694, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9864075183868408, 'reward_repeat_soft_std': 0.013196037150919437, 'reward_judge_quality_mean': 0.6487500071525574, 'reward_judge_quality_std': 0.2507951855659485, 'reward_total_composite_mean': 0.572398841381073, 'reward_total_composite_std': 0.21716076135635376, 'epoch': 0.04}

 13%|█▎        | 441/3300 [1:06:09<5:44:16,  7.23s/it]INFO 04-13 08:27:07 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 442/3300 [1:06:15<5:33:39,  7.00s/it]2026-04-13 08:27:13,756 | INFO | train_grpo_train | metrics_logged mode=train step=442

                                                      
{'loss': -0.0043, 'grad_norm': 15.964285850524902, 'learning_rate': 8.663636363636363e-06, 'num_tokens': 800289.0, 'completions/mean_length': 34.5, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7339317798614502, 'rewards/meter/std': 0.3718337118625641, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9568861722946167, 'rewards/repeat_soft/std': 0.06399570405483246, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5337520837783813, 'rewards/total_composite/std': 0.1071075052022934, 'reward': 0.5337520837783813, 'reward_std': 0.1071075052022934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24122373759746552, 'sampling/sampling_logp_difference/max': 1.3552074432373047, 'sampling/importance_sampling_ratio/min': 0.25789380073547363, 'sampling/importance_sampling_ratio/mean': 1.0309560298919678, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.323012739419937, 'clip_ratio/low_mean': 0.07353896275162697, 'clip_ratio/low_min': 0.07353896275162697, 'clip_ratio/high_mean': 0.10821158112958074, 'clip_ratio/high_max': 0.10821158112958074, 'clip_ratio/region_mean': 0.1817505438812077, 'reward_total_mean': 0.5337520837783813, 'reward_meter_mean': 0.7339317798614502, 'reward_meter_std': 0.3718337118625641, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9568861722946167, 'reward_repeat_soft_std': 0.06399570405483246, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5337520837783813, 'reward_total_composite_std': 0.1071075052022934, 'epoch': 0.04}

 13%|█▎        | 442/3300 [1:06:15<5:33:39,  7.00s/it]INFO 04-13 08:27:14 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 443/3300 [1:06:24<5:52:14,  7.40s/it]2026-04-13 08:27:22,070 | INFO | train_grpo_train | metrics_logged mode=train step=443

                                                      
{'loss': 0.2345, 'grad_norm': 11.721576690673828, 'learning_rate': 8.660606060606062e-06, 'num_tokens': 802296.0, 'completions/mean_length': 74.875, 'completions/min_length': 46.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.875, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.5554511547088623, 'rewards/meter/std': 0.35469502210617065, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9882750511169434, 'rewards/repeat_soft/std': 0.011020518839359283, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.1256980448961258, 'rewards/total_composite/mean': 0.4977571368217468, 'rewards/total_composite/std': 0.10918775945901871, 'reward': 0.4977571368217468, 'reward_std': 0.10918774455785751, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23504464328289032, 'sampling/sampling_logp_difference/max': 1.4760236740112305, 'sampling/importance_sampling_ratio/min': 0.2285446673631668, 'sampling/importance_sampling_ratio/mean': 1.017876386642456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.2571473568677902, 'clip_ratio/low_mean': 0.041979849338531494, 'clip_ratio/low_min': 0.041979849338531494, 'clip_ratio/high_mean': 0.13672222942113876, 'clip_ratio/high_max': 0.13672222942113876, 'clip_ratio/region_mean': 0.17870207875967026, 'reward_total_mean': 0.4977571368217468, 'reward_meter_mean': 0.5554511547088623, 'reward_meter_std': 0.35469502210617065, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9882750511169434, 'reward_repeat_soft_std': 0.011020518839359283, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.1256980448961258, 'reward_total_composite_mean': 0.4977571368217468, 'reward_total_composite_std': 0.10918775945901871, 'epoch': 0.04}

 13%|█▎        | 443/3300 [1:06:24<5:52:14,  7.40s/it]INFO 04-13 08:27:22 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 444/3300 [1:06:36<6:58:43,  8.80s/it]2026-04-13 08:27:34,131 | INFO | train_grpo_train | metrics_logged mode=train step=444

                                                      
{'loss': -0.1005, 'grad_norm': 4.826160907745361, 'learning_rate': 8.657575757575758e-06, 'num_tokens': 804366.0, 'completions/mean_length': 132.75, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 78.5714340209961, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.8388128280639648, 'rewards/meter/std': 0.30288782715797424, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.962822437286377, 'rewards/repeat_soft/std': 0.02336599864065647, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.4479786157608032, 'rewards/total_composite/std': 0.27719706296920776, 'reward': 0.4479786157608032, 'reward_std': 0.27719706296920776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21294830739498138, 'sampling/sampling_logp_difference/max': 1.9244985580444336, 'sampling/importance_sampling_ratio/min': 0.14594891667366028, 'sampling/importance_sampling_ratio/mean': 1.0390417575836182, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0412132292985916, 'clip_ratio/low_mean': 0.023148147389292717, 'clip_ratio/low_min': 0.023148147389292717, 'clip_ratio/high_mean': 0.15710760466754436, 'clip_ratio/high_max': 0.15710760466754436, 'clip_ratio/region_mean': 0.18025575205683708, 'reward_total_mean': 0.4479786157608032, 'reward_meter_mean': 0.8388128280639648, 'reward_meter_std': 0.30288782715797424, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.962822437286377, 'reward_repeat_soft_std': 0.02336599864065647, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.4479786157608032, 'reward_total_composite_std': 0.27719706296920776, 'epoch': 0.04}

 13%|█▎        | 444/3300 [1:06:36<6:58:43,  8.80s/it]INFO 04-13 08:27:34 [block_pool.py:378] Successfully reset prefix cache

 13%|█▎        | 445/3300 [1:06:44<6:54:34,  8.71s/it]2026-04-13 08:27:42,647 | INFO | train_grpo_train | metrics_logged mode=train step=445

                                                      
{'loss': 0.0424, 'grad_norm': 8.42081069946289, 'learning_rate': 8.654545454545455e-06, 'num_tokens': 806720.0, 'completions/mean_length': 111.25, 'completions/min_length': 98.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.25, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9921140670776367, 'rewards/meter/std': 0.005870350170880556, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8212906718254089, 'rewards/repeat_soft/std': 0.18632890284061432, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.1524970829486847, 'rewards/total_composite/mean': 0.5564401149749756, 'rewards/total_composite/std': 0.10963989049196243, 'reward': 0.5564401149749756, 'reward_std': 0.10963989049196243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1865592896938324, 'sampling/sampling_logp_difference/max': 1.4757966995239258, 'sampling/importance_sampling_ratio/min': 0.2285965532064438, 'sampling/importance_sampling_ratio/mean': 1.0599757432937622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.430830806493759, 'clip_ratio/low_mean': 0.050699902698397636, 'clip_ratio/low_min': 0.050699902698397636, 'clip_ratio/high_mean': 0.09075277671217918, 'clip_ratio/high_max': 0.09075277671217918, 'clip_ratio/region_mean': 0.14145267941057682, 'reward_total_mean': 0.5564401149749756, 'reward_meter_mean': 0.9921140670776367, 'reward_meter_std': 0.005870350170880556, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8212906718254089, 'reward_repeat_soft_std': 0.18632890284061432, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.1524970829486847, 'reward_total_composite_mean': 0.5564401149749756, 'reward_total_composite_std': 0.10963989049196243, 'epoch': 0.04}

 13%|█▎        | 445/3300 [1:06:44<6:54:34,  8.71s/it]INFO 04-13 08:27:42 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 446/3300 [1:06:56<7:41:44,  9.71s/it]2026-04-13 08:27:54,678 | INFO | train_grpo_train | metrics_logged mode=train step=446

                                                      
{'loss': -0.0893, 'grad_norm': 5.285683631896973, 'learning_rate': 8.651515151515152e-06, 'num_tokens': 808322.0, 'completions/mean_length': 97.25, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.3324628472328186, 'rewards/meter/std': 0.3397580683231354, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.995673656463623, 'rewards/repeat_soft/std': 0.004901529755443335, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.276082307100296, 'rewards/total_composite/mean': 0.45673373341560364, 'rewards/total_composite/std': 0.24027307331562042, 'reward': 0.45673373341560364, 'reward_std': 0.24027305841445923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1777818351984024, 'sampling/sampling_logp_difference/max': 1.8305916786193848, 'sampling/importance_sampling_ratio/min': 0.16031868755817413, 'sampling/importance_sampling_ratio/mean': 1.0151317119598389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9830478876829147, 'clip_ratio/low_mean': 0.07306061871349812, 'clip_ratio/low_min': 0.07306061871349812, 'clip_ratio/high_mean': 0.08636224828660488, 'clip_ratio/high_max': 0.08636224828660488, 'clip_ratio/region_mean': 0.159422867000103, 'reward_total_mean': 0.45673373341560364, 'reward_meter_mean': 0.3324628472328186, 'reward_meter_std': 0.3397580683231354, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.995673656463623, 'reward_repeat_soft_std': 0.004901529755443335, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.276082307100296, 'reward_total_composite_mean': 0.45673373341560364, 'reward_total_composite_std': 0.24027307331562042, 'epoch': 0.04}

 14%|█▎        | 446/3300 [1:06:56<7:41:44,  9.71s/it]INFO 04-13 08:27:54 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 447/3300 [1:07:04<7:05:09,  8.94s/it]2026-04-13 08:28:01,832 | INFO | train_grpo_train | metrics_logged mode=train step=447

                                                      
{'loss': 0.0263, 'grad_norm': 16.588058471679688, 'learning_rate': 8.64848484848485e-06, 'num_tokens': 810082.0, 'completions/mean_length': 37.0, 'completions/min_length': 33.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6714106798171997, 'rewards/meter/std': 0.3686300814151764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9947012066841125, 'rewards/repeat_soft/std': 0.005062811076641083, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.21931305527687073, 'rewards/total_composite/mean': 0.5521893501281738, 'rewards/total_composite/std': 0.18018938601016998, 'reward': 0.5521893501281738, 'reward_std': 0.1801893711090088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23548410832881927, 'sampling/sampling_logp_difference/max': 1.2349090576171875, 'sampling/importance_sampling_ratio/min': 0.29086118936538696, 'sampling/importance_sampling_ratio/mean': 1.0399689674377441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.829873412847519, 'clip_ratio/low_mean': 0.1163807213306427, 'clip_ratio/low_min': 0.1163807213306427, 'clip_ratio/high_mean': 0.12436641566455364, 'clip_ratio/high_max': 0.12436641566455364, 'clip_ratio/region_mean': 0.24074713699519634, 'reward_total_mean': 0.5521893501281738, 'reward_meter_mean': 0.6714106798171997, 'reward_meter_std': 0.3686300814151764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9947012066841125, 'reward_repeat_soft_std': 0.005062811076641083, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.21931305527687073, 'reward_total_composite_mean': 0.5521893501281738, 'reward_total_composite_std': 0.18018938601016998, 'epoch': 0.04}

 14%|█▎        | 447/3300 [1:07:04<7:05:09,  8.94s/it]INFO 04-13 08:28:02 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 448/3300 [1:07:11<6:37:14,  8.36s/it]2026-04-13 08:28:08,828 | INFO | train_grpo_train | metrics_logged mode=train step=448

                                                      
{'loss': 0.0682, 'grad_norm': 18.4505615234375, 'learning_rate': 8.645454545454545e-06, 'num_tokens': 811625.0, 'completions/mean_length': 33.875, 'completions/min_length': 15.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8039917945861816, 'rewards/meter/std': 0.26347020268440247, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9875937700271606, 'rewards/repeat_soft/std': 0.015741875395178795, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5737026929855347, 'rewards/total_composite/std': 0.10885868966579437, 'reward': 0.5737026929855347, 'reward_std': 0.10885868221521378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22540470957756042, 'sampling/sampling_logp_difference/max': 1.8954334259033203, 'sampling/importance_sampling_ratio/min': 0.15025319159030914, 'sampling/importance_sampling_ratio/mean': 1.0656899213790894, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4711096584796906, 'clip_ratio/low_mean': 0.07114621624350548, 'clip_ratio/low_min': 0.07114621624350548, 'clip_ratio/high_mean': 0.08506806381046772, 'clip_ratio/high_max': 0.08506806381046772, 'clip_ratio/region_mean': 0.1562142800539732, 'reward_total_mean': 0.5737026929855347, 'reward_meter_mean': 0.8039917945861816, 'reward_meter_std': 0.26347020268440247, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9875937700271606, 'reward_repeat_soft_std': 0.015741875395178795, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5737026929855347, 'reward_total_composite_std': 0.10885868966579437, 'epoch': 0.05}

 14%|█▎        | 448/3300 [1:07:11<6:37:14,  8.36s/it]INFO 04-13 08:28:09 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 449/3300 [1:07:18<6:20:28,  8.01s/it]2026-04-13 08:28:16,015 | INFO | train_grpo_train | metrics_logged mode=train step=449

                                                      
{'loss': 0.0277, 'grad_norm': 11.545234680175781, 'learning_rate': 8.642424242424242e-06, 'num_tokens': 813078.0, 'completions/mean_length': 28.625, 'completions/min_length': 23.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.6695095896720886, 'rewards/meter/std': 0.3916354775428772, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9931644797325134, 'rewards/repeat_soft/std': 0.012399251572787762, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.21413865685462952, 'rewards/total_composite/mean': 0.587734580039978, 'rewards/total_composite/std': 0.20412905514240265, 'reward': 0.587734580039978, 'reward_std': 0.20412907004356384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10341665893793106, 'sampling/sampling_logp_difference/max': 1.1627860069274902, 'sampling/importance_sampling_ratio/min': 0.3126140236854553, 'sampling/importance_sampling_ratio/mean': 1.013893961906433, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7491747736930847, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.06982870260253549, 'clip_ratio/high_max': 0.06982870260253549, 'clip_ratio/region_mean': 0.08024536957964301, 'reward_total_mean': 0.587734580039978, 'reward_meter_mean': 0.6695095896720886, 'reward_meter_std': 0.3916354775428772, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9931644797325134, 'reward_repeat_soft_std': 0.012399251572787762, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.21413865685462952, 'reward_total_composite_mean': 0.587734580039978, 'reward_total_composite_std': 0.20412905514240265, 'epoch': 0.05}

 14%|█▎        | 449/3300 [1:07:18<6:20:28,  8.01s/it]INFO 04-13 08:28:16 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 450/3300 [1:07:25<6:09:15,  7.77s/it]2026-04-13 08:28:23,246 | INFO | train_grpo_train | metrics_logged mode=train step=450

                                                      
{'loss': 0.0364, 'grad_norm': 23.196592330932617, 'learning_rate': 8.63939393939394e-06, 'num_tokens': 814546.0, 'completions/mean_length': 33.5, 'completions/min_length': 28.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8268553018569946, 'rewards/meter/std': 0.3144039213657379, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9936370253562927, 'rewards/repeat_soft/std': 0.006234763655811548, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.11310552060604095, 'rewards/total_composite/mean': 0.6005456447601318, 'rewards/total_composite/std': 0.09474692493677139, 'reward': 0.6005456447601318, 'reward_std': 0.09474693238735199, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2116003930568695, 'sampling/sampling_logp_difference/max': 1.4157276153564453, 'sampling/importance_sampling_ratio/min': 0.2427489310503006, 'sampling/importance_sampling_ratio/mean': 1.039547324180603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1985170394182205, 'clip_ratio/low_mean': 0.01953125, 'clip_ratio/low_min': 0.01953125, 'clip_ratio/high_mean': 0.1654675779864192, 'clip_ratio/high_max': 0.1654675779864192, 'clip_ratio/region_mean': 0.1849988279864192, 'reward_total_mean': 0.6005456447601318, 'reward_meter_mean': 0.8268553018569946, 'reward_meter_std': 0.3144039213657379, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9936370253562927, 'reward_repeat_soft_std': 0.006234763655811548, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.11310552060604095, 'reward_total_composite_mean': 0.6005456447601318, 'reward_total_composite_std': 0.09474692493677139, 'epoch': 0.05}

 14%|█▎        | 450/3300 [1:07:25<6:09:15,  7.77s/it]INFO 04-13 08:28:23 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:10<00:40,  5.08s/it]

 30%|███       | 3/10 [00:19<00:49,  7.01s/it]

 40%|████      | 4/10 [00:29<00:47,  7.92s/it]

 50%|█████     | 5/10 [00:35<00:36,  7.20s/it]

 60%|██████    | 6/10 [00:44<00:31,  7.94s/it]

 70%|███████   | 7/10 [00:50<00:22,  7.37s/it]

 80%|████████  | 8/10 [00:55<00:13,  6.54s/it]

 90%|█████████ | 9/10 [01:05<00:07,  7.50s/it]

100%|██████████| 10/10 [01:09<00:00,  6.49s/it]2026-04-13 08:29:42,702 | INFO | train_grpo_train | metrics_logged mode=eval step=450

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 79.4198, 'eval_samples_per_second': 1.007, 'eval_steps_per_second': 0.126, 'eval_num_tokens': 814546.0, 'eval_completions/mean_length': 103.1, 'eval_completions/min_length': 26.1, 'eval_completions/max_length': 361.3, 'eval_completions/clipped_ratio': 0.0875, 'eval_completions/mean_terminated_length': 63.70952453613281, 'eval_completions/min_terminated_length': 26.1, 'eval_completions/max_terminated_length': 134.8, 'eval_rewards/meter/mean': 0.5790894836187362, 'eval_rewards/meter/std': 0.3777293890714645, 'eval_rewards/count_adherence/mean': 0.9304166555404663, 'eval_rewards/count_adherence/std': 0.12106326185166835, 'eval_rewards/hard_gate/mean': 0.875, 'eval_rewards/hard_gate/std': 0.28029437065124513, 'eval_rewards/repeat_soft/mean': 0.9575423061847687, 'eval_rewards/repeat_soft/std': 0.08738601845689117, 'eval_rewards/judge_quality/mean': 0.4236250013113022, 'eval_rewards/judge_quality/std': 0.19241945892572404, 'eval_rewards/total_composite/mean': 0.4662149131298065, 'eval_rewards/total_composite/std': 0.22698013633489608, 'eval_reward': 0.4662149131298065, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.15265945792198182, 'eval_sampling/sampling_logp_difference/max': 1.12666597366333, 'eval_sampling/importance_sampling_ratio/min': 0.32656060755252836, 'eval_sampling/importance_sampling_ratio/mean': 1.0445361137390137, 'eval_sampling/importance_sampling_ratio/max': 1.7058947801589965, 'eval_entropy': 2.8295164823532106, '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': 0.4662149131298065, 'eval_reward_meter_mean': 0.5790894836187362, 'eval_reward_meter_std': 0.3777293890714645, 'eval_reward_count_adherence_mean': 0.9304166555404663, 'eval_reward_count_adherence_std': 0.12106326185166835, 'eval_reward_hard_gate_mean': 0.875, 'eval_reward_hard_gate_std': 0.28029437065124513, 'eval_reward_repeat_soft_mean': 0.9575423061847687, 'eval_reward_repeat_soft_std': 0.08738601845689117, 'eval_reward_judge_quality_mean': 0.4236250013113022, 'eval_reward_judge_quality_std': 0.19241945892572404, 'eval_reward_total_composite_mean': 0.4662149131298065, 'eval_reward_total_composite_std': 0.22698013633489608, 'epoch': 0.05}

 14%|█▎        | 450/3300 [1:08:44<6:09:15,  7.77s/it]

100%|██████████| 10/10 [01:09<00:00,  6.49s/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 [1:08:55<25:40:27, 32.44s/it]2026-04-13 08:29:53,250 | INFO | train_grpo_train | metrics_logged mode=train step=451

                                                       
{'loss': 0.0657, 'grad_norm': 14.971288681030273, 'learning_rate': 8.636363636363637e-06, 'num_tokens': 816493.0, 'completions/mean_length': 61.375, 'completions/min_length': 57.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7840032577514648, 'rewards/meter/std': 0.33735308051109314, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9903982281684875, 'rewards/repeat_soft/std': 0.009481220506131649, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6165653467178345, 'rewards/total_composite/std': 0.1414032131433487, 'reward': 0.6165653467178345, 'reward_std': 0.1414031982421875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23850522935390472, 'sampling/sampling_logp_difference/max': 1.3375978469848633, 'sampling/importance_sampling_ratio/min': 0.2624754309654236, 'sampling/importance_sampling_ratio/mean': 1.0306938886642456, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.7483876645565033, 'clip_ratio/low_mean': 0.1360711269080639, 'clip_ratio/low_min': 0.1360711269080639, 'clip_ratio/high_mean': 0.08889990113675594, 'clip_ratio/high_max': 0.08889990113675594, 'clip_ratio/region_mean': 0.22497102804481983, 'reward_total_mean': 0.6165653467178345, 'reward_meter_mean': 0.7840032577514648, 'reward_meter_std': 0.33735308051109314, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9903982281684875, 'reward_repeat_soft_std': 0.009481220506131649, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6165653467178345, 'reward_total_composite_std': 0.1414032131433487, 'epoch': 0.05}

 14%|█▎        | 451/3300 [1:08:55<25:40:27, 32.44s/it]INFO 04-13 08:29:53 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 452/3300 [1:09:08<20:59:31, 26.53s/it]2026-04-13 08:30:06,001 | INFO | train_grpo_train | metrics_logged mode=train step=452

                                                       
{'loss': -0.1426, 'grad_norm': 4.067929267883301, 'learning_rate': 8.633333333333334e-06, 'num_tokens': 818955.0, 'completions/mean_length': 154.75, 'completions/min_length': 97.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 103.71428680419922, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.45677509903907776, 'rewards/meter/std': 0.34570378065109253, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.946021556854248, 'rewards/repeat_soft/std': 0.0689612403512001, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.4367862939834595, 'rewards/total_composite/std': 0.22509154677391052, 'reward': 0.4367862939834595, 'reward_std': 0.22509156167507172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19598762691020966, 'sampling/sampling_logp_difference/max': 1.459304928779602, 'sampling/importance_sampling_ratio/min': 0.23239775002002716, 'sampling/importance_sampling_ratio/mean': 1.031680941581726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5597435683012009, 'clip_ratio/low_mean': 0.05722092464566231, 'clip_ratio/low_min': 0.05722092464566231, 'clip_ratio/high_mean': 0.0950970258563757, 'clip_ratio/high_max': 0.0950970258563757, 'clip_ratio/region_mean': 0.152317950502038, 'reward_total_mean': 0.4367862939834595, 'reward_meter_mean': 0.45677509903907776, 'reward_meter_std': 0.34570378065109253, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.946021556854248, 'reward_repeat_soft_std': 0.0689612403512001, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.4367862939834595, 'reward_total_composite_std': 0.22509154677391052, 'epoch': 0.05}

 14%|█▎        | 452/3300 [1:09:08<20:59:31, 26.53s/it]INFO 04-13 08:30:06 [block_pool.py:378] Successfully reset prefix cache

 14%|█▎        | 453/3300 [1:09:20<17:32:01, 22.17s/it]2026-04-13 08:30:17,991 | INFO | train_grpo_train | metrics_logged mode=train step=453

                                                       
{'loss': -0.0817, 'grad_norm': 3.1986804008483887, 'learning_rate': 8.630303030303032e-06, 'num_tokens': 820538.0, 'completions/mean_length': 160.875, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 43.833335876464844, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7236465811729431, 'rewards/meter/std': 0.4409694969654083, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9389361143112183, 'rewards/repeat_soft/std': 0.0779365822672844, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.2564002275466919, 'rewards/total_composite/mean': 0.4048718810081482, 'rewards/total_composite/std': 0.3453499972820282, 'reward': 0.4048718810081482, 'reward_std': 0.3453499972820282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1954086422920227, 'sampling/sampling_logp_difference/max': 1.4350799322128296, 'sampling/importance_sampling_ratio/min': 0.23809632658958435, 'sampling/importance_sampling_ratio/mean': 1.0326827764511108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5149724185466766, 'clip_ratio/low_mean': 0.027027027681469917, 'clip_ratio/low_min': 0.027027027681469917, 'clip_ratio/high_mean': 0.1101987762376666, 'clip_ratio/high_max': 0.1101987762376666, 'clip_ratio/region_mean': 0.13722580391913652, 'reward_total_mean': 0.4048718810081482, 'reward_meter_mean': 0.7236465811729431, 'reward_meter_std': 0.4409694969654083, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9389361143112183, 'reward_repeat_soft_std': 0.0779365822672844, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.2564002275466919, 'reward_total_composite_mean': 0.4048718810081482, 'reward_total_composite_std': 0.3453499972820282, 'epoch': 0.05}

 14%|█▎        | 453/3300 [1:09:20<17:32:01, 22.17s/it]INFO 04-13 08:30:18 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 454/3300 [1:09:27<14:07:08, 17.86s/it]2026-04-13 08:30:25,788 | INFO | train_grpo_train | metrics_logged mode=train step=454

                                                       
{'loss': 0.0339, 'grad_norm': 15.092330932617188, 'learning_rate': 8.627272727272727e-06, 'num_tokens': 822099.0, 'completions/mean_length': 31.125, 'completions/min_length': 29.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9551911950111389, 'rewards/meter/std': 0.04419620335102081, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9948406219482422, 'rewards/repeat_soft/std': 0.004915553145110607, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.21822334825992584, 'rewards/total_composite/mean': 0.690136194229126, 'rewards/total_composite/std': 0.14133980870246887, 'reward': 0.690136194229126, 'reward_std': 0.14133977890014648, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18606992065906525, 'sampling/sampling_logp_difference/max': 1.713277816772461, 'sampling/importance_sampling_ratio/min': 0.18027392029762268, 'sampling/importance_sampling_ratio/mean': 1.0348671674728394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6780478358268738, 'clip_ratio/low_mean': 0.12285635527223349, 'clip_ratio/low_min': 0.12285635527223349, 'clip_ratio/high_mean': 0.04359878972172737, 'clip_ratio/high_max': 0.04359878972172737, 'clip_ratio/region_mean': 0.16645514499396086, 'reward_total_mean': 0.690136194229126, 'reward_meter_mean': 0.9551911950111389, 'reward_meter_std': 0.04419620335102081, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9948406219482422, 'reward_repeat_soft_std': 0.004915553145110607, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.21822334825992584, 'reward_total_composite_mean': 0.690136194229126, 'reward_total_composite_std': 0.14133980870246887, 'epoch': 0.05}

 14%|█▍        | 454/3300 [1:09:28<14:07:08, 17.86s/it]INFO 04-13 08:30:26 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 455/3300 [1:09:40<12:52:49, 16.30s/it]2026-04-13 08:30:38,445 | INFO | train_grpo_train | metrics_logged mode=train step=455

                                                       
{'loss': -0.1641, 'grad_norm': 2.153820514678955, 'learning_rate': 8.624242424242424e-06, 'num_tokens': 824073.0, 'completions/mean_length': 126.75, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 71.71428680419922, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.8764752745628357, 'rewards/meter/std': 0.22269082069396973, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9060655832290649, 'rewards/repeat_soft/std': 0.10061728209257126, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.5059380531311035, 'rewards/total_composite/std': 0.20810282230377197, 'reward': 0.5059380531311035, 'reward_std': 0.20810280740261078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20393435657024384, 'sampling/sampling_logp_difference/max': 1.2285990715026855, 'sampling/importance_sampling_ratio/min': 0.2927023470401764, 'sampling/importance_sampling_ratio/mean': 1.0643165111541748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.230049356818199, 'clip_ratio/low_mean': 0.017473118379712105, 'clip_ratio/low_min': 0.017473118379712105, 'clip_ratio/high_mean': 0.15428759716451168, 'clip_ratio/high_max': 0.15428759716451168, 'clip_ratio/region_mean': 0.17176071554422379, 'reward_total_mean': 0.5059380531311035, 'reward_meter_mean': 0.8764752745628357, 'reward_meter_std': 0.22269082069396973, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9060655832290649, 'reward_repeat_soft_std': 0.10061728209257126, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.5059380531311035, 'reward_total_composite_std': 0.20810282230377197, 'epoch': 0.05}

 14%|█▍        | 455/3300 [1:09:40<12:52:49, 16.30s/it]INFO 04-13 08:30:38 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 456/3300 [1:09:48<10:51:38, 13.75s/it]2026-04-13 08:30:46,244 | INFO | train_grpo_train | metrics_logged mode=train step=456

                                                       
{'loss': -0.0041, 'grad_norm': 15.241849899291992, 'learning_rate': 8.621212121212122e-06, 'num_tokens': 825872.0, 'completions/mean_length': 50.875, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5705286264419556, 'rewards/meter/std': 0.4435427486896515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980649948120117, 'rewards/repeat_soft/std': 0.0014921160181984305, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.6546416282653809, 'rewards/total_composite/std': 0.25483405590057373, 'reward': 0.6546416282653809, 'reward_std': 0.25483405590057373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10343720763921738, 'sampling/sampling_logp_difference/max': 1.5640950202941895, 'sampling/importance_sampling_ratio/min': 0.20927731692790985, 'sampling/importance_sampling_ratio/mean': 0.9993636608123779, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4827927704900503, 'clip_ratio/low_mean': 0.06564394384622574, 'clip_ratio/low_min': 0.06564394384622574, 'clip_ratio/high_mean': 0.03059012140147388, 'clip_ratio/high_max': 0.03059012140147388, 'clip_ratio/region_mean': 0.09623406524769962, 'reward_total_mean': 0.6546416282653809, 'reward_meter_mean': 0.5705286264419556, 'reward_meter_std': 0.4435427486896515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980649948120117, 'reward_repeat_soft_std': 0.0014921160181984305, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.6546416282653809, 'reward_total_composite_std': 0.25483405590057373, 'epoch': 0.05}

 14%|█▍        | 456/3300 [1:09:48<10:51:38, 13.75s/it]INFO 04-13 08:30:46 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 457/3300 [1:10:00<10:27:49, 13.25s/it]2026-04-13 08:30:58,332 | INFO | train_grpo_train | metrics_logged mode=train step=457

                                                       
{'loss': -0.0919, 'grad_norm': 3.6754684448242188, 'learning_rate': 8.618181818181819e-06, 'num_tokens': 827262.0, 'completions/mean_length': 93.75, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5208476185798645, 'rewards/meter/std': 0.33431369066238403, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.991050124168396, 'rewards/repeat_soft/std': 0.012516842223703861, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.2665520906448364, 'rewards/total_composite/mean': 0.44626879692077637, 'rewards/total_composite/std': 0.20548522472381592, 'reward': 0.44626879692077637, 'reward_std': 0.20548520982265472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2468198984861374, 'sampling/sampling_logp_difference/max': 1.9478073120117188, 'sampling/importance_sampling_ratio/min': 0.1425863802433014, 'sampling/importance_sampling_ratio/mean': 1.025658130645752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.4353190809488297, 'clip_ratio/low_mean': 0.04750503972172737, 'clip_ratio/low_min': 0.04750503972172737, 'clip_ratio/high_mean': 0.12348518334329128, 'clip_ratio/high_max': 0.12348518334329128, 'clip_ratio/region_mean': 0.17099022306501865, 'reward_total_mean': 0.44626879692077637, 'reward_meter_mean': 0.5208476185798645, 'reward_meter_std': 0.33431369066238403, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.991050124168396, 'reward_repeat_soft_std': 0.012516842223703861, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.2665520906448364, 'reward_total_composite_mean': 0.44626879692077637, 'reward_total_composite_std': 0.20548522472381592, 'epoch': 0.05}

 14%|█▍        | 457/3300 [1:10:00<10:27:49, 13.25s/it]INFO 04-13 08:30:58 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 458/3300 [1:10:07<8:53:58, 11.27s/it] 2026-04-13 08:31:04,989 | INFO | train_grpo_train | metrics_logged mode=train step=458

                                                      
{'loss': -0.0244, 'grad_norm': 16.54827308654785, 'learning_rate': 8.615151515151516e-06, 'num_tokens': 828726.0, 'completions/mean_length': 26.0, 'completions/min_length': 18.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6205868124961853, 'rewards/meter/std': 0.37930798530578613, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9410344362258911, 'rewards/repeat_soft/std': 0.03412774205207825, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.5913739204406738, 'rewards/total_composite/std': 0.22341004014015198, 'reward': 0.5913739204406738, 'reward_std': 0.22341004014015198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1890912503004074, 'sampling/sampling_logp_difference/max': 0.9961767196655273, 'sampling/importance_sampling_ratio/min': 0.3692886531352997, 'sampling/importance_sampling_ratio/mean': 1.0250221490859985, 'sampling/importance_sampling_ratio/max': 1.89702570438385, 'entropy': 1.8906628489494324, 'clip_ratio/low_mean': 0.09209030866622925, 'clip_ratio/low_min': 0.09209030866622925, 'clip_ratio/high_mean': 0.06419101729989052, 'clip_ratio/high_max': 0.06419101729989052, 'clip_ratio/region_mean': 0.15628132596611977, 'reward_total_mean': 0.5913739204406738, 'reward_meter_mean': 0.6205868124961853, 'reward_meter_std': 0.37930798530578613, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9410344362258911, 'reward_repeat_soft_std': 0.03412774205207825, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.5913739204406738, 'reward_total_composite_std': 0.22341004014015198, 'epoch': 0.05}

 14%|█▍        | 458/3300 [1:10:07<8:53:58, 11.27s/it]INFO 04-13 08:31:05 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 459/3300 [1:10:15<8:07:10, 10.29s/it]2026-04-13 08:31:12,981 | INFO | train_grpo_train | metrics_logged mode=train step=459

                                                      
{'loss': 0.1777, 'grad_norm': 17.388628005981445, 'learning_rate': 8.612121212121213e-06, 'num_tokens': 830349.0, 'completions/mean_length': 39.875, 'completions/min_length': 30.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9704955816268921, 'rewards/meter/std': 0.03484657034277916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9952265024185181, 'rewards/repeat_soft/std': 0.007804577238857746, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.19111983478069305, 'rewards/total_composite/mean': 0.7075623273849487, 'rewards/total_composite/std': 0.12064624577760696, 'reward': 0.7075623273849487, 'reward_std': 0.12064624577760696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21925927698612213, 'sampling/sampling_logp_difference/max': 1.3226232528686523, 'sampling/importance_sampling_ratio/min': 0.2664354741573334, 'sampling/importance_sampling_ratio/mean': 1.0799548625946045, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.6135420203208923, 'clip_ratio/low_mean': 0.11544453538954258, 'clip_ratio/low_min': 0.11544453538954258, 'clip_ratio/high_mean': 0.06963716261088848, 'clip_ratio/high_max': 0.06963716261088848, 'clip_ratio/region_mean': 0.18508169800043106, 'reward_total_mean': 0.7075623273849487, 'reward_meter_mean': 0.9704955816268921, 'reward_meter_std': 0.03484657034277916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9952265024185181, 'reward_repeat_soft_std': 0.007804577238857746, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.19111983478069305, 'reward_total_composite_mean': 0.7075623273849487, 'reward_total_composite_std': 0.12064624577760696, 'epoch': 0.05}

 14%|█▍        | 459/3300 [1:10:15<8:07:10, 10.29s/it]INFO 04-13 08:31:13 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 460/3300 [1:10:23<7:36:07,  9.64s/it]2026-04-13 08:31:21,108 | INFO | train_grpo_train | metrics_logged mode=train step=460

                                                      
{'loss': -0.0999, 'grad_norm': 13.907397270202637, 'learning_rate': 8.60909090909091e-06, 'num_tokens': 832391.0, 'completions/mean_length': 71.25, 'completions/min_length': 37.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.5582280158996582, 'rewards/meter/std': 0.28584811091423035, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957082271575928, 'rewards/repeat_soft/std': 0.004323306959122419, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.49182212352752686, 'rewards/total_composite/std': 0.09495563805103302, 'reward': 0.49182212352752686, 'reward_std': 0.09495565295219421, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1721809357404709, 'sampling/sampling_logp_difference/max': 1.770277976989746, 'sampling/importance_sampling_ratio/min': 0.1702856421470642, 'sampling/importance_sampling_ratio/mean': 1.0064765214920044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.007186632603407, 'clip_ratio/low_mean': 0.06141394283622503, 'clip_ratio/low_min': 0.06141394283622503, 'clip_ratio/high_mean': 0.10252820141613483, 'clip_ratio/high_max': 0.10252820141613483, 'clip_ratio/region_mean': 0.16394214425235987, 'reward_total_mean': 0.49182212352752686, 'reward_meter_mean': 0.5582280158996582, 'reward_meter_std': 0.28584811091423035, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957082271575928, 'reward_repeat_soft_std': 0.004323306959122419, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.49182212352752686, 'reward_total_composite_std': 0.09495563805103302, 'epoch': 0.05}

 14%|█▍        | 460/3300 [1:10:23<7:36:07,  9.64s/it]INFO 04-13 08:31:21 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 461/3300 [1:10:30<7:05:37,  9.00s/it]2026-04-13 08:31:28,599 | INFO | train_grpo_train | metrics_logged mode=train step=461

                                                      
{'loss': -0.057, 'grad_norm': 25.16307258605957, 'learning_rate': 8.606060606060606e-06, 'num_tokens': 833928.0, 'completions/mean_length': 28.125, 'completions/min_length': 21.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7145506739616394, 'rewards/meter/std': 0.358488529920578, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995843231678009, 'rewards/repeat_soft/std': 0.0051807500422000885, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5824846029281616, 'rewards/total_composite/std': 0.16257049143314362, 'reward': 0.5824846029281616, 'reward_std': 0.16257049143314362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19297046959400177, 'sampling/sampling_logp_difference/max': 1.3607921600341797, 'sampling/importance_sampling_ratio/min': 0.3609359860420227, 'sampling/importance_sampling_ratio/mean': 1.037373661994934, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3269875720143318, 'clip_ratio/low_mean': 0.13090106844902039, 'clip_ratio/low_min': 0.13090106844902039, 'clip_ratio/high_mean': 0.0990843316540122, 'clip_ratio/high_max': 0.0990843316540122, 'clip_ratio/region_mean': 0.2299854001030326, 'reward_total_mean': 0.5824846029281616, 'reward_meter_mean': 0.7145506739616394, 'reward_meter_std': 0.358488529920578, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995843231678009, 'reward_repeat_soft_std': 0.0051807500422000885, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5824846029281616, 'reward_total_composite_std': 0.16257049143314362, 'epoch': 0.05}

 14%|█▍        | 461/3300 [1:10:30<7:05:37,  9.00s/it]INFO 04-13 08:31:28 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 462/3300 [1:10:38<6:47:34,  8.62s/it]2026-04-13 08:31:36,329 | INFO | train_grpo_train | metrics_logged mode=train step=462

                                                      
{'loss': 0.0167, 'grad_norm': 11.881519317626953, 'learning_rate': 8.603030303030303e-06, 'num_tokens': 835639.0, 'completions/mean_length': 50.875, 'completions/min_length': 47.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8157943487167358, 'rewards/meter/std': 0.2549947202205658, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9945453405380249, 'rewards/repeat_soft/std': 0.005802301224321127, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6300034523010254, 'rewards/total_composite/std': 0.11230958998203278, 'reward': 0.6300034523010254, 'reward_std': 0.11230959743261337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17807799577713013, 'sampling/sampling_logp_difference/max': 1.3732337951660156, 'sampling/importance_sampling_ratio/min': 0.25328654050827026, 'sampling/importance_sampling_ratio/mean': 1.0447531938552856, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8934835344552994, 'clip_ratio/low_mean': 0.10798620246350765, 'clip_ratio/low_min': 0.10798620246350765, 'clip_ratio/high_mean': 0.02238159440457821, 'clip_ratio/high_max': 0.02238159440457821, 'clip_ratio/region_mean': 0.13036779686808586, 'reward_total_mean': 0.6300034523010254, 'reward_meter_mean': 0.8157943487167358, 'reward_meter_std': 0.2549947202205658, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9945453405380249, 'reward_repeat_soft_std': 0.005802301224321127, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6300034523010254, 'reward_total_composite_std': 0.11230958998203278, 'epoch': 0.05}

 14%|█▍        | 462/3300 [1:10:38<6:47:34,  8.62s/it]INFO 04-13 08:31:36 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 463/3300 [1:10:45<6:25:23,  8.15s/it]2026-04-13 08:31:43,394 | INFO | train_grpo_train | metrics_logged mode=train step=463

                                                      
{'loss': -0.0119, 'grad_norm': 16.91208839416504, 'learning_rate': 8.6e-06, 'num_tokens': 837146.0, 'completions/mean_length': 35.375, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6957108974456787, 'rewards/meter/std': 0.32157692313194275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9874091148376465, 'rewards/repeat_soft/std': 0.015234079211950302, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.7096598148345947, 'rewards/total_composite/std': 0.20204918086528778, 'reward': 0.7096598148345947, 'reward_std': 0.2020491659641266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20317187905311584, 'sampling/sampling_logp_difference/max': 1.7107229232788086, 'sampling/importance_sampling_ratio/min': 0.18073508143424988, 'sampling/importance_sampling_ratio/mean': 1.0086177587509155, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5891913622617722, 'clip_ratio/low_mean': 0.03721730224788189, 'clip_ratio/low_min': 0.03721730224788189, 'clip_ratio/high_mean': 0.10852282680571079, 'clip_ratio/high_max': 0.10852282680571079, 'clip_ratio/region_mean': 0.14574012905359268, 'reward_total_mean': 0.7096598148345947, 'reward_meter_mean': 0.6957108974456787, 'reward_meter_std': 0.32157692313194275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9874091148376465, 'reward_repeat_soft_std': 0.015234079211950302, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.7096598148345947, 'reward_total_composite_std': 0.20204918086528778, 'epoch': 0.05}

 14%|█▍        | 463/3300 [1:10:45<6:25:23,  8.15s/it]INFO 04-13 08:31:43 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 464/3300 [1:10:52<6:13:01,  7.89s/it]2026-04-13 08:31:50,681 | INFO | train_grpo_train | metrics_logged mode=train step=464

                                                      
{'loss': 0.054, 'grad_norm': 12.949996948242188, 'learning_rate': 8.596969696969698e-06, 'num_tokens': 838720.0, 'completions/mean_length': 32.75, 'completions/min_length': 31.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5962459444999695, 'rewards/meter/std': 0.33634284138679504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.964966893196106, 'rewards/repeat_soft/std': 0.011403634212911129, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5142271518707275, 'rewards/total_composite/std': 0.08487378060817719, 'reward': 0.5142271518707275, 'reward_std': 0.08487378060817719, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09167084097862244, 'sampling/sampling_logp_difference/max': 1.1023716926574707, 'sampling/importance_sampling_ratio/min': 0.3320825397968292, 'sampling/importance_sampling_ratio/mean': 1.03141450881958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7570228986442089, 'clip_ratio/low_mean': 0.030455619795247912, 'clip_ratio/low_min': 0.030455619795247912, 'clip_ratio/high_mean': 0.05427376227453351, 'clip_ratio/high_max': 0.05427376227453351, 'clip_ratio/region_mean': 0.08472938206978142, 'reward_total_mean': 0.5142271518707275, 'reward_meter_mean': 0.5962459444999695, 'reward_meter_std': 0.33634284138679504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.964966893196106, 'reward_repeat_soft_std': 0.011403634212911129, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5142271518707275, 'reward_total_composite_std': 0.08487378060817719, 'epoch': 0.05}

 14%|█▍        | 464/3300 [1:10:52<6:13:01,  7.89s/it]INFO 04-13 08:31:50 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 465/3300 [1:10:59<5:58:36,  7.59s/it]2026-04-13 08:31:57,566 | INFO | train_grpo_train | metrics_logged mode=train step=465

                                                      
{'loss': 0.0338, 'grad_norm': 19.132905960083008, 'learning_rate': 8.593939393939395e-06, 'num_tokens': 839982.0, 'completions/mean_length': 20.75, 'completions/min_length': 17.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.7611384391784668, 'rewards/meter/std': 0.33914369344711304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9602519273757935, 'rewards/repeat_soft/std': 0.006358357612043619, 'rewards/judge_quality/mean': 0.36374998092651367, 'rewards/judge_quality/std': 0.09500939399003983, 'rewards/total_composite/mean': 0.5303007960319519, 'rewards/total_composite/std': 0.09380699694156647, 'reward': 0.5303007960319519, 'reward_std': 0.09380699694156647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20365390181541443, 'sampling/sampling_logp_difference/max': 0.8794517517089844, 'sampling/importance_sampling_ratio/min': 0.41501039266586304, 'sampling/importance_sampling_ratio/mean': 1.028915524482727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1659615635871887, 'clip_ratio/low_mean': 0.11045639030635357, 'clip_ratio/low_min': 0.11045639030635357, 'clip_ratio/high_mean': 0.1121031790971756, 'clip_ratio/high_max': 0.1121031790971756, 'clip_ratio/region_mean': 0.22255956940352917, 'reward_total_mean': 0.5303007960319519, 'reward_meter_mean': 0.7611384391784668, 'reward_meter_std': 0.33914369344711304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9602519273757935, 'reward_repeat_soft_std': 0.006358357612043619, 'reward_judge_quality_mean': 0.36374998092651367, 'reward_judge_quality_std': 0.09500939399003983, 'reward_total_composite_mean': 0.5303007960319519, 'reward_total_composite_std': 0.09380699694156647, 'epoch': 0.05}

 14%|█▍        | 465/3300 [1:10:59<5:58:36,  7.59s/it]INFO 04-13 08:31:57 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 466/3300 [1:11:08<6:10:36,  7.85s/it]2026-04-13 08:32:06,009 | INFO | train_grpo_train | metrics_logged mode=train step=466

                                                      
{'loss': -0.0183, 'grad_norm': 14.528681755065918, 'learning_rate': 8.590909090909092e-06, 'num_tokens': 841861.0, 'completions/mean_length': 69.875, 'completions/min_length': 59.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7092288136482239, 'rewards/meter/std': 0.3216937780380249, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.981717586517334, 'rewards/repeat_soft/std': 0.01015958096832037, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.2133910059928894, 'rewards/total_composite/mean': 0.6228625178337097, 'rewards/total_composite/std': 0.11276643723249435, 'reward': 0.6228625178337097, 'reward_std': 0.11276643723249435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19361798465251923, 'sampling/sampling_logp_difference/max': 1.257155418395996, 'sampling/importance_sampling_ratio/min': 0.284462034702301, 'sampling/importance_sampling_ratio/mean': 1.029369831085205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9872581958770752, 'clip_ratio/low_mean': 0.1122061163187027, 'clip_ratio/low_min': 0.1122061163187027, 'clip_ratio/high_mean': 0.042305586859583855, 'clip_ratio/high_max': 0.042305586859583855, 'clip_ratio/region_mean': 0.15451170317828655, 'reward_total_mean': 0.6228625178337097, 'reward_meter_mean': 0.7092288136482239, 'reward_meter_std': 0.3216937780380249, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.981717586517334, 'reward_repeat_soft_std': 0.01015958096832037, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.2133910059928894, 'reward_total_composite_mean': 0.6228625178337097, 'reward_total_composite_std': 0.11276643723249435, 'epoch': 0.05}

 14%|█▍        | 466/3300 [1:11:08<6:10:36,  7.85s/it]INFO 04-13 08:32:06 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 467/3300 [1:11:14<5:54:50,  7.52s/it]2026-04-13 08:32:12,755 | INFO | train_grpo_train | metrics_logged mode=train step=467

                                                      
{'loss': 0.1651, 'grad_norm': 28.204408645629883, 'learning_rate': 8.587878787878788e-06, 'num_tokens': 843422.0, 'completions/mean_length': 30.125, 'completions/min_length': 23.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.125, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6777348518371582, 'rewards/meter/std': 0.3601137399673462, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9606072902679443, 'rewards/repeat_soft/std': 0.037384357303380966, 'rewards/judge_quality/mean': 0.5887500047683716, 'rewards/judge_quality/std': 0.22183892130851746, 'rewards/total_composite/mean': 0.5091491937637329, 'rewards/total_composite/std': 0.26229557394981384, 'reward': 0.5091491937637329, 'reward_std': 0.26229560375213623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2445119023323059, 'sampling/sampling_logp_difference/max': 2.081390380859375, 'sampling/importance_sampling_ratio/min': 0.12475662678480148, 'sampling/importance_sampling_ratio/mean': 1.021946668624878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9585850909352303, 'clip_ratio/low_mean': 0.08880617469549179, 'clip_ratio/low_min': 0.08880617469549179, 'clip_ratio/high_mean': 0.15226692333817482, 'clip_ratio/high_max': 0.15226692333817482, 'clip_ratio/region_mean': 0.2410730980336666, 'reward_total_mean': 0.5091491937637329, 'reward_meter_mean': 0.6777348518371582, 'reward_meter_std': 0.3601137399673462, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9606072902679443, 'reward_repeat_soft_std': 0.037384357303380966, 'reward_judge_quality_mean': 0.5887500047683716, 'reward_judge_quality_std': 0.22183892130851746, 'reward_total_composite_mean': 0.5091491937637329, 'reward_total_composite_std': 0.26229557394981384, 'epoch': 0.05}

 14%|█▍        | 467/3300 [1:11:14<5:54:50,  7.52s/it]INFO 04-13 08:32:13 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 468/3300 [1:11:22<5:49:33,  7.41s/it]2026-04-13 08:32:19,906 | INFO | train_grpo_train | metrics_logged mode=train step=468

                                                      
{'loss': 0.0665, 'grad_norm': 21.026941299438477, 'learning_rate': 8.584848484848485e-06, 'num_tokens': 845212.0, 'completions/mean_length': 48.75, 'completions/min_length': 46.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7666157484054565, 'rewards/meter/std': 0.28943267464637756, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9710464477539062, 'rewards/repeat_soft/std': 0.027212951332330704, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5549430847167969, 'rewards/total_composite/std': 0.07763629406690598, 'reward': 0.5549430847167969, 'reward_std': 0.07763627916574478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20255707204341888, 'sampling/sampling_logp_difference/max': 1.8685169219970703, 'sampling/importance_sampling_ratio/min': 0.15435241162776947, 'sampling/importance_sampling_ratio/mean': 0.9967963099479675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.243774227797985, 'clip_ratio/low_mean': 0.04402872361242771, 'clip_ratio/low_min': 0.04402872361242771, 'clip_ratio/high_mean': 0.1832888089120388, 'clip_ratio/high_max': 0.1832888089120388, 'clip_ratio/region_mean': 0.22731753252446651, 'reward_total_mean': 0.5549430847167969, 'reward_meter_mean': 0.7666157484054565, 'reward_meter_std': 0.28943267464637756, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9710464477539062, 'reward_repeat_soft_std': 0.027212951332330704, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5549430847167969, 'reward_total_composite_std': 0.07763629406690598, 'epoch': 0.05}

 14%|█▍        | 468/3300 [1:11:22<5:49:33,  7.41s/it]INFO 04-13 08:32:20 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 469/3300 [1:11:29<5:47:34,  7.37s/it]2026-04-13 08:32:27,178 | INFO | train_grpo_train | metrics_logged mode=train step=469

                                                      
{'loss': 0.0522, 'grad_norm': 15.375267028808594, 'learning_rate': 8.581818181818183e-06, 'num_tokens': 846747.0, 'completions/mean_length': 35.875, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.594794511795044, 'rewards/meter/std': 0.4424631595611572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9962560534477234, 'rewards/repeat_soft/std': 0.00541492085903883, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.5491325855255127, 'rewards/total_composite/std': 0.1585705280303955, 'reward': 0.5491325855255127, 'reward_std': 0.1585705429315567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23154234886169434, 'sampling/sampling_logp_difference/max': 1.4478940963745117, 'sampling/importance_sampling_ratio/min': 0.2350647896528244, 'sampling/importance_sampling_ratio/mean': 1.0410370826721191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.5242821127176285, 'clip_ratio/low_mean': 0.09705086797475815, 'clip_ratio/low_min': 0.09705086797475815, 'clip_ratio/high_mean': 0.1299495492130518, 'clip_ratio/high_max': 0.1299495492130518, 'clip_ratio/region_mean': 0.22700041718780994, 'reward_total_mean': 0.5491325855255127, 'reward_meter_mean': 0.594794511795044, 'reward_meter_std': 0.4424631595611572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9962560534477234, 'reward_repeat_soft_std': 0.00541492085903883, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.5491325855255127, 'reward_total_composite_std': 0.1585705280303955, 'epoch': 0.05}

 14%|█▍        | 469/3300 [1:11:29<5:47:34,  7.37s/it]INFO 04-13 08:32:27 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 470/3300 [1:11:36<5:43:07,  7.27s/it]2026-04-13 08:32:34,237 | INFO | train_grpo_train | metrics_logged mode=train step=470

                                                      
{'loss': 0.1063, 'grad_norm': 18.91813850402832, 'learning_rate': 8.57878787878788e-06, 'num_tokens': 848342.0, 'completions/mean_length': 34.375, 'completions/min_length': 31.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6956475377082825, 'rewards/meter/std': 0.3264574110507965, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9633575081825256, 'rewards/repeat_soft/std': 0.0264133308082819, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5591296553611755, 'rewards/total_composite/std': 0.09579839557409286, 'reward': 0.5591296553611755, 'reward_std': 0.09579840302467346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19767968356609344, 'sampling/sampling_logp_difference/max': 1.4106063842773438, 'sampling/importance_sampling_ratio/min': 0.2439952939748764, 'sampling/importance_sampling_ratio/mean': 1.0326085090637207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9085746258497238, 'clip_ratio/low_mean': 0.078125, 'clip_ratio/low_min': 0.078125, 'clip_ratio/high_mean': 0.10592788551002741, 'clip_ratio/high_max': 0.10592788551002741, 'clip_ratio/region_mean': 0.1840528855100274, 'reward_total_mean': 0.5591296553611755, 'reward_meter_mean': 0.6956475377082825, 'reward_meter_std': 0.3264574110507965, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9633575081825256, 'reward_repeat_soft_std': 0.0264133308082819, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5591296553611755, 'reward_total_composite_std': 0.09579839557409286, 'epoch': 0.05}

 14%|█▍        | 470/3300 [1:11:36<5:43:07,  7.27s/it]INFO 04-13 08:32:34 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 471/3300 [1:11:43<5:37:56,  7.17s/it]2026-04-13 08:32:41,154 | INFO | train_grpo_train | metrics_logged mode=train step=471

                                                      
{'loss': 0.0715, 'grad_norm': 12.654136657714844, 'learning_rate': 8.575757575757575e-06, 'num_tokens': 850102.0, 'completions/mean_length': 57.0, 'completions/min_length': 50.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.5279563665390015, 'rewards/meter/std': 0.3748530149459839, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.922204852104187, 'rewards/repeat_soft/std': 0.09402043372392654, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906257808208466, 'rewards/total_composite/mean': 0.44840341806411743, 'rewards/total_composite/std': 0.22141452133655548, 'reward': 0.44840341806411743, 'reward_std': 0.2214145064353943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21207599341869354, 'sampling/sampling_logp_difference/max': 1.8050222396850586, 'sampling/importance_sampling_ratio/min': 0.16447080671787262, 'sampling/importance_sampling_ratio/mean': 1.0279196500778198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.282945603132248, 'clip_ratio/low_mean': 0.07176537904888391, 'clip_ratio/low_min': 0.07176537904888391, 'clip_ratio/high_mean': 0.1051339004188776, 'clip_ratio/high_max': 0.1051339004188776, 'clip_ratio/region_mean': 0.17689927946776152, 'reward_total_mean': 0.44840341806411743, 'reward_meter_mean': 0.5279563665390015, 'reward_meter_std': 0.3748530149459839, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.922204852104187, 'reward_repeat_soft_std': 0.09402043372392654, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906257808208466, 'reward_total_composite_mean': 0.44840341806411743, 'reward_total_composite_std': 0.22141452133655548, 'epoch': 0.05}

 14%|█▍        | 471/3300 [1:11:43<5:37:56,  7.17s/it]INFO 04-13 08:32:41 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 472/3300 [1:11:50<5:37:51,  7.17s/it]2026-04-13 08:32:48,326 | INFO | train_grpo_train | metrics_logged mode=train step=472

                                                      
{'loss': -0.0355, 'grad_norm': 12.875102043151855, 'learning_rate': 8.572727272727274e-06, 'num_tokens': 851964.0, 'completions/mean_length': 57.75, 'completions/min_length': 49.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9655599594116211, 'rewards/meter/std': 0.04402674362063408, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8224166631698608, 'rewards/repeat_soft/std': 0.2796517312526703, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.16446885466575623, 'rewards/total_composite/mean': 0.5757567882537842, 'rewards/total_composite/std': 0.13585293292999268, 'reward': 0.5757567882537842, 'reward_std': 0.13585293292999268, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17359167337417603, 'sampling/sampling_logp_difference/max': 1.4312171936035156, 'sampling/importance_sampling_ratio/min': 0.23901782929897308, 'sampling/importance_sampling_ratio/mean': 1.0380702018737793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.810102954506874, 'clip_ratio/low_mean': 0.017087023705244064, 'clip_ratio/low_min': 0.017087023705244064, 'clip_ratio/high_mean': 0.14845621213316917, 'clip_ratio/high_max': 0.14845621213316917, 'clip_ratio/region_mean': 0.16554323583841324, 'reward_total_mean': 0.5757567882537842, 'reward_meter_mean': 0.9655599594116211, 'reward_meter_std': 0.04402674362063408, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8224166631698608, 'reward_repeat_soft_std': 0.2796517312526703, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.16446885466575623, 'reward_total_composite_mean': 0.5757567882537842, 'reward_total_composite_std': 0.13585293292999268, 'epoch': 0.05}

 14%|█▍        | 472/3300 [1:11:50<5:37:51,  7.17s/it]INFO 04-13 08:32:48 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 473/3300 [1:12:02<6:51:20,  8.73s/it]2026-04-13 08:33:00,702 | INFO | train_grpo_train | metrics_logged mode=train step=473

                                                      
{'loss': -0.0469, 'grad_norm': 7.2138872146606445, 'learning_rate': 8.56969696969697e-06, 'num_tokens': 853783.0, 'completions/mean_length': 124.375, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.45798543095588684, 'rewards/meter/std': 0.24643538892269135, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9681723713874817, 'rewards/repeat_soft/std': 0.033738333731889725, 'rewards/judge_quality/mean': 0.39750000834465027, 'rewards/judge_quality/std': 0.19031928479671478, 'rewards/total_composite/mean': 0.3047120273113251, 'rewards/total_composite/std': 0.26664313673973083, 'reward': 0.3047120273113251, 'reward_std': 0.26664313673973083, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21132715046405792, 'sampling/sampling_logp_difference/max': 1.4810504913330078, 'sampling/importance_sampling_ratio/min': 0.22739869356155396, 'sampling/importance_sampling_ratio/mean': 1.033980131149292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6572274565696716, 'clip_ratio/low_mean': 0.056581297889351845, 'clip_ratio/low_min': 0.056581297889351845, 'clip_ratio/high_mean': 0.12632135301828384, 'clip_ratio/high_max': 0.12632135301828384, 'clip_ratio/region_mean': 0.1829026509076357, 'reward_total_mean': 0.3047120273113251, 'reward_meter_mean': 0.45798543095588684, 'reward_meter_std': 0.24643538892269135, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9681723713874817, 'reward_repeat_soft_std': 0.033738333731889725, 'reward_judge_quality_mean': 0.39750000834465027, 'reward_judge_quality_std': 0.19031928479671478, 'reward_total_composite_mean': 0.3047120273113251, 'reward_total_composite_std': 0.26664313673973083, 'epoch': 0.05}

 14%|█▍        | 473/3300 [1:12:02<6:51:20,  8.73s/it]INFO 04-13 08:33:00 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 474/3300 [1:12:12<7:01:51,  8.96s/it]2026-04-13 08:33:10,187 | INFO | train_grpo_train | metrics_logged mode=train step=474

                                                      
{'loss': 1.3768, 'grad_norm': 20.031064987182617, 'learning_rate': 8.566666666666667e-06, 'num_tokens': 855436.0, 'completions/mean_length': 45.625, 'completions/min_length': 16.0, 'completions/max_length': 237.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 237.0, 'rewards/meter/mean': 0.7345542907714844, 'rewards/meter/std': 0.43264904618263245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9424101114273071, 'rewards/repeat_soft/std': 0.049144160002470016, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.15315140783786774, 'rewards/total_composite/mean': 0.4785913825035095, 'rewards/total_composite/std': 0.21766385436058044, 'reward': 0.4785913825035095, 'reward_std': 0.21766385436058044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17347753047943115, 'sampling/sampling_logp_difference/max': 0.9444046020507812, 'sampling/importance_sampling_ratio/min': 0.38891109824180603, 'sampling/importance_sampling_ratio/mean': 1.0217217206954956, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.500089704990387, 'clip_ratio/low_mean': 0.0179412798024714, 'clip_ratio/low_min': 0.0179412798024714, 'clip_ratio/high_mean': 0.1697557345032692, 'clip_ratio/high_max': 0.1697557345032692, 'clip_ratio/region_mean': 0.1876970143057406, 'reward_total_mean': 0.4785913825035095, 'reward_meter_mean': 0.7345542907714844, 'reward_meter_std': 0.43264904618263245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9424101114273071, 'reward_repeat_soft_std': 0.049144160002470016, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.15315140783786774, 'reward_total_composite_mean': 0.4785913825035095, 'reward_total_composite_std': 0.21766385436058044, 'epoch': 0.05}

 14%|█▍        | 474/3300 [1:12:12<7:01:51,  8.96s/it]INFO 04-13 08:33:10 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 475/3300 [1:12:18<6:27:12,  8.22s/it]2026-04-13 08:33:16,703 | INFO | train_grpo_train | metrics_logged mode=train step=475

                                                      
{'loss': 0.0029, 'grad_norm': 30.126554489135742, 'learning_rate': 8.563636363636364e-06, 'num_tokens': 856853.0, 'completions/mean_length': 32.125, 'completions/min_length': 29.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5700655579566956, 'rewards/meter/std': 0.41669371724128723, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.993262767791748, 'rewards/repeat_soft/std': 0.013746356591582298, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5526953339576721, 'rewards/total_composite/std': 0.1694902777671814, 'reward': 0.5526953339576721, 'reward_std': 0.1694902628660202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20681506395339966, 'sampling/sampling_logp_difference/max': 1.572995662689209, 'sampling/importance_sampling_ratio/min': 0.2074228674173355, 'sampling/importance_sampling_ratio/mean': 1.050066590309143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.492697849869728, 'clip_ratio/low_mean': 0.11188212595880032, 'clip_ratio/low_min': 0.11188212595880032, 'clip_ratio/high_mean': 0.08080914337188005, 'clip_ratio/high_max': 0.08080914337188005, 'clip_ratio/region_mean': 0.19269126933068037, 'reward_total_mean': 0.5526953339576721, 'reward_meter_mean': 0.5700655579566956, 'reward_meter_std': 0.41669371724128723, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.993262767791748, 'reward_repeat_soft_std': 0.013746356591582298, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5526953339576721, 'reward_total_composite_std': 0.1694902777671814, 'epoch': 0.05}

 14%|█▍        | 475/3300 [1:12:18<6:27:12,  8.22s/it]INFO 04-13 08:33:16 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 476/3300 [1:12:26<6:14:17,  7.95s/it]2026-04-13 08:33:24,020 | INFO | train_grpo_train | metrics_logged mode=train step=476

                                                      
{'loss': -0.0005, 'grad_norm': 19.27696418762207, 'learning_rate': 8.560606060606062e-06, 'num_tokens': 858189.0, 'completions/mean_length': 18.0, 'completions/min_length': 15.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9814795851707458, 'rewards/meter/std': 0.011831186711788177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9000933170318604, 'rewards/repeat_soft/std': 0.06073679402470589, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.3004015386104584, 'rewards/total_composite/mean': 0.6974913477897644, 'rewards/total_composite/std': 0.18535955250263214, 'reward': 0.6974913477897644, 'reward_std': 0.18535956740379333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18992812931537628, 'sampling/sampling_logp_difference/max': 1.2975273132324219, 'sampling/importance_sampling_ratio/min': 0.273206502199173, 'sampling/importance_sampling_ratio/mean': 1.0402055978775024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2534800693392754, 'clip_ratio/low_mean': 0.08107639150694013, 'clip_ratio/low_min': 0.08107639150694013, 'clip_ratio/high_mean': 0.04687001742422581, 'clip_ratio/high_max': 0.04687001742422581, 'clip_ratio/region_mean': 0.12794640893116593, 'reward_total_mean': 0.6974913477897644, 'reward_meter_mean': 0.9814795851707458, 'reward_meter_std': 0.011831186711788177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9000933170318604, 'reward_repeat_soft_std': 0.06073679402470589, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.3004015386104584, 'reward_total_composite_mean': 0.6974913477897644, 'reward_total_composite_std': 0.18535955250263214, 'epoch': 0.05}

 14%|█▍        | 476/3300 [1:12:26<6:14:17,  7.95s/it]INFO 04-13 08:33:24 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 477/3300 [1:12:33<6:00:19,  7.66s/it]2026-04-13 08:33:30,993 | INFO | train_grpo_train | metrics_logged mode=train step=477

                                                      
{'loss': 0.0299, 'grad_norm': 18.905426025390625, 'learning_rate': 8.557575757575757e-06, 'num_tokens': 859789.0, 'completions/mean_length': 34.0, 'completions/min_length': 29.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7084445953369141, 'rewards/meter/std': 0.3654259741306305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9225499629974365, 'rewards/repeat_soft/std': 0.11385763436555862, 'rewards/judge_quality/mean': 0.6612499952316284, 'rewards/judge_quality/std': 0.20883607864379883, 'rewards/total_composite/mean': 0.671148419380188, 'rewards/total_composite/std': 0.21911805868148804, 'reward': 0.671148419380188, 'reward_std': 0.21911807358264923, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2158888727426529, 'sampling/sampling_logp_difference/max': 1.3481683731079102, 'sampling/importance_sampling_ratio/min': 0.2597155272960663, 'sampling/importance_sampling_ratio/mean': 0.998117983341217, 'sampling/importance_sampling_ratio/max': 1.9984525442123413, 'entropy': 1.6007494032382965, 'clip_ratio/low_mean': 0.05719418544322252, 'clip_ratio/low_min': 0.05719418544322252, 'clip_ratio/high_mean': 0.09562562964856625, 'clip_ratio/high_max': 0.09562562964856625, 'clip_ratio/region_mean': 0.15281981509178877, 'reward_total_mean': 0.671148419380188, 'reward_meter_mean': 0.7084445953369141, 'reward_meter_std': 0.3654259741306305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9225499629974365, 'reward_repeat_soft_std': 0.11385763436555862, 'reward_judge_quality_mean': 0.6612499952316284, 'reward_judge_quality_std': 0.20883607864379883, 'reward_total_composite_mean': 0.671148419380188, 'reward_total_composite_std': 0.21911805868148804, 'epoch': 0.05}

 14%|█▍        | 477/3300 [1:12:33<6:00:19,  7.66s/it]INFO 04-13 08:33:31 [block_pool.py:378] Successfully reset prefix cache

 14%|█▍        | 478/3300 [1:12:45<7:04:33,  9.03s/it]2026-04-13 08:33:43,220 | INFO | train_grpo_train | metrics_logged mode=train step=478

                                                      
{'loss': -0.0867, 'grad_norm': 3.2877233028411865, 'learning_rate': 8.554545454545456e-06, 'num_tokens': 861386.0, 'completions/mean_length': 165.625, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 50.16666793823242, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.684908390045166, 'rewards/meter/std': 0.3519260287284851, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9943232536315918, 'rewards/repeat_soft/std': 0.0065365200862288475, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.22025957703590393, 'rewards/total_composite/mean': 0.34827280044555664, 'rewards/total_composite/std': 0.2925608158111572, 'reward': 0.34827280044555664, 'reward_std': 0.2925608158111572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23292741179466248, 'sampling/sampling_logp_difference/max': 0.9896097183227539, 'sampling/importance_sampling_ratio/min': 0.3717217445373535, 'sampling/importance_sampling_ratio/mean': 1.0913057327270508, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.8434442281723022, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.15885165706276894, 'clip_ratio/high_max': 0.15885165706276894, 'clip_ratio/region_mean': 0.17670880071818829, 'reward_total_mean': 0.34827280044555664, 'reward_meter_mean': 0.684908390045166, 'reward_meter_std': 0.3519260287284851, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9943232536315918, 'reward_repeat_soft_std': 0.0065365200862288475, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.22025957703590393, 'reward_total_composite_mean': 0.34827280044555664, 'reward_total_composite_std': 0.2925608158111572, 'epoch': 0.05}

 14%|█▍        | 478/3300 [1:12:45<7:04:33,  9.03s/it]INFO 04-13 08:33:43 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 479/3300 [1:12:57<7:52:59, 10.06s/it]2026-04-13 08:33:55,679 | INFO | train_grpo_train | metrics_logged mode=train step=479

                                                      
{'loss': -0.0661, 'grad_norm': 3.211437225341797, 'learning_rate': 8.551515151515152e-06, 'num_tokens': 862782.0, 'completions/mean_length': 82.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 21.142858505249023, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5668131709098816, 'rewards/meter/std': 0.41561824083328247, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4437500238418579, 'rewards/judge_quality/std': 0.23427319526672363, 'rewards/total_composite/mean': 0.4611922800540924, 'rewards/total_composite/std': 0.21047601103782654, 'reward': 0.4611922800540924, 'reward_std': 0.21047599613666534, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19073869287967682, 'sampling/sampling_logp_difference/max': 0.8275690078735352, 'sampling/importance_sampling_ratio/min': 0.43711063265800476, 'sampling/importance_sampling_ratio/mean': 1.0617198944091797, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8035916537046432, 'clip_ratio/low_mean': 0.08925875462591648, 'clip_ratio/low_min': 0.08925875462591648, 'clip_ratio/high_mean': 0.06343954289332032, 'clip_ratio/high_max': 0.06343954289332032, 'clip_ratio/region_mean': 0.1526982975192368, 'reward_total_mean': 0.4611922800540924, 'reward_meter_mean': 0.5668131709098816, 'reward_meter_std': 0.41561824083328247, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4437500238418579, 'reward_judge_quality_std': 0.23427319526672363, 'reward_total_composite_mean': 0.4611922800540924, 'reward_total_composite_std': 0.21047601103782654, 'epoch': 0.05}

 15%|█▍        | 479/3300 [1:12:57<7:52:59, 10.06s/it]INFO 04-13 08:33:55 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 480/3300 [1:13:09<8:14:24, 10.52s/it]2026-04-13 08:34:07,270 | INFO | train_grpo_train | metrics_logged mode=train step=480

                                                      
{'loss': -0.0958, 'grad_norm': 1.992726445198059, 'learning_rate': 8.548484848484849e-06, 'num_tokens': 864478.0, 'completions/mean_length': 161.0, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7057309150695801, 'rewards/meter/std': 0.42240098118782043, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9811805486679077, 'rewards/repeat_soft/std': 0.02527131699025631, 'rewards/judge_quality/mean': 0.30000001192092896, 'rewards/judge_quality/std': 0.18007934093475342, 'rewards/total_composite/mean': 0.41488438844680786, 'rewards/total_composite/std': 0.27114206552505493, 'reward': 0.41488438844680786, 'reward_std': 0.27114206552505493, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1979491263628006, 'sampling/sampling_logp_difference/max': 1.4906492233276367, 'sampling/importance_sampling_ratio/min': 0.22522638738155365, 'sampling/importance_sampling_ratio/mean': 1.0441219806671143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7134811282157898, 'clip_ratio/low_mean': 0.012755102477967739, 'clip_ratio/low_min': 0.012755102477967739, 'clip_ratio/high_mean': 0.10755876265466213, 'clip_ratio/high_max': 0.10755876265466213, 'clip_ratio/region_mean': 0.12031386513262987, 'reward_total_mean': 0.41488438844680786, 'reward_meter_mean': 0.7057309150695801, 'reward_meter_std': 0.42240098118782043, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9811805486679077, 'reward_repeat_soft_std': 0.02527131699025631, 'reward_judge_quality_mean': 0.30000001192092896, 'reward_judge_quality_std': 0.18007934093475342, 'reward_total_composite_mean': 0.41488438844680786, 'reward_total_composite_std': 0.27114206552505493, 'epoch': 0.05}

 15%|█▍        | 480/3300 [1:13:09<8:14:24, 10.52s/it]INFO 04-13 08:34:07 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 481/3300 [1:13:20<8:23:41, 10.72s/it]2026-04-13 08:34:18,460 | INFO | train_grpo_train | metrics_logged mode=train step=481

                                                      
{'loss': -0.0458, 'grad_norm': 1.5926659107208252, 'learning_rate': 8.545454545454546e-06, 'num_tokens': 865680.0, 'completions/mean_length': 142.25, 'completions/min_length': 13.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7509875893592834, 'rewards/meter/std': 0.4451952874660492, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9718749523162842, 'rewards/repeat_soft/std': 0.017359137535095215, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.3384418189525604, 'rewards/total_composite/mean': 0.5301734209060669, 'rewards/total_composite/std': 0.36170145869255066, 'reward': 0.5301734209060669, 'reward_std': 0.36170145869255066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17140702903270721, 'sampling/sampling_logp_difference/max': 1.5347795486450195, 'sampling/importance_sampling_ratio/min': 0.21550320088863373, 'sampling/importance_sampling_ratio/mean': 1.048685073852539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.331137478351593, 'clip_ratio/low_mean': 0.028846153989434242, 'clip_ratio/low_min': 0.028846153989434242, 'clip_ratio/high_mean': 0.08085820684209466, 'clip_ratio/high_max': 0.08085820684209466, 'clip_ratio/region_mean': 0.1097043608315289, 'reward_total_mean': 0.5301734209060669, 'reward_meter_mean': 0.7509875893592834, 'reward_meter_std': 0.4451952874660492, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9718749523162842, 'reward_repeat_soft_std': 0.017359137535095215, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.3384418189525604, 'reward_total_composite_mean': 0.5301734209060669, 'reward_total_composite_std': 0.36170145869255066, 'epoch': 0.05}

 15%|█▍        | 481/3300 [1:13:20<8:23:41, 10.72s/it]INFO 04-13 08:34:18 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 482/3300 [1:13:34<9:01:17, 11.53s/it]2026-04-13 08:34:31,879 | INFO | train_grpo_train | metrics_logged mode=train step=482

                                                      
{'loss': -0.0911, 'grad_norm': 4.708141326904297, 'learning_rate': 8.542424242424243e-06, 'num_tokens': 867433.0, 'completions/mean_length': 104.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.85714340209961, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6382356882095337, 'rewards/meter/std': 0.3718057870864868, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9953064918518066, 'rewards/repeat_soft/std': 0.006628294009715319, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.22226113080978394, 'rewards/total_composite/mean': 0.5219759941101074, 'rewards/total_composite/std': 0.25120195746421814, 'reward': 0.5219759941101074, 'reward_std': 0.25120195746421814, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23645931482315063, 'sampling/sampling_logp_difference/max': 1.9882946014404297, 'sampling/importance_sampling_ratio/min': 0.1369287371635437, 'sampling/importance_sampling_ratio/mean': 1.039041519165039, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8812013119459152, 'clip_ratio/low_mean': 0.058726416900753975, 'clip_ratio/low_min': 0.058726416900753975, 'clip_ratio/high_mean': 0.1265664417296648, 'clip_ratio/high_max': 0.1265664417296648, 'clip_ratio/region_mean': 0.18529285863041878, 'reward_total_mean': 0.5219759941101074, 'reward_meter_mean': 0.6382356882095337, 'reward_meter_std': 0.3718057870864868, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9953064918518066, 'reward_repeat_soft_std': 0.006628294009715319, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.22226113080978394, 'reward_total_composite_mean': 0.5219759941101074, 'reward_total_composite_std': 0.25120195746421814, 'epoch': 0.05}

 15%|█▍        | 482/3300 [1:13:34<9:01:17, 11.53s/it]INFO 04-13 08:34:32 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 483/3300 [1:13:46<9:10:38, 11.73s/it]2026-04-13 08:34:44,076 | INFO | train_grpo_train | metrics_logged mode=train step=483

                                                      
{'loss': -0.0851, 'grad_norm': 1.8731340169906616, 'learning_rate': 8.539393939393939e-06, 'num_tokens': 869103.0, 'completions/mean_length': 151.75, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 31.666667938232422, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7622488737106323, 'rewards/meter/std': 0.3347495496273041, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9956107139587402, 'rewards/repeat_soft/std': 0.007884573191404343, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.1718803495168686, 'rewards/total_composite/mean': 0.44280362129211426, 'rewards/total_composite/std': 0.2757055163383484, 'reward': 0.44280362129211426, 'reward_std': 0.275705486536026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22553057968616486, 'sampling/sampling_logp_difference/max': 1.406686782836914, 'sampling/importance_sampling_ratio/min': 0.24495352804660797, 'sampling/importance_sampling_ratio/mean': 1.0452746152877808, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.158455014228821, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1746977586299181, 'clip_ratio/high_max': 0.1746977586299181, 'clip_ratio/region_mean': 0.1746977586299181, 'reward_total_mean': 0.44280362129211426, 'reward_meter_mean': 0.7622488737106323, 'reward_meter_std': 0.3347495496273041, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9956107139587402, 'reward_repeat_soft_std': 0.007884573191404343, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.1718803495168686, 'reward_total_composite_mean': 0.44280362129211426, 'reward_total_composite_std': 0.2757055163383484, 'epoch': 0.05}

 15%|█▍        | 483/3300 [1:13:46<9:10:38, 11.73s/it]INFO 04-13 08:34:44 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 484/3300 [1:13:58<9:13:07, 11.79s/it]2026-04-13 08:34:55,983 | INFO | train_grpo_train | metrics_logged mode=train step=484

                                                      
{'loss': -0.0574, 'grad_norm': 1.668796420097351, 'learning_rate': 8.536363636363636e-06, 'num_tokens': 870578.0, 'completions/mean_length': 275.375, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6802476048469543, 'rewards/meter/std': 0.34469830989837646, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.9809967875480652, 'rewards/repeat_soft/std': 0.05172518268227577, 'rewards/judge_quality/mean': 0.4637500047683716, 'rewards/judge_quality/std': 0.44580385088920593, 'rewards/total_composite/mean': 0.39847850799560547, 'rewards/total_composite/std': 0.44206559658050537, 'reward': 0.39847850799560547, 'reward_std': 0.44206562638282776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23351459205150604, 'sampling/sampling_logp_difference/max': 1.6386666297912598, 'sampling/importance_sampling_ratio/min': 0.19423885643482208, 'sampling/importance_sampling_ratio/mean': 1.0209335088729858, 'sampling/importance_sampling_ratio/max': 1.936795949935913, 'entropy': 1.0624631345272064, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.12084288150072098, 'clip_ratio/high_max': 0.12084288150072098, 'clip_ratio/region_mean': 0.12084288150072098, 'reward_total_mean': 0.39847850799560547, 'reward_meter_mean': 0.6802476048469543, 'reward_meter_std': 0.34469830989837646, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.9809967875480652, 'reward_repeat_soft_std': 0.05172518268227577, 'reward_judge_quality_mean': 0.4637500047683716, 'reward_judge_quality_std': 0.44580385088920593, 'reward_total_composite_mean': 0.39847850799560547, 'reward_total_composite_std': 0.44206559658050537, 'epoch': 0.05}

 15%|█▍        | 484/3300 [1:13:58<9:13:07, 11.79s/it]INFO 04-13 08:34:56 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 485/3300 [1:14:04<7:49:32, 10.01s/it]2026-04-13 08:35:01,842 | INFO | train_grpo_train | metrics_logged mode=train step=485

                                                      
{'loss': 0.0234, 'grad_norm': 26.57079315185547, 'learning_rate': 8.533333333333335e-06, 'num_tokens': 871947.0, 'completions/mean_length': 21.125, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9872746467590332, 'rewards/meter/std': 0.006401092279702425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9247238039970398, 'rewards/repeat_soft/std': 0.08560340106487274, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.596958339214325, 'rewards/total_composite/std': 0.05244814231991768, 'reward': 0.596958339214325, 'reward_std': 0.052448149770498276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.24062158167362213, 'sampling/sampling_logp_difference/max': 1.3762474060058594, 'sampling/importance_sampling_ratio/min': 0.25252440571784973, 'sampling/importance_sampling_ratio/mean': 1.1309151649475098, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.681212991476059, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.14498528745025396, 'clip_ratio/high_max': 0.14498528745025396, 'clip_ratio/region_mean': 0.14498528745025396, 'reward_total_mean': 0.596958339214325, 'reward_meter_mean': 0.9872746467590332, 'reward_meter_std': 0.006401092279702425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9247238039970398, 'reward_repeat_soft_std': 0.08560340106487274, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.596958339214325, 'reward_total_composite_std': 0.05244814231991768, 'epoch': 0.05}

 15%|█▍        | 485/3300 [1:14:04<7:49:32, 10.01s/it]INFO 04-13 08:35:02 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 486/3300 [1:14:09<6:46:45,  8.67s/it]2026-04-13 08:35:07,400 | INFO | train_grpo_train | metrics_logged mode=train step=486

                                                      
{'loss': 0.0409, 'grad_norm': 16.259777069091797, 'learning_rate': 8.53030303030303e-06, 'num_tokens': 873338.0, 'completions/mean_length': 21.875, 'completions/min_length': 18.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9103868007659912, 'rewards/meter/std': 0.2393653243780136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9541035294532776, 'rewards/repeat_soft/std': 0.02374878153204918, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.5367006063461304, 'rewards/total_composite/std': 0.07179273664951324, 'reward': 0.5367006063461304, 'reward_std': 0.07179275155067444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18916529417037964, 'sampling/sampling_logp_difference/max': 1.0688691139221191, 'sampling/importance_sampling_ratio/min': 0.3433966338634491, 'sampling/importance_sampling_ratio/mean': 1.0399225950241089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.394902676343918, 'clip_ratio/low_mean': 0.12793267704546452, 'clip_ratio/low_min': 0.12793267704546452, 'clip_ratio/high_mean': 0.08188131637871265, 'clip_ratio/high_max': 0.08188131637871265, 'clip_ratio/region_mean': 0.20981399342417717, 'reward_total_mean': 0.5367006063461304, 'reward_meter_mean': 0.9103868007659912, 'reward_meter_std': 0.2393653243780136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9541035294532776, 'reward_repeat_soft_std': 0.02374878153204918, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.5367006063461304, 'reward_total_composite_std': 0.07179273664951324, 'epoch': 0.05}

 15%|█▍        | 486/3300 [1:14:09<6:46:45,  8.67s/it]INFO 04-13 08:35:07 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 487/3300 [1:14:15<6:09:30,  7.88s/it]2026-04-13 08:35:13,437 | INFO | train_grpo_train | metrics_logged mode=train step=487

                                                      
{'loss': 0.0231, 'grad_norm': 16.082199096679688, 'learning_rate': 8.527272727272728e-06, 'num_tokens': 875051.0, 'completions/mean_length': 34.125, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8738746643066406, 'rewards/meter/std': 0.23040351271629333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9863909482955933, 'rewards/repeat_soft/std': 0.007166656665503979, 'rewards/judge_quality/mean': 0.65625, 'rewards/judge_quality/std': 0.13741882145404816, 'rewards/total_composite/mean': 0.717002809047699, 'rewards/total_composite/std': 0.1237366646528244, 'reward': 0.717002809047699, 'reward_std': 0.1237366646528244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18899743258953094, 'sampling/sampling_logp_difference/max': 2.695510149002075, 'sampling/importance_sampling_ratio/min': 0.06750793755054474, 'sampling/importance_sampling_ratio/mean': 1.0138459205627441, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2141579985618591, 'clip_ratio/low_mean': 0.08772321604192257, 'clip_ratio/low_min': 0.08772321604192257, 'clip_ratio/high_mean': 0.13157242350280285, 'clip_ratio/high_max': 0.13157242350280285, 'clip_ratio/region_mean': 0.21929563954472542, 'reward_total_mean': 0.717002809047699, 'reward_meter_mean': 0.8738746643066406, 'reward_meter_std': 0.23040351271629333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9863909482955933, 'reward_repeat_soft_std': 0.007166656665503979, 'reward_judge_quality_mean': 0.65625, 'reward_judge_quality_std': 0.13741882145404816, 'reward_total_composite_mean': 0.717002809047699, 'reward_total_composite_std': 0.1237366646528244, 'epoch': 0.05}

 15%|█▍        | 487/3300 [1:14:15<6:09:30,  7.88s/it]INFO 04-13 08:35:13 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 488/3300 [1:14:22<5:57:51,  7.64s/it]2026-04-13 08:35:20,506 | INFO | train_grpo_train | metrics_logged mode=train step=488

                                                      
{'loss': 0.0179, 'grad_norm': 8.090012550354004, 'learning_rate': 8.524242424242425e-06, 'num_tokens': 877412.0, 'completions/mean_length': 106.125, 'completions/min_length': 91.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.125, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9861431121826172, 'rewards/meter/std': 0.010033386759459972, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7791574597358704, 'rewards/repeat_soft/std': 0.14937090873718262, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5187814235687256, 'rewards/total_composite/std': 0.06136050075292587, 'reward': 0.5187814235687256, 'reward_std': 0.06136050820350647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15641728043556213, 'sampling/sampling_logp_difference/max': 2.371082305908203, 'sampling/importance_sampling_ratio/min': 0.09337960928678513, 'sampling/importance_sampling_ratio/mean': 1.0398927927017212, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7370524555444717, 'clip_ratio/low_mean': 0.038747772574424744, 'clip_ratio/low_min': 0.038747772574424744, 'clip_ratio/high_mean': 0.10292078740894794, 'clip_ratio/high_max': 0.10292078740894794, 'clip_ratio/region_mean': 0.1416685599833727, 'reward_total_mean': 0.5187814235687256, 'reward_meter_mean': 0.9861431121826172, 'reward_meter_std': 0.010033386759459972, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7791574597358704, 'reward_repeat_soft_std': 0.14937090873718262, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5187814235687256, 'reward_total_composite_std': 0.06136050075292587, 'epoch': 0.05}

 15%|█▍        | 488/3300 [1:14:22<5:57:51,  7.64s/it]INFO 04-13 08:35:20 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 489/3300 [1:14:29<5:48:35,  7.44s/it]2026-04-13 08:35:27,483 | INFO | train_grpo_train | metrics_logged mode=train step=489

                                                      
{'loss': -0.0457, 'grad_norm': 19.640453338623047, 'learning_rate': 8.521212121212123e-06, 'num_tokens': 879081.0, 'completions/mean_length': 39.625, 'completions/min_length': 33.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8429540395736694, 'rewards/meter/std': 0.17335128784179688, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9792447090148926, 'rewards/repeat_soft/std': 0.020537715405225754, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.2377273440361023, 'rewards/total_composite/mean': 0.6701964139938354, 'rewards/total_composite/std': 0.315605491399765, 'reward': 0.6701964139938354, 'reward_std': 0.315605491399765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1898537278175354, 'sampling/sampling_logp_difference/max': 1.5631322860717773, 'sampling/importance_sampling_ratio/min': 0.20947889983654022, 'sampling/importance_sampling_ratio/mean': 1.0006773471832275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.35911126434803, 'clip_ratio/low_mean': 0.09288949333131313, 'clip_ratio/low_min': 0.09288949333131313, 'clip_ratio/high_mean': 0.1005353033542633, 'clip_ratio/high_max': 0.1005353033542633, 'clip_ratio/region_mean': 0.19342479668557644, 'reward_total_mean': 0.6701964139938354, 'reward_meter_mean': 0.8429540395736694, 'reward_meter_std': 0.17335128784179688, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9792447090148926, 'reward_repeat_soft_std': 0.020537715405225754, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.2377273440361023, 'reward_total_composite_mean': 0.6701964139938354, 'reward_total_composite_std': 0.315605491399765, 'epoch': 0.05}

 15%|█▍        | 489/3300 [1:14:29<5:48:35,  7.44s/it]INFO 04-13 08:35:27 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 490/3300 [1:14:35<5:25:57,  6.96s/it]2026-04-13 08:35:33,323 | INFO | train_grpo_train | metrics_logged mode=train step=490

                                                      
{'loss': -0.0337, 'grad_norm': 21.66672134399414, 'learning_rate': 8.518181818181818e-06, 'num_tokens': 880446.0, 'completions/mean_length': 22.625, 'completions/min_length': 13.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.625, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7413045167922974, 'rewards/meter/std': 0.4184134006500244, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9648010730743408, 'rewards/repeat_soft/std': 0.015713289380073547, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/total_composite/mean': 0.5909124612808228, 'rewards/total_composite/std': 0.2900278866291046, 'reward': 0.5909124612808228, 'reward_std': 0.2900278568267822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19013434648513794, 'sampling/sampling_logp_difference/max': 0.9521064758300781, 'sampling/importance_sampling_ratio/min': 0.40101349353790283, 'sampling/importance_sampling_ratio/mean': 1.0587222576141357, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.216610014438629, 'clip_ratio/low_mean': 0.04273504298180342, 'clip_ratio/low_min': 0.04273504298180342, 'clip_ratio/high_mean': 0.16760481614619493, 'clip_ratio/high_max': 0.16760481614619493, 'clip_ratio/region_mean': 0.21033985912799835, 'reward_total_mean': 0.5909124612808228, 'reward_meter_mean': 0.7413045167922974, 'reward_meter_std': 0.4184134006500244, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9648010730743408, 'reward_repeat_soft_std': 0.015713289380073547, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.5909124612808228, 'reward_total_composite_std': 0.2900278866291046, 'epoch': 0.05}

 15%|█▍        | 490/3300 [1:14:35<5:25:57,  6.96s/it]INFO 04-13 08:35:33 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 491/3300 [1:14:41<5:13:54,  6.70s/it]2026-04-13 08:35:39,432 | INFO | train_grpo_train | metrics_logged mode=train step=491

                                                      
{'loss': 0.0707, 'grad_norm': 14.891180992126465, 'learning_rate': 8.515151515151517e-06, 'num_tokens': 882310.0, 'completions/mean_length': 56.0, 'completions/min_length': 48.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7073996663093567, 'rewards/meter/std': 0.27903494238853455, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.962126612663269, 'rewards/repeat_soft/std': 0.033207882195711136, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291433572769165, 'rewards/total_composite/mean': 0.5687705278396606, 'rewards/total_composite/std': 0.12742707133293152, 'reward': 0.5687705278396606, 'reward_std': 0.12742707133293152, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20946232974529266, 'sampling/sampling_logp_difference/max': 1.7851076126098633, 'sampling/importance_sampling_ratio/min': 0.16777901351451874, 'sampling/importance_sampling_ratio/mean': 1.0301470756530762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7885248810052872, 'clip_ratio/low_mean': 0.08910230919718742, 'clip_ratio/low_min': 0.08910230919718742, 'clip_ratio/high_mean': 0.12386363744735718, 'clip_ratio/high_max': 0.12386363744735718, 'clip_ratio/region_mean': 0.2129659466445446, 'reward_total_mean': 0.5687705278396606, 'reward_meter_mean': 0.7073996663093567, 'reward_meter_std': 0.27903494238853455, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.962126612663269, 'reward_repeat_soft_std': 0.033207882195711136, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291433572769165, 'reward_total_composite_mean': 0.5687705278396606, 'reward_total_composite_std': 0.12742707133293152, 'epoch': 0.05}

 15%|█▍        | 491/3300 [1:14:41<5:13:54,  6.70s/it]INFO 04-13 08:35:39 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 492/3300 [1:14:48<5:09:36,  6.62s/it]2026-04-13 08:35:45,842 | INFO | train_grpo_train | metrics_logged mode=train step=492

                                                      
{'loss': -0.0054, 'grad_norm': 11.34774112701416, 'learning_rate': 8.512121212121213e-06, 'num_tokens': 884214.0, 'completions/mean_length': 62.0, 'completions/min_length': 53.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7701433300971985, 'rewards/meter/std': 0.35311195254325867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9936911463737488, 'rewards/repeat_soft/std': 0.006063162814825773, 'rewards/judge_quality/mean': 0.4699999690055847, 'rewards/judge_quality/std': 0.1414213627576828, 'rewards/total_composite/mean': 0.5916240811347961, 'rewards/total_composite/std': 0.14878517389297485, 'reward': 0.5916240811347961, 'reward_std': 0.14878515899181366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20833709836006165, 'sampling/sampling_logp_difference/max': 1.5889415740966797, 'sampling/importance_sampling_ratio/min': 0.2041415572166443, 'sampling/importance_sampling_ratio/mean': 1.0326781272888184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.634389191865921, 'clip_ratio/low_mean': 0.08448945730924606, 'clip_ratio/low_min': 0.08448945730924606, 'clip_ratio/high_mean': 0.10614589788019657, 'clip_ratio/high_max': 0.10614589788019657, 'clip_ratio/region_mean': 0.19063535518944263, 'reward_total_mean': 0.5916240811347961, 'reward_meter_mean': 0.7701433300971985, 'reward_meter_std': 0.35311195254325867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9936911463737488, 'reward_repeat_soft_std': 0.006063162814825773, 'reward_judge_quality_mean': 0.4699999690055847, 'reward_judge_quality_std': 0.1414213627576828, 'reward_total_composite_mean': 0.5916240811347961, 'reward_total_composite_std': 0.14878517389297485, 'epoch': 0.05}

 15%|█▍        | 492/3300 [1:14:48<5:09:36,  6.62s/it]INFO 04-13 08:35:46 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 493/3300 [1:14:57<5:47:14,  7.42s/it]2026-04-13 08:35:55,144 | INFO | train_grpo_train | metrics_logged mode=train step=493

                                                      
{'loss': 0.1141, 'grad_norm': 19.325977325439453, 'learning_rate': 8.50909090909091e-06, 'num_tokens': 885525.0, 'completions/mean_length': 19.875, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7870991230010986, 'rewards/meter/std': 0.35019874572753906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.6802371740341187, 'rewards/total_composite/std': 0.21170316636562347, 'reward': 0.6802371740341187, 'reward_std': 0.21170316636562347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20156840980052948, 'sampling/sampling_logp_difference/max': 1.8572874069213867, 'sampling/importance_sampling_ratio/min': 0.1560954749584198, 'sampling/importance_sampling_ratio/mean': 1.022932767868042, 'sampling/importance_sampling_ratio/max': 1.8618160486221313, 'entropy': 1.7631626427173615, 'clip_ratio/low_mean': 0.10885781142860651, 'clip_ratio/low_min': 0.10885781142860651, 'clip_ratio/high_mean': 0.06308049615472555, 'clip_ratio/high_max': 0.06308049615472555, 'clip_ratio/region_mean': 0.17193830758333206, 'reward_total_mean': 0.6802371740341187, 'reward_meter_mean': 0.7870991230010986, 'reward_meter_std': 0.35019874572753906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.6802371740341187, 'reward_total_composite_std': 0.21170316636562347, 'epoch': 0.05}

 15%|█▍        | 493/3300 [1:14:57<5:47:14,  7.42s/it]INFO 04-13 08:35:55 [block_pool.py:378] Successfully reset prefix cache

 15%|█▍        | 494/3300 [1:15:08<6:38:31,  8.52s/it]2026-04-13 08:36:06,230 | INFO | train_grpo_train | metrics_logged mode=train step=494

                                                      
{'loss': -0.1293, 'grad_norm': 7.374967575073242, 'learning_rate': 8.506060606060607e-06, 'num_tokens': 887357.0, 'completions/mean_length': 128.0, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.14286041259766, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.7525448799133301, 'rewards/meter/std': 0.28220289945602417, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9910486936569214, 'rewards/repeat_soft/std': 0.007813452742993832, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.21625052392482758, 'rewards/total_composite/mean': 0.48917150497436523, 'rewards/total_composite/std': 0.31889182329177856, 'reward': 0.48917150497436523, 'reward_std': 0.31889182329177856, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2065785527229309, 'sampling/sampling_logp_difference/max': 2.0177111625671387, 'sampling/importance_sampling_ratio/min': 0.13295944035053253, 'sampling/importance_sampling_ratio/mean': 0.9862677454948425, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1765784621238708, 'clip_ratio/low_mean': 0.04494047723710537, 'clip_ratio/low_min': 0.04494047723710537, 'clip_ratio/high_mean': 0.11561843566596508, 'clip_ratio/high_max': 0.11561843566596508, 'clip_ratio/region_mean': 0.16055891290307045, 'reward_total_mean': 0.48917150497436523, 'reward_meter_mean': 0.7525448799133301, 'reward_meter_std': 0.28220289945602417, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9910486936569214, 'reward_repeat_soft_std': 0.007813452742993832, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.21625052392482758, 'reward_total_composite_mean': 0.48917150497436523, 'reward_total_composite_std': 0.31889182329177856, 'epoch': 0.05}

 15%|█▍        | 494/3300 [1:15:08<6:38:31,  8.52s/it]INFO 04-13 08:36:06 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 495/3300 [1:15:15<6:11:15,  7.94s/it]2026-04-13 08:36:12,818 | INFO | train_grpo_train | metrics_logged mode=train step=495

                                                      
{'loss': 0.0185, 'grad_norm': 17.67508888244629, 'learning_rate': 8.503030303030304e-06, 'num_tokens': 888939.0, 'completions/mean_length': 35.75, 'completions/min_length': 30.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8715434074401855, 'rewards/meter/std': 0.2754010856151581, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9766156077384949, 'rewards/repeat_soft/std': 0.023068033158779144, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6576051115989685, 'rewards/total_composite/std': 0.15992027521133423, 'reward': 0.6576051115989685, 'reward_std': 0.15992026031017303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15880848467350006, 'sampling/sampling_logp_difference/max': 1.0550956726074219, 'sampling/importance_sampling_ratio/min': 0.3481591045856476, 'sampling/importance_sampling_ratio/mean': 1.0528186559677124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5566628128290176, 'clip_ratio/low_mean': 0.15518195182085037, 'clip_ratio/low_min': 0.15518195182085037, 'clip_ratio/high_mean': 0.045036764815449715, 'clip_ratio/high_max': 0.045036764815449715, 'clip_ratio/region_mean': 0.2002187166363001, 'reward_total_mean': 0.6576051115989685, 'reward_meter_mean': 0.8715434074401855, 'reward_meter_std': 0.2754010856151581, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9766156077384949, 'reward_repeat_soft_std': 0.023068033158779144, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6576051115989685, 'reward_total_composite_std': 0.15992027521133423, 'epoch': 0.05}

 15%|█▌        | 495/3300 [1:15:15<6:11:15,  7.94s/it]INFO 04-13 08:36:13 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 496/3300 [1:15:20<5:42:30,  7.33s/it]2026-04-13 08:36:18,721 | INFO | train_grpo_train | metrics_logged mode=train step=496

                                                      
{'loss': 0.0684, 'grad_norm': 20.635723114013672, 'learning_rate': 8.5e-06, 'num_tokens': 890393.0, 'completions/mean_length': 19.75, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9725742340087891, 'rewards/meter/std': 0.040043070912361145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9284564256668091, 'rewards/repeat_soft/std': 0.08202844858169556, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.20860078930854797, 'rewards/total_composite/mean': 0.6181299090385437, 'rewards/total_composite/std': 0.13773126900196075, 'reward': 0.6181299090385437, 'reward_std': 0.13773126900196075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22871235013008118, 'sampling/sampling_logp_difference/max': 1.5547971725463867, 'sampling/importance_sampling_ratio/min': 0.21123221516609192, 'sampling/importance_sampling_ratio/mean': 1.041054129600525, 'sampling/importance_sampling_ratio/max': 1.9155250787734985, 'entropy': 2.313529849052429, 'clip_ratio/low_mean': 0.15082208067178726, 'clip_ratio/low_min': 0.15082208067178726, 'clip_ratio/high_mean': 0.013888888992369175, 'clip_ratio/high_max': 0.013888888992369175, 'clip_ratio/region_mean': 0.16471096966415644, 'reward_total_mean': 0.6181299090385437, 'reward_meter_mean': 0.9725742340087891, 'reward_meter_std': 0.040043070912361145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9284564256668091, 'reward_repeat_soft_std': 0.08202844858169556, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.20860078930854797, 'reward_total_composite_mean': 0.6181299090385437, 'reward_total_composite_std': 0.13773126900196075, 'epoch': 0.05}

 15%|█▌        | 496/3300 [1:15:20<5:42:30,  7.33s/it]INFO 04-13 08:36:18 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 497/3300 [1:15:29<5:55:17,  7.61s/it]2026-04-13 08:36:26,968 | INFO | train_grpo_train | metrics_logged mode=train step=497

                                                      
{'loss': 0.4003, 'grad_norm': 10.301472663879395, 'learning_rate': 8.496969696969697e-06, 'num_tokens': 892350.0, 'completions/mean_length': 88.625, 'completions/min_length': 55.0, 'completions/max_length': 251.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 251.0, 'rewards/meter/mean': 0.5521746873855591, 'rewards/meter/std': 0.4495016932487488, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9963110089302063, 'rewards/repeat_soft/std': 0.0030512267258018255, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.40128767490386963, 'rewards/total_composite/std': 0.28263911604881287, 'reward': 0.40128767490386963, 'reward_std': 0.28263911604881287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18699179589748383, 'sampling/sampling_logp_difference/max': 1.7110190391540527, 'sampling/importance_sampling_ratio/min': 0.18068158626556396, 'sampling/importance_sampling_ratio/mean': 1.0362626314163208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 3.036473363637924, 'clip_ratio/low_mean': 0.051028044894337654, 'clip_ratio/low_min': 0.051028044894337654, 'clip_ratio/high_mean': 0.09508397709578276, 'clip_ratio/high_max': 0.09508397709578276, 'clip_ratio/region_mean': 0.1461120219901204, 'reward_total_mean': 0.40128767490386963, 'reward_meter_mean': 0.5521746873855591, 'reward_meter_std': 0.4495016932487488, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9963110089302063, 'reward_repeat_soft_std': 0.0030512267258018255, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.40128767490386963, 'reward_total_composite_std': 0.28263911604881287, 'epoch': 0.05}

 15%|█▌        | 497/3300 [1:15:29<5:55:17,  7.61s/it]INFO 04-13 08:36:27 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 498/3300 [1:15:35<5:35:15,  7.18s/it]2026-04-13 08:36:33,154 | INFO | train_grpo_train | metrics_logged mode=train step=498

                                                      
{'loss': -0.0166, 'grad_norm': 19.435779571533203, 'learning_rate': 8.493939393939394e-06, 'num_tokens': 893804.0, 'completions/mean_length': 31.75, 'completions/min_length': 26.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.709601640701294, 'rewards/meter/std': 0.4097140431404114, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9933873414993286, 'rewards/repeat_soft/std': 0.0044791982509195805, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.19167962670326233, 'rewards/total_composite/mean': 0.6304831504821777, 'rewards/total_composite/std': 0.18362566828727722, 'reward': 0.6304831504821777, 'reward_std': 0.18362563848495483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19538196921348572, 'sampling/sampling_logp_difference/max': 1.905601978302002, 'sampling/importance_sampling_ratio/min': 0.14873307943344116, 'sampling/importance_sampling_ratio/mean': 1.0222121477127075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9627540111541748, 'clip_ratio/low_mean': 0.0757354386150837, 'clip_ratio/low_min': 0.0757354386150837, 'clip_ratio/high_mean': 0.07185769081115723, 'clip_ratio/high_max': 0.07185769081115723, 'clip_ratio/region_mean': 0.14759312942624092, 'reward_total_mean': 0.6304831504821777, 'reward_meter_mean': 0.709601640701294, 'reward_meter_std': 0.4097140431404114, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9933873414993286, 'reward_repeat_soft_std': 0.0044791982509195805, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.19167962670326233, 'reward_total_composite_mean': 0.6304831504821777, 'reward_total_composite_std': 0.18362566828727722, 'epoch': 0.05}

 15%|█▌        | 498/3300 [1:15:35<5:35:15,  7.18s/it]INFO 04-13 08:36:33 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 499/3300 [1:15:41<5:22:00,  6.90s/it]2026-04-13 08:36:39,395 | INFO | train_grpo_train | metrics_logged mode=train step=499

                                                      
{'loss': 0.0002, 'grad_norm': 10.80377197265625, 'learning_rate': 8.490909090909092e-06, 'num_tokens': 895363.0, 'completions/mean_length': 47.875, 'completions/min_length': 42.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8173096179962158, 'rewards/meter/std': 0.30853766202926636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9982814788818359, 'rewards/repeat_soft/std': 0.003359787864610553, 'rewards/judge_quality/mean': 0.6974999904632568, 'rewards/judge_quality/std': 0.2247379571199417, 'rewards/total_composite/mean': 0.6352838277816772, 'rewards/total_composite/std': 0.31333625316619873, 'reward': 0.6352838277816772, 'reward_std': 0.31333625316619873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15956829488277435, 'sampling/sampling_logp_difference/max': 2.1714234352111816, 'sampling/importance_sampling_ratio/min': 0.11401520669460297, 'sampling/importance_sampling_ratio/mean': 1.011419653892517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1996866837143898, 'clip_ratio/low_mean': 0.061936674639582634, 'clip_ratio/low_min': 0.061936674639582634, 'clip_ratio/high_mean': 0.10531785013154149, 'clip_ratio/high_max': 0.10531785013154149, 'clip_ratio/region_mean': 0.16725452477112412, 'reward_total_mean': 0.6352838277816772, 'reward_meter_mean': 0.8173096179962158, 'reward_meter_std': 0.30853766202926636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9982814788818359, 'reward_repeat_soft_std': 0.003359787864610553, 'reward_judge_quality_mean': 0.6974999904632568, 'reward_judge_quality_std': 0.2247379571199417, 'reward_total_composite_mean': 0.6352838277816772, 'reward_total_composite_std': 0.31333625316619873, 'epoch': 0.05}

 15%|█▌        | 499/3300 [1:15:41<5:22:00,  6.90s/it]INFO 04-13 08:36:39 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 500/3300 [1:15:53<6:30:19,  8.36s/it]2026-04-13 08:36:51,181 | INFO | train_grpo_train | metrics_logged mode=train step=500

                                                      
{'loss': -0.1111, 'grad_norm': 2.1623306274414062, 'learning_rate': 8.487878787878789e-06, 'num_tokens': 897051.0, 'completions/mean_length': 293.0, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.38441598415374756, 'rewards/meter/std': 0.41772931814193726, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.9976768493652344, 'rewards/repeat_soft/std': 0.003024688921868801, 'rewards/judge_quality/mean': 0.33500000834465027, 'rewards/judge_quality/std': 0.3443005383014679, 'rewards/total_composite/mean': 0.32037678360939026, 'rewards/total_composite/std': 0.3601384162902832, 'reward': 0.32037678360939026, 'reward_std': 0.3601383864879608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1918148249387741, 'sampling/sampling_logp_difference/max': 1.6068658828735352, 'sampling/importance_sampling_ratio/min': 0.20051506161689758, 'sampling/importance_sampling_ratio/mean': 1.045210838317871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9770064651966095, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08429248817265034, 'clip_ratio/high_max': 0.08429248817265034, 'clip_ratio/region_mean': 0.08429248817265034, 'reward_total_mean': 0.32037678360939026, 'reward_meter_mean': 0.38441598415374756, 'reward_meter_std': 0.41772931814193726, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.9976768493652344, 'reward_repeat_soft_std': 0.003024688921868801, 'reward_judge_quality_mean': 0.33500000834465027, 'reward_judge_quality_std': 0.3443005383014679, 'reward_total_composite_mean': 0.32037678360939026, 'reward_total_composite_std': 0.3601384162902832, 'epoch': 0.05}

 15%|█▌        | 500/3300 [1:15:53<6:30:19,  8.36s/it]INFO 04-13 08:36:51 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.80s/it]

 30%|███       | 3/10 [00:08<00:20,  2.99s/it]

 40%|████      | 4/10 [00:11<00:19,  3.17s/it]

 50%|█████     | 5/10 [00:16<00:18,  3.68s/it]

 60%|██████    | 6/10 [00:27<00:24,  6.02s/it]

 70%|███████   | 7/10 [00:31<00:16,  5.48s/it]

 80%|████████  | 8/10 [00:35<00:10,  5.12s/it]

 90%|█████████ | 9/10 [00:44<00:06,  6.30s/it]

100%|██████████| 10/10 [00:47<00:00,  5.37s/it]2026-04-13 08:37:43,626 | INFO | train_grpo_train | metrics_logged mode=eval step=500

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 52.4101, 'eval_samples_per_second': 1.526, 'eval_steps_per_second': 0.191, 'eval_num_tokens': 897051.0, 'eval_completions/mean_length': 73.825, 'eval_completions/min_length': 27.2, 'eval_completions/max_length': 198.2, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 62.648214721679686, 'eval_completions/min_terminated_length': 27.2, 'eval_completions/max_terminated_length': 119.9, 'eval_rewards/meter/mean': 0.6945112824440003, 'eval_rewards/meter/std': 0.3352689057588577, 'eval_rewards/count_adherence/mean': 0.9570833265781402, 'eval_rewards/count_adherence/std': 0.07772159017622471, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.11700168251991272, 'eval_rewards/repeat_soft/mean': 0.9513344824314117, 'eval_rewards/repeat_soft/std': 0.06777058504521846, 'eval_rewards/judge_quality/mean': 0.4666249960660934, 'eval_rewards/judge_quality/std': 0.15483680814504625, 'eval_rewards/total_composite/mean': 0.5227623641490936, 'eval_rewards/total_composite/std': 0.15929170995950698, 'eval_reward': 0.5227623641490936, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.1298849441111088, 'eval_sampling/sampling_logp_difference/max': 1.1411276817321778, 'eval_sampling/importance_sampling_ratio/min': 0.32457238137722016, 'eval_sampling/importance_sampling_ratio/mean': 1.0405490040779113, 'eval_sampling/importance_sampling_ratio/max': 1.5930978775024414, 'eval_entropy': 1.8423508882522583, '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': 0.5227623641490936, 'eval_reward_meter_mean': 0.6945112824440003, 'eval_reward_meter_std': 0.3352689057588577, 'eval_reward_count_adherence_mean': 0.9570833265781402, 'eval_reward_count_adherence_std': 0.07772159017622471, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.11700168251991272, 'eval_reward_repeat_soft_mean': 0.9513344824314117, 'eval_reward_repeat_soft_std': 0.06777058504521846, 'eval_reward_judge_quality_mean': 0.4666249960660934, 'eval_reward_judge_quality_std': 0.15483680814504625, 'eval_reward_total_composite_mean': 0.5227623641490936, 'eval_reward_total_composite_std': 0.15929170995950698, 'epoch': 0.05}

 15%|█▌        | 500/3300 [1:16:45<6:30:19,  8.36s/it]

100%|██████████| 10/10 [00:48<00:00,  5.37s/it]

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

 15%|█▌        | 501/3300 [1:16:54<18:55:17, 24.34s/it]2026-04-13 08:37:52,792 | INFO | train_grpo_train | metrics_logged mode=train step=501

                                                       
{'loss': 0.0167, 'grad_norm': 16.002286911010742, 'learning_rate': 8.484848484848486e-06, 'num_tokens': 898809.0, 'completions/mean_length': 50.75, 'completions/min_length': 45.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7950263023376465, 'rewards/meter/std': 0.26878464221954346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9539934992790222, 'rewards/repeat_soft/std': 0.03938141465187073, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5769858360290527, 'rewards/total_composite/std': 0.0846138745546341, 'reward': 0.5769858360290527, 'reward_std': 0.0846138596534729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1991541087627411, 'sampling/sampling_logp_difference/max': 1.085914134979248, 'sampling/importance_sampling_ratio/min': 0.34946075081825256, 'sampling/importance_sampling_ratio/mean': 1.047680139541626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8200058937072754, 'clip_ratio/low_mean': 0.043792517855763435, 'clip_ratio/low_min': 0.043792517855763435, 'clip_ratio/high_mean': 0.12483904603868723, 'clip_ratio/high_max': 0.12483904603868723, 'clip_ratio/region_mean': 0.16863156389445066, 'reward_total_mean': 0.5769858360290527, 'reward_meter_mean': 0.7950263023376465, 'reward_meter_std': 0.26878464221954346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9539934992790222, 'reward_repeat_soft_std': 0.03938141465187073, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5769858360290527, 'reward_total_composite_std': 0.0846138745546341, 'epoch': 0.05}

 15%|█▌        | 501/3300 [1:16:55<18:55:17, 24.34s/it]INFO 04-13 08:37:53 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 502/3300 [1:17:02<14:53:32, 19.16s/it]2026-04-13 08:37:59,875 | INFO | train_grpo_train | metrics_logged mode=train step=502

                                                       
{'loss': 0.0449, 'grad_norm': 9.340448379516602, 'learning_rate': 8.481818181818182e-06, 'num_tokens': 901101.0, 'completions/mean_length': 98.5, 'completions/min_length': 83.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.5, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.6727849841117859, 'rewards/meter/std': 0.3345721662044525, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9954895973205566, 'rewards/repeat_soft/std': 0.005467571318149567, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.1524970829486847, 'rewards/total_composite/mean': 0.5365538001060486, 'rewards/total_composite/std': 0.1377352625131607, 'reward': 0.5365538001060486, 'reward_std': 0.1377352476119995, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21152883768081665, 'sampling/sampling_logp_difference/max': 2.0427350997924805, 'sampling/importance_sampling_ratio/min': 0.1296735554933548, 'sampling/importance_sampling_ratio/mean': 1.0404530763626099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.467586889863014, 'clip_ratio/low_mean': 0.10863708704710007, 'clip_ratio/low_min': 0.10863708704710007, 'clip_ratio/high_mean': 0.08720150776207447, 'clip_ratio/high_max': 0.08720150776207447, 'clip_ratio/region_mean': 0.19583859480917454, 'reward_total_mean': 0.5365538001060486, 'reward_meter_mean': 0.6727849841117859, 'reward_meter_std': 0.3345721662044525, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9954895973205566, 'reward_repeat_soft_std': 0.005467571318149567, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.1524970829486847, 'reward_total_composite_mean': 0.5365538001060486, 'reward_total_composite_std': 0.1377352625131607, 'epoch': 0.05}

 15%|█▌        | 502/3300 [1:17:02<14:53:32, 19.16s/it]INFO 04-13 08:38:00 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 503/3300 [1:17:13<13:01:52, 16.77s/it]2026-04-13 08:38:11,072 | INFO | train_grpo_train | metrics_logged mode=train step=503

                                                       
{'loss': -0.0118, 'grad_norm': 5.307713985443115, 'learning_rate': 8.478787878787879e-06, 'num_tokens': 903201.0, 'completions/mean_length': 134.5, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.5714340209961, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8535789847373962, 'rewards/meter/std': 0.24601230025291443, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9460170269012451, 'rewards/repeat_soft/std': 0.045325759798288345, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.24744409322738647, 'rewards/total_composite/mean': 0.45327916741371155, 'rewards/total_composite/std': 0.29592910408973694, 'reward': 0.45327916741371155, 'reward_std': 0.29592907428741455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2000032216310501, 'sampling/sampling_logp_difference/max': 1.9902896881103516, 'sampling/importance_sampling_ratio/min': 0.13665582239627838, 'sampling/importance_sampling_ratio/mean': 1.0321557521820068, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1861378848552704, 'clip_ratio/low_mean': 0.008858268149197102, 'clip_ratio/low_min': 0.008858268149197102, 'clip_ratio/high_mean': 0.13697792310267687, 'clip_ratio/high_max': 0.13697792310267687, 'clip_ratio/region_mean': 0.14583619125187397, 'reward_total_mean': 0.45327916741371155, 'reward_meter_mean': 0.8535789847373962, 'reward_meter_std': 0.24601230025291443, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9460170269012451, 'reward_repeat_soft_std': 0.045325759798288345, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.24744409322738647, 'reward_total_composite_mean': 0.45327916741371155, 'reward_total_composite_std': 0.29592910408973694, 'epoch': 0.05}

 15%|█▌        | 503/3300 [1:17:13<13:01:52, 16.77s/it]INFO 04-13 08:38:11 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 504/3300 [1:17:24<11:40:23, 15.03s/it]2026-04-13 08:38:22,036 | INFO | train_grpo_train | metrics_logged mode=train step=504

                                                       
{'loss': -0.0667, 'grad_norm': 1.7402321100234985, 'learning_rate': 8.475757575757576e-06, 'num_tokens': 904632.0, 'completions/mean_length': 81.875, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 20.428571701049805, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8614223599433899, 'rewards/meter/std': 0.3481668531894684, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9605586528778076, 'rewards/repeat_soft/std': 0.024460788816213608, 'rewards/judge_quality/mean': 0.35999998450279236, 'rewards/judge_quality/std': 0.140813946723938, 'rewards/total_composite/mean': 0.5268834829330444, 'rewards/total_composite/std': 0.21758151054382324, 'reward': 0.5268834829330444, 'reward_std': 0.21758146584033966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1638796031475067, 'sampling/sampling_logp_difference/max': 0.9651508331298828, 'sampling/importance_sampling_ratio/min': 0.3809257447719574, 'sampling/importance_sampling_ratio/mean': 1.017634630203247, 'sampling/importance_sampling_ratio/max': 1.5614029169082642, 'entropy': 1.234602004289627, 'clip_ratio/low_mean': 0.004464285913854837, 'clip_ratio/low_min': 0.004464285913854837, 'clip_ratio/high_mean': 0.09794001141563058, 'clip_ratio/high_max': 0.09794001141563058, 'clip_ratio/region_mean': 0.10240429732948542, 'reward_total_mean': 0.5268834829330444, 'reward_meter_mean': 0.8614223599433899, 'reward_meter_std': 0.3481668531894684, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9605586528778076, 'reward_repeat_soft_std': 0.024460788816213608, 'reward_judge_quality_mean': 0.35999998450279236, 'reward_judge_quality_std': 0.140813946723938, 'reward_total_composite_mean': 0.5268834829330444, 'reward_total_composite_std': 0.21758151054382324, 'epoch': 0.05}

 15%|█▌        | 504/3300 [1:17:24<11:40:23, 15.03s/it]INFO 04-13 08:38:22 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 505/3300 [1:17:30<9:34:15, 12.33s/it] 2026-04-13 08:38:28,057 | INFO | train_grpo_train | metrics_logged mode=train step=505

                                                      
{'loss': 0.1209, 'grad_norm': 16.707841873168945, 'learning_rate': 8.472727272727274e-06, 'num_tokens': 906189.0, 'completions/mean_length': 39.625, 'completions/min_length': 35.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9801276326179504, 'rewards/meter/std': 0.013514342717826366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9287209510803223, 'rewards/repeat_soft/std': 0.08914603292942047, 'rewards/judge_quality/mean': 0.42499998211860657, 'rewards/judge_quality/std': 0.0707106739282608, 'rewards/total_composite/mean': 0.6100842952728271, 'rewards/total_composite/std': 0.04736431688070297, 'reward': 0.6100842952728271, 'reward_std': 0.047364309430122375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21045367419719696, 'sampling/sampling_logp_difference/max': 1.5575942993164062, 'sampling/importance_sampling_ratio/min': 0.21064220368862152, 'sampling/importance_sampling_ratio/mean': 1.0381056070327759, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.1052683740854263, 'clip_ratio/low_mean': 0.03557233791798353, 'clip_ratio/low_min': 0.03557233791798353, 'clip_ratio/high_mean': 0.16790770925581455, 'clip_ratio/high_max': 0.16790770925581455, 'clip_ratio/region_mean': 0.20348004717379808, 'reward_total_mean': 0.6100842952728271, 'reward_meter_mean': 0.9801276326179504, 'reward_meter_std': 0.013514342717826366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9287209510803223, 'reward_repeat_soft_std': 0.08914603292942047, 'reward_judge_quality_mean': 0.42499998211860657, 'reward_judge_quality_std': 0.0707106739282608, 'reward_total_composite_mean': 0.6100842952728271, 'reward_total_composite_std': 0.04736431688070297, 'epoch': 0.05}

 15%|█▌        | 505/3300 [1:17:30<9:34:15, 12.33s/it]INFO 04-13 08:38:28 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 506/3300 [1:17:36<8:07:08, 10.46s/it]2026-04-13 08:38:34,164 | INFO | train_grpo_train | metrics_logged mode=train step=506

                                                      
{'loss': 0.0806, 'grad_norm': 13.86902141571045, 'learning_rate': 8.46969696969697e-06, 'num_tokens': 907717.0, 'completions/mean_length': 41.0, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6281594038009644, 'rewards/meter/std': 0.2802545726299286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969707131385803, 'rewards/repeat_soft/std': 0.007184124551713467, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.7251849174499512, 'rewards/total_composite/std': 0.16820569336414337, 'reward': 0.7251849174499512, 'reward_std': 0.16820570826530457, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11823517829179764, 'sampling/sampling_logp_difference/max': 1.3563261032104492, 'sampling/importance_sampling_ratio/min': 0.2576054632663727, 'sampling/importance_sampling_ratio/mean': 1.0046759843826294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5806138329207897, 'clip_ratio/low_mean': 0.06011569872498512, 'clip_ratio/low_min': 0.06011569872498512, 'clip_ratio/high_mean': 0.06972717307507992, 'clip_ratio/high_max': 0.06972717307507992, 'clip_ratio/region_mean': 0.12984287180006504, 'reward_total_mean': 0.7251849174499512, 'reward_meter_mean': 0.6281594038009644, 'reward_meter_std': 0.2802545726299286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969707131385803, 'reward_repeat_soft_std': 0.007184124551713467, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.7251849174499512, 'reward_total_composite_std': 0.16820569336414337, 'epoch': 0.05}

 15%|█▌        | 506/3300 [1:17:36<8:07:08, 10.46s/it]INFO 04-13 08:38:34 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 507/3300 [1:17:47<8:13:58, 10.61s/it]2026-04-13 08:38:45,125 | INFO | train_grpo_train | metrics_logged mode=train step=507

                                                      
{'loss': -0.0453, 'grad_norm': 4.813800811767578, 'learning_rate': 8.466666666666668e-06, 'num_tokens': 909083.0, 'completions/mean_length': 86.75, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.000001907348633, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.8087725639343262, 'rewards/meter/std': 0.29176250100135803, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9967178106307983, 'rewards/repeat_soft/std': 0.004181791562587023, 'rewards/judge_quality/mean': 0.48124998807907104, 'rewards/judge_quality/std': 0.3015880584716797, 'rewards/total_composite/mean': 0.5315700173377991, 'rewards/total_composite/std': 0.3606009781360626, 'reward': 0.5315700173377991, 'reward_std': 0.36060091853141785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19166681170463562, 'sampling/sampling_logp_difference/max': 0.9522361755371094, 'sampling/importance_sampling_ratio/min': 0.38587716221809387, 'sampling/importance_sampling_ratio/mean': 1.077354907989502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2864388674497604, 'clip_ratio/low_mean': 0.05902777798473835, 'clip_ratio/low_min': 0.05902777798473835, 'clip_ratio/high_mean': 0.07762877829372883, 'clip_ratio/high_max': 0.07762877829372883, 'clip_ratio/region_mean': 0.13665655627846718, 'reward_total_mean': 0.5315700173377991, 'reward_meter_mean': 0.8087725639343262, 'reward_meter_std': 0.29176250100135803, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9967178106307983, 'reward_repeat_soft_std': 0.004181791562587023, 'reward_judge_quality_mean': 0.48124998807907104, 'reward_judge_quality_std': 0.3015880584716797, 'reward_total_composite_mean': 0.5315700173377991, 'reward_total_composite_std': 0.3606009781360626, 'epoch': 0.05}

 15%|█▌        | 507/3300 [1:17:47<8:13:58, 10.61s/it]INFO 04-13 08:38:45 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 508/3300 [1:17:58<8:20:57, 10.77s/it]2026-04-13 08:38:56,251 | INFO | train_grpo_train | metrics_logged mode=train step=508

                                                      
{'loss': -0.0888, 'grad_norm': 4.174500942230225, 'learning_rate': 8.463636363636364e-06, 'num_tokens': 910795.0, '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': 70.0, 'rewards/meter/mean': 0.4167751967906952, 'rewards/meter/std': 0.3632891774177551, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9934234619140625, 'rewards/repeat_soft/std': 0.006900710053741932, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.29731839895248413, 'rewards/total_composite/mean': 0.4950295090675354, 'rewards/total_composite/std': 0.2799857556819916, 'reward': 0.4950295090675354, 'reward_std': 0.2799857258796692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20668399333953857, 'sampling/sampling_logp_difference/max': 1.6211118698120117, 'sampling/importance_sampling_ratio/min': 0.19767877459526062, 'sampling/importance_sampling_ratio/mean': 0.9983342289924622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8299167305231094, 'clip_ratio/low_mean': 0.10491517372429371, 'clip_ratio/low_min': 0.10491517372429371, 'clip_ratio/high_mean': 0.05370689695701003, 'clip_ratio/high_max': 0.05370689695701003, 'clip_ratio/region_mean': 0.15862207068130374, 'reward_total_mean': 0.4950295090675354, 'reward_meter_mean': 0.4167751967906952, 'reward_meter_std': 0.3632891774177551, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9934234619140625, 'reward_repeat_soft_std': 0.006900710053741932, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.29731839895248413, 'reward_total_composite_mean': 0.4950295090675354, 'reward_total_composite_std': 0.2799857556819916, 'epoch': 0.05}

 15%|█▌        | 508/3300 [1:17:58<8:20:57, 10.77s/it]INFO 04-13 08:38:56 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 509/3300 [1:18:04<7:19:03,  9.44s/it]2026-04-13 08:39:02,594 | INFO | train_grpo_train | metrics_logged mode=train step=509

                                                      
{'loss': 0.215, 'grad_norm': 16.09877586364746, 'learning_rate': 8.460606060606061e-06, 'num_tokens': 912375.0, 'completions/mean_length': 45.5, 'completions/min_length': 36.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.7594577074050903, 'rewards/meter/std': 0.399588018655777, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9873208999633789, 'rewards/repeat_soft/std': 0.025526583194732666, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.22385823726654053, 'rewards/total_composite/mean': 0.6417040824890137, 'rewards/total_composite/std': 0.2031925767660141, 'reward': 0.6417040824890137, 'reward_std': 0.2031925767660141, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1875120997428894, 'sampling/sampling_logp_difference/max': 2.4239516258239746, 'sampling/importance_sampling_ratio/min': 0.30619949102401733, 'sampling/importance_sampling_ratio/mean': 1.038735032081604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.102756232023239, 'clip_ratio/low_mean': 0.12419511005282402, 'clip_ratio/low_min': 0.12419511005282402, 'clip_ratio/high_mean': 0.070777028799057, 'clip_ratio/high_max': 0.070777028799057, 'clip_ratio/region_mean': 0.19497213885188103, 'reward_total_mean': 0.6417040824890137, 'reward_meter_mean': 0.7594577074050903, 'reward_meter_std': 0.399588018655777, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9873208999633789, 'reward_repeat_soft_std': 0.025526583194732666, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.22385823726654053, 'reward_total_composite_mean': 0.6417040824890137, 'reward_total_composite_std': 0.2031925767660141, 'epoch': 0.05}

 15%|█▌        | 509/3300 [1:18:04<7:19:03,  9.44s/it]INFO 04-13 08:39:02 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 510/3300 [1:18:10<6:31:13,  8.41s/it]2026-04-13 08:39:08,614 | INFO | train_grpo_train | metrics_logged mode=train step=510

                                                      
{'loss': 0.0228, 'grad_norm': 16.135772705078125, 'learning_rate': 8.457575757575758e-06, 'num_tokens': 914171.0, 'completions/mean_length': 56.5, 'completions/min_length': 49.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.4396429657936096, 'rewards/meter/std': 0.3085101842880249, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9942453503608704, 'rewards/repeat_soft/std': 0.00546035123988986, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.4842175245285034, 'rewards/total_composite/std': 0.0867922455072403, 'reward': 0.4842175245285034, 'reward_std': 0.0867922455072403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22289490699768066, 'sampling/sampling_logp_difference/max': 2.671037197113037, 'sampling/importance_sampling_ratio/min': 0.0691804364323616, 'sampling/importance_sampling_ratio/mean': 1.0352829694747925, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.149636998772621, 'clip_ratio/low_mean': 0.0742317233234644, 'clip_ratio/low_min': 0.0742317233234644, 'clip_ratio/high_mean': 0.10805025603622198, 'clip_ratio/high_max': 0.10805025603622198, 'clip_ratio/region_mean': 0.18228197935968637, 'reward_total_mean': 0.4842175245285034, 'reward_meter_mean': 0.4396429657936096, 'reward_meter_std': 0.3085101842880249, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9942453503608704, 'reward_repeat_soft_std': 0.00546035123988986, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.4842175245285034, 'reward_total_composite_std': 0.0867922455072403, 'epoch': 0.05}

 15%|█▌        | 510/3300 [1:18:10<6:31:13,  8.41s/it]INFO 04-13 08:39:08 [block_pool.py:378] Successfully reset prefix cache

 15%|█▌        | 511/3300 [1:18:21<7:07:08,  9.19s/it]2026-04-13 08:39:19,614 | INFO | train_grpo_train | metrics_logged mode=train step=511

                                                      
{'loss': -0.1018, 'grad_norm': 3.400689125061035, 'learning_rate': 8.454545454545455e-06, 'num_tokens': 915678.0, 'completions/mean_length': 94.375, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.71428680419922, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8086185455322266, 'rewards/meter/std': 0.3364562392234802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9483927488327026, 'rewards/repeat_soft/std': 0.0429542101919651, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.28965190052986145, 'rewards/total_composite/mean': 0.5976336002349854, 'rewards/total_composite/std': 0.28357553482055664, 'reward': 0.5976336002349854, 'reward_std': 0.28357553482055664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18736304342746735, 'sampling/sampling_logp_difference/max': 1.2291431427001953, 'sampling/importance_sampling_ratio/min': 0.29254311323165894, 'sampling/importance_sampling_ratio/mean': 1.0303345918655396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3929908722639084, 'clip_ratio/low_mean': 0.042258523404598236, 'clip_ratio/low_min': 0.042258523404598236, 'clip_ratio/high_mean': 0.10468585602939129, 'clip_ratio/high_max': 0.10468585602939129, 'clip_ratio/region_mean': 0.14694437943398952, 'reward_total_mean': 0.5976336002349854, 'reward_meter_mean': 0.8086185455322266, 'reward_meter_std': 0.3364562392234802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9483927488327026, 'reward_repeat_soft_std': 0.0429542101919651, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.28965190052986145, 'reward_total_composite_mean': 0.5976336002349854, 'reward_total_composite_std': 0.28357553482055664, 'epoch': 0.05}

 15%|█▌        | 511/3300 [1:18:21<7:07:08,  9.19s/it]INFO 04-13 08:39:19 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 512/3300 [1:18:28<6:29:42,  8.39s/it]2026-04-13 08:39:26,127 | INFO | train_grpo_train | metrics_logged mode=train step=512

                                                      
{'loss': 0.0273, 'grad_norm': 34.653568267822266, 'learning_rate': 8.451515151515151e-06, 'num_tokens': 917114.0, 'completions/mean_length': 22.5, 'completions/min_length': 21.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8293523192405701, 'rewards/meter/std': 0.3166838586330414, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9123008251190186, 'rewards/repeat_soft/std': 0.15004493296146393, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5746053457260132, 'rewards/total_composite/std': 0.0899825245141983, 'reward': 0.5746053457260132, 'reward_std': 0.0899825170636177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1600390523672104, 'sampling/sampling_logp_difference/max': 3.9742989540100098, 'sampling/importance_sampling_ratio/min': 0.018792470917105675, 'sampling/importance_sampling_ratio/mean': 0.9933913350105286, 'sampling/importance_sampling_ratio/max': 1.8240867853164673, 'entropy': 0.7740676701068878, 'clip_ratio/low_mean': 0.05046583851799369, 'clip_ratio/low_min': 0.05046583851799369, 'clip_ratio/high_mean': 0.0725137647241354, 'clip_ratio/high_max': 0.0725137647241354, 'clip_ratio/region_mean': 0.12297960324212909, 'reward_total_mean': 0.5746053457260132, 'reward_meter_mean': 0.8293523192405701, 'reward_meter_std': 0.3166838586330414, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9123008251190186, 'reward_repeat_soft_std': 0.15004493296146393, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5746053457260132, 'reward_total_composite_std': 0.0899825245141983, 'epoch': 0.05}

 16%|█▌        | 512/3300 [1:18:28<6:29:42,  8.39s/it]INFO 04-13 08:39:26 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 513/3300 [1:18:39<7:07:21,  9.20s/it]2026-04-13 08:39:37,227 | INFO | train_grpo_train | metrics_logged mode=train step=513

                                                      
{'loss': -0.1581, 'grad_norm': 3.4269769191741943, 'learning_rate': 8.44848484848485e-06, 'num_tokens': 919291.0, 'completions/mean_length': 207.125, 'completions/min_length': 94.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 105.5, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9621840715408325, 'rewards/meter/std': 0.035620398819446564, 'rewards/count_adherence/mean': 0.675000011920929, 'rewards/count_adherence/std': 0.23754701018333435, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9443340301513672, 'rewards/repeat_soft/std': 0.05840669572353363, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.22238561511039734, 'rewards/total_composite/mean': 0.36305707693099976, 'rewards/total_composite/std': 0.3129875063896179, 'reward': 0.36305707693099976, 'reward_std': 0.3129875063896179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1959700733423233, 'sampling/sampling_logp_difference/max': 1.6656246185302734, 'sampling/importance_sampling_ratio/min': 0.18907251954078674, 'sampling/importance_sampling_ratio/mean': 1.042391061782837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8112002611160278, 'clip_ratio/low_mean': 0.026595745235681534, 'clip_ratio/low_min': 0.026595745235681534, 'clip_ratio/high_mean': 0.10398740135133266, 'clip_ratio/high_max': 0.10398740135133266, 'clip_ratio/region_mean': 0.1305831465870142, 'reward_total_mean': 0.36305707693099976, 'reward_meter_mean': 0.9621840715408325, 'reward_meter_std': 0.035620398819446564, 'reward_count_adherence_mean': 0.675000011920929, 'reward_count_adherence_std': 0.23754701018333435, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9443340301513672, 'reward_repeat_soft_std': 0.05840669572353363, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.22238561511039734, 'reward_total_composite_mean': 0.36305707693099976, 'reward_total_composite_std': 0.3129875063896179, 'epoch': 0.05}

 16%|█▌        | 513/3300 [1:18:39<7:07:21,  9.20s/it]INFO 04-13 08:39:37 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 514/3300 [1:18:45<6:19:00,  8.16s/it]2026-04-13 08:39:42,968 | INFO | train_grpo_train | metrics_logged mode=train step=514

                                                      
{'loss': 0.1274, 'grad_norm': 28.19949722290039, 'learning_rate': 8.445454545454547e-06, 'num_tokens': 920785.0, 'completions/mean_length': 28.75, 'completions/min_length': 24.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.470637708902359, 'rewards/meter/std': 0.3691936433315277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973574280738831, 'rewards/repeat_soft/std': 0.004874487407505512, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.47808772325515747, 'rewards/total_composite/std': 0.10119297355413437, 'reward': 0.47808772325515747, 'reward_std': 0.10119297355413437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21324360370635986, 'sampling/sampling_logp_difference/max': 2.2050304412841797, 'sampling/importance_sampling_ratio/min': 0.1102471649646759, 'sampling/importance_sampling_ratio/mean': 1.0146220922470093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.515910029411316, 'clip_ratio/low_mean': 0.06988866627216339, 'clip_ratio/low_min': 0.06988866627216339, 'clip_ratio/high_mean': 0.13314206153154373, 'clip_ratio/high_max': 0.13314206153154373, 'clip_ratio/region_mean': 0.20303072780370712, 'reward_total_mean': 0.47808772325515747, 'reward_meter_mean': 0.470637708902359, 'reward_meter_std': 0.3691936433315277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973574280738831, 'reward_repeat_soft_std': 0.004874487407505512, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.47808772325515747, 'reward_total_composite_std': 0.10119297355413437, 'epoch': 0.05}

 16%|█▌        | 514/3300 [1:18:45<6:19:00,  8.16s/it]INFO 04-13 08:39:43 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 515/3300 [1:18:50<5:44:12,  7.42s/it]2026-04-13 08:39:48,638 | INFO | train_grpo_train | metrics_logged mode=train step=515

                                                      
{'loss': 0.0781, 'grad_norm': 26.827844619750977, 'learning_rate': 8.442424242424243e-06, 'num_tokens': 922203.0, 'completions/mean_length': 26.25, 'completions/min_length': 22.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.518803596496582, 'rewards/meter/std': 0.39614447951316833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9638574719429016, 'rewards/repeat_soft/std': 0.049367066472768784, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.48019248247146606, 'rewards/total_composite/std': 0.11675118654966354, 'reward': 0.48019248247146606, 'reward_std': 0.11675118654966354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2694185972213745, 'sampling/sampling_logp_difference/max': 2.053173065185547, 'sampling/importance_sampling_ratio/min': 0.12832707166671753, 'sampling/importance_sampling_ratio/mean': 1.0390163660049438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8183883130550385, 'clip_ratio/low_mean': 0.08647186495363712, 'clip_ratio/low_min': 0.08647186495363712, 'clip_ratio/high_mean': 0.11229076609015465, 'clip_ratio/high_max': 0.11229076609015465, 'clip_ratio/region_mean': 0.19876263104379177, 'reward_total_mean': 0.48019248247146606, 'reward_meter_mean': 0.518803596496582, 'reward_meter_std': 0.39614447951316833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9638574719429016, 'reward_repeat_soft_std': 0.049367066472768784, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.48019248247146606, 'reward_total_composite_std': 0.11675118654966354, 'epoch': 0.05}

 16%|█▌        | 515/3300 [1:18:50<5:44:12,  7.42s/it]INFO 04-13 08:39:48 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 516/3300 [1:19:01<6:34:19,  8.50s/it]2026-04-13 08:39:59,666 | INFO | train_grpo_train | metrics_logged mode=train step=516

                                                      
{'loss': -0.1341, 'grad_norm': 2.6788177490234375, 'learning_rate': 8.43939393939394e-06, 'num_tokens': 923869.0, 'completions/mean_length': 106.25, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.28571701049805, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8491782546043396, 'rewards/meter/std': 0.22291876375675201, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9589000344276428, 'rewards/repeat_soft/std': 0.02780892513692379, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.5105176568031311, 'rewards/total_composite/std': 0.2141517996788025, 'reward': 0.5105176568031311, 'reward_std': 0.2141517698764801, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20677360892295837, 'sampling/sampling_logp_difference/max': 1.553481101989746, 'sampling/importance_sampling_ratio/min': 0.21151040494441986, 'sampling/importance_sampling_ratio/mean': 1.0398728847503662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6627289354801178, 'clip_ratio/low_mean': 0.005434782709926367, 'clip_ratio/low_min': 0.005434782709926367, 'clip_ratio/high_mean': 0.13930488843470812, 'clip_ratio/high_max': 0.13930488843470812, 'clip_ratio/region_mean': 0.14473967114463449, 'reward_total_mean': 0.5105176568031311, 'reward_meter_mean': 0.8491782546043396, 'reward_meter_std': 0.22291876375675201, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9589000344276428, 'reward_repeat_soft_std': 0.02780892513692379, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.5105176568031311, 'reward_total_composite_std': 0.2141517996788025, 'epoch': 0.05}

 16%|█▌        | 516/3300 [1:19:01<6:34:19,  8.50s/it]INFO 04-13 08:39:59 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 517/3300 [1:19:12<7:09:08,  9.25s/it]2026-04-13 08:40:10,676 | INFO | train_grpo_train | metrics_logged mode=train step=517

                                                      
{'loss': -0.1557, 'grad_norm': 4.7905707359313965, 'learning_rate': 8.436363636363637e-06, 'num_tokens': 925718.0, 'completions/mean_length': 135.125, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.28572082519531, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.3914533853530884, 'rewards/meter/std': 0.37040913105010986, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9313338994979858, 'rewards/repeat_soft/std': 0.10688944905996323, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.18845234811306, 'rewards/total_composite/mean': 0.3853323459625244, 'rewards/total_composite/std': 0.18787869811058044, 'reward': 0.3853323459625244, 'reward_std': 0.18787869811058044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19800002872943878, 'sampling/sampling_logp_difference/max': 1.8448104858398438, 'sampling/importance_sampling_ratio/min': 0.15805527567863464, 'sampling/importance_sampling_ratio/mean': 1.0412434339523315, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.553493708372116, 'clip_ratio/low_mean': 0.06119980476796627, 'clip_ratio/low_min': 0.06119980476796627, 'clip_ratio/high_mean': 0.08065548166632652, 'clip_ratio/high_max': 0.08065548166632652, 'clip_ratio/region_mean': 0.1418552864342928, 'reward_total_mean': 0.3853323459625244, 'reward_meter_mean': 0.3914533853530884, 'reward_meter_std': 0.37040913105010986, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9313338994979858, 'reward_repeat_soft_std': 0.10688944905996323, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.18845234811306, 'reward_total_composite_mean': 0.3853323459625244, 'reward_total_composite_std': 0.18787869811058044, 'epoch': 0.05}

 16%|█▌        | 517/3300 [1:19:12<7:09:08,  9.25s/it]INFO 04-13 08:40:10 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 518/3300 [1:19:20<6:50:50,  8.86s/it]2026-04-13 08:40:18,623 | INFO | train_grpo_train | metrics_logged mode=train step=518

                                                      
{'loss': 0.3621, 'grad_norm': 10.242082595825195, 'learning_rate': 8.433333333333334e-06, 'num_tokens': 927843.0, 'completions/mean_length': 97.625, 'completions/min_length': 67.0, 'completions/max_length': 206.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 206.0, 'rewards/meter/mean': 0.6179324388504028, 'rewards/meter/std': 0.2663954794406891, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.36443448066711426, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8473659753799438, 'rewards/repeat_soft/std': 0.1457204520702362, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.48461443185806274, 'rewards/total_composite/std': 0.16915346682071686, 'reward': 0.48461443185806274, 'reward_std': 0.16915346682071686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14129699766635895, 'sampling/sampling_logp_difference/max': 1.7774124145507812, 'sampling/importance_sampling_ratio/min': 0.16907508671283722, 'sampling/importance_sampling_ratio/mean': 1.019058346748352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0807285755872726, 'clip_ratio/low_mean': 0.057006788440048695, 'clip_ratio/low_min': 0.057006788440048695, 'clip_ratio/high_mean': 0.11320867296308279, 'clip_ratio/high_max': 0.11320867296308279, 'clip_ratio/region_mean': 0.17021546140313148, 'reward_total_mean': 0.48461443185806274, 'reward_meter_mean': 0.6179324388504028, 'reward_meter_std': 0.2663954794406891, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.36443448066711426, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8473659753799438, 'reward_repeat_soft_std': 0.1457204520702362, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.48461443185806274, 'reward_total_composite_std': 0.16915346682071686, 'epoch': 0.05}

 16%|█▌        | 518/3300 [1:19:20<6:50:50,  8.86s/it]INFO 04-13 08:40:18 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 519/3300 [1:19:26<6:07:50,  7.94s/it]2026-04-13 08:40:24,402 | INFO | train_grpo_train | metrics_logged mode=train step=519

                                                      
{'loss': 0.0702, 'grad_norm': 24.148468017578125, 'learning_rate': 8.43030303030303e-06, 'num_tokens': 929414.0, 'completions/mean_length': 20.375, 'completions/min_length': 14.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8721743822097778, 'rewards/meter/std': 0.3301948606967926, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9354029893875122, 'rewards/repeat_soft/std': 0.031347643584012985, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6613031625747681, 'rewards/total_composite/std': 0.18917421996593475, 'reward': 0.6613031625747681, 'reward_std': 0.18917423486709595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1783645898103714, 'sampling/sampling_logp_difference/max': 1.3344464302062988, 'sampling/importance_sampling_ratio/min': 0.2633039057254791, 'sampling/importance_sampling_ratio/mean': 0.9993137121200562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5200989693403244, 'clip_ratio/low_mean': 0.1105669941753149, 'clip_ratio/low_min': 0.1105669941753149, 'clip_ratio/high_mean': 0.057528410106897354, 'clip_ratio/high_max': 0.057528410106897354, 'clip_ratio/region_mean': 0.16809540428221226, 'reward_total_mean': 0.6613031625747681, 'reward_meter_mean': 0.8721743822097778, 'reward_meter_std': 0.3301948606967926, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9354029893875122, 'reward_repeat_soft_std': 0.031347643584012985, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6613031625747681, 'reward_total_composite_std': 0.18917421996593475, 'epoch': 0.05}

 16%|█▌        | 519/3300 [1:19:26<6:07:50,  7.94s/it]INFO 04-13 08:40:24 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 520/3300 [1:19:32<5:36:57,  7.27s/it]2026-04-13 08:40:30,127 | INFO | train_grpo_train | metrics_logged mode=train step=520

                                                      
{'loss': 0.0834, 'grad_norm': 20.35475730895996, 'learning_rate': 8.427272727272729e-06, 'num_tokens': 930876.0, 'completions/mean_length': 30.75, 'completions/min_length': 27.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.75, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.737522304058075, 'rewards/meter/std': 0.14284822344779968, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9978242516517639, 'rewards/repeat_soft/std': 0.004353147931396961, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.11444743722677231, 'rewards/total_composite/mean': 0.5755347013473511, 'rewards/total_composite/std': 0.07672619819641113, 'reward': 0.5755347013473511, 'reward_std': 0.07672620564699173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16212056577205658, 'sampling/sampling_logp_difference/max': 1.6107200384140015, 'sampling/importance_sampling_ratio/min': 0.19974373281002045, 'sampling/importance_sampling_ratio/mean': 1.0059845447540283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.65384092181921, 'clip_ratio/low_mean': 0.09329796768724918, 'clip_ratio/low_min': 0.09329796768724918, 'clip_ratio/high_mean': 0.03525246400386095, 'clip_ratio/high_max': 0.03525246400386095, 'clip_ratio/region_mean': 0.12855043169111013, 'reward_total_mean': 0.5755347013473511, 'reward_meter_mean': 0.737522304058075, 'reward_meter_std': 0.14284822344779968, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9978242516517639, 'reward_repeat_soft_std': 0.004353147931396961, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.11444743722677231, 'reward_total_composite_mean': 0.5755347013473511, 'reward_total_composite_std': 0.07672619819641113, 'epoch': 0.05}

 16%|█▌        | 520/3300 [1:19:32<5:36:57,  7.27s/it]INFO 04-13 08:40:30 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 521/3300 [1:19:38<5:21:40,  6.95s/it]2026-04-13 08:40:36,307 | INFO | train_grpo_train | metrics_logged mode=train step=521

                                                      
{'loss': 0.0146, 'grad_norm': 11.828908920288086, 'learning_rate': 8.424242424242425e-06, 'num_tokens': 932838.0, 'completions/mean_length': 67.25, 'completions/min_length': 60.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9194561243057251, 'rewards/meter/std': 0.1151917576789856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8541878461837769, 'rewards/repeat_soft/std': 0.06426283717155457, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5657186508178711, 'rewards/total_composite/std': 0.04520780220627785, 'reward': 0.5657186508178711, 'reward_std': 0.04520779103040695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1585492193698883, 'sampling/sampling_logp_difference/max': 1.5798873901367188, 'sampling/importance_sampling_ratio/min': 0.2059982866048813, 'sampling/importance_sampling_ratio/mean': 1.0087366104125977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3772781193256378, 'clip_ratio/low_mean': 0.03209841717034578, 'clip_ratio/low_min': 0.03209841717034578, 'clip_ratio/high_mean': 0.14448866993188858, 'clip_ratio/high_max': 0.14448866993188858, 'clip_ratio/region_mean': 0.17658708710223436, 'reward_total_mean': 0.5657186508178711, 'reward_meter_mean': 0.9194561243057251, 'reward_meter_std': 0.1151917576789856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8541878461837769, 'reward_repeat_soft_std': 0.06426283717155457, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5657186508178711, 'reward_total_composite_std': 0.04520780220627785, 'epoch': 0.05}

 16%|█▌        | 521/3300 [1:19:38<5:21:40,  6.95s/it]INFO 04-13 08:40:36 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 522/3300 [1:19:44<5:12:53,  6.76s/it]2026-04-13 08:40:42,629 | INFO | train_grpo_train | metrics_logged mode=train step=522

                                                      
{'loss': 0.0427, 'grad_norm': 13.171037673950195, 'learning_rate': 8.421212121212122e-06, 'num_tokens': 934492.0, 'completions/mean_length': 44.75, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.9922068119049072, 'rewards/meter/std': 0.0037105989176779985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8781986236572266, 'rewards/repeat_soft/std': 0.1043878048658371, 'rewards/judge_quality/mean': 0.4649999737739563, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.6316136717796326, 'rewards/total_composite/std': 0.13218602538108826, 'reward': 0.6316136717796326, 'reward_std': 0.13218601047992706, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2231953889131546, 'sampling/sampling_logp_difference/max': 1.8242545127868652, 'sampling/importance_sampling_ratio/min': 0.16133788228034973, 'sampling/importance_sampling_ratio/mean': 1.0236644744873047, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.74095219373703, 'clip_ratio/low_mean': 0.13583752140402794, 'clip_ratio/low_min': 0.13583752140402794, 'clip_ratio/high_mean': 0.06309650093317032, 'clip_ratio/high_max': 0.06309650093317032, 'clip_ratio/region_mean': 0.19893402233719826, 'reward_total_mean': 0.6316136717796326, 'reward_meter_mean': 0.9922068119049072, 'reward_meter_std': 0.0037105989176779985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8781986236572266, 'reward_repeat_soft_std': 0.1043878048658371, 'reward_judge_quality_mean': 0.4649999737739563, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.6316136717796326, 'reward_total_composite_std': 0.13218602538108826, 'epoch': 0.05}

 16%|█▌        | 522/3300 [1:19:44<5:12:53,  6.76s/it]INFO 04-13 08:40:42 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 523/3300 [1:19:52<5:19:48,  6.91s/it]2026-04-13 08:40:49,893 | INFO | train_grpo_train | metrics_logged mode=train step=523

                                                      
{'loss': 0.035, 'grad_norm': 22.766992568969727, 'learning_rate': 8.418181818181819e-06, 'num_tokens': 936371.0, 'completions/mean_length': 69.875, 'completions/min_length': 53.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.4473544955253601, 'rewards/meter/std': 0.38985496759414673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9960489869117737, 'rewards/repeat_soft/std': 0.001926091150380671, 'rewards/judge_quality/mean': 0.5737500190734863, 'rewards/judge_quality/std': 0.2672044634819031, 'rewards/total_composite/mean': 0.5015914440155029, 'rewards/total_composite/std': 0.136424720287323, 'reward': 0.5015914440155029, 'reward_std': 0.136424720287323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21991561353206635, 'sampling/sampling_logp_difference/max': 2.5158519744873047, 'sampling/importance_sampling_ratio/min': 0.08079405128955841, 'sampling/importance_sampling_ratio/mean': 0.9984143972396851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8166926428675652, 'clip_ratio/low_mean': 0.09622285235673189, 'clip_ratio/low_min': 0.09622285235673189, 'clip_ratio/high_mean': 0.07709643803536892, 'clip_ratio/high_max': 0.07709643803536892, 'clip_ratio/region_mean': 0.1733192903921008, 'reward_total_mean': 0.5015914440155029, 'reward_meter_mean': 0.4473544955253601, 'reward_meter_std': 0.38985496759414673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9960489869117737, 'reward_repeat_soft_std': 0.001926091150380671, 'reward_judge_quality_mean': 0.5737500190734863, 'reward_judge_quality_std': 0.2672044634819031, 'reward_total_composite_mean': 0.5015914440155029, 'reward_total_composite_std': 0.136424720287323, 'epoch': 0.05}

 16%|█▌        | 523/3300 [1:19:52<5:19:48,  6.91s/it]INFO 04-13 08:40:50 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 524/3300 [1:19:57<5:04:18,  6.58s/it]2026-04-13 08:40:55,694 | INFO | train_grpo_train | metrics_logged mode=train step=524

                                                      
{'loss': 0.0343, 'grad_norm': 37.11225128173828, 'learning_rate': 8.415151515151516e-06, 'num_tokens': 937775.0, 'completions/mean_length': 24.5, 'completions/min_length': 24.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.7918415069580078, 'rewards/meter/std': 0.22908450663089752, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.988369345664978, 'rewards/repeat_soft/std': 0.008541061542928219, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.7832573652267456, 'rewards/total_composite/std': 0.14881549775600433, 'reward': 0.7832573652267456, 'reward_std': 0.14881551265716553, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11703936755657196, 'sampling/sampling_logp_difference/max': 1.1861143112182617, 'sampling/importance_sampling_ratio/min': 0.3054056763648987, 'sampling/importance_sampling_ratio/mean': 1.0091761350631714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4233832322061062, 'clip_ratio/low_mean': 0.03041666653007269, 'clip_ratio/low_min': 0.03041666653007269, 'clip_ratio/high_mean': 0.04541666619479656, 'clip_ratio/high_max': 0.04541666619479656, 'clip_ratio/region_mean': 0.07583333272486925, 'reward_total_mean': 0.7832573652267456, 'reward_meter_mean': 0.7918415069580078, 'reward_meter_std': 0.22908450663089752, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.988369345664978, 'reward_repeat_soft_std': 0.008541061542928219, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.7832573652267456, 'reward_total_composite_std': 0.14881549775600433, 'epoch': 0.05}

 16%|█▌        | 524/3300 [1:19:57<5:04:18,  6.58s/it]INFO 04-13 08:40:55 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 525/3300 [1:20:03<4:55:43,  6.39s/it]2026-04-13 08:41:01,661 | INFO | train_grpo_train | metrics_logged mode=train step=525

                                                      
{'loss': -0.0101, 'grad_norm': 21.21489143371582, 'learning_rate': 8.412121212121212e-06, 'num_tokens': 939377.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5493050813674927, 'rewards/meter/std': 0.40603959560394287, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989762306213379, 'rewards/repeat_soft/std': 0.0016704278532415628, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/total_composite/mean': 0.5870556235313416, 'rewards/total_composite/std': 0.1870722472667694, 'reward': 0.5870556235313416, 'reward_std': 0.18707223236560822, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22771434485912323, 'sampling/sampling_logp_difference/max': 1.3903067111968994, 'sampling/importance_sampling_ratio/min': 0.2798469662666321, 'sampling/importance_sampling_ratio/mean': 1.0537641048431396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8079886883497238, 'clip_ratio/low_mean': 0.13587182201445103, 'clip_ratio/low_min': 0.13587182201445103, 'clip_ratio/high_mean': 0.13082965649664402, 'clip_ratio/high_max': 0.13082965649664402, 'clip_ratio/region_mean': 0.26670147851109505, 'reward_total_mean': 0.5870556235313416, 'reward_meter_mean': 0.5493050813674927, 'reward_meter_std': 0.40603959560394287, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989762306213379, 'reward_repeat_soft_std': 0.0016704278532415628, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.5870556235313416, 'reward_total_composite_std': 0.1870722472667694, 'epoch': 0.05}

 16%|█▌        | 525/3300 [1:20:03<4:55:43,  6.39s/it]INFO 04-13 08:41:01 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 526/3300 [1:20:10<4:52:26,  6.33s/it]2026-04-13 08:41:07,827 | INFO | train_grpo_train | metrics_logged mode=train step=526

                                                      
{'loss': 0.0285, 'grad_norm': 37.78364181518555, 'learning_rate': 8.40909090909091e-06, 'num_tokens': 940866.0, 'completions/mean_length': 28.125, 'completions/min_length': 22.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6110023260116577, 'rewards/meter/std': 0.3921199440956116, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9955050349235535, 'rewards/repeat_soft/std': 0.011455883271992207, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.23256337642669678, 'rewards/total_composite/mean': 0.5320631265640259, 'rewards/total_composite/std': 0.3082868456840515, 'reward': 0.5320631265640259, 'reward_std': 0.3082868456840515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23938944935798645, 'sampling/sampling_logp_difference/max': 2.1711440086364746, 'sampling/importance_sampling_ratio/min': 0.11404707282781601, 'sampling/importance_sampling_ratio/mean': 0.9849367737770081, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9804778918623924, 'clip_ratio/low_mean': 0.07969757076352835, 'clip_ratio/low_min': 0.07969757076352835, 'clip_ratio/high_mean': 0.10590091813355684, 'clip_ratio/high_max': 0.10590091813355684, 'clip_ratio/region_mean': 0.1855984888970852, 'reward_total_mean': 0.5320631265640259, 'reward_meter_mean': 0.6110023260116577, 'reward_meter_std': 0.3921199440956116, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9955050349235535, 'reward_repeat_soft_std': 0.011455883271992207, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.23256337642669678, 'reward_total_composite_mean': 0.5320631265640259, 'reward_total_composite_std': 0.3082868456840515, 'epoch': 0.05}

 16%|█▌        | 526/3300 [1:20:10<4:52:26,  6.33s/it]INFO 04-13 08:41:08 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 527/3300 [1:20:20<5:50:55,  7.59s/it]2026-04-13 08:41:18,376 | INFO | train_grpo_train | metrics_logged mode=train step=527

                                                      
{'loss': -0.0521, 'grad_norm': 13.999894142150879, 'learning_rate': 8.406060606060606e-06, 'num_tokens': 942629.0, 'completions/mean_length': 58.375, 'completions/min_length': 45.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.5185929536819458, 'rewards/meter/std': 0.3680461347103119, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9829244613647461, 'rewards/repeat_soft/std': 0.03222934156656265, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.514051079750061, 'rewards/total_composite/std': 0.1065848097205162, 'reward': 0.514051079750061, 'reward_std': 0.10658480226993561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19776305556297302, 'sampling/sampling_logp_difference/max': 2.6942830085754395, 'sampling/importance_sampling_ratio/min': 0.06759082525968552, 'sampling/importance_sampling_ratio/mean': 1.0025206804275513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.344873622059822, 'clip_ratio/low_mean': 0.09406970255076885, 'clip_ratio/low_min': 0.09406970255076885, 'clip_ratio/high_mean': 0.10103962756693363, 'clip_ratio/high_max': 0.10103962756693363, 'clip_ratio/region_mean': 0.19510933011770248, 'reward_total_mean': 0.514051079750061, 'reward_meter_mean': 0.5185929536819458, 'reward_meter_std': 0.3680461347103119, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9829244613647461, 'reward_repeat_soft_std': 0.03222934156656265, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.514051079750061, 'reward_total_composite_std': 0.1065848097205162, 'epoch': 0.05}

 16%|█▌        | 527/3300 [1:20:20<5:50:55,  7.59s/it]INFO 04-13 08:41:18 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 528/3300 [1:20:29<6:09:39,  8.00s/it]2026-04-13 08:41:27,330 | INFO | train_grpo_train | metrics_logged mode=train step=528

                                                      
{'loss': 0.0579, 'grad_norm': 13.73117733001709, 'learning_rate': 8.403030303030304e-06, 'num_tokens': 944334.0, 'completions/mean_length': 55.125, 'completions/min_length': 44.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8622069358825684, 'rewards/meter/std': 0.23428656160831451, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9304995536804199, 'rewards/repeat_soft/std': 0.04984993487596512, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.635810136795044, 'rewards/total_composite/std': 0.1618661880493164, 'reward': 0.635810136795044, 'reward_std': 0.1618662029504776, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19811634719371796, 'sampling/sampling_logp_difference/max': 1.4720497131347656, 'sampling/importance_sampling_ratio/min': 0.22945468127727509, 'sampling/importance_sampling_ratio/mean': 1.0227909088134766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2130194753408432, 'clip_ratio/low_mean': 0.11467535234987736, 'clip_ratio/low_min': 0.11467535234987736, 'clip_ratio/high_mean': 0.06414141319692135, 'clip_ratio/high_max': 0.06414141319692135, 'clip_ratio/region_mean': 0.1788167655467987, 'reward_total_mean': 0.635810136795044, 'reward_meter_mean': 0.8622069358825684, 'reward_meter_std': 0.23428656160831451, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9304995536804199, 'reward_repeat_soft_std': 0.04984993487596512, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.635810136795044, 'reward_total_composite_std': 0.1618661880493164, 'epoch': 0.05}

 16%|█▌        | 528/3300 [1:20:29<6:09:39,  8.00s/it]INFO 04-13 08:41:27 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 529/3300 [1:20:36<5:59:33,  7.79s/it]2026-04-13 08:41:34,612 | INFO | train_grpo_train | metrics_logged mode=train step=529

                                                      
{'loss': 0.1271, 'grad_norm': 21.89297103881836, 'learning_rate': 8.400000000000001e-06, 'num_tokens': 945994.0, 'completions/mean_length': 37.5, 'completions/min_length': 27.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6194216012954712, 'rewards/meter/std': 0.3172769546508789, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9978482723236084, 'rewards/repeat_soft/std': 0.0031347365584224463, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.5457921028137207, 'rewards/total_composite/std': 0.07250598818063736, 'reward': 0.5457921028137207, 'reward_std': 0.07250598073005676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23090067505836487, 'sampling/sampling_logp_difference/max': 2.173288345336914, 'sampling/importance_sampling_ratio/min': 0.11380278319120407, 'sampling/importance_sampling_ratio/mean': 1.012063980102539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.971572533249855, 'clip_ratio/low_mean': 0.1034488994628191, 'clip_ratio/low_min': 0.1034488994628191, 'clip_ratio/high_mean': 0.06478441413491964, 'clip_ratio/high_max': 0.06478441413491964, 'clip_ratio/region_mean': 0.16823331359773874, 'reward_total_mean': 0.5457921028137207, 'reward_meter_mean': 0.6194216012954712, 'reward_meter_std': 0.3172769546508789, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9978482723236084, 'reward_repeat_soft_std': 0.0031347365584224463, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.5457921028137207, 'reward_total_composite_std': 0.07250598818063736, 'epoch': 0.05}

 16%|█▌        | 529/3300 [1:20:36<5:59:33,  7.79s/it]INFO 04-13 08:41:34 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 530/3300 [1:20:44<5:57:13,  7.74s/it]2026-04-13 08:41:42,237 | INFO | train_grpo_train | metrics_logged mode=train step=530

                                                      
{'loss': 0.06, 'grad_norm': 11.848499298095703, 'learning_rate': 8.396969696969698e-06, 'num_tokens': 947820.0, 'completions/mean_length': 60.25, 'completions/min_length': 54.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7378818988800049, 'rewards/meter/std': 0.32036516070365906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.996192455291748, 'rewards/repeat_soft/std': 0.004050884861499071, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.09941796213388443, 'rewards/total_composite/mean': 0.5040265917778015, 'rewards/total_composite/std': 0.23453491926193237, 'reward': 0.5040265917778015, 'reward_std': 0.23453491926193237, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20574524998664856, 'sampling/sampling_logp_difference/max': 2.756950855255127, 'sampling/importance_sampling_ratio/min': 0.0634850487112999, 'sampling/importance_sampling_ratio/mean': 1.0227644443511963, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5768088102340698, 'clip_ratio/low_mean': 0.06443452462553978, 'clip_ratio/low_min': 0.06443452462553978, 'clip_ratio/high_mean': 0.12275873497128487, 'clip_ratio/high_max': 0.12275873497128487, 'clip_ratio/region_mean': 0.18719325959682465, 'reward_total_mean': 0.5040265917778015, 'reward_meter_mean': 0.7378818988800049, 'reward_meter_std': 0.32036516070365906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.996192455291748, 'reward_repeat_soft_std': 0.004050884861499071, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.09941796213388443, 'reward_total_composite_mean': 0.5040265917778015, 'reward_total_composite_std': 0.23453491926193237, 'epoch': 0.05}

 16%|█▌        | 530/3300 [1:20:44<5:57:13,  7.74s/it]INFO 04-13 08:41:42 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 531/3300 [1:20:50<5:39:43,  7.36s/it]2026-04-13 08:41:48,722 | INFO | train_grpo_train | metrics_logged mode=train step=531

                                                      
{'loss': -0.0127, 'grad_norm': 24.54881477355957, 'learning_rate': 8.393939393939394e-06, 'num_tokens': 949306.0, 'completions/mean_length': 34.75, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.4521898031234741, 'rewards/meter/std': 0.4134586751461029, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9783536195755005, 'rewards/repeat_soft/std': 0.023981817066669464, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.18372632563114166, 'rewards/total_composite/mean': 0.5186299085617065, 'rewards/total_composite/std': 0.19654682278633118, 'reward': 0.5186299085617065, 'reward_std': 0.19654680788516998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22910979390144348, 'sampling/sampling_logp_difference/max': 2.3857579231262207, 'sampling/importance_sampling_ratio/min': 0.0920192152261734, 'sampling/importance_sampling_ratio/mean': 1.021804690361023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0811530873179436, 'clip_ratio/low_mean': 0.12342437170445919, 'clip_ratio/low_min': 0.12342437170445919, 'clip_ratio/high_mean': 0.09848485141992569, 'clip_ratio/high_max': 0.09848485141992569, 'clip_ratio/region_mean': 0.22190922312438488, 'reward_total_mean': 0.5186299085617065, 'reward_meter_mean': 0.4521898031234741, 'reward_meter_std': 0.4134586751461029, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9783536195755005, 'reward_repeat_soft_std': 0.023981817066669464, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.18372632563114166, 'reward_total_composite_mean': 0.5186299085617065, 'reward_total_composite_std': 0.19654682278633118, 'epoch': 0.05}

 16%|█▌        | 531/3300 [1:20:50<5:39:43,  7.36s/it]INFO 04-13 08:41:48 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 532/3300 [1:20:56<5:20:41,  6.95s/it]2026-04-13 08:41:54,717 | INFO | train_grpo_train | metrics_logged mode=train step=532

                                                      
{'loss': 0.0368, 'grad_norm': 21.143571853637695, 'learning_rate': 8.390909090909091e-06, 'num_tokens': 950966.0, 'completions/mean_length': 41.5, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7355077266693115, 'rewards/meter/std': 0.373830646276474, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9728670120239258, 'rewards/repeat_soft/std': 0.027256207540631294, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.25150617957115173, 'rewards/total_composite/mean': 0.535254716873169, 'rewards/total_composite/std': 0.2600457966327667, 'reward': 0.535254716873169, 'reward_std': 0.2600457966327667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21611906588077545, 'sampling/sampling_logp_difference/max': 1.5225722789764404, 'sampling/importance_sampling_ratio/min': 0.2181500345468521, 'sampling/importance_sampling_ratio/mean': 1.009087085723877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7905576527118683, 'clip_ratio/low_mean': 0.06814480666071177, 'clip_ratio/low_min': 0.06814480666071177, 'clip_ratio/high_mean': 0.13873954489827156, 'clip_ratio/high_max': 0.13873954489827156, 'clip_ratio/region_mean': 0.20688435155898333, 'reward_total_mean': 0.535254716873169, 'reward_meter_mean': 0.7355077266693115, 'reward_meter_std': 0.373830646276474, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9728670120239258, 'reward_repeat_soft_std': 0.027256207540631294, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.25150617957115173, 'reward_total_composite_mean': 0.535254716873169, 'reward_total_composite_std': 0.2600457966327667, 'epoch': 0.05}

 16%|█▌        | 532/3300 [1:20:56<5:20:41,  6.95s/it]INFO 04-13 08:41:54 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 533/3300 [1:21:03<5:08:56,  6.70s/it]2026-04-13 08:42:00,827 | INFO | train_grpo_train | metrics_logged mode=train step=533

                                                      
{'loss': 0.0203, 'grad_norm': 16.195524215698242, 'learning_rate': 8.387878787878788e-06, 'num_tokens': 952456.0, 'completions/mean_length': 33.25, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9594668745994568, 'rewards/meter/std': 0.041511066257953644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921087622642517, 'rewards/repeat_soft/std': 0.007358473259955645, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.19668686389923096, 'rewards/total_composite/mean': 0.7268799543380737, 'rewards/total_composite/std': 0.11619677394628525, 'reward': 0.7268799543380737, 'reward_std': 0.11619676649570465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17709314823150635, 'sampling/sampling_logp_difference/max': 4.454348564147949, 'sampling/importance_sampling_ratio/min': 0.011627892963588238, 'sampling/importance_sampling_ratio/mean': 1.0329965353012085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2515765875577927, 'clip_ratio/low_mean': 0.07639688905328512, 'clip_ratio/low_min': 0.07639688905328512, 'clip_ratio/high_mean': 0.05112999910488725, 'clip_ratio/high_max': 0.05112999910488725, 'clip_ratio/region_mean': 0.12752688815817237, 'reward_total_mean': 0.7268799543380737, 'reward_meter_mean': 0.9594668745994568, 'reward_meter_std': 0.041511066257953644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921087622642517, 'reward_repeat_soft_std': 0.007358473259955645, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.19668686389923096, 'reward_total_composite_mean': 0.7268799543380737, 'reward_total_composite_std': 0.11619677394628525, 'epoch': 0.05}

 16%|█▌        | 533/3300 [1:21:03<5:08:56,  6.70s/it]INFO 04-13 08:42:01 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 534/3300 [1:21:21<7:48:05, 10.15s/it]2026-04-13 08:42:19,042 | INFO | train_grpo_train | metrics_logged mode=train step=534

                                                      
{'loss': -0.0745, 'grad_norm': 1.7643216848373413, 'learning_rate': 8.384848484848485e-06, 'num_tokens': 954003.0, 'completions/mean_length': 83.375, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.142858505249023, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.838414192199707, 'rewards/meter/std': 0.3414279520511627, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.5166328549385071, 'rewards/total_composite/std': 0.21202123165130615, 'reward': 0.5166328549385071, 'reward_std': 0.21202121675014496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.23563916981220245, 'sampling/sampling_logp_difference/max': 1.946211814880371, 'sampling/importance_sampling_ratio/min': 0.1428140550851822, 'sampling/importance_sampling_ratio/mean': 1.001214623451233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4632943868637085, 'clip_ratio/low_mean': 0.019999999552965164, 'clip_ratio/low_min': 0.019999999552965164, 'clip_ratio/high_mean': 0.18262881226837635, 'clip_ratio/high_max': 0.18262881226837635, 'clip_ratio/region_mean': 0.20262881182134151, 'reward_total_mean': 0.5166328549385071, 'reward_meter_mean': 0.838414192199707, 'reward_meter_std': 0.3414279520511627, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.5166328549385071, 'reward_total_composite_std': 0.21202123165130615, 'epoch': 0.05}

 16%|█▌        | 534/3300 [1:21:21<7:48:05, 10.15s/it]INFO 04-13 08:42:19 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 535/3300 [1:21:28<7:14:21,  9.43s/it]2026-04-13 08:42:26,768 | INFO | train_grpo_train | metrics_logged mode=train step=535

                                                      
{'loss': 0.0078, 'grad_norm': 10.34975528717041, 'learning_rate': 8.381818181818183e-06, 'num_tokens': 956000.0, 'completions/mean_length': 74.625, 'completions/min_length': 65.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.8910391330718994, 'rewards/meter/std': 0.240591362118721, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9675438404083252, 'rewards/repeat_soft/std': 0.030771438032388687, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.45899564027786255, 'rewards/total_composite/std': 0.2833832800388336, 'reward': 0.45899564027786255, 'reward_std': 0.2833832800388336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19376224279403687, 'sampling/sampling_logp_difference/max': 1.599329948425293, 'sampling/importance_sampling_ratio/min': 0.20203185081481934, 'sampling/importance_sampling_ratio/mean': 1.0360289812088013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9693769812583923, 'clip_ratio/low_mean': 0.05228365398943424, 'clip_ratio/low_min': 0.05228365398943424, 'clip_ratio/high_mean': 0.15215030871331692, 'clip_ratio/high_max': 0.15215030871331692, 'clip_ratio/region_mean': 0.20443396270275116, 'reward_total_mean': 0.45899564027786255, 'reward_meter_mean': 0.8910391330718994, 'reward_meter_std': 0.240591362118721, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9675438404083252, 'reward_repeat_soft_std': 0.030771438032388687, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.45899564027786255, 'reward_total_composite_std': 0.2833832800388336, 'epoch': 0.05}

 16%|█▌        | 535/3300 [1:21:28<7:14:21,  9.43s/it]INFO 04-13 08:42:27 [block_pool.py:378] Successfully reset prefix cache

 16%|█▌        | 536/3300 [1:21:35<6:36:28,  8.61s/it]2026-04-13 08:42:33,464 | INFO | train_grpo_train | metrics_logged mode=train step=536

                                                      
{'loss': 0.058, 'grad_norm': 12.88140869140625, 'learning_rate': 8.37878787878788e-06, 'num_tokens': 958141.0, 'completions/mean_length': 86.625, 'completions/min_length': 81.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.625, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.5977858901023865, 'rewards/meter/std': 0.26751548051834106, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9242637157440186, 'rewards/repeat_soft/std': 0.12277748435735703, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.21540246903896332, 'rewards/total_composite/mean': 0.5444878339767456, 'rewards/total_composite/std': 0.1356063038110733, 'reward': 0.5444878339767456, 'reward_std': 0.1356062889099121, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17371222376823425, 'sampling/sampling_logp_difference/max': 3.1495916843414307, 'sampling/importance_sampling_ratio/min': 0.042869627475738525, 'sampling/importance_sampling_ratio/mean': 0.9901180267333984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7599944397807121, 'clip_ratio/low_mean': 0.05059726629406214, 'clip_ratio/low_min': 0.05059726629406214, 'clip_ratio/high_mean': 0.08251654729247093, 'clip_ratio/high_max': 0.08251654729247093, 'clip_ratio/region_mean': 0.13311381358653307, 'reward_total_mean': 0.5444878339767456, 'reward_meter_mean': 0.5977858901023865, 'reward_meter_std': 0.26751548051834106, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9242637157440186, 'reward_repeat_soft_std': 0.12277748435735703, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.21540246903896332, 'reward_total_composite_mean': 0.5444878339767456, 'reward_total_composite_std': 0.1356063038110733, 'epoch': 0.05}

 16%|█▌        | 536/3300 [1:21:35<6:36:28,  8.61s/it]INFO 04-13 08:42:33 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 537/3300 [1:21:41<5:57:20,  7.76s/it]2026-04-13 08:42:39,248 | INFO | train_grpo_train | metrics_logged mode=train step=537

                                                      
{'loss': 0.0314, 'grad_norm': 20.37936019897461, 'learning_rate': 8.375757575757576e-06, 'num_tokens': 959775.0, 'completions/mean_length': 33.25, 'completions/min_length': 26.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7705623507499695, 'rewards/meter/std': 0.3330211043357849, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9400973320007324, 'rewards/repeat_soft/std': 0.04098653420805931, 'rewards/judge_quality/mean': 0.6200000047683716, 'rewards/judge_quality/std': 0.22677870094776154, 'rewards/total_composite/mean': 0.6590745449066162, 'rewards/total_composite/std': 0.20637114346027374, 'reward': 0.6590745449066162, 'reward_std': 0.20637112855911255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1917899250984192, 'sampling/sampling_logp_difference/max': 2.235095977783203, 'sampling/importance_sampling_ratio/min': 0.10698185861110687, 'sampling/importance_sampling_ratio/mean': 1.0382293462753296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1163994669914246, 'clip_ratio/low_mean': 0.09055245015770197, 'clip_ratio/low_min': 0.09055245015770197, 'clip_ratio/high_mean': 0.06938631366938353, 'clip_ratio/high_max': 0.06938631366938353, 'clip_ratio/region_mean': 0.1599387638270855, 'reward_total_mean': 0.6590745449066162, 'reward_meter_mean': 0.7705623507499695, 'reward_meter_std': 0.3330211043357849, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9400973320007324, 'reward_repeat_soft_std': 0.04098653420805931, 'reward_judge_quality_mean': 0.6200000047683716, 'reward_judge_quality_std': 0.22677870094776154, 'reward_total_composite_mean': 0.6590745449066162, 'reward_total_composite_std': 0.20637114346027374, 'epoch': 0.05}

 16%|█▋        | 537/3300 [1:21:41<5:57:20,  7.76s/it]INFO 04-13 08:42:39 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 538/3300 [1:21:48<5:41:34,  7.42s/it]2026-04-13 08:42:45,876 | INFO | train_grpo_train | metrics_logged mode=train step=538

                                                      
{'loss': 0.0148, 'grad_norm': 11.401845932006836, 'learning_rate': 8.372727272727273e-06, 'num_tokens': 961714.0, 'completions/mean_length': 65.375, 'completions/min_length': 53.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.375, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9534566402435303, 'rewards/meter/std': 0.08393868803977966, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9308726787567139, 'rewards/repeat_soft/std': 0.05148974433541298, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.17062386870384216, 'rewards/total_composite/mean': 0.7569006085395813, 'rewards/total_composite/std': 0.11870553344488144, 'reward': 0.7569006085395813, 'reward_std': 0.11870554834604263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17962917685508728, 'sampling/sampling_logp_difference/max': 1.9633567333221436, 'sampling/importance_sampling_ratio/min': 0.14038638770580292, 'sampling/importance_sampling_ratio/mean': 1.026145577430725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2660537511110306, 'clip_ratio/low_mean': 0.07088575605303049, 'clip_ratio/low_min': 0.07088575605303049, 'clip_ratio/high_mean': 0.09831744618713856, 'clip_ratio/high_max': 0.09831744618713856, 'clip_ratio/region_mean': 0.16920320224016905, 'reward_total_mean': 0.7569006085395813, 'reward_meter_mean': 0.9534566402435303, 'reward_meter_std': 0.08393868803977966, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9308726787567139, 'reward_repeat_soft_std': 0.05148974433541298, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.17062386870384216, 'reward_total_composite_mean': 0.7569006085395813, 'reward_total_composite_std': 0.11870553344488144, 'epoch': 0.05}

 16%|█▋        | 538/3300 [1:21:48<5:41:34,  7.42s/it]INFO 04-13 08:42:46 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 539/3300 [1:21:54<5:28:18,  7.13s/it]2026-04-13 08:42:52,344 | INFO | train_grpo_train | metrics_logged mode=train step=539

                                                      
{'loss': -0.0456, 'grad_norm': 15.343780517578125, 'learning_rate': 8.36969696969697e-06, 'num_tokens': 963359.0, 'completions/mean_length': 37.625, 'completions/min_length': 33.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.3659902811050415, 'rewards/meter/std': 0.3842555284500122, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9993736743927002, 'rewards/repeat_soft/std': 0.0009022135054692626, 'rewards/judge_quality/mean': 0.8600000143051147, 'rewards/judge_quality/std': 0.11122693121433258, 'rewards/total_composite/mean': 0.539921760559082, 'rewards/total_composite/std': 0.19357290863990784, 'reward': 0.539921760559082, 'reward_std': 0.19357289373874664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12317245453596115, 'sampling/sampling_logp_difference/max': 1.6021738052368164, 'sampling/importance_sampling_ratio/min': 0.20145811140537262, 'sampling/importance_sampling_ratio/mean': 1.0202888250350952, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7020215913653374, 'clip_ratio/low_mean': 0.09025133214890957, 'clip_ratio/low_min': 0.09025133214890957, 'clip_ratio/high_mean': 0.038580391090363264, 'clip_ratio/high_max': 0.038580391090363264, 'clip_ratio/region_mean': 0.12883172323927283, 'reward_total_mean': 0.539921760559082, 'reward_meter_mean': 0.3659902811050415, 'reward_meter_std': 0.3842555284500122, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9993736743927002, 'reward_repeat_soft_std': 0.0009022135054692626, 'reward_judge_quality_mean': 0.8600000143051147, 'reward_judge_quality_std': 0.11122693121433258, 'reward_total_composite_mean': 0.539921760559082, 'reward_total_composite_std': 0.19357290863990784, 'epoch': 0.05}

 16%|█▋        | 539/3300 [1:21:54<5:28:18,  7.13s/it]INFO 04-13 08:42:52 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 540/3300 [1:22:05<6:22:38,  8.32s/it]2026-04-13 08:43:03,426 | INFO | train_grpo_train | metrics_logged mode=train step=540

                                                      
{'loss': -0.0766, 'grad_norm': 3.216494083404541, 'learning_rate': 8.366666666666667e-06, 'num_tokens': 964777.0, 'completions/mean_length': 87.25, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.571430206298828, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.8376508951187134, 'rewards/meter/std': 0.30644145607948303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.6862500309944153, 'rewards/judge_quality/std': 0.3422170579433441, 'rewards/total_composite/mean': 0.7103989124298096, 'rewards/total_composite/std': 0.3222620189189911, 'reward': 0.7103989124298096, 'reward_std': 0.3222619891166687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13616713881492615, 'sampling/sampling_logp_difference/max': 1.0162763595581055, 'sampling/importance_sampling_ratio/min': 0.3619401752948761, 'sampling/importance_sampling_ratio/mean': 1.0602952241897583, 'sampling/importance_sampling_ratio/max': 1.8277971744537354, 'entropy': 0.8669724240899086, 'clip_ratio/low_mean': 0.04556958540342748, 'clip_ratio/low_min': 0.04556958540342748, 'clip_ratio/high_mean': 0.045931439846754074, 'clip_ratio/high_max': 0.045931439846754074, 'clip_ratio/region_mean': 0.09150102525018156, 'reward_total_mean': 0.7103989124298096, 'reward_meter_mean': 0.8376508951187134, 'reward_meter_std': 0.30644145607948303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.6862500309944153, 'reward_judge_quality_std': 0.3422170579433441, 'reward_total_composite_mean': 0.7103989124298096, 'reward_total_composite_std': 0.3222620189189911, 'epoch': 0.05}

 16%|█▋        | 540/3300 [1:22:05<6:22:38,  8.32s/it]INFO 04-13 08:43:03 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 541/3300 [1:22:12<5:56:09,  7.75s/it]2026-04-13 08:43:09,834 | INFO | train_grpo_train | metrics_logged mode=train step=541

                                                      
{'loss': 0.0156, 'grad_norm': 14.095969200134277, 'learning_rate': 8.363636363636365e-06, 'num_tokens': 966520.0, 'completions/mean_length': 63.875, 'completions/min_length': 50.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5517494678497314, 'rewards/meter/std': 0.2106204777956009, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985389709472656, 'rewards/repeat_soft/std': 0.0013374953996390104, 'rewards/judge_quality/mean': 0.7862499952316284, 'rewards/judge_quality/std': 0.21967104077339172, 'rewards/total_composite/mean': 0.6565507650375366, 'rewards/total_composite/std': 0.15947522222995758, 'reward': 0.6565507650375366, 'reward_std': 0.15947522222995758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14598749577999115, 'sampling/sampling_logp_difference/max': 2.278921365737915, 'sampling/importance_sampling_ratio/min': 0.10239458829164505, 'sampling/importance_sampling_ratio/mean': 1.0148061513900757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7033660002052784, 'clip_ratio/low_mean': 0.050267857033759356, 'clip_ratio/low_min': 0.050267857033759356, 'clip_ratio/high_mean': 0.08420956321060658, 'clip_ratio/high_max': 0.08420956321060658, 'clip_ratio/region_mean': 0.13447742024436593, 'reward_total_mean': 0.6565507650375366, 'reward_meter_mean': 0.5517494678497314, 'reward_meter_std': 0.2106204777956009, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985389709472656, 'reward_repeat_soft_std': 0.0013374953996390104, 'reward_judge_quality_mean': 0.7862499952316284, 'reward_judge_quality_std': 0.21967104077339172, 'reward_total_composite_mean': 0.6565507650375366, 'reward_total_composite_std': 0.15947522222995758, 'epoch': 0.05}

 16%|█▋        | 541/3300 [1:22:12<5:56:09,  7.75s/it]INFO 04-13 08:43:10 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 542/3300 [1:22:17<5:31:05,  7.20s/it]2026-04-13 08:43:15,770 | INFO | train_grpo_train | metrics_logged mode=train step=542

                                                      
{'loss': 0.0043, 'grad_norm': 17.3575439453125, 'learning_rate': 8.360606060606062e-06, 'num_tokens': 968089.0, 'completions/mean_length': 33.125, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8381178379058838, 'rewards/meter/std': 0.2982902526855469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9539903998374939, 'rewards/repeat_soft/std': 0.06638079881668091, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.67268967628479, 'rewards/total_composite/std': 0.18456526100635529, 'reward': 0.67268967628479, 'reward_std': 0.1845652461051941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1786205917596817, 'sampling/sampling_logp_difference/max': 1.7740802764892578, 'sampling/importance_sampling_ratio/min': 0.16963939368724823, 'sampling/importance_sampling_ratio/mean': 1.026566505432129, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6051879823207855, 'clip_ratio/low_mean': 0.11004703678190708, 'clip_ratio/low_min': 0.11004703678190708, 'clip_ratio/high_mean': 0.08839285932481289, 'clip_ratio/high_max': 0.08839285932481289, 'clip_ratio/region_mean': 0.19843989610671997, 'reward_total_mean': 0.67268967628479, 'reward_meter_mean': 0.8381178379058838, 'reward_meter_std': 0.2982902526855469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9539903998374939, 'reward_repeat_soft_std': 0.06638079881668091, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.67268967628479, 'reward_total_composite_std': 0.18456526100635529, 'epoch': 0.05}

 16%|█▋        | 542/3300 [1:22:17<5:31:05,  7.20s/it]INFO 04-13 08:43:15 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 543/3300 [1:22:29<6:26:05,  8.40s/it]2026-04-13 08:43:26,972 | INFO | train_grpo_train | metrics_logged mode=train step=543

                                                      
{'loss': -0.0796, 'grad_norm': 3.910327434539795, 'learning_rate': 8.357575757575759e-06, 'num_tokens': 969548.0, 'completions/mean_length': 88.375, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.85714340209961, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.45423436164855957, 'rewards/meter/std': 0.38455843925476074, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9979463815689087, 'rewards/repeat_soft/std': 0.005131691228598356, 'rewards/judge_quality/mean': 0.36374998092651367, 'rewards/judge_quality/std': 0.14302222430706024, 'rewards/total_composite/mean': 0.4281260073184967, 'rewards/total_composite/std': 0.20449328422546387, 'reward': 0.4281260073184967, 'reward_std': 0.20449328422546387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20579834282398224, 'sampling/sampling_logp_difference/max': 1.9135963916778564, 'sampling/importance_sampling_ratio/min': 0.14754877984523773, 'sampling/importance_sampling_ratio/mean': 1.0137248039245605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2950005009770393, 'clip_ratio/low_mean': 0.06727272644639015, 'clip_ratio/low_min': 0.06727272644639015, 'clip_ratio/high_mean': 0.09855741448700428, 'clip_ratio/high_max': 0.09855741448700428, 'clip_ratio/region_mean': 0.16583014093339443, 'reward_total_mean': 0.4281260073184967, 'reward_meter_mean': 0.45423436164855957, 'reward_meter_std': 0.38455843925476074, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9979463815689087, 'reward_repeat_soft_std': 0.005131691228598356, 'reward_judge_quality_mean': 0.36374998092651367, 'reward_judge_quality_std': 0.14302222430706024, 'reward_total_composite_mean': 0.4281260073184967, 'reward_total_composite_std': 0.20449328422546387, 'epoch': 0.05}

 16%|█▋        | 543/3300 [1:22:29<6:26:05,  8.40s/it]INFO 04-13 08:43:27 [block_pool.py:378] Successfully reset prefix cache

 16%|█▋        | 544/3300 [1:22:35<5:52:07,  7.67s/it]2026-04-13 08:43:32,921 | INFO | train_grpo_train | metrics_logged mode=train step=544

                                                      
{'loss': 0.0019, 'grad_norm': 27.745807647705078, 'learning_rate': 8.354545454545455e-06, 'num_tokens': 971066.0, 'completions/mean_length': 25.75, 'completions/min_length': 20.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7400469183921814, 'rewards/meter/std': 0.35674533247947693, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9558712244033813, 'rewards/repeat_soft/std': 0.018749024718999863, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.6075124740600586, 'rewards/total_composite/std': 0.15838244557380676, 'reward': 0.6075124740600586, 'reward_std': 0.15838246047496796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1736089140176773, 'sampling/sampling_logp_difference/max': 1.2797205448150635, 'sampling/importance_sampling_ratio/min': 0.3145306706428528, 'sampling/importance_sampling_ratio/mean': 1.0237181186676025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1577944606542587, 'clip_ratio/low_mean': 0.06673119682818651, 'clip_ratio/low_min': 0.06673119682818651, 'clip_ratio/high_mean': 0.05760630592703819, 'clip_ratio/high_max': 0.05760630592703819, 'clip_ratio/region_mean': 0.1243375027552247, 'reward_total_mean': 0.6075124740600586, 'reward_meter_mean': 0.7400469183921814, 'reward_meter_std': 0.35674533247947693, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9558712244033813, 'reward_repeat_soft_std': 0.018749024718999863, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.6075124740600586, 'reward_total_composite_std': 0.15838244557380676, 'epoch': 0.05}

 16%|█▋        | 544/3300 [1:22:35<5:52:07,  7.67s/it]INFO 04-13 08:43:33 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 545/3300 [1:22:41<5:30:22,  7.19s/it]2026-04-13 08:43:39,016 | INFO | train_grpo_train | metrics_logged mode=train step=545

                                                      
{'loss': 0.0132, 'grad_norm': 14.433030128479004, 'learning_rate': 8.351515151515152e-06, 'num_tokens': 972727.0, 'completions/mean_length': 43.625, 'completions/min_length': 40.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7088823318481445, 'rewards/meter/std': 0.25771865248680115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.996931791305542, 'rewards/repeat_soft/std': 0.003609251230955124, 'rewards/judge_quality/mean': 0.8362500667572021, 'rewards/judge_quality/std': 0.17104199528694153, 'rewards/total_composite/mean': 0.7405744791030884, 'rewards/total_composite/std': 0.18008670210838318, 'reward': 0.7405744791030884, 'reward_std': 0.18008668720722198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1720448136329651, 'sampling/sampling_logp_difference/max': 1.806600570678711, 'sampling/importance_sampling_ratio/min': 0.1642114222049713, 'sampling/importance_sampling_ratio/mean': 1.0251959562301636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3306350335478783, 'clip_ratio/low_mean': 0.05352247226983309, 'clip_ratio/low_min': 0.05352247226983309, 'clip_ratio/high_mean': 0.10940993577241898, 'clip_ratio/high_max': 0.10940993577241898, 'clip_ratio/region_mean': 0.16293240804225206, 'reward_total_mean': 0.7405744791030884, 'reward_meter_mean': 0.7088823318481445, 'reward_meter_std': 0.25771865248680115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.996931791305542, 'reward_repeat_soft_std': 0.003609251230955124, 'reward_judge_quality_mean': 0.8362500667572021, 'reward_judge_quality_std': 0.17104199528694153, 'reward_total_composite_mean': 0.7405744791030884, 'reward_total_composite_std': 0.18008670210838318, 'epoch': 0.05}

 17%|█▋        | 545/3300 [1:22:41<5:30:22,  7.19s/it]INFO 04-13 08:43:39 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 546/3300 [1:22:47<5:13:46,  6.84s/it]2026-04-13 08:43:45,014 | INFO | train_grpo_train | metrics_logged mode=train step=546

                                                      
{'loss': 0.0764, 'grad_norm': 22.39914321899414, 'learning_rate': 8.348484848484849e-06, 'num_tokens': 974133.0, 'completions/mean_length': 24.75, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9132776856422424, 'rewards/meter/std': 0.14427873492240906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5974857807159424, 'rewards/total_composite/std': 0.037712667137384415, 'reward': 0.5974857807159424, 'reward_std': 0.03771267458796501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13518214225769043, 'sampling/sampling_logp_difference/max': 0.8918828964233398, 'sampling/importance_sampling_ratio/min': 0.4098832607269287, 'sampling/importance_sampling_ratio/mean': 1.0276554822921753, 'sampling/importance_sampling_ratio/max': 1.8256620168685913, 'entropy': 1.02891393750906, 'clip_ratio/low_mean': 0.013392857741564512, 'clip_ratio/low_min': 0.013392857741564512, 'clip_ratio/high_mean': 0.10959415882825851, 'clip_ratio/high_max': 0.10959415882825851, 'clip_ratio/region_mean': 0.12298701656982303, 'reward_total_mean': 0.5974857807159424, 'reward_meter_mean': 0.9132776856422424, 'reward_meter_std': 0.14427873492240906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5974857807159424, 'reward_total_composite_std': 0.037712667137384415, 'epoch': 0.05}

 17%|█▋        | 546/3300 [1:22:47<5:13:46,  6.84s/it]INFO 04-13 08:43:45 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 547/3300 [1:22:53<4:59:59,  6.54s/it]2026-04-13 08:43:50,858 | INFO | train_grpo_train | metrics_logged mode=train step=547

                                                      
{'loss': 0.0325, 'grad_norm': 15.003937721252441, 'learning_rate': 8.345454545454546e-06, 'num_tokens': 975818.0, 'completions/mean_length': 33.625, 'completions/min_length': 26.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9554383158683777, 'rewards/meter/std': 0.062374863773584366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9542199969291687, 'rewards/repeat_soft/std': 0.04472750797867775, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.22663691639900208, 'rewards/total_composite/mean': 0.6425566673278809, 'rewards/total_composite/std': 0.2965072989463806, 'reward': 0.6425566673278809, 'reward_std': 0.2965072989463806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16646623611450195, 'sampling/sampling_logp_difference/max': 1.8456273078918457, 'sampling/importance_sampling_ratio/min': 0.15792621672153473, 'sampling/importance_sampling_ratio/mean': 1.013169288635254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1770803853869438, 'clip_ratio/low_mean': 0.12836424820125103, 'clip_ratio/low_min': 0.12836424820125103, 'clip_ratio/high_mean': 0.06548554822802544, 'clip_ratio/high_max': 0.06548554822802544, 'clip_ratio/region_mean': 0.19384979642927647, 'reward_total_mean': 0.6425566673278809, 'reward_meter_mean': 0.9554383158683777, 'reward_meter_std': 0.062374863773584366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9542199969291687, 'reward_repeat_soft_std': 0.04472750797867775, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.22663691639900208, 'reward_total_composite_mean': 0.6425566673278809, 'reward_total_composite_std': 0.2965072989463806, 'epoch': 0.05}

 17%|█▋        | 547/3300 [1:22:53<4:59:59,  6.54s/it]INFO 04-13 08:43:51 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 548/3300 [1:22:58<4:50:53,  6.34s/it]2026-04-13 08:43:56,742 | INFO | train_grpo_train | metrics_logged mode=train step=548

                                                      
{'loss': 0.0166, 'grad_norm': 18.18979835510254, 'learning_rate': 8.342424242424244e-06, 'num_tokens': 977363.0, 'completions/mean_length': 35.125, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9609639644622803, 'rewards/meter/std': 0.060940779745578766, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969215393066406, 'rewards/repeat_soft/std': 0.005694178864359856, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.8910194635391235, 'rewards/total_composite/std': 0.12894363701343536, 'reward': 0.8910194635391235, 'reward_std': 0.12894362211227417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16594471037387848, 'sampling/sampling_logp_difference/max': 1.630638599395752, 'sampling/importance_sampling_ratio/min': 0.19580449163913727, 'sampling/importance_sampling_ratio/mean': 1.0251747369766235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8938365876674652, 'clip_ratio/low_mean': 0.025735294446349144, 'clip_ratio/low_min': 0.025735294446349144, 'clip_ratio/high_mean': 0.07592684868723154, 'clip_ratio/high_max': 0.07592684868723154, 'clip_ratio/region_mean': 0.10166214313358068, 'reward_total_mean': 0.8910194635391235, 'reward_meter_mean': 0.9609639644622803, 'reward_meter_std': 0.060940779745578766, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969215393066406, 'reward_repeat_soft_std': 0.005694178864359856, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.8910194635391235, 'reward_total_composite_std': 0.12894363701343536, 'epoch': 0.06}

 17%|█▋        | 548/3300 [1:22:58<4:50:53,  6.34s/it]INFO 04-13 08:43:56 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 549/3300 [1:23:06<5:00:52,  6.56s/it]2026-04-13 08:44:03,816 | INFO | train_grpo_train | metrics_logged mode=train step=549

                                                      
{'loss': 0.0707, 'grad_norm': 17.993812561035156, 'learning_rate': 8.339393939393941e-06, 'num_tokens': 979548.0, 'completions/mean_length': 89.125, 'completions/min_length': 61.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.125, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.5475062727928162, 'rewards/meter/std': 0.3193122446537018, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9620498418807983, 'rewards/repeat_soft/std': 0.03260951116681099, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.18071287870407104, 'rewards/total_composite/mean': 0.4460676908493042, 'rewards/total_composite/std': 0.2205701619386673, 'reward': 0.4460676908493042, 'reward_std': 0.2205701470375061, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1919461488723755, 'sampling/sampling_logp_difference/max': 2.2864503860473633, 'sampling/importance_sampling_ratio/min': 0.10162656009197235, 'sampling/importance_sampling_ratio/mean': 1.0069726705551147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2636125013232231, 'clip_ratio/low_mean': 0.0745668588206172, 'clip_ratio/low_min': 0.0745668588206172, 'clip_ratio/high_mean': 0.10123595409095287, 'clip_ratio/high_max': 0.10123595409095287, 'clip_ratio/region_mean': 0.17580281291157007, 'reward_total_mean': 0.4460676908493042, 'reward_meter_mean': 0.5475062727928162, 'reward_meter_std': 0.3193122446537018, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9620498418807983, 'reward_repeat_soft_std': 0.03260951116681099, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.18071287870407104, 'reward_total_composite_mean': 0.4460676908493042, 'reward_total_composite_std': 0.2205701619386673, 'epoch': 0.06}

 17%|█▋        | 549/3300 [1:23:06<5:00:52,  6.56s/it]INFO 04-13 08:44:04 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 550/3300 [1:23:12<4:57:46,  6.50s/it]2026-04-13 08:44:10,163 | INFO | train_grpo_train | metrics_logged mode=train step=550

                                                      
{'loss': -0.0036, 'grad_norm': 10.08622932434082, 'learning_rate': 8.336363636363636e-06, 'num_tokens': 981377.0, 'completions/mean_length': 59.625, 'completions/min_length': 51.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9905789494514465, 'rewards/meter/std': 0.004395305644720793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7257529497146606, 'rewards/repeat_soft/std': 0.16920576989650726, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.6059386134147644, 'rewards/total_composite/std': 0.07042774558067322, 'reward': 0.6059386134147644, 'reward_std': 0.07042776048183441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16200944781303406, 'sampling/sampling_logp_difference/max': 2.120972156524658, 'sampling/importance_sampling_ratio/min': 0.11991500109434128, 'sampling/importance_sampling_ratio/mean': 1.020952820777893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2838521227240562, 'clip_ratio/low_mean': 0.056965377647429705, 'clip_ratio/low_min': 0.056965377647429705, 'clip_ratio/high_mean': 0.09082790184766054, 'clip_ratio/high_max': 0.09082790184766054, 'clip_ratio/region_mean': 0.14779327949509025, 'reward_total_mean': 0.6059386134147644, 'reward_meter_mean': 0.9905789494514465, 'reward_meter_std': 0.004395305644720793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7257529497146606, 'reward_repeat_soft_std': 0.16920576989650726, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.6059386134147644, 'reward_total_composite_std': 0.07042774558067322, 'epoch': 0.06}

 17%|█▋        | 550/3300 [1:23:12<4:57:46,  6.50s/it]INFO 04-13 08:44:10 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:13,  1.67s/it]

 30%|███       | 3/10 [00:12<00:32,  4.59s/it]

 40%|████      | 4/10 [00:15<00:24,  4.11s/it]

 50%|█████     | 5/10 [00:20<00:21,  4.34s/it]

 60%|██████    | 6/10 [00:24<00:16,  4.25s/it]

 70%|███████   | 7/10 [00:28<00:12,  4.28s/it]

 80%|████████  | 8/10 [00:32<00:08,  4.18s/it]

 90%|█████████ | 9/10 [00:36<00:04,  4.12s/it]

100%|██████████| 10/10 [00:39<00:00,  3.93s/it]2026-04-13 08:44:53,871 | INFO | train_grpo_train | metrics_logged mode=eval step=550

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 43.6753, 'eval_samples_per_second': 1.832, 'eval_steps_per_second': 0.229, 'eval_num_tokens': 981377.0, 'eval_completions/mean_length': 71.7875, 'eval_completions/min_length': 26.6, 'eval_completions/max_length': 143.4, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 60.645833587646486, 'eval_completions/min_terminated_length': 26.6, 'eval_completions/max_terminated_length': 102.9, 'eval_rewards/meter/mean': 0.6610803008079529, 'eval_rewards/meter/std': 0.33115275800228117, 'eval_rewards/count_adherence/mean': 0.9912499904632568, 'eval_rewards/count_adherence/std': 0.02474873922765255, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.19317627549171448, 'eval_rewards/repeat_soft/mean': 0.9651135087013245, 'eval_rewards/repeat_soft/std': 0.03794628819450736, 'eval_rewards/judge_quality/mean': 0.5327500224113464, 'eval_rewards/judge_quality/std': 0.20415999740362167, 'eval_rewards/total_composite/mean': 0.5292426437139511, 'eval_rewards/total_composite/std': 0.19960086941719055, 'eval_reward': 0.5292426437139511, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.10629716143012047, 'eval_sampling/sampling_logp_difference/max': 1.1720905303955078, 'eval_sampling/importance_sampling_ratio/min': 0.3158396989107132, 'eval_sampling/importance_sampling_ratio/mean': 1.023800778388977, 'eval_sampling/importance_sampling_ratio/max': 1.4582764148712157, 'eval_entropy': 1.327412235736847, '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': 0.5292426437139511, 'eval_reward_meter_mean': 0.6610803008079529, 'eval_reward_meter_std': 0.33115275800228117, 'eval_reward_count_adherence_mean': 0.9912499904632568, 'eval_reward_count_adherence_std': 0.02474873922765255, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.19317627549171448, 'eval_reward_repeat_soft_mean': 0.9651135087013245, 'eval_reward_repeat_soft_std': 0.03794628819450736, 'eval_reward_judge_quality_mean': 0.5327500224113464, 'eval_reward_judge_quality_std': 0.20415999740362167, 'eval_reward_total_composite_mean': 0.5292426437139511, 'eval_reward_total_composite_std': 0.19960086941719055, 'epoch': 0.06}

 17%|█▋        | 550/3300 [1:23:56<4:57:46,  6.50s/it]

100%|██████████| 10/10 [00:39<00:00,  3.93s/it]

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

 17%|█▋        | 551/3300 [1:24:06<15:52:24, 20.79s/it]2026-04-13 08:45:04,294 | INFO | train_grpo_train | metrics_logged mode=train step=551

                                                       
{'loss': 0.1879, 'grad_norm': 20.56439971923828, 'learning_rate': 8.333333333333334e-06, 'num_tokens': 982994.0, 'completions/mean_length': 40.125, 'completions/min_length': 31.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6993616223335266, 'rewards/meter/std': 0.37903064489364624, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9672762155532837, 'rewards/repeat_soft/std': 0.0380360409617424, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.18431341648101807, 'rewards/total_composite/mean': 0.5190733671188354, 'rewards/total_composite/std': 0.21756958961486816, 'reward': 0.5190733671188354, 'reward_std': 0.21756958961486816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20431751012802124, 'sampling/sampling_logp_difference/max': 1.978684425354004, 'sampling/importance_sampling_ratio/min': 0.13825100660324097, 'sampling/importance_sampling_ratio/mean': 1.0076302289962769, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4213807880878448, 'clip_ratio/low_mean': 0.044801464304327965, 'clip_ratio/low_min': 0.044801464304327965, 'clip_ratio/high_mean': 0.12536935694515705, 'clip_ratio/high_max': 0.12536935694515705, 'clip_ratio/region_mean': 0.17017082124948502, 'reward_total_mean': 0.5190733671188354, 'reward_meter_mean': 0.6993616223335266, 'reward_meter_std': 0.37903064489364624, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9672762155532837, 'reward_repeat_soft_std': 0.0380360409617424, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.18431341648101807, 'reward_total_composite_mean': 0.5190733671188354, 'reward_total_composite_std': 0.21756958961486816, 'epoch': 0.06}

 17%|█▋        | 551/3300 [1:24:06<15:52:24, 20.79s/it]INFO 04-13 08:45:04 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 552/3300 [1:24:12<12:35:22, 16.49s/it]2026-04-13 08:45:10,769 | INFO | train_grpo_train | metrics_logged mode=train step=552

                                                       
{'loss': -0.0518, 'grad_norm': 18.777772903442383, 'learning_rate': 8.330303030303031e-06, 'num_tokens': 984750.0, 'completions/mean_length': 39.5, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8659785389900208, 'rewards/meter/std': 0.17113295197486877, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9889838099479675, 'rewards/repeat_soft/std': 0.008318226784467697, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.6291760802268982, 'rewards/total_composite/std': 0.12751486897468567, 'reward': 0.6291760802268982, 'reward_std': 0.12751485407352448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1923200637102127, 'sampling/sampling_logp_difference/max': 1.5239753723144531, 'sampling/importance_sampling_ratio/min': 0.217844158411026, 'sampling/importance_sampling_ratio/mean': 1.0168901681900024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4269536957144737, 'clip_ratio/low_mean': 0.15119895059615374, 'clip_ratio/low_min': 0.15119895059615374, 'clip_ratio/high_mean': 0.012500000186264515, 'clip_ratio/high_max': 0.012500000186264515, 'clip_ratio/region_mean': 0.16369895078241825, 'reward_total_mean': 0.6291760802268982, 'reward_meter_mean': 0.8659785389900208, 'reward_meter_std': 0.17113295197486877, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9889838099479675, 'reward_repeat_soft_std': 0.008318226784467697, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.6291760802268982, 'reward_total_composite_std': 0.12751486897468567, 'epoch': 0.06}

 17%|█▋        | 552/3300 [1:24:12<12:35:22, 16.49s/it]INFO 04-13 08:45:11 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 553/3300 [1:24:19<10:17:24, 13.49s/it]2026-04-13 08:45:17,233 | INFO | train_grpo_train | metrics_logged mode=train step=553

                                                       
{'loss': -0.0032, 'grad_norm': 19.589412689208984, 'learning_rate': 8.327272727272728e-06, 'num_tokens': 986289.0, 'completions/mean_length': 33.375, 'completions/min_length': 30.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.435761958360672, 'rewards/meter/std': 0.35650891065597534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9923810958862305, 'rewards/repeat_soft/std': 0.008589569479227066, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.5534482002258301, 'rewards/total_composite/std': 0.17337222397327423, 'reward': 0.5534482002258301, 'reward_std': 0.17337222397327423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1868068426847458, 'sampling/sampling_logp_difference/max': 2.0306060314178467, 'sampling/importance_sampling_ratio/min': 0.13125595450401306, 'sampling/importance_sampling_ratio/mean': 1.0006626844406128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8828173168003559, 'clip_ratio/low_mean': 0.0615382962860167, 'clip_ratio/low_min': 0.0615382962860167, 'clip_ratio/high_mean': 0.10045306198298931, 'clip_ratio/high_max': 0.10045306198298931, 'clip_ratio/region_mean': 0.161991358269006, 'reward_total_mean': 0.5534482002258301, 'reward_meter_mean': 0.435761958360672, 'reward_meter_std': 0.35650891065597534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9923810958862305, 'reward_repeat_soft_std': 0.008589569479227066, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.5534482002258301, 'reward_total_composite_std': 0.17337222397327423, 'epoch': 0.06}

 17%|█▋        | 553/3300 [1:24:19<10:17:24, 13.49s/it]INFO 04-13 08:45:17 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 554/3300 [1:24:31<9:53:00, 12.96s/it] 2026-04-13 08:45:28,957 | INFO | train_grpo_train | metrics_logged mode=train step=554

                                                      
{'loss': -0.1029, 'grad_norm': 3.118800163269043, 'learning_rate': 8.324242424242425e-06, 'num_tokens': 987937.0, 'completions/mean_length': 101.0, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.28571701049805, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.777889609336853, 'rewards/meter/std': 0.38723862171173096, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9913482069969177, 'rewards/repeat_soft/std': 0.013020490296185017, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5406044125556946, 'rewards/total_composite/std': 0.2407016158103943, 'reward': 0.5406044125556946, 'reward_std': 0.2407016158103943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21402639150619507, 'sampling/sampling_logp_difference/max': 1.3933115005493164, 'sampling/importance_sampling_ratio/min': 0.24825185537338257, 'sampling/importance_sampling_ratio/mean': 1.00910222530365, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4998129308223724, 'clip_ratio/low_mean': 0.021276595070958138, 'clip_ratio/low_min': 0.021276595070958138, 'clip_ratio/high_mean': 0.15984027367085218, 'clip_ratio/high_max': 0.15984027367085218, 'clip_ratio/region_mean': 0.18111686874181032, 'reward_total_mean': 0.5406044125556946, 'reward_meter_mean': 0.777889609336853, 'reward_meter_std': 0.38723862171173096, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9913482069969177, 'reward_repeat_soft_std': 0.013020490296185017, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5406044125556946, 'reward_total_composite_std': 0.2407016158103943, 'epoch': 0.06}

 17%|█▋        | 554/3300 [1:24:31<9:53:00, 12.96s/it]INFO 04-13 08:45:29 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 555/3300 [1:24:37<8:22:40, 10.99s/it]2026-04-13 08:45:35,351 | INFO | train_grpo_train | metrics_logged mode=train step=555

                                                      
{'loss': 0.0479, 'grad_norm': 13.309117317199707, 'learning_rate': 8.321212121212123e-06, 'num_tokens': 989836.0, 'completions/mean_length': 64.375, 'completions/min_length': 44.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.6408752202987671, 'rewards/meter/std': 0.3405615985393524, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9968665242195129, 'rewards/repeat_soft/std': 0.002312359632924199, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6169316172599792, 'rewards/total_composite/std': 0.16860747337341309, 'reward': 0.6169316172599792, 'reward_std': 0.16860747337341309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19001272320747375, 'sampling/sampling_logp_difference/max': 1.2530326843261719, 'sampling/importance_sampling_ratio/min': 0.3489038050174713, 'sampling/importance_sampling_ratio/mean': 1.0523494482040405, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7197429314255714, 'clip_ratio/low_mean': 0.11605823412537575, 'clip_ratio/low_min': 0.11605823412537575, 'clip_ratio/high_mean': 0.07755850814282894, 'clip_ratio/high_max': 0.07755850814282894, 'clip_ratio/region_mean': 0.1936167422682047, 'reward_total_mean': 0.6169316172599792, 'reward_meter_mean': 0.6408752202987671, 'reward_meter_std': 0.3405615985393524, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9968665242195129, 'reward_repeat_soft_std': 0.002312359632924199, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6169316172599792, 'reward_total_composite_std': 0.16860747337341309, 'epoch': 0.06}

 17%|█▋        | 555/3300 [1:24:37<8:22:40, 10.99s/it]INFO 04-13 08:45:35 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 556/3300 [1:24:43<7:19:41,  9.61s/it]2026-04-13 08:45:41,762 | INFO | train_grpo_train | metrics_logged mode=train step=556

                                                      
{'loss': 0.1001, 'grad_norm': 18.60330581665039, 'learning_rate': 8.318181818181818e-06, 'num_tokens': 991329.0, 'completions/mean_length': 33.625, 'completions/min_length': 28.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9354280233383179, 'rewards/meter/std': 0.11490907520055771, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9618691205978394, 'rewards/repeat_soft/std': 0.03599108010530472, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.25176167488098145, 'rewards/total_composite/mean': 0.681156575679779, 'rewards/total_composite/std': 0.13293671607971191, 'reward': 0.681156575679779, 'reward_std': 0.13293671607971191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17382174730300903, 'sampling/sampling_logp_difference/max': 1.3681821823120117, 'sampling/importance_sampling_ratio/min': 0.2602751553058624, 'sampling/importance_sampling_ratio/mean': 1.0161211490631104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2629192024469376, 'clip_ratio/low_mean': 0.08075824566185474, 'clip_ratio/low_min': 0.08075824566185474, 'clip_ratio/high_mean': 0.05999729596078396, 'clip_ratio/high_max': 0.05999729596078396, 'clip_ratio/region_mean': 0.1407555416226387, 'reward_total_mean': 0.681156575679779, 'reward_meter_mean': 0.9354280233383179, 'reward_meter_std': 0.11490907520055771, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9618691205978394, 'reward_repeat_soft_std': 0.03599108010530472, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.25176167488098145, 'reward_total_composite_mean': 0.681156575679779, 'reward_total_composite_std': 0.13293671607971191, 'epoch': 0.06}

 17%|█▋        | 556/3300 [1:24:43<7:19:41,  9.61s/it]INFO 04-13 08:45:42 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 557/3300 [1:24:50<6:32:17,  8.58s/it]2026-04-13 08:45:47,932 | INFO | train_grpo_train | metrics_logged mode=train step=557

                                                      
{'loss': 0.0789, 'grad_norm': 22.79527473449707, 'learning_rate': 8.315151515151516e-06, 'num_tokens': 992845.0, 'completions/mean_length': 28.5, 'completions/min_length': 27.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.45790940523147583, 'rewards/meter/std': 0.3044460117816925, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.986666202545166, 'rewards/repeat_soft/std': 0.014813925139605999, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.5273362398147583, 'rewards/total_composite/std': 0.10774970799684525, 'reward': 0.5273362398147583, 'reward_std': 0.10774971544742584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1532248854637146, 'sampling/sampling_logp_difference/max': 1.1305646896362305, 'sampling/importance_sampling_ratio/min': 0.32285091280937195, 'sampling/importance_sampling_ratio/mean': 1.024466872215271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8794346153736115, 'clip_ratio/low_mean': 0.07793672196567059, 'clip_ratio/low_min': 0.07793672196567059, 'clip_ratio/high_mean': 0.12040343880653381, 'clip_ratio/high_max': 0.12040343880653381, 'clip_ratio/region_mean': 0.1983401607722044, 'reward_total_mean': 0.5273362398147583, 'reward_meter_mean': 0.45790940523147583, 'reward_meter_std': 0.3044460117816925, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.986666202545166, 'reward_repeat_soft_std': 0.014813925139605999, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.5273362398147583, 'reward_total_composite_std': 0.10774970799684525, 'epoch': 0.06}

 17%|█▋        | 557/3300 [1:24:50<6:32:17,  8.58s/it]INFO 04-13 08:45:48 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 558/3300 [1:24:56<6:03:56,  7.96s/it]2026-04-13 08:45:54,455 | INFO | train_grpo_train | metrics_logged mode=train step=558

                                                      
{'loss': 0.0572, 'grad_norm': 18.1457462310791, 'learning_rate': 8.312121212121213e-06, 'num_tokens': 994267.0, 'completions/mean_length': 26.75, 'completions/min_length': 20.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.7564411163330078, 'rewards/meter/std': 0.42362114787101746, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9554247856140137, 'rewards/repeat_soft/std': 0.017965074628591537, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5618064403533936, 'rewards/total_composite/std': 0.12254803627729416, 'reward': 0.5618064403533936, 'reward_std': 0.12254803627729416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17637909948825836, 'sampling/sampling_logp_difference/max': 1.416147232055664, 'sampling/importance_sampling_ratio/min': 0.24264706671237946, 'sampling/importance_sampling_ratio/mean': 1.0523557662963867, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3914649784564972, 'clip_ratio/low_mean': 0.04214015230536461, 'clip_ratio/low_min': 0.04214015230536461, 'clip_ratio/high_mean': 0.1335457293316722, 'clip_ratio/high_max': 0.1335457293316722, 'clip_ratio/region_mean': 0.1756858816370368, 'reward_total_mean': 0.5618064403533936, 'reward_meter_mean': 0.7564411163330078, 'reward_meter_std': 0.42362114787101746, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9554247856140137, 'reward_repeat_soft_std': 0.017965074628591537, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5618064403533936, 'reward_total_composite_std': 0.12254803627729416, 'epoch': 0.06}

 17%|█▋        | 558/3300 [1:24:56<6:03:56,  7.96s/it]INFO 04-13 08:45:54 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 559/3300 [1:25:03<5:49:38,  7.65s/it]2026-04-13 08:46:01,392 | INFO | train_grpo_train | metrics_logged mode=train step=559

                                                      
{'loss': -0.1657, 'grad_norm': 23.551433563232422, 'learning_rate': 8.30909090909091e-06, 'num_tokens': 995834.0, 'completions/mean_length': 38.875, 'completions/min_length': 9.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 9.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6795017719268799, 'rewards/meter/std': 0.3432086408138275, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9923096299171448, 'rewards/repeat_soft/std': 0.006110474467277527, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.26042547821998596, 'rewards/total_composite/mean': 0.4991290867328644, 'rewards/total_composite/std': 0.3228200376033783, 'reward': 0.4991290867328644, 'reward_std': 0.3228200376033783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1697164624929428, 'sampling/sampling_logp_difference/max': 1.7782261371612549, 'sampling/importance_sampling_ratio/min': 0.1689375638961792, 'sampling/importance_sampling_ratio/mean': 1.0004527568817139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.101341262459755, 'clip_ratio/low_mean': 0.0486111119389534, 'clip_ratio/low_min': 0.0486111119389534, 'clip_ratio/high_mean': 0.13180590840056539, 'clip_ratio/high_max': 0.13180590840056539, 'clip_ratio/region_mean': 0.18041702033951879, 'reward_total_mean': 0.4991290867328644, 'reward_meter_mean': 0.6795017719268799, 'reward_meter_std': 0.3432086408138275, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9923096299171448, 'reward_repeat_soft_std': 0.006110474467277527, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.26042547821998596, 'reward_total_composite_mean': 0.4991290867328644, 'reward_total_composite_std': 0.3228200376033783, 'epoch': 0.06}

 17%|█▋        | 559/3300 [1:25:03<5:49:38,  7.65s/it]INFO 04-13 08:46:01 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 560/3300 [1:25:09<5:24:06,  7.10s/it]2026-04-13 08:46:07,194 | INFO | train_grpo_train | metrics_logged mode=train step=560

                                                      
{'loss': 0.0135, 'grad_norm': 22.024646759033203, 'learning_rate': 8.306060606060606e-06, 'num_tokens': 997180.0, 'completions/mean_length': 19.25, 'completions/min_length': 15.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.8734430074691772, 'rewards/meter/std': 0.31486058235168457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9593343734741211, 'rewards/repeat_soft/std': 0.008953613229095936, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6650569438934326, 'rewards/total_composite/std': 0.1872844099998474, 'reward': 0.6650569438934326, 'reward_std': 0.1872844099998474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16088703274726868, 'sampling/sampling_logp_difference/max': 0.7771531343460083, 'sampling/importance_sampling_ratio/min': 0.4651177227497101, 'sampling/importance_sampling_ratio/mean': 1.0268609523773193, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4013171046972275, 'clip_ratio/low_mean': 0.10809837467968464, 'clip_ratio/low_min': 0.10809837467968464, 'clip_ratio/high_mean': 0.03910427913069725, 'clip_ratio/high_max': 0.03910427913069725, 'clip_ratio/region_mean': 0.1472026538103819, 'reward_total_mean': 0.6650569438934326, 'reward_meter_mean': 0.8734430074691772, 'reward_meter_std': 0.31486058235168457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9593343734741211, 'reward_repeat_soft_std': 0.008953613229095936, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6650569438934326, 'reward_total_composite_std': 0.1872844099998474, 'epoch': 0.06}

 17%|█▋        | 560/3300 [1:25:09<5:24:06,  7.10s/it]INFO 04-13 08:46:07 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 561/3300 [1:25:21<6:26:17,  8.46s/it]2026-04-13 08:46:18,831 | INFO | train_grpo_train | metrics_logged mode=train step=561

                                                      
{'loss': -0.1958, 'grad_norm': 3.2702996730804443, 'learning_rate': 8.303030303030305e-06, 'num_tokens': 999803.0, 'completions/mean_length': 183.875, 'completions/min_length': 101.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 137.0, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.8589916229248047, 'rewards/meter/std': 0.22314368188381195, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9383134841918945, 'rewards/repeat_soft/std': 0.03509976714849472, 'rewards/judge_quality/mean': 0.32499998807907104, 'rewards/judge_quality/std': 0.15212775766849518, 'rewards/total_composite/mean': 0.45598551630973816, 'rewards/total_composite/std': 0.2027304619550705, 'reward': 0.45598551630973816, 'reward_std': 0.2027304619550705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19196419417858124, 'sampling/sampling_logp_difference/max': 1.3713769912719727, 'sampling/importance_sampling_ratio/min': 0.2537572979927063, 'sampling/importance_sampling_ratio/mean': 1.0575730800628662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.8707079887390137, 'clip_ratio/low_mean': 0.018103448674082756, 'clip_ratio/low_min': 0.018103448674082756, 'clip_ratio/high_mean': 0.12402686849236488, 'clip_ratio/high_max': 0.12402686849236488, 'clip_ratio/region_mean': 0.14213031716644764, 'reward_total_mean': 0.45598551630973816, 'reward_meter_mean': 0.8589916229248047, 'reward_meter_std': 0.22314368188381195, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9383134841918945, 'reward_repeat_soft_std': 0.03509976714849472, 'reward_judge_quality_mean': 0.32499998807907104, 'reward_judge_quality_std': 0.15212775766849518, 'reward_total_composite_mean': 0.45598551630973816, 'reward_total_composite_std': 0.2027304619550705, 'epoch': 0.06}

 17%|█▋        | 561/3300 [1:25:21<6:26:17,  8.46s/it]INFO 04-13 08:46:19 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 562/3300 [1:25:28<6:07:48,  8.06s/it]2026-04-13 08:46:26,253 | INFO | train_grpo_train | metrics_logged mode=train step=562

                                                      
{'loss': 0.0926, 'grad_norm': 15.954366683959961, 'learning_rate': 8.3e-06, 'num_tokens': 1001398.0, 'completions/mean_length': 42.375, 'completions/min_length': 37.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8383921384811401, 'rewards/meter/std': 0.14216502010822296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9863839149475098, 'rewards/repeat_soft/std': 0.02003294602036476, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7530303001403809, 'rewards/total_composite/std': 0.1667768359184265, 'reward': 0.7530303001403809, 'reward_std': 0.1667768657207489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18183524906635284, 'sampling/sampling_logp_difference/max': 1.2515754699707031, 'sampling/importance_sampling_ratio/min': 0.28605377674102783, 'sampling/importance_sampling_ratio/mean': 1.0057904720306396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3422640785574913, 'clip_ratio/low_mean': 0.04274665657430887, 'clip_ratio/low_min': 0.04274665657430887, 'clip_ratio/high_mean': 0.09359593316912651, 'clip_ratio/high_max': 0.09359593316912651, 'clip_ratio/region_mean': 0.13634258974343538, 'reward_total_mean': 0.7530303001403809, 'reward_meter_mean': 0.8383921384811401, 'reward_meter_std': 0.14216502010822296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9863839149475098, 'reward_repeat_soft_std': 0.02003294602036476, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7530303001403809, 'reward_total_composite_std': 0.1667768359184265, 'epoch': 0.06}

 17%|█▋        | 562/3300 [1:25:28<6:07:48,  8.06s/it]INFO 04-13 08:46:26 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 563/3300 [1:25:35<5:57:29,  7.84s/it]2026-04-13 08:46:33,269 | INFO | train_grpo_train | metrics_logged mode=train step=563

                                                      
{'loss': 0.093, 'grad_norm': 13.501004219055176, 'learning_rate': 8.296969696969697e-06, 'num_tokens': 1003244.0, 'completions/mean_length': 56.75, 'completions/min_length': 49.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8068620562553406, 'rewards/meter/std': 0.31526026129722595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9576163291931152, 'rewards/repeat_soft/std': 0.03529730811715126, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.48893821239471436, 'rewards/total_composite/std': 0.21414178609848022, 'reward': 0.48893821239471436, 'reward_std': 0.21414178609848022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16479219496250153, 'sampling/sampling_logp_difference/max': 1.5833408832550049, 'sampling/importance_sampling_ratio/min': 0.2052880972623825, 'sampling/importance_sampling_ratio/mean': 1.0237668752670288, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0530619770288467, 'clip_ratio/low_mean': 0.06251307297497988, 'clip_ratio/low_min': 0.06251307297497988, 'clip_ratio/high_mean': 0.10389220993965864, 'clip_ratio/high_max': 0.10389220993965864, 'clip_ratio/region_mean': 0.16640528291463852, 'reward_total_mean': 0.48893821239471436, 'reward_meter_mean': 0.8068620562553406, 'reward_meter_std': 0.31526026129722595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9576163291931152, 'reward_repeat_soft_std': 0.03529730811715126, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.48893821239471436, 'reward_total_composite_std': 0.21414178609848022, 'epoch': 0.06}

 17%|█▋        | 563/3300 [1:25:35<5:57:29,  7.84s/it]INFO 04-13 08:46:33 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 564/3300 [1:25:41<5:30:38,  7.25s/it]2026-04-13 08:46:39,152 | INFO | train_grpo_train | metrics_logged mode=train step=564

                                                      
{'loss': 0.0644, 'grad_norm': 20.114315032958984, 'learning_rate': 8.293939393939395e-06, 'num_tokens': 1004749.0, 'completions/mean_length': 22.125, 'completions/min_length': 19.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9016860127449036, 'rewards/meter/std': 0.14367784559726715, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9148127436637878, 'rewards/repeat_soft/std': 0.06743112951517105, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7272079586982727, 'rewards/total_composite/std': 0.1561218947172165, 'reward': 0.7272079586982727, 'reward_std': 0.15612190961837769, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16825152933597565, 'sampling/sampling_logp_difference/max': 1.327385663986206, 'sampling/importance_sampling_ratio/min': 0.3427193760871887, 'sampling/importance_sampling_ratio/mean': 1.029595971107483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.017693430185318, 'clip_ratio/low_mean': 0.08651043241843581, 'clip_ratio/low_min': 0.08651043241843581, 'clip_ratio/high_mean': 0.0468017878010869, 'clip_ratio/high_max': 0.0468017878010869, 'clip_ratio/region_mean': 0.13331222021952271, 'reward_total_mean': 0.7272079586982727, 'reward_meter_mean': 0.9016860127449036, 'reward_meter_std': 0.14367784559726715, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9148127436637878, 'reward_repeat_soft_std': 0.06743112951517105, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7272079586982727, 'reward_total_composite_std': 0.1561218947172165, 'epoch': 0.06}

 17%|█▋        | 564/3300 [1:25:41<5:30:38,  7.25s/it]INFO 04-13 08:46:39 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 565/3300 [1:25:47<5:14:51,  6.91s/it]2026-04-13 08:46:45,260 | INFO | train_grpo_train | metrics_logged mode=train step=565

                                                      
{'loss': -0.0434, 'grad_norm': 18.368539810180664, 'learning_rate': 8.290909090909092e-06, 'num_tokens': 1006144.0, 'completions/mean_length': 19.375, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8119484186172485, 'rewards/meter/std': 0.3543795049190521, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9605007171630859, 'rewards/repeat_soft/std': 0.004598783794790506, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.524174690246582, 'rewards/total_composite/std': 0.2167847901582718, 'reward': 0.524174690246582, 'reward_std': 0.2167847752571106, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19802412390708923, 'sampling/sampling_logp_difference/max': 1.334218978881836, 'sampling/importance_sampling_ratio/min': 0.263363778591156, 'sampling/importance_sampling_ratio/mean': 0.9816744923591614, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5238390415906906, 'clip_ratio/low_mean': 0.03851540666073561, 'clip_ratio/low_min': 0.03851540666073561, 'clip_ratio/high_mean': 0.1667059948667884, 'clip_ratio/high_max': 0.1667059948667884, 'clip_ratio/region_mean': 0.205221401527524, 'reward_total_mean': 0.524174690246582, 'reward_meter_mean': 0.8119484186172485, 'reward_meter_std': 0.3543795049190521, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9605007171630859, 'reward_repeat_soft_std': 0.004598783794790506, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.524174690246582, 'reward_total_composite_std': 0.2167847901582718, 'epoch': 0.06}

 17%|█▋        | 565/3300 [1:25:47<5:14:51,  6.91s/it]INFO 04-13 08:46:45 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 566/3300 [1:25:53<5:07:46,  6.75s/it]2026-04-13 08:46:51,655 | INFO | train_grpo_train | metrics_logged mode=train step=566

                                                      
{'loss': 0.1211, 'grad_norm': 20.236555099487305, 'learning_rate': 8.287878787878787e-06, 'num_tokens': 1007630.0, 'completions/mean_length': 41.75, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8665573596954346, 'rewards/meter/std': 0.3077116906642914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9985834360122681, 'rewards/repeat_soft/std': 0.003662696573883295, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.22937417030334473, 'rewards/total_composite/mean': 0.6672229170799255, 'rewards/total_composite/std': 0.18799370527267456, 'reward': 0.6672229170799255, 'reward_std': 0.18799370527267456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19202497601509094, 'sampling/sampling_logp_difference/max': 1.9503707885742188, 'sampling/importance_sampling_ratio/min': 0.1422213315963745, 'sampling/importance_sampling_ratio/mean': 1.0313491821289062, 'sampling/importance_sampling_ratio/max': 1.9711854457855225, 'entropy': 1.863529920578003, 'clip_ratio/low_mean': 0.12671365309506655, 'clip_ratio/low_min': 0.12671365309506655, 'clip_ratio/high_mean': 0.038888889364898205, 'clip_ratio/high_max': 0.038888889364898205, 'clip_ratio/region_mean': 0.16560254245996475, 'reward_total_mean': 0.6672229170799255, 'reward_meter_mean': 0.8665573596954346, 'reward_meter_std': 0.3077116906642914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9985834360122681, 'reward_repeat_soft_std': 0.003662696573883295, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.22937417030334473, 'reward_total_composite_mean': 0.6672229170799255, 'reward_total_composite_std': 0.18799370527267456, 'epoch': 0.06}

 17%|█▋        | 566/3300 [1:25:53<5:07:46,  6.75s/it]INFO 04-13 08:46:51 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 567/3300 [1:25:59<4:58:08,  6.55s/it]2026-04-13 08:46:57,713 | INFO | train_grpo_train | metrics_logged mode=train step=567

                                                      
{'loss': 0.0495, 'grad_norm': 18.38676643371582, 'learning_rate': 8.284848484848486e-06, 'num_tokens': 1009144.0, 'completions/mean_length': 38.25, 'completions/min_length': 34.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9538601636886597, 'rewards/meter/std': 0.10505902767181396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.980604887008667, 'rewards/repeat_soft/std': 0.019677551463246346, 'rewards/judge_quality/mean': 0.7487500309944153, 'rewards/judge_quality/std': 0.21853001415729523, 'rewards/total_composite/mean': 0.8193771243095398, 'rewards/total_composite/std': 0.16228778660297394, 'reward': 0.8193771243095398, 'reward_std': 0.16228780150413513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1814090460538864, 'sampling/sampling_logp_difference/max': 1.8432121276855469, 'sampling/importance_sampling_ratio/min': 0.15830810368061066, 'sampling/importance_sampling_ratio/mean': 1.0160726308822632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1473611667752266, 'clip_ratio/low_mean': 0.06210801098495722, 'clip_ratio/low_min': 0.06210801098495722, 'clip_ratio/high_mean': 0.1015858231112361, 'clip_ratio/high_max': 0.1015858231112361, 'clip_ratio/region_mean': 0.1636938340961933, 'reward_total_mean': 0.8193771243095398, 'reward_meter_mean': 0.9538601636886597, 'reward_meter_std': 0.10505902767181396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.980604887008667, 'reward_repeat_soft_std': 0.019677551463246346, 'reward_judge_quality_mean': 0.7487500309944153, 'reward_judge_quality_std': 0.21853001415729523, 'reward_total_composite_mean': 0.8193771243095398, 'reward_total_composite_std': 0.16228778660297394, 'epoch': 0.06}

 17%|█▋        | 567/3300 [1:25:59<4:58:08,  6.55s/it]INFO 04-13 08:46:57 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 568/3300 [1:26:06<4:57:46,  6.54s/it]2026-04-13 08:47:04,240 | INFO | train_grpo_train | metrics_logged mode=train step=568

                                                      
{'loss': 0.0749, 'grad_norm': 15.373639106750488, 'learning_rate': 8.281818181818182e-06, 'num_tokens': 1010964.0, 'completions/mean_length': 48.5, 'completions/min_length': 42.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9488092660903931, 'rewards/meter/std': 0.10913003236055374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.937246561050415, 'rewards/repeat_soft/std': 0.07722450792789459, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.585860550403595, 'rewards/total_composite/std': 0.053448330610990524, 'reward': 0.585860550403595, 'reward_std': 0.05344833433628082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20093071460723877, 'sampling/sampling_logp_difference/max': 1.6151256561279297, 'sampling/importance_sampling_ratio/min': 0.19886568188667297, 'sampling/importance_sampling_ratio/mean': 0.9998076558113098, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3005739152431488, 'clip_ratio/low_mean': 0.04811224527657032, 'clip_ratio/low_min': 0.04811224527657032, 'clip_ratio/high_mean': 0.1577774751931429, 'clip_ratio/high_max': 0.1577774751931429, 'clip_ratio/region_mean': 0.2058897204697132, 'reward_total_mean': 0.585860550403595, 'reward_meter_mean': 0.9488092660903931, 'reward_meter_std': 0.10913003236055374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.937246561050415, 'reward_repeat_soft_std': 0.07722450792789459, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.585860550403595, 'reward_total_composite_std': 0.053448330610990524, 'epoch': 0.06}

 17%|█▋        | 568/3300 [1:26:06<4:57:46,  6.54s/it]INFO 04-13 08:47:04 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 569/3300 [1:26:12<4:49:20,  6.36s/it]2026-04-13 08:47:10,169 | INFO | train_grpo_train | metrics_logged mode=train step=569

                                                      
{'loss': 0.0922, 'grad_norm': 23.371784210205078, 'learning_rate': 8.27878787878788e-06, 'num_tokens': 1012403.0, 'completions/mean_length': 29.875, 'completions/min_length': 23.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.6001070141792297, 'rewards/meter/std': 0.4020649492740631, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9339101314544678, 'rewards/repeat_soft/std': 0.06879226118326187, 'rewards/judge_quality/mean': 0.8650000095367432, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/total_composite/mean': 0.6974748969078064, 'rewards/total_composite/std': 0.24909807741641998, 'reward': 0.6974748969078064, 'reward_std': 0.2490980625152588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15942718088626862, 'sampling/sampling_logp_difference/max': 1.461019515991211, 'sampling/importance_sampling_ratio/min': 0.23199963569641113, 'sampling/importance_sampling_ratio/mean': 0.9991790652275085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7846294343471527, 'clip_ratio/low_mean': 0.039762435015290976, 'clip_ratio/low_min': 0.039762435015290976, 'clip_ratio/high_mean': 0.06495859334245324, 'clip_ratio/high_max': 0.06495859334245324, 'clip_ratio/region_mean': 0.10472102835774422, 'reward_total_mean': 0.6974748969078064, 'reward_meter_mean': 0.6001070141792297, 'reward_meter_std': 0.4020649492740631, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9339101314544678, 'reward_repeat_soft_std': 0.06879226118326187, 'reward_judge_quality_mean': 0.8650000095367432, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.6974748969078064, 'reward_total_composite_std': 0.24909807741641998, 'epoch': 0.06}

 17%|█▋        | 569/3300 [1:26:12<4:49:20,  6.36s/it]INFO 04-13 08:47:10 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 570/3300 [1:26:18<4:43:07,  6.22s/it]2026-04-13 08:47:16,080 | INFO | train_grpo_train | metrics_logged mode=train step=570

                                                      
{'loss': 0.0473, 'grad_norm': 17.427688598632812, 'learning_rate': 8.275757575757577e-06, 'num_tokens': 1013981.0, 'completions/mean_length': 40.25, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5616979598999023, 'rewards/meter/std': 0.3486446738243103, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9733009338378906, 'rewards/repeat_soft/std': 0.028935855254530907, 'rewards/judge_quality/mean': 0.543749988079071, 'rewards/judge_quality/std': 0.14647649228572845, 'rewards/total_composite/mean': 0.48278260231018066, 'rewards/total_composite/std': 0.2439933717250824, 'reward': 0.48278260231018066, 'reward_std': 0.2439933717250824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16998045146465302, 'sampling/sampling_logp_difference/max': 1.2940139770507812, 'sampling/importance_sampling_ratio/min': 0.2741680443286896, 'sampling/importance_sampling_ratio/mean': 1.0163642168045044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3805529028177261, 'clip_ratio/low_mean': 0.06506410427391529, 'clip_ratio/low_min': 0.06506410427391529, 'clip_ratio/high_mean': 0.10038667172193527, 'clip_ratio/high_max': 0.10038667172193527, 'clip_ratio/region_mean': 0.16545077599585056, 'reward_total_mean': 0.48278260231018066, 'reward_meter_mean': 0.5616979598999023, 'reward_meter_std': 0.3486446738243103, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9733009338378906, 'reward_repeat_soft_std': 0.028935855254530907, 'reward_judge_quality_mean': 0.543749988079071, 'reward_judge_quality_std': 0.14647649228572845, 'reward_total_composite_mean': 0.48278260231018066, 'reward_total_composite_std': 0.2439933717250824, 'epoch': 0.06}

 17%|█▋        | 570/3300 [1:26:18<4:43:07,  6.22s/it]INFO 04-13 08:47:16 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 571/3300 [1:26:29<5:54:28,  7.79s/it]2026-04-13 08:47:27,537 | INFO | train_grpo_train | metrics_logged mode=train step=571

                                                      
{'loss': -0.1987, 'grad_norm': 3.877699613571167, 'learning_rate': 8.272727272727274e-06, 'num_tokens': 1016407.0, 'completions/mean_length': 154.25, 'completions/min_length': 84.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 103.14286041259766, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.49068641662597656, 'rewards/meter/std': 0.2176356166601181, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.993622899055481, 'rewards/repeat_soft/std': 0.003882135497406125, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.2121320515871048, 'rewards/total_composite/mean': 0.4497496485710144, 'rewards/total_composite/std': 0.20662787556648254, 'reward': 0.4497496485710144, 'reward_std': 0.20662787556648254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19983239471912384, 'sampling/sampling_logp_difference/max': 2.1510086059570312, 'sampling/importance_sampling_ratio/min': 0.11636672168970108, 'sampling/importance_sampling_ratio/mean': 1.033197045326233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.37043796479702, 'clip_ratio/low_mean': 0.06433553248643875, 'clip_ratio/low_min': 0.06433553248643875, 'clip_ratio/high_mean': 0.08608468901365995, 'clip_ratio/high_max': 0.08608468901365995, 'clip_ratio/region_mean': 0.1504202215000987, 'reward_total_mean': 0.4497496485710144, 'reward_meter_mean': 0.49068641662597656, 'reward_meter_std': 0.2176356166601181, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.993622899055481, 'reward_repeat_soft_std': 0.003882135497406125, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.2121320515871048, 'reward_total_composite_mean': 0.4497496485710144, 'reward_total_composite_std': 0.20662787556648254, 'epoch': 0.06}

 17%|█▋        | 571/3300 [1:26:29<5:54:28,  7.79s/it]INFO 04-13 08:47:27 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 572/3300 [1:26:36<5:36:15,  7.40s/it]2026-04-13 08:47:34,005 | INFO | train_grpo_train | metrics_logged mode=train step=572

                                                      
{'loss': 0.1492, 'grad_norm': 16.63024139404297, 'learning_rate': 8.269696969696971e-06, 'num_tokens': 1018101.0, 'completions/mean_length': 59.75, 'completions/min_length': 46.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7083966135978699, 'rewards/meter/std': 0.357369601726532, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9095284342765808, 'rewards/repeat_soft/std': 0.0766533613204956, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5992577075958252, 'rewards/total_composite/std': 0.1747995764017105, 'reward': 0.5992577075958252, 'reward_std': 0.17479956150054932, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16766788065433502, 'sampling/sampling_logp_difference/max': 1.4486650228500366, 'sampling/importance_sampling_ratio/min': 0.23488366603851318, 'sampling/importance_sampling_ratio/mean': 1.0143681764602661, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0160523355007172, 'clip_ratio/low_mean': 0.06509399600327015, 'clip_ratio/low_min': 0.06509399600327015, 'clip_ratio/high_mean': 0.09081595111638308, 'clip_ratio/high_max': 0.09081595111638308, 'clip_ratio/region_mean': 0.15590994711965322, 'reward_total_mean': 0.5992577075958252, 'reward_meter_mean': 0.7083966135978699, 'reward_meter_std': 0.357369601726532, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9095284342765808, 'reward_repeat_soft_std': 0.0766533613204956, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5992577075958252, 'reward_total_composite_std': 0.1747995764017105, 'epoch': 0.06}

 17%|█▋        | 572/3300 [1:26:36<5:36:15,  7.40s/it]INFO 04-13 08:47:34 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 573/3300 [1:26:47<6:27:03,  8.52s/it]2026-04-13 08:47:45,137 | INFO | train_grpo_train | metrics_logged mode=train step=573

                                                      
{'loss': -0.0601, 'grad_norm': 5.129828929901123, 'learning_rate': 8.266666666666667e-06, 'num_tokens': 1019906.0, 'completions/mean_length': 110.625, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.28571701049805, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8684149980545044, 'rewards/meter/std': 0.3483780026435852, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.7412621974945068, 'rewards/repeat_soft/std': 0.20249877870082855, 'rewards/judge_quality/mean': 0.33125001192092896, 'rewards/judge_quality/std': 0.13715866208076477, 'rewards/total_composite/mean': 0.3450530469417572, 'rewards/total_composite/std': 0.2909436523914337, 'reward': 0.3450530469417572, 'reward_std': 0.29094362258911133, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1584600806236267, 'sampling/sampling_logp_difference/max': 1.4114131927490234, 'sampling/importance_sampling_ratio/min': 0.24379850924015045, 'sampling/importance_sampling_ratio/mean': 1.0394682884216309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.247895561158657, 'clip_ratio/low_mean': 0.054653484374284744, 'clip_ratio/low_min': 0.054653484374284744, 'clip_ratio/high_mean': 0.07300596591085196, 'clip_ratio/high_max': 0.07300596591085196, 'clip_ratio/region_mean': 0.1276594502851367, 'reward_total_mean': 0.3450530469417572, 'reward_meter_mean': 0.8684149980545044, 'reward_meter_std': 0.3483780026435852, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.7412621974945068, 'reward_repeat_soft_std': 0.20249877870082855, 'reward_judge_quality_mean': 0.33125001192092896, 'reward_judge_quality_std': 0.13715866208076477, 'reward_total_composite_mean': 0.3450530469417572, 'reward_total_composite_std': 0.2909436523914337, 'epoch': 0.06}

 17%|█▋        | 573/3300 [1:26:47<6:27:03,  8.52s/it]INFO 04-13 08:47:45 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 574/3300 [1:26:53<5:52:15,  7.75s/it]2026-04-13 08:47:51,108 | INFO | train_grpo_train | metrics_logged mode=train step=574

                                                      
{'loss': 0.0174, 'grad_norm': 33.07956314086914, 'learning_rate': 8.263636363636366e-06, 'num_tokens': 1021418.0, 'completions/mean_length': 41.0, 'completions/min_length': 30.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6174425482749939, 'rewards/meter/std': 0.37483736872673035, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9980881810188293, 'rewards/repeat_soft/std': 0.005338655319064856, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.1277204155921936, 'rewards/total_composite/mean': 0.5354558229446411, 'rewards/total_composite/std': 0.09790372103452682, 'reward': 0.5354558229446411, 'reward_std': 0.09790373593568802, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17352846264839172, 'sampling/sampling_logp_difference/max': 1.3274812698364258, 'sampling/importance_sampling_ratio/min': 0.2651442587375641, 'sampling/importance_sampling_ratio/mean': 0.9985684156417847, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8166610077023506, 'clip_ratio/low_mean': 0.08509136270731688, 'clip_ratio/low_min': 0.08509136270731688, 'clip_ratio/high_mean': 0.08918650913983583, 'clip_ratio/high_max': 0.08918650913983583, 'clip_ratio/region_mean': 0.1742778718471527, 'reward_total_mean': 0.5354558229446411, 'reward_meter_mean': 0.6174425482749939, 'reward_meter_std': 0.37483736872673035, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9980881810188293, 'reward_repeat_soft_std': 0.005338655319064856, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.1277204155921936, 'reward_total_composite_mean': 0.5354558229446411, 'reward_total_composite_std': 0.09790372103452682, 'epoch': 0.06}

 17%|█▋        | 574/3300 [1:26:53<5:52:15,  7.75s/it]INFO 04-13 08:47:51 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 575/3300 [1:26:59<5:27:48,  7.22s/it]2026-04-13 08:47:57,079 | INFO | train_grpo_train | metrics_logged mode=train step=575

                                                      
{'loss': 0.0678, 'grad_norm': 20.26004409790039, 'learning_rate': 8.260606060606061e-06, 'num_tokens': 1023041.0, 'completions/mean_length': 32.875, 'completions/min_length': 25.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.830098032951355, 'rewards/meter/std': 0.3257444500923157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9855596423149109, 'rewards/repeat_soft/std': 0.018953589722514153, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.27994900941848755, 'rewards/total_composite/mean': 0.6754432320594788, 'rewards/total_composite/std': 0.20469191670417786, 'reward': 0.6754432320594788, 'reward_std': 0.20469191670417786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16457249224185944, 'sampling/sampling_logp_difference/max': 1.3743727207183838, 'sampling/importance_sampling_ratio/min': 0.2529982328414917, 'sampling/importance_sampling_ratio/mean': 0.9803752303123474, 'sampling/importance_sampling_ratio/max': 1.7883014678955078, 'entropy': 1.0091527700424194, 'clip_ratio/low_mean': 0.03716179495677352, 'clip_ratio/low_min': 0.03716179495677352, 'clip_ratio/high_mean': 0.036587481619790196, 'clip_ratio/high_max': 0.036587481619790196, 'clip_ratio/region_mean': 0.07374927657656372, 'reward_total_mean': 0.6754432320594788, 'reward_meter_mean': 0.830098032951355, 'reward_meter_std': 0.3257444500923157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9855596423149109, 'reward_repeat_soft_std': 0.018953589722514153, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.27994900941848755, 'reward_total_composite_mean': 0.6754432320594788, 'reward_total_composite_std': 0.20469191670417786, 'epoch': 0.06}

 17%|█▋        | 575/3300 [1:26:59<5:27:48,  7.22s/it]INFO 04-13 08:47:57 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 576/3300 [1:27:05<5:12:06,  6.87s/it]2026-04-13 08:48:03,152 | INFO | train_grpo_train | metrics_logged mode=train step=576

                                                      
{'loss': 0.0122, 'grad_norm': 29.855863571166992, 'learning_rate': 8.257575757575758e-06, 'num_tokens': 1024508.0, 'completions/mean_length': 32.375, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8965328931808472, 'rewards/meter/std': 0.14603206515312195, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9723629951477051, 'rewards/repeat_soft/std': 0.042240433394908905, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5822746157646179, 'rewards/total_composite/std': 0.04208853840827942, 'reward': 0.5822746157646179, 'reward_std': 0.042088523507118225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17073535919189453, 'sampling/sampling_logp_difference/max': 2.2079977989196777, 'sampling/importance_sampling_ratio/min': 0.10992051661014557, 'sampling/importance_sampling_ratio/mean': 1.0111312866210938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9421104788780212, 'clip_ratio/low_mean': 0.03553427383303642, 'clip_ratio/low_min': 0.03553427383303642, 'clip_ratio/high_mean': 0.10378700867295265, 'clip_ratio/high_max': 0.10378700867295265, 'clip_ratio/region_mean': 0.13932128250598907, 'reward_total_mean': 0.5822746157646179, 'reward_meter_mean': 0.8965328931808472, 'reward_meter_std': 0.14603206515312195, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9723629951477051, 'reward_repeat_soft_std': 0.042240433394908905, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5822746157646179, 'reward_total_composite_std': 0.04208853840827942, 'epoch': 0.06}

 17%|█▋        | 576/3300 [1:27:05<5:12:06,  6.87s/it]INFO 04-13 08:48:03 [block_pool.py:378] Successfully reset prefix cache

 17%|█▋        | 577/3300 [1:27:16<6:14:42,  8.26s/it]2026-04-13 08:48:14,633 | INFO | train_grpo_train | metrics_logged mode=train step=577

                                                      
{'loss': -0.1619, 'grad_norm': 4.448294162750244, 'learning_rate': 8.254545454545456e-06, 'num_tokens': 1026542.0, 'completions/mean_length': 127.25, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 72.28572082519531, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.7502434253692627, 'rewards/meter/std': 0.2678214907646179, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9235777854919434, 'rewards/repeat_soft/std': 0.09846273064613342, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.2105392962694168, 'rewards/total_composite/mean': 0.5139748454093933, 'rewards/total_composite/std': 0.2346297651529312, 'reward': 0.5139748454093933, 'reward_std': 0.2346297651529312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20501717925071716, 'sampling/sampling_logp_difference/max': 1.633583664894104, 'sampling/importance_sampling_ratio/min': 0.1952286809682846, 'sampling/importance_sampling_ratio/mean': 1.032249093055725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.357079640030861, 'clip_ratio/low_mean': 0.0470130518078804, 'clip_ratio/low_min': 0.0470130518078804, 'clip_ratio/high_mean': 0.11227748077362776, 'clip_ratio/high_max': 0.11227748077362776, 'clip_ratio/region_mean': 0.15929053258150816, 'reward_total_mean': 0.5139748454093933, 'reward_meter_mean': 0.7502434253692627, 'reward_meter_std': 0.2678214907646179, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9235777854919434, 'reward_repeat_soft_std': 0.09846273064613342, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.2105392962694168, 'reward_total_composite_mean': 0.5139748454093933, 'reward_total_composite_std': 0.2346297651529312, 'epoch': 0.06}

 17%|█▋        | 577/3300 [1:27:16<6:14:42,  8.26s/it]INFO 04-13 08:48:14 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 578/3300 [1:27:22<5:45:43,  7.62s/it]2026-04-13 08:48:20,772 | INFO | train_grpo_train | metrics_logged mode=train step=578

                                                      
{'loss': 0.0497, 'grad_norm': 16.69968032836914, 'learning_rate': 8.251515151515153e-06, 'num_tokens': 1028078.0, 'completions/mean_length': 35.0, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9882268905639648, 'rewards/meter/std': 0.004939976613968611, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9579628705978394, 'rewards/repeat_soft/std': 0.053920794278383255, 'rewards/judge_quality/mean': 0.5974999666213989, 'rewards/judge_quality/std': 0.220891073346138, 'rewards/total_composite/mean': 0.7276318073272705, 'rewards/total_composite/std': 0.14510180056095123, 'reward': 0.7276318073272705, 'reward_std': 0.14510181546211243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2057153284549713, 'sampling/sampling_logp_difference/max': 1.1017704010009766, 'sampling/importance_sampling_ratio/min': 0.33228227496147156, 'sampling/importance_sampling_ratio/mean': 1.0183676481246948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7146768867969513, 'clip_ratio/low_mean': 0.08770505990833044, 'clip_ratio/low_min': 0.08770505990833044, 'clip_ratio/high_mean': 0.07052139192819595, 'clip_ratio/high_max': 0.07052139192819595, 'clip_ratio/region_mean': 0.1582264518365264, 'reward_total_mean': 0.7276318073272705, 'reward_meter_mean': 0.9882268905639648, 'reward_meter_std': 0.004939976613968611, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9579628705978394, 'reward_repeat_soft_std': 0.053920794278383255, 'reward_judge_quality_mean': 0.5974999666213989, 'reward_judge_quality_std': 0.220891073346138, 'reward_total_composite_mean': 0.7276318073272705, 'reward_total_composite_std': 0.14510180056095123, 'epoch': 0.06}

 18%|█▊        | 578/3300 [1:27:23<5:45:43,  7.62s/it]INFO 04-13 08:48:21 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 579/3300 [1:27:29<5:28:56,  7.25s/it]2026-04-13 08:48:27,165 | INFO | train_grpo_train | metrics_logged mode=train step=579

                                                      
{'loss': -0.0024, 'grad_norm': 20.66975212097168, 'learning_rate': 8.248484848484848e-06, 'num_tokens': 1029624.0, 'completions/mean_length': 37.25, 'completions/min_length': 33.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.26896071434020996, 'rewards/meter/std': 0.26846253871917725, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975525736808777, 'rewards/repeat_soft/std': 0.0028212887700647116, 'rewards/judge_quality/mean': 0.6012499928474426, 'rewards/judge_quality/std': 0.22350697219371796, 'rewards/total_composite/mean': 0.4431477189064026, 'rewards/total_composite/std': 0.08015746623277664, 'reward': 0.4431477189064026, 'reward_std': 0.08015745133161545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17363011837005615, 'sampling/sampling_logp_difference/max': 1.3714125156402588, 'sampling/importance_sampling_ratio/min': 0.2537482678890228, 'sampling/importance_sampling_ratio/mean': 1.0310235023498535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.264290452003479, 'clip_ratio/low_mean': 0.08566673472523689, 'clip_ratio/low_min': 0.08566673472523689, 'clip_ratio/high_mean': 0.052020203322172165, 'clip_ratio/high_max': 0.052020203322172165, 'clip_ratio/region_mean': 0.13768693804740906, 'reward_total_mean': 0.4431477189064026, 'reward_meter_mean': 0.26896071434020996, 'reward_meter_std': 0.26846253871917725, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975525736808777, 'reward_repeat_soft_std': 0.0028212887700647116, 'reward_judge_quality_mean': 0.6012499928474426, 'reward_judge_quality_std': 0.22350697219371796, 'reward_total_composite_mean': 0.4431477189064026, 'reward_total_composite_std': 0.08015746623277664, 'epoch': 0.06}

 18%|█▊        | 579/3300 [1:27:29<5:28:56,  7.25s/it]INFO 04-13 08:48:27 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 580/3300 [1:27:36<5:27:25,  7.22s/it]2026-04-13 08:48:34,317 | INFO | train_grpo_train | metrics_logged mode=train step=580

                                                      
{'loss': 0.0927, 'grad_norm': 9.298893928527832, 'learning_rate': 8.245454545454546e-06, 'num_tokens': 1031667.0, 'completions/mean_length': 76.375, 'completions/min_length': 61.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.375, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.963345468044281, 'rewards/meter/std': 0.05591030418872833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6659365892410278, 'rewards/repeat_soft/std': 0.18439270555973053, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.17266297340393066, 'rewards/total_composite/mean': 0.5362281799316406, 'rewards/total_composite/std': 0.12325633317232132, 'reward': 0.5362281799316406, 'reward_std': 0.12325632572174072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11173097044229507, 'sampling/sampling_logp_difference/max': 1.7862787246704102, 'sampling/importance_sampling_ratio/min': 0.1675826460123062, 'sampling/importance_sampling_ratio/mean': 1.0131728649139404, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8593889139592648, 'clip_ratio/low_mean': 0.02751857414841652, 'clip_ratio/low_min': 0.02751857414841652, 'clip_ratio/high_mean': 0.07695895247161388, 'clip_ratio/high_max': 0.07695895247161388, 'clip_ratio/region_mean': 0.1044775266200304, 'reward_total_mean': 0.5362281799316406, 'reward_meter_mean': 0.963345468044281, 'reward_meter_std': 0.05591030418872833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6659365892410278, 'reward_repeat_soft_std': 0.18439270555973053, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.17266297340393066, 'reward_total_composite_mean': 0.5362281799316406, 'reward_total_composite_std': 0.12325633317232132, 'epoch': 0.06}

 18%|█▊        | 580/3300 [1:27:36<5:27:25,  7.22s/it]INFO 04-13 08:48:34 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 581/3300 [1:27:42<5:12:58,  6.91s/it]2026-04-13 08:48:40,485 | INFO | train_grpo_train | metrics_logged mode=train step=581

                                                      
{'loss': -0.0275, 'grad_norm': 17.0106143951416, 'learning_rate': 8.242424242424243e-06, 'num_tokens': 1033521.0, 'completions/mean_length': 49.75, 'completions/min_length': 38.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.838412344455719, 'rewards/meter/std': 0.3349061608314514, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9712759256362915, 'rewards/repeat_soft/std': 0.0544203482568264, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.57071852684021, 'rewards/total_composite/std': 0.35998836159706116, 'reward': 0.57071852684021, 'reward_std': 0.35998836159706116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1976938396692276, 'sampling/sampling_logp_difference/max': 1.3331255912780762, 'sampling/importance_sampling_ratio/min': 0.26365190744400024, 'sampling/importance_sampling_ratio/mean': 1.0169309377670288, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2996100634336472, 'clip_ratio/low_mean': 0.04820086061954498, 'clip_ratio/low_min': 0.04820086061954498, 'clip_ratio/high_mean': 0.1340278284624219, 'clip_ratio/high_max': 0.1340278284624219, 'clip_ratio/region_mean': 0.18222868908196688, 'reward_total_mean': 0.57071852684021, 'reward_meter_mean': 0.838412344455719, 'reward_meter_std': 0.3349061608314514, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9712759256362915, 'reward_repeat_soft_std': 0.0544203482568264, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.57071852684021, 'reward_total_composite_std': 0.35998836159706116, 'epoch': 0.06}

 18%|█▊        | 581/3300 [1:27:42<5:12:58,  6.91s/it]INFO 04-13 08:48:40 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 582/3300 [1:27:48<5:01:56,  6.67s/it]2026-04-13 08:48:46,590 | INFO | train_grpo_train | metrics_logged mode=train step=582

                                                      
{'loss': 0.1068, 'grad_norm': 25.380718231201172, 'learning_rate': 8.23939393939394e-06, 'num_tokens': 1035309.0, 'completions/mean_length': 48.5, 'completions/min_length': 46.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7526752352714539, 'rewards/meter/std': 0.32828620076179504, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8937437534332275, 'rewards/repeat_soft/std': 0.08044855296611786, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.611682653427124, 'rewards/total_composite/std': 0.1384030431509018, 'reward': 0.611682653427124, 'reward_std': 0.1384030282497406, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18637044727802277, 'sampling/sampling_logp_difference/max': 1.5626764297485352, 'sampling/importance_sampling_ratio/min': 0.20957441627979279, 'sampling/importance_sampling_ratio/mean': 1.0096079111099243, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1540735960006714, 'clip_ratio/low_mean': 0.08840708434581757, 'clip_ratio/low_min': 0.08840708434581757, 'clip_ratio/high_mean': 0.06679530814290047, 'clip_ratio/high_max': 0.06679530814290047, 'clip_ratio/region_mean': 0.15520239248871803, 'reward_total_mean': 0.611682653427124, 'reward_meter_mean': 0.7526752352714539, 'reward_meter_std': 0.32828620076179504, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8937437534332275, 'reward_repeat_soft_std': 0.08044855296611786, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.611682653427124, 'reward_total_composite_std': 0.1384030431509018, 'epoch': 0.06}

 18%|█▊        | 582/3300 [1:27:48<5:01:56,  6.67s/it]INFO 04-13 08:48:46 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 583/3300 [1:27:54<4:52:00,  6.45s/it]2026-04-13 08:48:52,531 | INFO | train_grpo_train | metrics_logged mode=train step=583

                                                      
{'loss': 0.016, 'grad_norm': 20.738035202026367, 'learning_rate': 8.236363636363637e-06, 'num_tokens': 1036800.0, 'completions/mean_length': 28.375, 'completions/min_length': 23.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.375, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.8921211957931519, 'rewards/meter/std': 0.17309178411960602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.933234691619873, 'rewards/repeat_soft/std': 0.04354261979460716, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.22937417030334473, 'rewards/total_composite/mean': 0.6594759821891785, 'rewards/total_composite/std': 0.14793506264686584, 'reward': 0.6594759821891785, 'reward_std': 0.14793504774570465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20247289538383484, 'sampling/sampling_logp_difference/max': 1.9466551542282104, 'sampling/importance_sampling_ratio/min': 0.14275075495243073, 'sampling/importance_sampling_ratio/mean': 0.9954608678817749, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0263854786753654, 'clip_ratio/low_mean': 0.12533949594944715, 'clip_ratio/low_min': 0.12533949594944715, 'clip_ratio/high_mean': 0.03525641094893217, 'clip_ratio/high_max': 0.03525641094893217, 'clip_ratio/region_mean': 0.16059590689837933, 'reward_total_mean': 0.6594759821891785, 'reward_meter_mean': 0.8921211957931519, 'reward_meter_std': 0.17309178411960602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.933234691619873, 'reward_repeat_soft_std': 0.04354261979460716, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.22937417030334473, 'reward_total_composite_mean': 0.6594759821891785, 'reward_total_composite_std': 0.14793506264686584, 'epoch': 0.06}

 18%|█▊        | 583/3300 [1:27:54<4:52:00,  6.45s/it]INFO 04-13 08:48:52 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 584/3300 [1:28:06<6:03:01,  8.02s/it]2026-04-13 08:49:04,216 | INFO | train_grpo_train | metrics_logged mode=train step=584

                                                      
{'loss': -0.1672, 'grad_norm': 3.1960246562957764, 'learning_rate': 8.233333333333335e-06, 'num_tokens': 1038841.0, 'completions/mean_length': 139.125, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 85.85714721679688, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.7035621404647827, 'rewards/meter/std': 0.4134138822555542, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.2828427255153656, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8031180500984192, 'rewards/repeat_soft/std': 0.1281014382839203, 'rewards/judge_quality/mean': 0.3687499761581421, 'rewards/judge_quality/std': 0.1940866857767105, 'rewards/total_composite/mean': 0.4458889067173004, 'rewards/total_composite/std': 0.20599308609962463, 'reward': 0.4458889067173004, 'reward_std': 0.20599307119846344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15093861520290375, 'sampling/sampling_logp_difference/max': 2.0359926223754883, 'sampling/importance_sampling_ratio/min': 0.1305508315563202, 'sampling/importance_sampling_ratio/mean': 1.0371010303497314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3707344383001328, 'clip_ratio/low_mean': 0.04888310004025698, 'clip_ratio/low_min': 0.04888310004025698, 'clip_ratio/high_mean': 0.08071568608283997, 'clip_ratio/high_max': 0.08071568608283997, 'clip_ratio/region_mean': 0.12959878612309694, 'reward_total_mean': 0.4458889067173004, 'reward_meter_mean': 0.7035621404647827, 'reward_meter_std': 0.4134138822555542, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.2828427255153656, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8031180500984192, 'reward_repeat_soft_std': 0.1281014382839203, 'reward_judge_quality_mean': 0.3687499761581421, 'reward_judge_quality_std': 0.1940866857767105, 'reward_total_composite_mean': 0.4458889067173004, 'reward_total_composite_std': 0.20599308609962463, 'epoch': 0.06}

 18%|█▊        | 584/3300 [1:28:06<6:03:01,  8.02s/it]INFO 04-13 08:49:04 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 585/3300 [1:28:12<5:37:08,  7.45s/it]2026-04-13 08:49:10,339 | INFO | train_grpo_train | metrics_logged mode=train step=585

                                                      
{'loss': -0.0034, 'grad_norm': 24.6851806640625, 'learning_rate': 8.23030303030303e-06, 'num_tokens': 1040186.0, 'completions/mean_length': 17.125, 'completions/min_length': 15.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.125, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.4121133089065552, 'rewards/meter/std': 0.42796626687049866, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4591677784919739, 'rewards/total_composite/std': 0.11875561624765396, 'reward': 0.4591677784919739, 'reward_std': 0.11875561624765396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09549973160028458, 'sampling/sampling_logp_difference/max': 0.8919789791107178, 'sampling/importance_sampling_ratio/min': 0.40984389185905457, 'sampling/importance_sampling_ratio/mean': 1.0326519012451172, 'sampling/importance_sampling_ratio/max': 1.9693711996078491, 'entropy': 0.662678062915802, 'clip_ratio/low_mean': 0.014756944496184587, 'clip_ratio/low_min': 0.014756944496184587, 'clip_ratio/high_mean': 0.0350694446824491, 'clip_ratio/high_max': 0.0350694446824491, 'clip_ratio/region_mean': 0.04982638917863369, 'reward_total_mean': 0.4591677784919739, 'reward_meter_mean': 0.4121133089065552, 'reward_meter_std': 0.42796626687049866, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4591677784919739, 'reward_total_composite_std': 0.11875561624765396, 'epoch': 0.06}

 18%|█▊        | 585/3300 [1:28:12<5:37:08,  7.45s/it]INFO 04-13 08:49:10 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 586/3300 [1:28:24<6:36:15,  8.76s/it]2026-04-13 08:49:22,157 | INFO | train_grpo_train | metrics_logged mode=train step=586

                                                      
{'loss': -0.0906, 'grad_norm': 5.4155659675598145, 'learning_rate': 8.227272727272728e-06, 'num_tokens': 1041915.0, 'completions/mean_length': 119.125, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.000003814697266, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5183769464492798, 'rewards/meter/std': 0.3440532386302948, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9439934492111206, 'rewards/repeat_soft/std': 0.08475925028324127, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.27102649211883545, 'rewards/total_composite/mean': 0.43112117052078247, 'rewards/total_composite/std': 0.28696581721305847, 'reward': 0.43112117052078247, 'reward_std': 0.28696581721305847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19397583603858948, 'sampling/sampling_logp_difference/max': 1.7971467971801758, 'sampling/importance_sampling_ratio/min': 0.16577118635177612, 'sampling/importance_sampling_ratio/mean': 1.0238008499145508, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4494723081588745, 'clip_ratio/low_mean': 0.03227124270051718, 'clip_ratio/low_min': 0.03227124270051718, 'clip_ratio/high_mean': 0.12011519446969032, 'clip_ratio/high_max': 0.12011519446969032, 'clip_ratio/region_mean': 0.1523864371702075, 'reward_total_mean': 0.43112117052078247, 'reward_meter_mean': 0.5183769464492798, 'reward_meter_std': 0.3440532386302948, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9439934492111206, 'reward_repeat_soft_std': 0.08475925028324127, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.27102649211883545, 'reward_total_composite_mean': 0.43112117052078247, 'reward_total_composite_std': 0.28696581721305847, 'epoch': 0.06}

 18%|█▊        | 586/3300 [1:28:24<6:36:15,  8.76s/it]INFO 04-13 08:49:22 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 587/3300 [1:28:30<6:01:28,  7.99s/it]2026-04-13 08:49:28,365 | INFO | train_grpo_train | metrics_logged mode=train step=587

                                                      
{'loss': 0.0502, 'grad_norm': 22.71309471130371, 'learning_rate': 8.224242424242425e-06, 'num_tokens': 1043235.0, 'completions/mean_length': 16.0, 'completions/min_length': 14.0, 'completions/max_length': 19.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.0, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 19.0, 'rewards/meter/mean': 0.7268494367599487, 'rewards/meter/std': 0.36402255296707153, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.39249998331069946, 'rewards/judge_quality/std': 0.08892211318016052, 'rewards/total_composite/mean': 0.4814658463001251, 'rewards/total_composite/std': 0.2159145176410675, 'reward': 0.4814658463001251, 'reward_std': 0.2159145176410675, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18526047468185425, 'sampling/sampling_logp_difference/max': 1.2848373651504517, 'sampling/importance_sampling_ratio/min': 0.27669557929039, 'sampling/importance_sampling_ratio/mean': 1.00920832157135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5635647252202034, 'clip_ratio/low_mean': 0.05330882407724857, 'clip_ratio/low_min': 0.05330882407724857, 'clip_ratio/high_mean': 0.08062988566234708, 'clip_ratio/high_max': 0.08062988566234708, 'clip_ratio/region_mean': 0.13393870973959565, 'reward_total_mean': 0.4814658463001251, 'reward_meter_mean': 0.7268494367599487, 'reward_meter_std': 0.36402255296707153, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.39249998331069946, 'reward_judge_quality_std': 0.08892211318016052, 'reward_total_composite_mean': 0.4814658463001251, 'reward_total_composite_std': 0.2159145176410675, 'epoch': 0.06}

 18%|█▊        | 587/3300 [1:28:30<6:01:28,  7.99s/it]INFO 04-13 08:49:28 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 588/3300 [1:28:37<5:42:59,  7.59s/it]2026-04-13 08:49:35,005 | INFO | train_grpo_train | metrics_logged mode=train step=588

                                                      
{'loss': 0.0346, 'grad_norm': 16.27345085144043, 'learning_rate': 8.221212121212122e-06, 'num_tokens': 1044755.0, 'completions/mean_length': 29.0, 'completions/min_length': 24.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.36250847578048706, 'rewards/meter/std': 0.3368504047393799, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9591778516769409, 'rewards/repeat_soft/std': 0.037539176642894745, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/total_composite/mean': 0.4934357702732086, 'rewards/total_composite/std': 0.18997235596179962, 'reward': 0.4934357702732086, 'reward_std': 0.18997234106063843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19533571600914001, 'sampling/sampling_logp_difference/max': 1.8965137004852295, 'sampling/importance_sampling_ratio/min': 0.15009097754955292, 'sampling/importance_sampling_ratio/mean': 1.0173364877700806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.109391801059246, 'clip_ratio/low_mean': 0.09431891050189734, 'clip_ratio/low_min': 0.09431891050189734, 'clip_ratio/high_mean': 0.0604330450296402, 'clip_ratio/high_max': 0.0604330450296402, 'clip_ratio/region_mean': 0.15475195553153753, 'reward_total_mean': 0.4934357702732086, 'reward_meter_mean': 0.36250847578048706, 'reward_meter_std': 0.3368504047393799, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9591778516769409, 'reward_repeat_soft_std': 0.037539176642894745, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.4934357702732086, 'reward_total_composite_std': 0.18997235596179962, 'epoch': 0.06}

 18%|█▊        | 588/3300 [1:28:37<5:42:59,  7.59s/it]INFO 04-13 08:49:35 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 589/3300 [1:28:43<5:23:16,  7.15s/it]2026-04-13 08:49:41,148 | INFO | train_grpo_train | metrics_logged mode=train step=589

                                                      
{'loss': -0.0357, 'grad_norm': 15.193470001220703, 'learning_rate': 8.21818181818182e-06, 'num_tokens': 1046080.0, 'completions/mean_length': 18.625, 'completions/min_length': 16.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.775652289390564, 'rewards/meter/std': 0.3668343424797058, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5973180532455444, 'rewards/total_composite/std': 0.15951067209243774, 'reward': 0.5973180532455444, 'reward_std': 0.15951068699359894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10052339732646942, 'sampling/sampling_logp_difference/max': 1.0645933151245117, 'sampling/importance_sampling_ratio/min': 0.34486809372901917, 'sampling/importance_sampling_ratio/mean': 1.0147958993911743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6124249398708344, 'clip_ratio/low_mean': 0.020955882500857115, 'clip_ratio/low_min': 0.020955882500857115, 'clip_ratio/high_mean': 0.051119307056069374, 'clip_ratio/high_max': 0.051119307056069374, 'clip_ratio/region_mean': 0.07207518955692649, 'reward_total_mean': 0.5973180532455444, 'reward_meter_mean': 0.775652289390564, 'reward_meter_std': 0.3668343424797058, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5973180532455444, 'reward_total_composite_std': 0.15951067209243774, 'epoch': 0.06}

 18%|█▊        | 589/3300 [1:28:43<5:23:16,  7.15s/it]INFO 04-13 08:49:41 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 590/3300 [1:28:49<5:12:19,  6.92s/it]2026-04-13 08:49:47,503 | INFO | train_grpo_train | metrics_logged mode=train step=590

                                                      
{'loss': 0.0741, 'grad_norm': 25.138277053833008, 'learning_rate': 8.215151515151517e-06, 'num_tokens': 1047456.0, 'completions/mean_length': 19.0, 'completions/min_length': 16.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9850324392318726, 'rewards/meter/std': 0.014376808889210224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8506903648376465, 'rewards/repeat_soft/std': 0.13034997880458832, 'rewards/judge_quality/mean': 0.5862500667572021, 'rewards/judge_quality/std': 0.2822834253311157, 'rewards/total_composite/mean': 0.7037005424499512, 'rewards/total_composite/std': 0.19040586054325104, 'reward': 0.7037005424499512, 'reward_std': 0.19040584564208984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19414843618869781, 'sampling/sampling_logp_difference/max': 1.3941915035247803, 'sampling/importance_sampling_ratio/min': 0.24981577694416046, 'sampling/importance_sampling_ratio/mean': 1.0324383974075317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6125606670975685, 'clip_ratio/low_mean': 0.10060129128396511, 'clip_ratio/low_min': 0.10060129128396511, 'clip_ratio/high_mean': 0.05801083752885461, 'clip_ratio/high_max': 0.05801083752885461, 'clip_ratio/region_mean': 0.15861212881281972, 'reward_total_mean': 0.7037005424499512, 'reward_meter_mean': 0.9850324392318726, 'reward_meter_std': 0.014376808889210224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8506903648376465, 'reward_repeat_soft_std': 0.13034997880458832, 'reward_judge_quality_mean': 0.5862500667572021, 'reward_judge_quality_std': 0.2822834253311157, 'reward_total_composite_mean': 0.7037005424499512, 'reward_total_composite_std': 0.19040586054325104, 'epoch': 0.06}

 18%|█▊        | 590/3300 [1:28:49<5:12:19,  6.92s/it]INFO 04-13 08:49:47 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 591/3300 [1:28:56<5:17:10,  7.02s/it]2026-04-13 08:49:54,785 | INFO | train_grpo_train | metrics_logged mode=train step=591

                                                      
{'loss': 0.0241, 'grad_norm': 22.737226486206055, 'learning_rate': 8.212121212121212e-06, 'num_tokens': 1049289.0, 'completions/mean_length': 51.125, 'completions/min_length': 47.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8187070488929749, 'rewards/meter/std': 0.1686120331287384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9087092876434326, 'rewards/repeat_soft/std': 0.0625695139169693, 'rewards/judge_quality/mean': 0.7074999809265137, 'rewards/judge_quality/std': 0.2474873960018158, 'rewards/total_composite/mean': 0.7206875085830688, 'rewards/total_composite/std': 0.17156639695167542, 'reward': 0.7206875085830688, 'reward_std': 0.17156638205051422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1752532571554184, 'sampling/sampling_logp_difference/max': 1.7380281686782837, 'sampling/importance_sampling_ratio/min': 0.17586684226989746, 'sampling/importance_sampling_ratio/mean': 1.0200912952423096, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2432586774230003, 'clip_ratio/low_mean': 0.07388498820364475, 'clip_ratio/low_min': 0.07388498820364475, 'clip_ratio/high_mean': 0.06151804141700268, 'clip_ratio/high_max': 0.06151804141700268, 'clip_ratio/region_mean': 0.13540302962064743, 'reward_total_mean': 0.7206875085830688, 'reward_meter_mean': 0.8187070488929749, 'reward_meter_std': 0.1686120331287384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9087092876434326, 'reward_repeat_soft_std': 0.0625695139169693, 'reward_judge_quality_mean': 0.7074999809265137, 'reward_judge_quality_std': 0.2474873960018158, 'reward_total_composite_mean': 0.7206875085830688, 'reward_total_composite_std': 0.17156639695167542, 'epoch': 0.06}

 18%|█▊        | 591/3300 [1:28:57<5:17:10,  7.02s/it]INFO 04-13 08:49:55 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 592/3300 [1:29:08<6:20:48,  8.44s/it]2026-04-13 08:50:06,516 | INFO | train_grpo_train | metrics_logged mode=train step=592

                                                      
{'loss': -0.149, 'grad_norm': 3.940671443939209, 'learning_rate': 8.20909090909091e-06, 'num_tokens': 1051196.0, 'completions/mean_length': 125.375, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.14286041259766, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8566869497299194, 'rewards/meter/std': 0.31202420592308044, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.864543080329895, 'rewards/repeat_soft/std': 0.11763402819633484, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5127476453781128, 'rewards/total_composite/std': 0.2380359172821045, 'reward': 0.5127476453781128, 'reward_std': 0.2380359172821045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16044767200946808, 'sampling/sampling_logp_difference/max': 1.304091453552246, 'sampling/importance_sampling_ratio/min': 0.2714190185070038, 'sampling/importance_sampling_ratio/mean': 1.037603735923767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3022792339324951, 'clip_ratio/low_mean': 0.02142857201397419, 'clip_ratio/low_min': 0.02142857201397419, 'clip_ratio/high_mean': 0.1300391824916005, 'clip_ratio/high_max': 0.1300391824916005, 'clip_ratio/region_mean': 0.1514677545055747, 'reward_total_mean': 0.5127476453781128, 'reward_meter_mean': 0.8566869497299194, 'reward_meter_std': 0.31202420592308044, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.864543080329895, 'reward_repeat_soft_std': 0.11763402819633484, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5127476453781128, 'reward_total_composite_std': 0.2380359172821045, 'epoch': 0.06}

 18%|█▊        | 592/3300 [1:29:08<6:20:48,  8.44s/it]INFO 04-13 08:50:06 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 593/3300 [1:29:15<6:01:19,  8.01s/it]2026-04-13 08:50:13,544 | INFO | train_grpo_train | metrics_logged mode=train step=593

                                                      
{'loss': 0.0705, 'grad_norm': 24.125486373901367, 'learning_rate': 8.206060606060607e-06, 'num_tokens': 1052987.0, 'completions/mean_length': 62.875, 'completions/min_length': 55.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.5702065229415894, 'rewards/meter/std': 0.37652307748794556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9546473026275635, 'rewards/repeat_soft/std': 0.052884023636579514, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.5126515030860901, 'rewards/total_composite/std': 0.09028740972280502, 'reward': 0.5126515030860901, 'reward_std': 0.09028739482164383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19030006229877472, 'sampling/sampling_logp_difference/max': 1.6490859985351562, 'sampling/importance_sampling_ratio/min': 0.19222553074359894, 'sampling/importance_sampling_ratio/mean': 1.0347181558609009, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7659383863210678, 'clip_ratio/low_mean': 0.09444891475141048, 'clip_ratio/low_min': 0.09444891475141048, 'clip_ratio/high_mean': 0.08395612798631191, 'clip_ratio/high_max': 0.08395612798631191, 'clip_ratio/region_mean': 0.1784050427377224, 'reward_total_mean': 0.5126515030860901, 'reward_meter_mean': 0.5702065229415894, 'reward_meter_std': 0.37652307748794556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9546473026275635, 'reward_repeat_soft_std': 0.052884023636579514, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.5126515030860901, 'reward_total_composite_std': 0.09028740972280502, 'epoch': 0.06}

 18%|█▊        | 593/3300 [1:29:15<6:01:19,  8.01s/it]INFO 04-13 08:50:13 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 594/3300 [1:29:22<5:48:13,  7.72s/it]2026-04-13 08:50:20,578 | INFO | train_grpo_train | metrics_logged mode=train step=594

                                                      
{'loss': 0.0226, 'grad_norm': 24.708345413208008, 'learning_rate': 8.203030303030304e-06, 'num_tokens': 1054725.0, 'completions/mean_length': 54.25, 'completions/min_length': 47.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.25, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.5746891498565674, 'rewards/meter/std': 0.33918896317481995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9928379058837891, 'rewards/repeat_soft/std': 0.0032845570240169764, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/total_composite/mean': 0.5154507160186768, 'rewards/total_composite/std': 0.26690253615379333, 'reward': 0.5154507160186768, 'reward_std': 0.26690250635147095, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20852459967136383, 'sampling/sampling_logp_difference/max': 1.890106201171875, 'sampling/importance_sampling_ratio/min': 0.15105575323104858, 'sampling/importance_sampling_ratio/mean': 1.010902762413025, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0982920974493027, 'clip_ratio/low_mean': 0.09287556447088718, 'clip_ratio/low_min': 0.09287556447088718, 'clip_ratio/high_mean': 0.09945436753332615, 'clip_ratio/high_max': 0.09945436753332615, 'clip_ratio/region_mean': 0.19232993200421333, 'reward_total_mean': 0.5154507160186768, 'reward_meter_mean': 0.5746891498565674, 'reward_meter_std': 0.33918896317481995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9928379058837891, 'reward_repeat_soft_std': 0.0032845570240169764, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.5154507160186768, 'reward_total_composite_std': 0.26690253615379333, 'epoch': 0.06}

 18%|█▊        | 594/3300 [1:29:22<5:48:13,  7.72s/it]INFO 04-13 08:50:20 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 595/3300 [1:29:28<5:19:44,  7.09s/it]2026-04-13 08:50:26,203 | INFO | train_grpo_train | metrics_logged mode=train step=595

                                                      
{'loss': 0.0578, 'grad_norm': 20.776994705200195, 'learning_rate': 8.2e-06, 'num_tokens': 1056083.0, 'completions/mean_length': 19.75, 'completions/min_length': 19.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7944017648696899, 'rewards/meter/std': 0.340792715549469, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.8193592429161072, 'rewards/total_composite/std': 0.20397496223449707, 'reward': 0.8193592429161072, 'reward_std': 0.20397497713565826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15076832473278046, 'sampling/sampling_logp_difference/max': 1.104593276977539, 'sampling/importance_sampling_ratio/min': 0.41009634733200073, 'sampling/importance_sampling_ratio/mean': 1.0511397123336792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1086817681789398, 'clip_ratio/low_mean': 0.053977273404598236, 'clip_ratio/low_min': 0.053977273404598236, 'clip_ratio/high_mean': 0.08421052573248744, 'clip_ratio/high_max': 0.08421052573248744, 'clip_ratio/region_mean': 0.13818779913708568, 'reward_total_mean': 0.8193592429161072, 'reward_meter_mean': 0.7944017648696899, 'reward_meter_std': 0.340792715549469, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.8193592429161072, 'reward_total_composite_std': 0.20397496223449707, 'epoch': 0.06}

 18%|█▊        | 595/3300 [1:29:28<5:19:44,  7.09s/it]INFO 04-13 08:50:26 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 596/3300 [1:29:34<5:10:29,  6.89s/it]2026-04-13 08:50:32,617 | INFO | train_grpo_train | metrics_logged mode=train step=596

                                                      
{'loss': 0.0182, 'grad_norm': 16.731706619262695, 'learning_rate': 8.196969696969698e-06, 'num_tokens': 1057927.0, 'completions/mean_length': 55.5, 'completions/min_length': 47.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7734062671661377, 'rewards/meter/std': 0.2578055262565613, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9969630837440491, 'rewards/repeat_soft/std': 0.002231556223705411, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.47790324687957764, 'rewards/total_composite/std': 0.20325395464897156, 'reward': 0.47790324687957764, 'reward_std': 0.20325393974781036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19447480142116547, 'sampling/sampling_logp_difference/max': 1.691053867340088, 'sampling/importance_sampling_ratio/min': 0.1843251734972, 'sampling/importance_sampling_ratio/mean': 1.010617971420288, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2968737334012985, 'clip_ratio/low_mean': 0.054579597897827625, 'clip_ratio/low_min': 0.054579597897827625, 'clip_ratio/high_mean': 0.14666733145713806, 'clip_ratio/high_max': 0.14666733145713806, 'clip_ratio/region_mean': 0.2012469293549657, 'reward_total_mean': 0.47790324687957764, 'reward_meter_mean': 0.7734062671661377, 'reward_meter_std': 0.2578055262565613, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9969630837440491, 'reward_repeat_soft_std': 0.002231556223705411, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.47790324687957764, 'reward_total_composite_std': 0.20325395464897156, 'epoch': 0.06}

 18%|█▊        | 596/3300 [1:29:34<5:10:29,  6.89s/it]INFO 04-13 08:50:32 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 597/3300 [1:29:45<6:06:50,  8.14s/it]2026-04-13 08:50:43,685 | INFO | train_grpo_train | metrics_logged mode=train step=597

                                                      
{'loss': -0.1452, 'grad_norm': 2.848376750946045, 'learning_rate': 8.193939393939394e-06, 'num_tokens': 1059752.0, 'completions/mean_length': 121.125, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.28572082519531, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7684168815612793, 'rewards/meter/std': 0.3329595625400543, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8443343043327332, 'rewards/repeat_soft/std': 0.1572190523147583, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.24483230710029602, 'rewards/total_composite/mean': 0.5488306879997253, 'rewards/total_composite/std': 0.27638494968414307, 'reward': 0.5488306879997253, 'reward_std': 0.27638494968414307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14185303449630737, 'sampling/sampling_logp_difference/max': 1.3351480960845947, 'sampling/importance_sampling_ratio/min': 0.2631192207336426, 'sampling/importance_sampling_ratio/mean': 1.0276942253112793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9028343670070171, 'clip_ratio/low_mean': 0.0386904776096344, 'clip_ratio/low_min': 0.0386904776096344, 'clip_ratio/high_mean': 0.07202116353437304, 'clip_ratio/high_max': 0.07202116353437304, 'clip_ratio/region_mean': 0.11071164114400744, 'reward_total_mean': 0.5488306879997253, 'reward_meter_mean': 0.7684168815612793, 'reward_meter_std': 0.3329595625400543, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8443343043327332, 'reward_repeat_soft_std': 0.1572190523147583, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.24483230710029602, 'reward_total_composite_mean': 0.5488306879997253, 'reward_total_composite_std': 0.27638494968414307, 'epoch': 0.06}

 18%|█▊        | 597/3300 [1:29:45<6:06:50,  8.14s/it]INFO 04-13 08:50:43 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 598/3300 [1:29:51<5:38:36,  7.52s/it]2026-04-13 08:50:49,750 | INFO | train_grpo_train | metrics_logged mode=train step=598

                                                      
{'loss': 0.0874, 'grad_norm': 15.479145050048828, 'learning_rate': 8.190909090909091e-06, 'num_tokens': 1061264.0, 'completions/mean_length': 33.0, 'completions/min_length': 29.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6340094208717346, 'rewards/meter/std': 0.44285714626312256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9069277048110962, 'rewards/repeat_soft/std': 0.10584849864244461, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5099337100982666, 'rewards/total_composite/std': 0.11327037215232849, 'reward': 0.5099337100982666, 'reward_std': 0.1132703647017479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19748561084270477, 'sampling/sampling_logp_difference/max': 1.3864812850952148, 'sampling/importance_sampling_ratio/min': 0.2499532848596573, 'sampling/importance_sampling_ratio/mean': 1.0355310440063477, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6433321237564087, 'clip_ratio/low_mean': 0.10831448249518871, 'clip_ratio/low_min': 0.10831448249518871, 'clip_ratio/high_mean': 0.09299219027161598, 'clip_ratio/high_max': 0.09299219027161598, 'clip_ratio/region_mean': 0.2013066727668047, 'reward_total_mean': 0.5099337100982666, 'reward_meter_mean': 0.6340094208717346, 'reward_meter_std': 0.44285714626312256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9069277048110962, 'reward_repeat_soft_std': 0.10584849864244461, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5099337100982666, 'reward_total_composite_std': 0.11327037215232849, 'epoch': 0.06}

 18%|█▊        | 598/3300 [1:29:51<5:38:36,  7.52s/it]INFO 04-13 08:50:49 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 599/3300 [1:30:03<6:32:57,  8.73s/it]2026-04-13 08:51:01,302 | INFO | train_grpo_train | metrics_logged mode=train step=599

                                                      
{'loss': -0.1053, 'grad_norm': 2.855733871459961, 'learning_rate': 8.187878787878788e-06, 'num_tokens': 1062902.0, 'completions/mean_length': 98.75, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.71428680419922, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5551939010620117, 'rewards/meter/std': 0.4094555675983429, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.972247302532196, 'rewards/repeat_soft/std': 0.04996117576956749, 'rewards/judge_quality/mean': 0.42250001430511475, 'rewards/judge_quality/std': 0.24423348903656006, 'rewards/total_composite/mean': 0.44813311100006104, 'rewards/total_composite/std': 0.20157425105571747, 'reward': 0.44813311100006104, 'reward_std': 0.20157423615455627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1727154403924942, 'sampling/sampling_logp_difference/max': 1.759019374847412, 'sampling/importance_sampling_ratio/min': 0.17221365869045258, 'sampling/importance_sampling_ratio/mean': 1.0211820602416992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3214404955506325, 'clip_ratio/low_mean': 0.02302631549537182, 'clip_ratio/low_min': 0.02302631549537182, 'clip_ratio/high_mean': 0.12357525061815977, 'clip_ratio/high_max': 0.12357525061815977, 'clip_ratio/region_mean': 0.1466015661135316, 'reward_total_mean': 0.44813311100006104, 'reward_meter_mean': 0.5551939010620117, 'reward_meter_std': 0.4094555675983429, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.972247302532196, 'reward_repeat_soft_std': 0.04996117576956749, 'reward_judge_quality_mean': 0.42250001430511475, 'reward_judge_quality_std': 0.24423348903656006, 'reward_total_composite_mean': 0.44813311100006104, 'reward_total_composite_std': 0.20157425105571747, 'epoch': 0.06}

 18%|█▊        | 599/3300 [1:30:03<6:32:57,  8.73s/it]INFO 04-13 08:51:01 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 600/3300 [1:30:09<5:52:56,  7.84s/it]2026-04-13 08:51:07,079 | INFO | train_grpo_train | metrics_logged mode=train step=600

                                                      
{'loss': 0.1465, 'grad_norm': 15.054155349731445, 'learning_rate': 8.184848484848486e-06, 'num_tokens': 1064284.0, 'completions/mean_length': 23.75, 'completions/min_length': 22.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.950312077999115, 'rewards/meter/std': 0.06284411251544952, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9781278371810913, 'rewards/repeat_soft/std': 0.014061374589800835, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6061543822288513, 'rewards/total_composite/std': 0.018617939203977585, 'reward': 0.6061543822288513, 'reward_std': 0.018617935478687286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06382521986961365, 'sampling/sampling_logp_difference/max': 1.6088030338287354, 'sampling/importance_sampling_ratio/min': 0.2001270055770874, 'sampling/importance_sampling_ratio/mean': 1.0066806077957153, 'sampling/importance_sampling_ratio/max': 1.8192110061645508, 'entropy': 0.23449560813605785, 'clip_ratio/low_mean': 0.018939394503831863, 'clip_ratio/low_min': 0.018939394503831863, 'clip_ratio/high_mean': 0.0437870561145246, 'clip_ratio/high_max': 0.0437870561145246, 'clip_ratio/region_mean': 0.06272645061835647, 'reward_total_mean': 0.6061543822288513, 'reward_meter_mean': 0.950312077999115, 'reward_meter_std': 0.06284411251544952, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9781278371810913, 'reward_repeat_soft_std': 0.014061374589800835, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6061543822288513, 'reward_total_composite_std': 0.018617939203977585, 'epoch': 0.06}

 18%|█▊        | 600/3300 [1:30:09<5:52:56,  7.84s/it]INFO 04-13 08:51:07 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.79s/it]

 30%|███       | 3/10 [00:12<00:32,  4.63s/it]

 40%|████      | 4/10 [00:15<00:24,  4.06s/it]

 50%|█████     | 5/10 [00:19<00:20,  4.04s/it]

 60%|██████    | 6/10 [00:28<00:22,  5.74s/it]

 70%|███████   | 7/10 [00:37<00:20,  6.72s/it]

 80%|████████  | 8/10 [00:40<00:11,  5.69s/it]

 90%|█████████ | 9/10 [00:44<00:05,  5.10s/it]

100%|██████████| 10/10 [00:47<00:00,  4.48s/it]2026-04-13 08:52:03,947 | INFO | train_grpo_train | metrics_logged mode=eval step=600

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 56.8265, 'eval_samples_per_second': 1.408, 'eval_steps_per_second': 0.176, 'eval_num_tokens': 1064284.0, 'eval_completions/mean_length': 84.6, 'eval_completions/min_length': 30.9, 'eval_completions/max_length': 260.2, 'eval_completions/clipped_ratio': 0.0625, 'eval_completions/mean_terminated_length': 56.774406051635744, 'eval_completions/min_terminated_length': 30.9, 'eval_completions/max_terminated_length': 97.4, 'eval_rewards/meter/mean': 0.5933304131031036, 'eval_rewards/meter/std': 0.3523423582315445, 'eval_rewards/count_adherence/mean': 0.9745833277702332, 'eval_rewards/count_adherence/std': 0.07188918963074684, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.19864802658557892, 'eval_rewards/repeat_soft/mean': 0.9164949417114258, 'eval_rewards/repeat_soft/std': 0.11432811766862869, 'eval_rewards/judge_quality/mean': 0.5162499994039536, 'eval_rewards/judge_quality/std': 0.23027257323265077, 'eval_rewards/total_composite/mean': 0.5080294042825699, 'eval_rewards/total_composite/std': 0.20953540951013566, 'eval_reward': 0.5080294042825699, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.09997781440615654, 'eval_sampling/sampling_logp_difference/max': 0.9178894281387329, 'eval_sampling/importance_sampling_ratio/min': 0.403290730714798, 'eval_sampling/importance_sampling_ratio/mean': 1.0318517208099365, 'eval_sampling/importance_sampling_ratio/max': 1.5025453090667724, 'eval_entropy': 1.2897954106330871, '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': 0.5080294042825699, 'eval_reward_meter_mean': 0.5933304131031036, 'eval_reward_meter_std': 0.3523423582315445, 'eval_reward_count_adherence_mean': 0.9745833277702332, 'eval_reward_count_adherence_std': 0.07188918963074684, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.19864802658557892, 'eval_reward_repeat_soft_mean': 0.9164949417114258, 'eval_reward_repeat_soft_std': 0.11432811766862869, 'eval_reward_judge_quality_mean': 0.5162499994039536, 'eval_reward_judge_quality_std': 0.23027257323265077, 'eval_reward_total_composite_mean': 0.5080294042825699, 'eval_reward_total_composite_std': 0.20953540951013566, 'epoch': 0.06}

 18%|█▊        | 600/3300 [1:31:06<5:52:56,  7.84s/it]

100%|██████████| 10/10 [00:47<00:00,  4.48s/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(

 18%|█▊        | 601/3300 [1:31:15<18:59:46, 25.34s/it]2026-04-13 08:52:13,244 | INFO | train_grpo_train | metrics_logged mode=train step=601

                                                       
{'loss': 0.0111, 'grad_norm': 9.83389663696289, 'learning_rate': 8.181818181818183e-06, 'num_tokens': 1066702.0, 'completions/mean_length': 69.25, 'completions/min_length': 63.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.25, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9074545502662659, 'rewards/meter/std': 0.14883460104465485, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9192219972610474, 'rewards/repeat_soft/std': 0.0344962514936924, 'rewards/judge_quality/mean': 0.4987499713897705, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.6348594427108765, 'rewards/total_composite/std': 0.16552099585533142, 'reward': 0.6348594427108765, 'reward_std': 0.16552099585533142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1714843213558197, 'sampling/sampling_logp_difference/max': 1.6988840103149414, 'sampling/importance_sampling_ratio/min': 0.1828875094652176, 'sampling/importance_sampling_ratio/mean': 1.02165949344635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6118394806981087, 'clip_ratio/low_mean': 0.10102782025933266, 'clip_ratio/low_min': 0.10102782025933266, 'clip_ratio/high_mean': 0.040490858256816864, 'clip_ratio/high_max': 0.040490858256816864, 'clip_ratio/region_mean': 0.14151867851614952, 'reward_total_mean': 0.6348594427108765, 'reward_meter_mean': 0.9074545502662659, 'reward_meter_std': 0.14883460104465485, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9192219972610474, 'reward_repeat_soft_std': 0.0344962514936924, 'reward_judge_quality_mean': 0.4987499713897705, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.6348594427108765, 'reward_total_composite_std': 0.16552099585533142, 'epoch': 0.06}

 18%|█▊        | 601/3300 [1:31:15<18:59:46, 25.34s/it]INFO 04-13 08:52:13 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 602/3300 [1:31:23<15:02:41, 20.07s/it]2026-04-13 08:52:21,030 | INFO | train_grpo_train | metrics_logged mode=train step=602

                                                       
{'loss': 0.0697, 'grad_norm': 9.658855438232422, 'learning_rate': 8.17878787878788e-06, 'num_tokens': 1069606.0, 'completions/mean_length': 124.0, 'completions/min_length': 97.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.0, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.4799366593360901, 'rewards/meter/std': 0.302285373210907, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.918707549571991, 'rewards/repeat_soft/std': 0.09461511671543121, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.46457159519195557, 'rewards/total_composite/std': 0.0959501639008522, 'reward': 0.46457159519195557, 'reward_std': 0.0959501564502716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19479040801525116, 'sampling/sampling_logp_difference/max': 2.191174030303955, 'sampling/importance_sampling_ratio/min': 0.11178543418645859, 'sampling/importance_sampling_ratio/mean': 1.0257471799850464, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7584725245833397, 'clip_ratio/low_mean': 0.10992318764328957, 'clip_ratio/low_min': 0.10992318764328957, 'clip_ratio/high_mean': 0.04259421583265066, 'clip_ratio/high_max': 0.04259421583265066, 'clip_ratio/region_mean': 0.15251740347594023, 'reward_total_mean': 0.46457159519195557, 'reward_meter_mean': 0.4799366593360901, 'reward_meter_std': 0.302285373210907, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.918707549571991, 'reward_repeat_soft_std': 0.09461511671543121, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.46457159519195557, 'reward_total_composite_std': 0.0959501639008522, 'epoch': 0.06}

 18%|█▊        | 602/3300 [1:31:23<15:02:41, 20.07s/it]INFO 04-13 08:52:21 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 603/3300 [1:31:34<13:07:11, 17.51s/it]2026-04-13 08:52:32,569 | INFO | train_grpo_train | metrics_logged mode=train step=603

                                                       
{'loss': -0.2125, 'grad_norm': 2.0980961322784424, 'learning_rate': 8.175757575757577e-06, 'num_tokens': 1071893.0, 'completions/mean_length': 271.875, 'completions/min_length': 110.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 127.80000305175781, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 158.0, 'rewards/meter/mean': 0.726004421710968, 'rewards/meter/std': 0.292203813791275, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.8333075046539307, 'rewards/repeat_soft/std': 0.13805323839187622, 'rewards/judge_quality/mean': 0.22999998927116394, 'rewards/judge_quality/std': 0.1437259316444397, 'rewards/total_composite/mean': 0.2765088677406311, 'rewards/total_composite/std': 0.23182496428489685, 'reward': 0.2765088677406311, 'reward_std': 0.23182496428489685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13859088718891144, 'sampling/sampling_logp_difference/max': 1.422471046447754, 'sampling/importance_sampling_ratio/min': 0.2411174774169922, 'sampling/importance_sampling_ratio/mean': 1.0195611715316772, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7712185755372047, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0950105395168066, 'clip_ratio/high_max': 0.0950105395168066, 'clip_ratio/region_mean': 0.0950105395168066, 'reward_total_mean': 0.2765088677406311, 'reward_meter_mean': 0.726004421710968, 'reward_meter_std': 0.292203813791275, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.8333075046539307, 'reward_repeat_soft_std': 0.13805323839187622, 'reward_judge_quality_mean': 0.22999998927116394, 'reward_judge_quality_std': 0.1437259316444397, 'reward_total_composite_mean': 0.2765088677406311, 'reward_total_composite_std': 0.23182496428489685, 'epoch': 0.06}

 18%|█▊        | 603/3300 [1:31:34<13:07:11, 17.51s/it]INFO 04-13 08:52:32 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 604/3300 [1:31:45<11:39:02, 15.56s/it]2026-04-13 08:52:43,566 | INFO | train_grpo_train | metrics_logged mode=train step=604

                                                       
{'loss': -0.074, 'grad_norm': 1.7580572366714478, 'learning_rate': 8.172727272727273e-06, 'num_tokens': 1073313.0, 'completions/mean_length': 150.5, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8189277648925781, 'rewards/meter/std': 0.3430798351764679, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9567056894302368, 'rewards/repeat_soft/std': 0.05685315281152725, 'rewards/judge_quality/mean': 0.3137499988079071, 'rewards/judge_quality/std': 0.17492344975471497, 'rewards/total_composite/mean': 0.44049832224845886, 'rewards/total_composite/std': 0.2767917811870575, 'reward': 0.44049832224845886, 'reward_std': 0.2767917513847351, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18948166072368622, 'sampling/sampling_logp_difference/max': 1.9374151229858398, 'sampling/importance_sampling_ratio/min': 0.1440758854150772, 'sampling/importance_sampling_ratio/mean': 1.0077388286590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3183616027235985, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11339256633073092, 'clip_ratio/high_max': 0.11339256633073092, 'clip_ratio/region_mean': 0.11339256633073092, 'reward_total_mean': 0.44049832224845886, 'reward_meter_mean': 0.8189277648925781, 'reward_meter_std': 0.3430798351764679, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9567056894302368, 'reward_repeat_soft_std': 0.05685315281152725, 'reward_judge_quality_mean': 0.3137499988079071, 'reward_judge_quality_std': 0.17492344975471497, 'reward_total_composite_mean': 0.44049832224845886, 'reward_total_composite_std': 0.2767917811870575, 'epoch': 0.06}

 18%|█▊        | 604/3300 [1:31:45<11:39:02, 15.56s/it]INFO 04-13 08:52:43 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 605/3300 [1:31:51<9:28:28, 12.66s/it] 2026-04-13 08:52:49,449 | INFO | train_grpo_train | metrics_logged mode=train step=605

                                                      
{'loss': 0.0357, 'grad_norm': 15.67078685760498, 'learning_rate': 8.16969696969697e-06, 'num_tokens': 1074844.0, 'completions/mean_length': 30.375, 'completions/min_length': 26.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9137496948242188, 'rewards/meter/std': 0.09076227247714996, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8560148477554321, 'rewards/repeat_soft/std': 0.05153679475188255, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.622412919998169, 'rewards/total_composite/std': 0.12228536605834961, 'reward': 0.622412919998169, 'reward_std': 0.12228536605834961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17435169219970703, 'sampling/sampling_logp_difference/max': 2.398494005203247, 'sampling/importance_sampling_ratio/min': 0.09085467457771301, 'sampling/importance_sampling_ratio/mean': 1.010645866394043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.181313119828701, 'clip_ratio/low_mean': 0.11573188286274672, 'clip_ratio/low_min': 0.11573188286274672, 'clip_ratio/high_mean': 0.02083333395421505, 'clip_ratio/high_max': 0.02083333395421505, 'clip_ratio/region_mean': 0.13656521681696177, 'reward_total_mean': 0.622412919998169, 'reward_meter_mean': 0.9137496948242188, 'reward_meter_std': 0.09076227247714996, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8560148477554321, 'reward_repeat_soft_std': 0.05153679475188255, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.622412919998169, 'reward_total_composite_std': 0.12228536605834961, 'epoch': 0.06}

 18%|█▊        | 605/3300 [1:31:51<9:28:28, 12.66s/it]INFO 04-13 08:52:49 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 606/3300 [1:32:02<9:07:53, 12.20s/it]2026-04-13 08:53:00,592 | INFO | train_grpo_train | metrics_logged mode=train step=606

                                                      
{'loss': -0.084, 'grad_norm': 4.847084999084473, 'learning_rate': 8.166666666666668e-06, 'num_tokens': 1076268.0, 'completions/mean_length': 92.0, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.29814398288726807, 'rewards/meter/std': 0.35805410146713257, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9650765657424927, 'rewards/repeat_soft/std': 0.03833828493952751, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.37954479455947876, 'rewards/total_composite/std': 0.1818016916513443, 'reward': 0.37954479455947876, 'reward_std': 0.1818016916513443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1880868524312973, 'sampling/sampling_logp_difference/max': 1.379664421081543, 'sampling/importance_sampling_ratio/min': 0.2516629993915558, 'sampling/importance_sampling_ratio/mean': 1.0463781356811523, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.129436418414116, 'clip_ratio/low_mean': 0.08789344504475594, 'clip_ratio/low_min': 0.08789344504475594, 'clip_ratio/high_mean': 0.039434524485841393, 'clip_ratio/high_max': 0.039434524485841393, 'clip_ratio/region_mean': 0.12732796953059733, 'reward_total_mean': 0.37954479455947876, 'reward_meter_mean': 0.29814398288726807, 'reward_meter_std': 0.35805410146713257, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9650765657424927, 'reward_repeat_soft_std': 0.03833828493952751, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.37954479455947876, 'reward_total_composite_std': 0.1818016916513443, 'epoch': 0.06}

 18%|█▊        | 606/3300 [1:32:02<9:07:53, 12.20s/it]INFO 04-13 08:53:00 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 607/3300 [1:32:15<9:13:19, 12.33s/it]2026-04-13 08:53:13,213 | INFO | train_grpo_train | metrics_logged mode=train step=607

                                                      
{'loss': -0.1326, 'grad_norm': 2.402979612350464, 'learning_rate': 8.163636363636365e-06, 'num_tokens': 1077917.0, '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': 65.0, 'rewards/meter/mean': 0.6160660982131958, 'rewards/meter/std': 0.28426313400268555, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9823377132415771, 'rewards/repeat_soft/std': 0.02172774076461792, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.24991071224212646, 'rewards/total_composite/mean': 0.5163519382476807, 'rewards/total_composite/std': 0.21996428072452545, 'reward': 0.5163519382476807, 'reward_std': 0.21996428072452545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16341672837734222, 'sampling/sampling_logp_difference/max': 2.2435765266418457, 'sampling/importance_sampling_ratio/min': 0.10607843846082687, 'sampling/importance_sampling_ratio/mean': 1.002920389175415, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7133377343416214, 'clip_ratio/low_mean': 0.016509434208273888, 'clip_ratio/low_min': 0.016509434208273888, 'clip_ratio/high_mean': 0.11188606917858124, 'clip_ratio/high_max': 0.11188606917858124, 'clip_ratio/region_mean': 0.12839550338685513, 'reward_total_mean': 0.5163519382476807, 'reward_meter_mean': 0.6160660982131958, 'reward_meter_std': 0.28426313400268555, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9823377132415771, 'reward_repeat_soft_std': 0.02172774076461792, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.24991071224212646, 'reward_total_composite_mean': 0.5163519382476807, 'reward_total_composite_std': 0.21996428072452545, 'epoch': 0.06}

 18%|█▊        | 607/3300 [1:32:15<9:13:19, 12.33s/it]INFO 04-13 08:53:13 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 608/3300 [1:32:27<9:06:51, 12.19s/it]2026-04-13 08:53:25,077 | INFO | train_grpo_train | metrics_logged mode=train step=608

                                                      
{'loss': -0.0922, 'grad_norm': 1.6431916952133179, 'learning_rate': 8.16060606060606e-06, 'num_tokens': 1079286.0, 'completions/mean_length': 88.125, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 27.571430206298828, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.78997403383255, 'rewards/meter/std': 0.3272418975830078, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9565540552139282, 'rewards/repeat_soft/std': 0.04775778576731682, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.5152409076690674, 'rewards/total_composite/std': 0.20895171165466309, 'reward': 0.5152409076690674, 'reward_std': 0.20895171165466309, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1754993200302124, 'sampling/sampling_logp_difference/max': 2.666184663772583, 'sampling/importance_sampling_ratio/min': 0.06951694935560226, 'sampling/importance_sampling_ratio/mean': 1.0019959211349487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0098569318652153, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.17423375509679317, 'clip_ratio/high_max': 0.17423375509679317, 'clip_ratio/region_mean': 0.17423375509679317, 'reward_total_mean': 0.5152409076690674, 'reward_meter_mean': 0.78997403383255, 'reward_meter_std': 0.3272418975830078, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9565540552139282, 'reward_repeat_soft_std': 0.04775778576731682, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.5152409076690674, 'reward_total_composite_std': 0.20895171165466309, 'epoch': 0.06}

 18%|█▊        | 608/3300 [1:32:27<9:06:51, 12.19s/it]INFO 04-13 08:53:25 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 609/3300 [1:32:39<9:01:48, 12.08s/it]2026-04-13 08:53:36,905 | INFO | train_grpo_train | metrics_logged mode=train step=609

                                                      
{'loss': -0.0935, 'grad_norm': 3.18972110748291, 'learning_rate': 8.15757575757576e-06, 'num_tokens': 1080771.0, 'completions/mean_length': 93.625, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.85714340209961, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5269536972045898, 'rewards/meter/std': 0.34882891178131104, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9965314269065857, 'rewards/repeat_soft/std': 0.00570334866642952, 'rewards/judge_quality/mean': 0.3812500238418579, 'rewards/judge_quality/std': 0.13452960550785065, 'rewards/total_composite/mean': 0.4282902479171753, 'rewards/total_composite/std': 0.1965111643075943, 'reward': 0.4282902479171753, 'reward_std': 0.1965111643075943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18389181792736053, 'sampling/sampling_logp_difference/max': 1.4303655624389648, 'sampling/importance_sampling_ratio/min': 0.239221453666687, 'sampling/importance_sampling_ratio/mean': 1.0592796802520752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5589155703783035, 'clip_ratio/low_mean': 0.05129310209304094, 'clip_ratio/low_min': 0.05129310209304094, 'clip_ratio/high_mean': 0.09888686705380678, 'clip_ratio/high_max': 0.09888686705380678, 'clip_ratio/region_mean': 0.15017996914684772, 'reward_total_mean': 0.4282902479171753, 'reward_meter_mean': 0.5269536972045898, 'reward_meter_std': 0.34882891178131104, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9965314269065857, 'reward_repeat_soft_std': 0.00570334866642952, 'reward_judge_quality_mean': 0.3812500238418579, 'reward_judge_quality_std': 0.13452960550785065, 'reward_total_composite_mean': 0.4282902479171753, 'reward_total_composite_std': 0.1965111643075943, 'epoch': 0.06}

 18%|█▊        | 609/3300 [1:32:39<9:01:48, 12.08s/it]INFO 04-13 08:53:37 [block_pool.py:378] Successfully reset prefix cache

 18%|█▊        | 610/3300 [1:32:50<8:53:01, 11.89s/it]2026-04-13 08:53:48,350 | INFO | train_grpo_train | metrics_logged mode=train step=610

                                                      
{'loss': -0.0698, 'grad_norm': 1.857463002204895, 'learning_rate': 8.154545454545455e-06, 'num_tokens': 1082119.0, 'completions/mean_length': 150.5, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.23472517728805542, 'rewards/meter/std': 0.2359900027513504, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8774616718292236, 'rewards/repeat_soft/std': 0.08078308403491974, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.3728941082954407, 'rewards/total_composite/mean': 0.32463350892066956, 'rewards/total_composite/std': 0.22264796495437622, 'reward': 0.32463350892066956, 'reward_std': 0.22264795005321503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18068988621234894, 'sampling/sampling_logp_difference/max': 1.5548620223999023, 'sampling/importance_sampling_ratio/min': 0.21121850609779358, 'sampling/importance_sampling_ratio/mean': 1.006516456604004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7865127325057983, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11486065573990345, 'clip_ratio/high_max': 0.11486065573990345, 'clip_ratio/region_mean': 0.11486065573990345, 'reward_total_mean': 0.32463350892066956, 'reward_meter_mean': 0.23472517728805542, 'reward_meter_std': 0.2359900027513504, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8774616718292236, 'reward_repeat_soft_std': 0.08078308403491974, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.3728941082954407, 'reward_total_composite_mean': 0.32463350892066956, 'reward_total_composite_std': 0.22264796495437622, 'epoch': 0.06}

 18%|█▊        | 610/3300 [1:32:50<8:53:01, 11.89s/it]INFO 04-13 08:53:48 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 611/3300 [1:33:01<8:43:58, 11.69s/it]2026-04-13 08:53:59,581 | INFO | train_grpo_train | metrics_logged mode=train step=611

                                                      
{'loss': -0.0851, 'grad_norm': 3.469688653945923, 'learning_rate': 8.151515151515152e-06, 'num_tokens': 1083519.0, 'completions/mean_length': 94.0, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.28571701049805, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6625721454620361, 'rewards/meter/std': 0.40199291706085205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9902437925338745, 'rewards/repeat_soft/std': 0.00864100456237793, 'rewards/judge_quality/mean': 0.6862499713897705, 'rewards/judge_quality/std': 0.34221702814102173, 'rewards/total_composite/mean': 0.6244262456893921, 'rewards/total_composite/std': 0.32269027829170227, 'reward': 0.6244262456893921, 'reward_std': 0.32269027829170227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19811153411865234, 'sampling/sampling_logp_difference/max': 1.5171453952789307, 'sampling/importance_sampling_ratio/min': 0.2193371206521988, 'sampling/importance_sampling_ratio/mean': 0.9901535511016846, 'sampling/importance_sampling_ratio/max': 1.8197382688522339, 'entropy': 1.529385268688202, 'clip_ratio/low_mean': 0.051674107322469354, 'clip_ratio/low_min': 0.051674107322469354, 'clip_ratio/high_mean': 0.07405303046107292, 'clip_ratio/high_max': 0.07405303046107292, 'clip_ratio/region_mean': 0.12572713778354228, 'reward_total_mean': 0.6244262456893921, 'reward_meter_mean': 0.6625721454620361, 'reward_meter_std': 0.40199291706085205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9902437925338745, 'reward_repeat_soft_std': 0.00864100456237793, 'reward_judge_quality_mean': 0.6862499713897705, 'reward_judge_quality_std': 0.34221702814102173, 'reward_total_composite_mean': 0.6244262456893921, 'reward_total_composite_std': 0.32269027829170227, 'epoch': 0.06}

 19%|█▊        | 611/3300 [1:33:01<8:43:58, 11.69s/it]INFO 04-13 08:53:59 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 612/3300 [1:33:13<8:44:56, 11.72s/it]2026-04-13 08:54:11,359 | INFO | train_grpo_train | metrics_logged mode=train step=612

                                                      
{'loss': -0.1151, 'grad_norm': 2.0957794189453125, 'learning_rate': 8.14848484848485e-06, 'num_tokens': 1085081.0, 'completions/mean_length': 166.25, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.46912112832069397, 'rewards/meter/std': 0.415269672870636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9899865388870239, 'rewards/repeat_soft/std': 0.016494235023856163, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.27166154980659485, 'rewards/total_composite/mean': 0.3901972770690918, 'rewards/total_composite/std': 0.254092812538147, 'reward': 0.3901972770690918, 'reward_std': 0.254092812538147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18144994974136353, 'sampling/sampling_logp_difference/max': 1.5548648834228516, 'sampling/importance_sampling_ratio/min': 0.21121792495250702, 'sampling/importance_sampling_ratio/mean': 1.044118881225586, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3020441830158234, 'clip_ratio/low_mean': 0.013020833022892475, 'clip_ratio/low_min': 0.013020833022892475, 'clip_ratio/high_mean': 0.10932030715048313, 'clip_ratio/high_max': 0.10932030715048313, 'clip_ratio/region_mean': 0.1223411401733756, 'reward_total_mean': 0.3901972770690918, 'reward_meter_mean': 0.46912112832069397, 'reward_meter_std': 0.415269672870636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9899865388870239, 'reward_repeat_soft_std': 0.016494235023856163, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.27166154980659485, 'reward_total_composite_mean': 0.3901972770690918, 'reward_total_composite_std': 0.254092812538147, 'epoch': 0.06}

 19%|█▊        | 612/3300 [1:33:13<8:44:56, 11.72s/it]INFO 04-13 08:54:11 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 613/3300 [1:33:19<7:26:49,  9.98s/it]2026-04-13 08:54:17,282 | INFO | train_grpo_train | metrics_logged mode=train step=613

                                                      
{'loss': -0.1371, 'grad_norm': 21.545236587524414, 'learning_rate': 8.145454545454547e-06, 'num_tokens': 1086456.0, 'completions/mean_length': 14.875, 'completions/min_length': 9.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 14.875, 'completions/min_terminated_length': 9.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.2173997163772583, 'rewards/meter/std': 0.34037137031555176, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.974461019039154, 'rewards/repeat_soft/std': 0.021496828645467758, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.23366262018680573, 'rewards/total_composite/mean': 0.28171297907829285, 'rewards/total_composite/std': 0.24723458290100098, 'reward': 0.28171297907829285, 'reward_std': 0.2472345530986786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20753099024295807, 'sampling/sampling_logp_difference/max': 1.3933192491531372, 'sampling/importance_sampling_ratio/min': 0.24824993312358856, 'sampling/importance_sampling_ratio/mean': 0.975199818611145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1872248575091362, 'clip_ratio/low_mean': 0.02291666716337204, 'clip_ratio/low_min': 0.02291666716337204, 'clip_ratio/high_mean': 0.08686145581305027, 'clip_ratio/high_max': 0.08686145581305027, 'clip_ratio/region_mean': 0.10977812297642231, 'reward_total_mean': 0.28171297907829285, 'reward_meter_mean': 0.2173997163772583, 'reward_meter_std': 0.34037137031555176, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.974461019039154, 'reward_repeat_soft_std': 0.021496828645467758, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.23366262018680573, 'reward_total_composite_mean': 0.28171297907829285, 'reward_total_composite_std': 0.24723458290100098, 'epoch': 0.06}

 19%|█▊        | 613/3300 [1:33:19<7:26:49,  9.98s/it]INFO 04-13 08:54:17 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 614/3300 [1:33:25<6:37:48,  8.89s/it]2026-04-13 08:54:23,635 | INFO | train_grpo_train | metrics_logged mode=train step=614

                                                      
{'loss': 0.0372, 'grad_norm': 31.44270133972168, 'learning_rate': 8.142424242424242e-06, 'num_tokens': 1087872.0, 'completions/mean_length': 30.0, 'completions/min_length': 27.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9214865565299988, 'rewards/meter/std': 0.1623855084180832, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9567240476608276, 'rewards/repeat_soft/std': 0.03225353732705116, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.6277033090591431, 'rewards/total_composite/std': 0.08398663997650146, 'reward': 0.6277033090591431, 'reward_std': 0.08398663997650146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17209674417972565, 'sampling/sampling_logp_difference/max': 1.3161673545837402, 'sampling/importance_sampling_ratio/min': 0.2681611180305481, 'sampling/importance_sampling_ratio/mean': 1.024430513381958, 'sampling/importance_sampling_ratio/max': 1.84481942653656, 'entropy': 1.3749855011701584, 'clip_ratio/low_mean': 0.11485749203711748, 'clip_ratio/low_min': 0.11485749203711748, 'clip_ratio/high_mean': 0.03750000149011612, 'clip_ratio/high_max': 0.03750000149011612, 'clip_ratio/region_mean': 0.1523574935272336, 'reward_total_mean': 0.6277033090591431, 'reward_meter_mean': 0.9214865565299988, 'reward_meter_std': 0.1623855084180832, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9567240476608276, 'reward_repeat_soft_std': 0.03225353732705116, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.6277033090591431, 'reward_total_composite_std': 0.08398663997650146, 'epoch': 0.06}

 19%|█▊        | 614/3300 [1:33:25<6:37:48,  8.89s/it]INFO 04-13 08:54:24 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 615/3300 [1:33:38<7:22:40,  9.89s/it]2026-04-13 08:54:35,859 | INFO | train_grpo_train | metrics_logged mode=train step=615

                                                      
{'loss': -0.1034, 'grad_norm': 1.75832998752594, 'learning_rate': 8.139393939393941e-06, 'num_tokens': 1089672.0, 'completions/mean_length': 289.0, 'completions/min_length': 63.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 66.0, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7459532022476196, 'rewards/meter/std': 0.32093924283981323, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.893631637096405, 'rewards/repeat_soft/std': 0.12420381605625153, 'rewards/judge_quality/mean': 0.23499998450279236, 'rewards/judge_quality/std': 0.19777332246303558, 'rewards/total_composite/mean': 0.29106974601745605, 'rewards/total_composite/std': 0.31163278222084045, 'reward': 0.29106974601745605, 'reward_std': 0.31163278222084045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18158908188343048, 'sampling/sampling_logp_difference/max': 1.7608652114868164, 'sampling/importance_sampling_ratio/min': 0.1718960702419281, 'sampling/importance_sampling_ratio/mean': 1.0098347663879395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7075674682855606, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08472961373627186, 'clip_ratio/high_max': 0.08472961373627186, 'clip_ratio/region_mean': 0.08472961373627186, 'reward_total_mean': 0.29106974601745605, 'reward_meter_mean': 0.7459532022476196, 'reward_meter_std': 0.32093924283981323, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.893631637096405, 'reward_repeat_soft_std': 0.12420381605625153, 'reward_judge_quality_mean': 0.23499998450279236, 'reward_judge_quality_std': 0.19777332246303558, 'reward_total_composite_mean': 0.29106974601745605, 'reward_total_composite_std': 0.31163278222084045, 'epoch': 0.06}

 19%|█▊        | 615/3300 [1:33:38<7:22:40,  9.89s/it]INFO 04-13 08:54:36 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 616/3300 [1:33:43<6:29:22,  8.70s/it]2026-04-13 08:54:41,789 | INFO | train_grpo_train | metrics_logged mode=train step=616

                                                      
{'loss': 0.0415, 'grad_norm': 19.410985946655273, 'learning_rate': 8.136363636363637e-06, 'num_tokens': 1091114.0, 'completions/mean_length': 34.25, 'completions/min_length': 27.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.398084819316864, 'rewards/meter/std': 0.3238059878349304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9631776809692383, 'rewards/repeat_soft/std': 0.0329229012131691, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4093209207057953, 'rewards/total_composite/std': 0.18217498064041138, 'reward': 0.4093209207057953, 'reward_std': 0.18217498064041138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17289279401302338, 'sampling/sampling_logp_difference/max': 1.4119362831115723, 'sampling/importance_sampling_ratio/min': 0.2436710149049759, 'sampling/importance_sampling_ratio/mean': 1.024407982826233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2369627207517624, 'clip_ratio/low_mean': 0.06231678184121847, 'clip_ratio/low_min': 0.06231678184121847, 'clip_ratio/high_mean': 0.09158502984791994, 'clip_ratio/high_max': 0.09158502984791994, 'clip_ratio/region_mean': 0.1539018116891384, 'reward_total_mean': 0.4093209207057953, 'reward_meter_mean': 0.398084819316864, 'reward_meter_std': 0.3238059878349304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9631776809692383, 'reward_repeat_soft_std': 0.0329229012131691, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4093209207057953, 'reward_total_composite_std': 0.18217498064041138, 'epoch': 0.06}

 19%|█▊        | 616/3300 [1:33:44<6:29:22,  8.70s/it]INFO 04-13 08:54:42 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 617/3300 [1:33:49<5:49:57,  7.83s/it]2026-04-13 08:54:47,565 | INFO | train_grpo_train | metrics_logged mode=train step=617

                                                      
{'loss': 0.1614, 'grad_norm': 20.967464447021484, 'learning_rate': 8.133333333333334e-06, 'num_tokens': 1092504.0, 'completions/mean_length': 25.75, 'completions/min_length': 19.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.4370259642601013, 'rewards/meter/std': 0.3252226412296295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9890512228012085, 'rewards/repeat_soft/std': 0.011672230437397957, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4696791172027588, 'rewards/total_composite/std': 0.08837082237005234, 'reward': 0.4696791172027588, 'reward_std': 0.08837081491947174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15416434407234192, 'sampling/sampling_logp_difference/max': 1.711958885192871, 'sampling/importance_sampling_ratio/min': 0.18051183223724365, 'sampling/importance_sampling_ratio/mean': 0.9849514365196228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7167652025818825, 'clip_ratio/low_mean': 0.10505807586014271, 'clip_ratio/low_min': 0.10505807586014271, 'clip_ratio/high_mean': 0.07016194332391024, 'clip_ratio/high_max': 0.07016194332391024, 'clip_ratio/region_mean': 0.17522001918405294, 'reward_total_mean': 0.4696791172027588, 'reward_meter_mean': 0.4370259642601013, 'reward_meter_std': 0.3252226412296295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9890512228012085, 'reward_repeat_soft_std': 0.011672230437397957, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4696791172027588, 'reward_total_composite_std': 0.08837082237005234, 'epoch': 0.06}

 19%|█▊        | 617/3300 [1:33:49<5:49:57,  7.83s/it]INFO 04-13 08:54:47 [block_pool.py:378] Successfully reset prefix cache

 19%|█▊        | 618/3300 [1:34:01<6:37:53,  8.90s/it]2026-04-13 08:54:58,972 | INFO | train_grpo_train | metrics_logged mode=train step=618

                                                      
{'loss': -0.1143, 'grad_norm': 3.0757648944854736, 'learning_rate': 8.130303030303031e-06, 'num_tokens': 1094125.0, 'completions/mean_length': 103.625, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.28571701049805, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7123228311538696, 'rewards/meter/std': 0.4260020852088928, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8826066255569458, 'rewards/repeat_soft/std': 0.1023331731557846, 'rewards/judge_quality/mean': 0.2849999964237213, 'rewards/judge_quality/std': 0.15390163660049438, 'rewards/total_composite/mean': 0.446541428565979, 'rewards/total_composite/std': 0.20337434113025665, 'reward': 0.446541428565979, 'reward_std': 0.20337434113025665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18212325870990753, 'sampling/sampling_logp_difference/max': 1.821512222290039, 'sampling/importance_sampling_ratio/min': 0.1617809236049652, 'sampling/importance_sampling_ratio/mean': 1.0186419486999512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5973176509141922, 'clip_ratio/low_mean': 0.03890810161828995, 'clip_ratio/low_min': 0.03890810161828995, 'clip_ratio/high_mean': 0.131703345105052, 'clip_ratio/high_max': 0.131703345105052, 'clip_ratio/region_mean': 0.17061144672334194, 'reward_total_mean': 0.446541428565979, 'reward_meter_mean': 0.7123228311538696, 'reward_meter_std': 0.4260020852088928, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8826066255569458, 'reward_repeat_soft_std': 0.1023331731557846, 'reward_judge_quality_mean': 0.2849999964237213, 'reward_judge_quality_std': 0.15390163660049438, 'reward_total_composite_mean': 0.446541428565979, 'reward_total_composite_std': 0.20337434113025665, 'epoch': 0.06}

 19%|█▊        | 618/3300 [1:34:01<6:37:53,  8.90s/it]INFO 04-13 08:54:59 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 619/3300 [1:34:07<6:00:38,  8.07s/it]2026-04-13 08:55:05,110 | INFO | train_grpo_train | metrics_logged mode=train step=619

                                                      
{'loss': 0.1221, 'grad_norm': 28.908626556396484, 'learning_rate': 8.127272727272728e-06, 'num_tokens': 1095532.0, 'completions/mean_length': 28.875, 'completions/min_length': 24.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.45595115423202515, 'rewards/meter/std': 0.3782821595668793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9545157551765442, 'rewards/repeat_soft/std': 0.09621007740497589, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.5886853337287903, 'rewards/total_composite/std': 0.21713681519031525, 'reward': 0.5886853337287903, 'reward_std': 0.21713680028915405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13166476786136627, 'sampling/sampling_logp_difference/max': 1.7379283905029297, 'sampling/importance_sampling_ratio/min': 0.17588438093662262, 'sampling/importance_sampling_ratio/mean': 0.9892944097518921, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4909357912838459, 'clip_ratio/low_mean': 0.07224462367594242, 'clip_ratio/low_min': 0.07224462367594242, 'clip_ratio/high_mean': 0.041193182580173016, 'clip_ratio/high_max': 0.041193182580173016, 'clip_ratio/region_mean': 0.11343780625611544, 'reward_total_mean': 0.5886853337287903, 'reward_meter_mean': 0.45595115423202515, 'reward_meter_std': 0.3782821595668793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9545157551765442, 'reward_repeat_soft_std': 0.09621007740497589, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.5886853337287903, 'reward_total_composite_std': 0.21713681519031525, 'epoch': 0.06}

 19%|█▉        | 619/3300 [1:34:07<6:00:38,  8.07s/it]INFO 04-13 08:55:05 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 620/3300 [1:34:13<5:36:11,  7.53s/it]2026-04-13 08:55:11,365 | INFO | train_grpo_train | metrics_logged mode=train step=620

                                                      
{'loss': 0.0685, 'grad_norm': 12.25234317779541, 'learning_rate': 8.124242424242424e-06, 'num_tokens': 1097493.0, 'completions/mean_length': 57.125, 'completions/min_length': 51.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5478992462158203, 'rewards/meter/std': 0.46492108702659607, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6984058618545532, 'rewards/repeat_soft/std': 0.09894108027219772, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.434698224067688, 'rewards/total_composite/std': 0.1239008679986, 'reward': 0.434698224067688, 'reward_std': 0.1239008754491806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11980719119310379, 'sampling/sampling_logp_difference/max': 1.217726707458496, 'sampling/importance_sampling_ratio/min': 0.2959020733833313, 'sampling/importance_sampling_ratio/mean': 1.0205646753311157, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0054671689867973, 'clip_ratio/low_mean': 0.05324944853782654, 'clip_ratio/low_min': 0.05324944853782654, 'clip_ratio/high_mean': 0.09571819379925728, 'clip_ratio/high_max': 0.09571819379925728, 'clip_ratio/region_mean': 0.14896764233708382, 'reward_total_mean': 0.434698224067688, 'reward_meter_mean': 0.5478992462158203, 'reward_meter_std': 0.46492108702659607, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6984058618545532, 'reward_repeat_soft_std': 0.09894108027219772, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.434698224067688, 'reward_total_composite_std': 0.1239008679986, 'epoch': 0.06}

 19%|█▉        | 620/3300 [1:34:13<5:36:11,  7.53s/it]INFO 04-13 08:55:11 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 621/3300 [1:34:19<5:16:02,  7.08s/it]2026-04-13 08:55:17,398 | INFO | train_grpo_train | metrics_logged mode=train step=621

                                                      
{'loss': 0.0479, 'grad_norm': 17.410306930541992, 'learning_rate': 8.121212121212121e-06, 'num_tokens': 1099127.0, 'completions/mean_length': 35.25, 'completions/min_length': 33.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.3701145648956299, 'rewards/meter/std': 0.4021562933921814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.999520480632782, 'rewards/repeat_soft/std': 0.0008879249216988683, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.45440980792045593, 'rewards/total_composite/std': 0.10795284807682037, 'reward': 0.45440980792045593, 'reward_std': 0.10795284062623978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18046563863754272, 'sampling/sampling_logp_difference/max': 2.2352566719055176, 'sampling/importance_sampling_ratio/min': 0.10696467012166977, 'sampling/importance_sampling_ratio/mean': 1.031589388847351, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9421701580286026, 'clip_ratio/low_mean': 0.09570794692263007, 'clip_ratio/low_min': 0.09570794692263007, 'clip_ratio/high_mean': 0.05011655203998089, 'clip_ratio/high_max': 0.05011655203998089, 'clip_ratio/region_mean': 0.14582449896261096, 'reward_total_mean': 0.45440980792045593, 'reward_meter_mean': 0.3701145648956299, 'reward_meter_std': 0.4021562933921814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.999520480632782, 'reward_repeat_soft_std': 0.0008879249216988683, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.45440980792045593, 'reward_total_composite_std': 0.10795284807682037, 'epoch': 0.06}

 19%|█▉        | 621/3300 [1:34:19<5:16:02,  7.08s/it]INFO 04-13 08:55:17 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 622/3300 [1:34:30<6:11:01,  8.31s/it]2026-04-13 08:55:28,593 | INFO | train_grpo_train | metrics_logged mode=train step=622

                                                      
{'loss': -0.1313, 'grad_norm': 3.2621078491210938, 'learning_rate': 8.118181818181819e-06, 'num_tokens': 1100870.0, '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': 69.0, 'rewards/meter/mean': 0.6296329498291016, 'rewards/meter/std': 0.3438984155654907, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8968924283981323, 'rewards/repeat_soft/std': 0.08101747930049896, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.44532322883605957, 'rewards/total_composite/std': 0.20290635526180267, 'reward': 0.44532322883605957, 'reward_std': 0.20290635526180267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14242790639400482, 'sampling/sampling_logp_difference/max': 1.327103614807129, 'sampling/importance_sampling_ratio/min': 0.26524439454078674, 'sampling/importance_sampling_ratio/mean': 1.0022187232971191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9111338779330254, 'clip_ratio/low_mean': 0.019625603687018156, 'clip_ratio/low_min': 0.019625603687018156, 'clip_ratio/high_mean': 0.08964197617024183, 'clip_ratio/high_max': 0.08964197617024183, 'clip_ratio/region_mean': 0.10926757985725999, 'reward_total_mean': 0.44532322883605957, 'reward_meter_mean': 0.6296329498291016, 'reward_meter_std': 0.3438984155654907, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8968924283981323, 'reward_repeat_soft_std': 0.08101747930049896, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.44532322883605957, 'reward_total_composite_std': 0.20290635526180267, 'epoch': 0.06}

 19%|█▉        | 622/3300 [1:34:30<6:11:01,  8.31s/it]INFO 04-13 08:55:28 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 623/3300 [1:34:36<5:36:56,  7.55s/it]2026-04-13 08:55:34,366 | INFO | train_grpo_train | metrics_logged mode=train step=623

                                                      
{'loss': -0.0309, 'grad_norm': 17.54698371887207, 'learning_rate': 8.115151515151516e-06, 'num_tokens': 1102515.0, 'completions/mean_length': 31.625, 'completions/min_length': 27.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9827060699462891, 'rewards/meter/std': 0.012876130640506744, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.958416223526001, 'rewards/repeat_soft/std': 0.03597824275493622, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6940488815307617, 'rewards/total_composite/std': 0.14654476940631866, 'reward': 0.6940488815307617, 'reward_std': 0.14654478430747986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14996962249279022, 'sampling/sampling_logp_difference/max': 1.2179861068725586, 'sampling/importance_sampling_ratio/min': 0.295825332403183, 'sampling/importance_sampling_ratio/mean': 1.0346026420593262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0386882349848747, 'clip_ratio/low_mean': 0.11467740964144468, 'clip_ratio/low_min': 0.11467740964144468, 'clip_ratio/high_mean': 0.043174341320991516, 'clip_ratio/high_max': 0.043174341320991516, 'clip_ratio/region_mean': 0.1578517509624362, 'reward_total_mean': 0.6940488815307617, 'reward_meter_mean': 0.9827060699462891, 'reward_meter_std': 0.012876130640506744, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.958416223526001, 'reward_repeat_soft_std': 0.03597824275493622, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6940488815307617, 'reward_total_composite_std': 0.14654476940631866, 'epoch': 0.06}

 19%|█▉        | 623/3300 [1:34:36<5:36:56,  7.55s/it]INFO 04-13 08:55:34 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 624/3300 [1:34:42<5:18:25,  7.14s/it]2026-04-13 08:55:40,546 | INFO | train_grpo_train | metrics_logged mode=train step=624

                                                      
{'loss': 0.132, 'grad_norm': 30.478708267211914, 'learning_rate': 8.112121212121213e-06, 'num_tokens': 1103830.0, 'completions/mean_length': 17.375, 'completions/min_length': 15.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.6895481944084167, 'rewards/meter/std': 0.3371638059616089, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5350354313850403, 'rewards/total_composite/std': 0.09472057223320007, 'reward': 0.5350354313850403, 'reward_std': 0.09472055733203888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20352110266685486, 'sampling/sampling_logp_difference/max': 1.9299743175506592, 'sampling/importance_sampling_ratio/min': 0.14515192806720734, 'sampling/importance_sampling_ratio/mean': 0.9885327816009521, 'sampling/importance_sampling_ratio/max': 1.9450926780700684, 'entropy': 1.3484907150268555, 'clip_ratio/low_mean': 0.08521303255110979, 'clip_ratio/low_min': 0.08521303255110979, 'clip_ratio/high_mean': 0.09353554248809814, 'clip_ratio/high_max': 0.09353554248809814, 'clip_ratio/region_mean': 0.17874857503920794, 'reward_total_mean': 0.5350354313850403, 'reward_meter_mean': 0.6895481944084167, 'reward_meter_std': 0.3371638059616089, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5350354313850403, 'reward_total_composite_std': 0.09472057223320007, 'epoch': 0.06}

 19%|█▉        | 624/3300 [1:34:42<5:18:25,  7.14s/it]INFO 04-13 08:55:40 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 625/3300 [1:34:49<5:09:49,  6.95s/it]2026-04-13 08:55:47,051 | INFO | train_grpo_train | metrics_logged mode=train step=625

                                                      
{'loss': 0.0167, 'grad_norm': 11.994166374206543, 'learning_rate': 8.10909090909091e-06, 'num_tokens': 1105301.0, 'completions/mean_length': 30.875, 'completions/min_length': 29.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.8243540525436401, 'rewards/meter/std': 0.3199911117553711, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9536811113357544, 'rewards/repeat_soft/std': 0.05209203436970711, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.19111984968185425, 'rewards/total_composite/mean': 0.6183902025222778, 'rewards/total_composite/std': 0.10725264996290207, 'reward': 0.6183902025222778, 'reward_std': 0.10725263506174088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1609441339969635, 'sampling/sampling_logp_difference/max': 1.681645393371582, 'sampling/importance_sampling_ratio/min': 0.18606756627559662, 'sampling/importance_sampling_ratio/mean': 0.9962921738624573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0047850385308266, 'clip_ratio/low_mean': 0.07782866479828954, 'clip_ratio/low_min': 0.07782866479828954, 'clip_ratio/high_mean': 0.09491125121712685, 'clip_ratio/high_max': 0.09491125121712685, 'clip_ratio/region_mean': 0.17273991601541638, 'reward_total_mean': 0.6183902025222778, 'reward_meter_mean': 0.8243540525436401, 'reward_meter_std': 0.3199911117553711, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9536811113357544, 'reward_repeat_soft_std': 0.05209203436970711, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.19111984968185425, 'reward_total_composite_mean': 0.6183902025222778, 'reward_total_composite_std': 0.10725264996290207, 'epoch': 0.06}

 19%|█▉        | 625/3300 [1:34:49<5:09:49,  6.95s/it]INFO 04-13 08:55:47 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 626/3300 [1:34:55<4:58:15,  6.69s/it]2026-04-13 08:55:53,142 | INFO | train_grpo_train | metrics_logged mode=train step=626

                                                      
{'loss': 0.0544, 'grad_norm': 20.4157772064209, 'learning_rate': 8.106060606060606e-06, 'num_tokens': 1106803.0, 'completions/mean_length': 35.75, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7206422090530396, 'rewards/meter/std': 0.34192830324172974, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9348602890968323, 'rewards/repeat_soft/std': 0.10164881497621536, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.15297061204910278, 'rewards/total_composite/mean': 0.6152657270431519, 'rewards/total_composite/std': 0.1582135111093521, 'reward': 0.6152657270431519, 'reward_std': 0.15821349620819092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18463172018527985, 'sampling/sampling_logp_difference/max': 1.4099512100219727, 'sampling/importance_sampling_ratio/min': 0.24415519833564758, 'sampling/importance_sampling_ratio/mean': 1.0475335121154785, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6291310265660286, 'clip_ratio/low_mean': 0.1022753482684493, 'clip_ratio/low_min': 0.1022753482684493, 'clip_ratio/high_mean': 0.09265873208642006, 'clip_ratio/high_max': 0.09265873208642006, 'clip_ratio/region_mean': 0.19493408035486937, 'reward_total_mean': 0.6152657270431519, 'reward_meter_mean': 0.7206422090530396, 'reward_meter_std': 0.34192830324172974, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9348602890968323, 'reward_repeat_soft_std': 0.10164881497621536, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.15297061204910278, 'reward_total_composite_mean': 0.6152657270431519, 'reward_total_composite_std': 0.1582135111093521, 'epoch': 0.06}

 19%|█▉        | 626/3300 [1:34:55<4:58:15,  6.69s/it]INFO 04-13 08:55:53 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 627/3300 [1:35:01<4:57:42,  6.68s/it]2026-04-13 08:55:59,800 | INFO | train_grpo_train | metrics_logged mode=train step=627

                                                      
{'loss': 0.0083, 'grad_norm': 12.007946968078613, 'learning_rate': 8.103030303030303e-06, 'num_tokens': 1108759.0, 'completions/mean_length': 74.5, 'completions/min_length': 63.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.5, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.8114852905273438, 'rewards/meter/std': 0.2701057493686676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.868984580039978, 'rewards/repeat_soft/std': 0.06701181828975677, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5804606676101685, 'rewards/total_composite/std': 0.10382802784442902, 'reward': 0.5804606676101685, 'reward_std': 0.10382802039384842, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1396416574716568, 'sampling/sampling_logp_difference/max': 1.781585693359375, 'sampling/importance_sampling_ratio/min': 0.16837096214294434, 'sampling/importance_sampling_ratio/mean': 1.0330348014831543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.169243462383747, 'clip_ratio/low_mean': 0.028571429662406445, 'clip_ratio/low_min': 0.028571429662406445, 'clip_ratio/high_mean': 0.09981811698526144, 'clip_ratio/high_max': 0.09981811698526144, 'clip_ratio/region_mean': 0.12838954664766788, 'reward_total_mean': 0.5804606676101685, 'reward_meter_mean': 0.8114852905273438, 'reward_meter_std': 0.2701057493686676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.868984580039978, 'reward_repeat_soft_std': 0.06701181828975677, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5804606676101685, 'reward_total_composite_std': 0.10382802784442902, 'epoch': 0.06}

 19%|█▉        | 627/3300 [1:35:02<4:57:42,  6.68s/it]INFO 04-13 08:56:00 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 628/3300 [1:35:13<5:59:22,  8.07s/it]2026-04-13 08:56:11,111 | INFO | train_grpo_train | metrics_logged mode=train step=628

                                                      
{'loss': -0.193, 'grad_norm': 2.5380585193634033, 'learning_rate': 8.1e-06, 'num_tokens': 1110957.0, 'completions/mean_length': 156.75, 'completions/min_length': 99.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 106.00000762939453, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.8502894639968872, 'rewards/meter/std': 0.2771815061569214, 'rewards/count_adherence/mean': 0.9791666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.5566608309745789, 'rewards/repeat_soft/std': 0.11484144628047943, 'rewards/judge_quality/mean': 0.2212499976158142, 'rewards/judge_quality/std': 0.10842212289571762, 'rewards/total_composite/mean': 0.36367928981781006, 'rewards/total_composite/std': 0.1626027375459671, 'reward': 0.36367928981781006, 'reward_std': 0.1626027375459671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08519712090492249, 'sampling/sampling_logp_difference/max': 1.9697628021240234, 'sampling/importance_sampling_ratio/min': 0.13948993384838104, 'sampling/importance_sampling_ratio/mean': 1.0176167488098145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6992944739758968, 'clip_ratio/low_mean': 0.0071428571827709675, 'clip_ratio/low_min': 0.0071428571827709675, 'clip_ratio/high_mean': 0.051890133414417505, 'clip_ratio/high_max': 0.051890133414417505, 'clip_ratio/region_mean': 0.05903299059718847, 'reward_total_mean': 0.36367928981781006, 'reward_meter_mean': 0.8502894639968872, 'reward_meter_std': 0.2771815061569214, 'reward_count_adherence_mean': 0.9791666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.5566608309745789, 'reward_repeat_soft_std': 0.11484144628047943, 'reward_judge_quality_mean': 0.2212499976158142, 'reward_judge_quality_std': 0.10842212289571762, 'reward_total_composite_mean': 0.36367928981781006, 'reward_total_composite_std': 0.1626027375459671, 'epoch': 0.06}

 19%|█▉        | 628/3300 [1:35:13<5:59:22,  8.07s/it]INFO 04-13 08:56:11 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 629/3300 [1:35:19<5:33:32,  7.49s/it]2026-04-13 08:56:17,255 | INFO | train_grpo_train | metrics_logged mode=train step=629

                                                      
{'loss': 0.0398, 'grad_norm': 18.1601505279541, 'learning_rate': 8.096969696969698e-06, 'num_tokens': 1112426.0, 'completions/mean_length': 30.625, 'completions/min_length': 29.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.5472124814987183, 'rewards/meter/std': 0.3712937533855438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9427904486656189, 'rewards/repeat_soft/std': 0.07211591303348541, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.4785916209220886, 'rewards/total_composite/std': 0.09244726598262787, 'reward': 0.4785916209220886, 'reward_std': 0.09244726598262787, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18447519838809967, 'sampling/sampling_logp_difference/max': 1.5100622177124023, 'sampling/importance_sampling_ratio/min': 0.22089622914791107, 'sampling/importance_sampling_ratio/mean': 1.0286273956298828, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5972175747156143, 'clip_ratio/low_mean': 0.05895900260657072, 'clip_ratio/low_min': 0.05895900260657072, 'clip_ratio/high_mean': 0.10901000909507275, 'clip_ratio/high_max': 0.10901000909507275, 'clip_ratio/region_mean': 0.16796901170164347, 'reward_total_mean': 0.4785916209220886, 'reward_meter_mean': 0.5472124814987183, 'reward_meter_std': 0.3712937533855438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9427904486656189, 'reward_repeat_soft_std': 0.07211591303348541, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.4785916209220886, 'reward_total_composite_std': 0.09244726598262787, 'epoch': 0.06}

 19%|█▉        | 629/3300 [1:35:19<5:33:32,  7.49s/it]INFO 04-13 08:56:17 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 630/3300 [1:35:30<6:25:36,  8.67s/it]2026-04-13 08:56:28,659 | INFO | train_grpo_train | metrics_logged mode=train step=630

                                                      
{'loss': -0.0552, 'grad_norm': 1.3623456954956055, 'learning_rate': 8.093939393939395e-06, 'num_tokens': 1114077.0, 'completions/mean_length': 398.375, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.34584999084472656, 'rewards/meter/std': 0.3909425735473633, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.37533053755760193, 'rewards/hard_gate/mean': 0.25, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.987758457660675, 'rewards/repeat_soft/std': 0.016199786216020584, 'rewards/judge_quality/mean': 0.22625000774860382, 'rewards/judge_quality/std': 0.3057047426700592, 'rewards/total_composite/mean': 0.17963233590126038, 'rewards/total_composite/std': 0.33561691641807556, 'reward': 0.17963233590126038, 'reward_std': 0.33561691641807556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21470822393894196, 'sampling/sampling_logp_difference/max': 1.2154841423034668, 'sampling/importance_sampling_ratio/min': 0.2965663969516754, 'sampling/importance_sampling_ratio/mean': 1.0532341003417969, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37769442796707153, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.045674530789256096, 'clip_ratio/high_max': 0.045674530789256096, 'clip_ratio/region_mean': 0.045674530789256096, 'reward_total_mean': 0.17963233590126038, 'reward_meter_mean': 0.34584999084472656, 'reward_meter_std': 0.3909425735473633, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.37533053755760193, 'reward_hard_gate_mean': 0.25, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.987758457660675, 'reward_repeat_soft_std': 0.016199786216020584, 'reward_judge_quality_mean': 0.22625000774860382, 'reward_judge_quality_std': 0.3057047426700592, 'reward_total_composite_mean': 0.17963233590126038, 'reward_total_composite_std': 0.33561691641807556, 'epoch': 0.06}

 19%|█▉        | 630/3300 [1:35:30<6:25:36,  8.67s/it]INFO 04-13 08:56:28 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 631/3300 [1:35:37<5:58:13,  8.05s/it]2026-04-13 08:56:35,278 | INFO | train_grpo_train | metrics_logged mode=train step=631

                                                      
{'loss': 0.019, 'grad_norm': 16.864667892456055, 'learning_rate': 8.090909090909092e-06, 'num_tokens': 1115731.0, 'completions/mean_length': 49.75, 'completions/min_length': 36.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.3106288015842438, 'rewards/meter/std': 0.25897109508514404, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9419691562652588, 'rewards/repeat_soft/std': 0.05332150682806969, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.4362753629684448, 'rewards/total_composite/std': 0.0686321035027504, 'reward': 0.4362753629684448, 'reward_std': 0.0686321035027504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19420987367630005, 'sampling/sampling_logp_difference/max': 2.466726779937744, 'sampling/importance_sampling_ratio/min': 0.08486217260360718, 'sampling/importance_sampling_ratio/mean': 1.00111722946167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8492881581187248, 'clip_ratio/low_mean': 0.11858218722045422, 'clip_ratio/low_min': 0.11858218722045422, 'clip_ratio/high_mean': 0.0524417320266366, 'clip_ratio/high_max': 0.0524417320266366, 'clip_ratio/region_mean': 0.17102391924709082, 'reward_total_mean': 0.4362753629684448, 'reward_meter_mean': 0.3106288015842438, 'reward_meter_std': 0.25897109508514404, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9419691562652588, 'reward_repeat_soft_std': 0.05332150682806969, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.4362753629684448, 'reward_total_composite_std': 0.0686321035027504, 'epoch': 0.06}

 19%|█▉        | 631/3300 [1:35:37<5:58:13,  8.05s/it]INFO 04-13 08:56:35 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 632/3300 [1:35:48<6:41:51,  9.04s/it]2026-04-13 08:56:46,617 | INFO | train_grpo_train | metrics_logged mode=train step=632

                                                      
{'loss': -0.0898, 'grad_norm': 3.517648458480835, 'learning_rate': 8.08787878787879e-06, 'num_tokens': 1117377.0, 'completions/mean_length': 99.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.85714340209961, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8064794540405273, 'rewards/meter/std': 0.3122844099998474, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9868118762969971, 'rewards/repeat_soft/std': 0.025085289031267166, 'rewards/judge_quality/mean': 0.51500004529953, 'rewards/judge_quality/std': 0.2660827040672302, 'rewards/total_composite/mean': 0.5743173360824585, 'rewards/total_composite/std': 0.2833027243614197, 'reward': 0.5743173360824585, 'reward_std': 0.2833027243614197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1882508546113968, 'sampling/sampling_logp_difference/max': 1.2981586456298828, 'sampling/importance_sampling_ratio/min': 0.27303409576416016, 'sampling/importance_sampling_ratio/mean': 1.0191739797592163, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3905847817659378, 'clip_ratio/low_mean': 0.04161173291504383, 'clip_ratio/low_min': 0.04161173291504383, 'clip_ratio/high_mean': 0.10609741974622011, 'clip_ratio/high_max': 0.10609741974622011, 'clip_ratio/region_mean': 0.14770915266126394, 'reward_total_mean': 0.5743173360824585, 'reward_meter_mean': 0.8064794540405273, 'reward_meter_std': 0.3122844099998474, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9868118762969971, 'reward_repeat_soft_std': 0.025085289031267166, 'reward_judge_quality_mean': 0.51500004529953, 'reward_judge_quality_std': 0.2660827040672302, 'reward_total_composite_mean': 0.5743173360824585, 'reward_total_composite_std': 0.2833027243614197, 'epoch': 0.06}

 19%|█▉        | 632/3300 [1:35:48<6:41:51,  9.04s/it]INFO 04-13 08:56:46 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 633/3300 [1:35:56<6:18:56,  8.53s/it]2026-04-13 08:56:53,946 | INFO | train_grpo_train | metrics_logged mode=train step=633

                                                      
{'loss': 0.0163, 'grad_norm': 8.809235572814941, 'learning_rate': 8.084848484848485e-06, 'num_tokens': 1119738.0, 'completions/mean_length': 93.125, 'completions/min_length': 69.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.125, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9915623664855957, 'rewards/meter/std': 0.004384336993098259, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6404680013656616, 'rewards/repeat_soft/std': 0.13526977598667145, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.5077430009841919, 'rewards/total_composite/std': 0.079136922955513, 'reward': 0.5077430009841919, 'reward_std': 0.079136922955513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14437268674373627, 'sampling/sampling_logp_difference/max': 2.0607922077178955, 'sampling/importance_sampling_ratio/min': 0.12735304236412048, 'sampling/importance_sampling_ratio/mean': 1.0201661586761475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1471482813358307, 'clip_ratio/low_mean': 0.05252622161060572, 'clip_ratio/low_min': 0.05252622161060572, 'clip_ratio/high_mean': 0.08306257333606482, 'clip_ratio/high_max': 0.08306257333606482, 'clip_ratio/region_mean': 0.13558879494667053, 'reward_total_mean': 0.5077430009841919, 'reward_meter_mean': 0.9915623664855957, 'reward_meter_std': 0.004384336993098259, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6404680013656616, 'reward_repeat_soft_std': 0.13526977598667145, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.5077430009841919, 'reward_total_composite_std': 0.079136922955513, 'epoch': 0.06}

 19%|█▉        | 633/3300 [1:35:56<6:18:56,  8.53s/it]INFO 04-13 08:56:54 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 634/3300 [1:36:02<5:48:50,  7.85s/it]2026-04-13 08:57:00,222 | INFO | train_grpo_train | metrics_logged mode=train step=634

                                                      
{'loss': -0.088, 'grad_norm': 17.352680206298828, 'learning_rate': 8.081818181818182e-06, 'num_tokens': 1121285.0, 'completions/mean_length': 42.375, 'completions/min_length': 29.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.4564714729785919, 'rewards/meter/std': 0.3743678331375122, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9164788126945496, 'rewards/repeat_soft/std': 0.06086752563714981, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.4197845458984375, 'rewards/total_composite/std': 0.19420462846755981, 'reward': 0.4197845458984375, 'reward_std': 0.19420461356639862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1582125574350357, 'sampling/sampling_logp_difference/max': 2.0730929374694824, 'sampling/importance_sampling_ratio/min': 0.1257961094379425, 'sampling/importance_sampling_ratio/mean': 1.0250484943389893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0219487696886063, 'clip_ratio/low_mean': 0.06842627469450235, 'clip_ratio/low_min': 0.06842627469450235, 'clip_ratio/high_mean': 0.08998904563486576, 'clip_ratio/high_max': 0.08998904563486576, 'clip_ratio/region_mean': 0.15841532032936811, 'reward_total_mean': 0.4197845458984375, 'reward_meter_mean': 0.4564714729785919, 'reward_meter_std': 0.3743678331375122, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9164788126945496, 'reward_repeat_soft_std': 0.06086752563714981, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.4197845458984375, 'reward_total_composite_std': 0.19420462846755981, 'epoch': 0.06}

 19%|█▉        | 634/3300 [1:36:02<5:48:50,  7.85s/it]INFO 04-13 08:57:00 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 635/3300 [1:36:08<5:28:24,  7.39s/it]2026-04-13 08:57:06,550 | INFO | train_grpo_train | metrics_logged mode=train step=635

                                                      
{'loss': -0.0257, 'grad_norm': 13.63864517211914, 'learning_rate': 8.07878787878788e-06, 'num_tokens': 1122979.0, 'completions/mean_length': 51.75, 'completions/min_length': 37.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6257561445236206, 'rewards/meter/std': 0.3036348819732666, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7920382022857666, 'rewards/repeat_soft/std': 0.0841648131608963, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5136308073997498, 'rewards/total_composite/std': 0.09853413701057434, 'reward': 0.5136308073997498, 'reward_std': 0.09853413701057434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14734230935573578, 'sampling/sampling_logp_difference/max': 2.8847289085388184, 'sampling/importance_sampling_ratio/min': 0.0558699332177639, 'sampling/importance_sampling_ratio/mean': 1.001219630241394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6683292649686337, 'clip_ratio/low_mean': 0.024950592778623104, 'clip_ratio/low_min': 0.024950592778623104, 'clip_ratio/high_mean': 0.09042746387422085, 'clip_ratio/high_max': 0.09042746387422085, 'clip_ratio/region_mean': 0.11537805665284395, 'reward_total_mean': 0.5136308073997498, 'reward_meter_mean': 0.6257561445236206, 'reward_meter_std': 0.3036348819732666, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7920382022857666, 'reward_repeat_soft_std': 0.0841648131608963, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5136308073997498, 'reward_total_composite_std': 0.09853413701057434, 'epoch': 0.06}

 19%|█▉        | 635/3300 [1:36:08<5:28:24,  7.39s/it]INFO 04-13 08:57:06 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 636/3300 [1:36:20<6:21:44,  8.60s/it]2026-04-13 08:57:17,961 | INFO | train_grpo_train | metrics_logged mode=train step=636

                                                      
{'loss': -0.1102, 'grad_norm': 5.966771125793457, 'learning_rate': 8.075757575757577e-06, 'num_tokens': 1124740.0, 'completions/mean_length': 109.125, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.57143020629883, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.1405440866947174, 'rewards/meter/std': 0.17522665858268738, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9446536302566528, 'rewards/repeat_soft/std': 0.06525295972824097, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.27637478709220886, 'rewards/total_composite/std': 0.17110449075698853, 'reward': 0.27637478709220886, 'reward_std': 0.17110447585582733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.27482447028160095, 'sampling/sampling_logp_difference/max': 3.4003939628601074, 'sampling/importance_sampling_ratio/min': 0.03336012363433838, 'sampling/importance_sampling_ratio/mean': 0.9913470149040222, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8371185436844826, 'clip_ratio/low_mean': 0.03125, 'clip_ratio/low_min': 0.03125, 'clip_ratio/high_mean': 0.157029053196311, 'clip_ratio/high_max': 0.157029053196311, 'clip_ratio/region_mean': 0.188279053196311, 'reward_total_mean': 0.27637478709220886, 'reward_meter_mean': 0.1405440866947174, 'reward_meter_std': 0.17522665858268738, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9446536302566528, 'reward_repeat_soft_std': 0.06525295972824097, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.27637478709220886, 'reward_total_composite_std': 0.17110449075698853, 'epoch': 0.06}

 19%|█▉        | 636/3300 [1:36:20<6:21:44,  8.60s/it]INFO 04-13 08:57:18 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 637/3300 [1:36:26<5:52:00,  7.93s/it]2026-04-13 08:57:24,336 | INFO | train_grpo_train | metrics_logged mode=train step=637

                                                      
{'loss': 0.0412, 'grad_norm': 15.5023832321167, 'learning_rate': 8.072727272727274e-06, 'num_tokens': 1126466.0, 'completions/mean_length': 40.75, 'completions/min_length': 31.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5743567943572998, 'rewards/meter/std': 0.36496245861053467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9650839567184448, 'rewards/repeat_soft/std': 0.0464431531727314, 'rewards/judge_quality/mean': 0.668749988079071, 'rewards/judge_quality/std': 0.25842589139938354, 'rewards/total_composite/mean': 0.594261884689331, 'rewards/total_composite/std': 0.17329005897045135, 'reward': 0.594261884689331, 'reward_std': 0.17329005897045135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15703284740447998, 'sampling/sampling_logp_difference/max': 1.3353967666625977, 'sampling/importance_sampling_ratio/min': 0.2630538046360016, 'sampling/importance_sampling_ratio/mean': 0.9977946877479553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.869310550391674, 'clip_ratio/low_mean': 0.046516415663063526, 'clip_ratio/low_min': 0.046516415663063526, 'clip_ratio/high_mean': 0.0637398287653923, 'clip_ratio/high_max': 0.0637398287653923, 'clip_ratio/region_mean': 0.11025624442845583, 'reward_total_mean': 0.594261884689331, 'reward_meter_mean': 0.5743567943572998, 'reward_meter_std': 0.36496245861053467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9650839567184448, 'reward_repeat_soft_std': 0.0464431531727314, 'reward_judge_quality_mean': 0.668749988079071, 'reward_judge_quality_std': 0.25842589139938354, 'reward_total_composite_mean': 0.594261884689331, 'reward_total_composite_std': 0.17329005897045135, 'epoch': 0.06}

 19%|█▉        | 637/3300 [1:36:26<5:52:00,  7.93s/it]INFO 04-13 08:57:24 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 638/3300 [1:36:32<5:29:36,  7.43s/it]2026-04-13 08:57:30,593 | INFO | train_grpo_train | metrics_logged mode=train step=638

                                                      
{'loss': -0.0425, 'grad_norm': 12.020691871643066, 'learning_rate': 8.069696969696971e-06, 'num_tokens': 1128476.0, 'completions/mean_length': 58.25, 'completions/min_length': 46.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5777385234832764, 'rewards/meter/std': 0.3327678442001343, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7816901206970215, 'rewards/repeat_soft/std': 0.17234176397323608, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.483951210975647, 'rewards/total_composite/std': 0.14631693065166473, 'reward': 0.483951210975647, 'reward_std': 0.14631694555282593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15901510417461395, 'sampling/sampling_logp_difference/max': 1.4788970947265625, 'sampling/importance_sampling_ratio/min': 0.22788889706134796, 'sampling/importance_sampling_ratio/mean': 1.0201103687286377, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9441844969987869, 'clip_ratio/low_mean': 0.07208479847759008, 'clip_ratio/low_min': 0.07208479847759008, 'clip_ratio/high_mean': 0.07776338700205088, 'clip_ratio/high_max': 0.07776338700205088, 'clip_ratio/region_mean': 0.14984818547964096, 'reward_total_mean': 0.483951210975647, 'reward_meter_mean': 0.5777385234832764, 'reward_meter_std': 0.3327678442001343, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7816901206970215, 'reward_repeat_soft_std': 0.17234176397323608, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.483951210975647, 'reward_total_composite_std': 0.14631693065166473, 'epoch': 0.06}

 19%|█▉        | 638/3300 [1:36:32<5:29:36,  7.43s/it]INFO 04-13 08:57:30 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 639/3300 [1:36:44<6:22:26,  8.62s/it]2026-04-13 08:57:42,006 | INFO | train_grpo_train | metrics_logged mode=train step=639

                                                      
{'loss': -0.1101, 'grad_norm': 3.4394798278808594, 'learning_rate': 8.066666666666667e-06, 'num_tokens': 1130121.0, 'completions/mean_length': 97.625, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9434645771980286, 'rewards/meter/std': 0.0684240534901619, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9191547632217407, 'rewards/repeat_soft/std': 0.07343998551368713, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.2546110451221466, 'rewards/total_composite/mean': 0.580635666847229, 'rewards/total_composite/std': 0.2574915587902069, 'reward': 0.580635666847229, 'reward_std': 0.2574915587902069, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1892760843038559, 'sampling/sampling_logp_difference/max': 1.8836870193481445, 'sampling/importance_sampling_ratio/min': 0.15202854573726654, 'sampling/importance_sampling_ratio/mean': 1.0117923021316528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4847816228866577, 'clip_ratio/low_mean': 0.0243055559694767, 'clip_ratio/low_min': 0.0243055559694767, 'clip_ratio/high_mean': 0.1251265201717615, 'clip_ratio/high_max': 0.1251265201717615, 'clip_ratio/region_mean': 0.1494320761412382, 'reward_total_mean': 0.580635666847229, 'reward_meter_mean': 0.9434645771980286, 'reward_meter_std': 0.0684240534901619, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9191547632217407, 'reward_repeat_soft_std': 0.07343998551368713, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.2546110451221466, 'reward_total_composite_mean': 0.580635666847229, 'reward_total_composite_std': 0.2574915587902069, 'epoch': 0.06}

 19%|█▉        | 639/3300 [1:36:44<6:22:26,  8.62s/it]INFO 04-13 08:57:42 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 640/3300 [1:36:51<5:58:32,  8.09s/it]2026-04-13 08:57:48,842 | INFO | train_grpo_train | metrics_logged mode=train step=640

                                                      
{'loss': 0.0052, 'grad_norm': 9.830313682556152, 'learning_rate': 8.063636363636364e-06, 'num_tokens': 1132487.0, 'completions/mean_length': 92.75, 'completions/min_length': 83.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.8226418495178223, 'rewards/meter/std': 0.21657024323940277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8834623098373413, 'rewards/repeat_soft/std': 0.09841935336589813, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5209257006645203, 'rewards/total_composite/std': 0.23504388332366943, 'reward': 0.5209257006645203, 'reward_std': 0.23504388332366943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16420626640319824, 'sampling/sampling_logp_difference/max': 1.6748542785644531, 'sampling/importance_sampling_ratio/min': 0.18733547627925873, 'sampling/importance_sampling_ratio/mean': 1.0051250457763672, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0939573794603348, 'clip_ratio/low_mean': 0.05977832060307264, 'clip_ratio/low_min': 0.05977832060307264, 'clip_ratio/high_mean': 0.117831876501441, 'clip_ratio/high_max': 0.117831876501441, 'clip_ratio/region_mean': 0.17761019710451365, 'reward_total_mean': 0.5209257006645203, 'reward_meter_mean': 0.8226418495178223, 'reward_meter_std': 0.21657024323940277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8834623098373413, 'reward_repeat_soft_std': 0.09841935336589813, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5209257006645203, 'reward_total_composite_std': 0.23504388332366943, 'epoch': 0.06}

 19%|█▉        | 640/3300 [1:36:51<5:58:32,  8.09s/it]INFO 04-13 08:57:49 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 641/3300 [1:36:57<5:30:29,  7.46s/it]2026-04-13 08:57:54,825 | INFO | train_grpo_train | metrics_logged mode=train step=641

                                                      
{'loss': -0.0035, 'grad_norm': 15.511313438415527, 'learning_rate': 8.060606060606061e-06, 'num_tokens': 1134063.0, 'completions/mean_length': 26.0, 'completions/min_length': 23.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7776494026184082, 'rewards/meter/std': 0.30838698148727417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9664133191108704, 'rewards/repeat_soft/std': 0.046133968979120255, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6498969197273254, 'rewards/total_composite/std': 0.16817516088485718, 'reward': 0.6498969197273254, 'reward_std': 0.16817517578601837, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10660341382026672, 'sampling/sampling_logp_difference/max': 1.3756260871887207, 'sampling/importance_sampling_ratio/min': 0.25268134474754333, 'sampling/importance_sampling_ratio/mean': 1.0160659551620483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.819943368434906, 'clip_ratio/low_mean': 0.07618798967450857, 'clip_ratio/low_min': 0.07618798967450857, 'clip_ratio/high_mean': 0.02855603490024805, 'clip_ratio/high_max': 0.02855603490024805, 'clip_ratio/region_mean': 0.10474402457475662, 'reward_total_mean': 0.6498969197273254, 'reward_meter_mean': 0.7776494026184082, 'reward_meter_std': 0.30838698148727417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9664133191108704, 'reward_repeat_soft_std': 0.046133968979120255, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6498969197273254, 'reward_total_composite_std': 0.16817516088485718, 'epoch': 0.06}

 19%|█▉        | 641/3300 [1:36:57<5:30:29,  7.46s/it]INFO 04-13 08:57:55 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 642/3300 [1:37:03<5:20:08,  7.23s/it]2026-04-13 08:58:01,517 | INFO | train_grpo_train | metrics_logged mode=train step=642

                                                      
{'loss': -0.0551, 'grad_norm': 13.684636116027832, 'learning_rate': 8.057575757575759e-06, 'num_tokens': 1135927.0, 'completions/mean_length': 68.0, 'completions/min_length': 54.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.47336992621421814, 'rewards/meter/std': 0.39060527086257935, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8227493166923523, 'rewards/repeat_soft/std': 0.12701508402824402, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.35991573333740234, 'rewards/total_composite/std': 0.17331789433956146, 'reward': 0.35991573333740234, 'reward_std': 0.17331789433956146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1437559574842453, 'sampling/sampling_logp_difference/max': 2.008169174194336, 'sampling/importance_sampling_ratio/min': 0.1342342048883438, 'sampling/importance_sampling_ratio/mean': 1.015454649925232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.000231422483921, 'clip_ratio/low_mean': 0.09347231686115265, 'clip_ratio/low_min': 0.09347231686115265, 'clip_ratio/high_mean': 0.08196821622550488, 'clip_ratio/high_max': 0.08196821622550488, 'clip_ratio/region_mean': 0.17544053308665752, 'reward_total_mean': 0.35991573333740234, 'reward_meter_mean': 0.47336992621421814, 'reward_meter_std': 0.39060527086257935, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8227493166923523, 'reward_repeat_soft_std': 0.12701508402824402, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.35991573333740234, 'reward_total_composite_std': 0.17331789433956146, 'epoch': 0.06}

 19%|█▉        | 642/3300 [1:37:03<5:20:08,  7.23s/it]INFO 04-13 08:58:01 [block_pool.py:378] Successfully reset prefix cache

 19%|█▉        | 643/3300 [1:37:14<6:11:58,  8.40s/it]2026-04-13 08:58:12,653 | INFO | train_grpo_train | metrics_logged mode=train step=643

                                                      
{'loss': -0.0572, 'grad_norm': 3.3506882190704346, 'learning_rate': 8.054545454545454e-06, 'num_tokens': 1137310.0, 'completions/mean_length': 79.875, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 18.142858505249023, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.48889678716659546, 'rewards/meter/std': 0.41893470287323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9566287994384766, 'rewards/repeat_soft/std': 0.016606280580163002, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.4388635754585266, 'rewards/total_composite/std': 0.20644930005073547, 'reward': 0.4388635754585266, 'reward_std': 0.20644930005073547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1394256204366684, 'sampling/sampling_logp_difference/max': 1.1767903566360474, 'sampling/importance_sampling_ratio/min': 0.3082665801048279, 'sampling/importance_sampling_ratio/mean': 1.0233968496322632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8707086369395256, 'clip_ratio/low_mean': 0.055800653994083405, 'clip_ratio/low_min': 0.055800653994083405, 'clip_ratio/high_mean': 0.09032346680760384, 'clip_ratio/high_max': 0.09032346680760384, 'clip_ratio/region_mean': 0.14612412080168724, 'reward_total_mean': 0.4388635754585266, 'reward_meter_mean': 0.48889678716659546, 'reward_meter_std': 0.41893470287323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9566287994384766, 'reward_repeat_soft_std': 0.016606280580163002, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.4388635754585266, 'reward_total_composite_std': 0.20644930005073547, 'epoch': 0.06}

 19%|█▉        | 643/3300 [1:37:14<6:11:58,  8.40s/it]INFO 04-13 08:58:12 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 644/3300 [1:37:22<6:01:30,  8.17s/it]2026-04-13 08:58:20,273 | INFO | train_grpo_train | metrics_logged mode=train step=644

                                                      
{'loss': 0.1479, 'grad_norm': 12.533270835876465, 'learning_rate': 8.051515151515153e-06, 'num_tokens': 1139727.0, 'completions/mean_length': 114.125, 'completions/min_length': 86.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.125, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.6617348790168762, 'rewards/meter/std': 0.27322185039520264, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.27368009090423584, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7496739625930786, 'rewards/repeat_soft/std': 0.1311352699995041, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.48523104190826416, 'rewards/total_composite/std': 0.1867832988500595, 'reward': 0.48523104190826416, 'reward_std': 0.1867832988500595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1406610608100891, 'sampling/sampling_logp_difference/max': 3.864229202270508, 'sampling/importance_sampling_ratio/min': 0.02097908779978752, 'sampling/importance_sampling_ratio/mean': 0.9994625449180603, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6350109204649925, 'clip_ratio/low_mean': 0.06876762956380844, 'clip_ratio/low_min': 0.06876762956380844, 'clip_ratio/high_mean': 0.06263576168566942, 'clip_ratio/high_max': 0.06263576168566942, 'clip_ratio/region_mean': 0.13140339124947786, 'reward_total_mean': 0.48523104190826416, 'reward_meter_mean': 0.6617348790168762, 'reward_meter_std': 0.27322185039520264, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.27368009090423584, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7496739625930786, 'reward_repeat_soft_std': 0.1311352699995041, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.48523104190826416, 'reward_total_composite_std': 0.1867832988500595, 'epoch': 0.06}

 20%|█▉        | 644/3300 [1:37:22<6:01:30,  8.17s/it]INFO 04-13 08:58:20 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 645/3300 [1:37:29<5:43:08,  7.75s/it]2026-04-13 08:58:27,066 | INFO | train_grpo_train | metrics_logged mode=train step=645

                                                      
{'loss': 0.0746, 'grad_norm': 10.910260200500488, 'learning_rate': 8.048484848484849e-06, 'num_tokens': 1141484.0, 'completions/mean_length': 53.625, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.625, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.27481338381767273, 'rewards/meter/std': 0.3332115411758423, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7180003523826599, 'rewards/repeat_soft/std': 0.19509495794773102, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.40990525484085083, 'rewards/total_composite/std': 0.16834187507629395, 'reward': 0.40990525484085083, 'reward_std': 0.16834186017513275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1587858498096466, 'sampling/sampling_logp_difference/max': 1.695112705230713, 'sampling/importance_sampling_ratio/min': 0.18357853591442108, 'sampling/importance_sampling_ratio/mean': 1.024529218673706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0420835763216019, 'clip_ratio/low_mean': 0.0652546074707061, 'clip_ratio/low_min': 0.0652546074707061, 'clip_ratio/high_mean': 0.07155228778719902, 'clip_ratio/high_max': 0.07155228778719902, 'clip_ratio/region_mean': 0.13680689525790513, 'reward_total_mean': 0.40990525484085083, 'reward_meter_mean': 0.27481338381767273, 'reward_meter_std': 0.3332115411758423, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7180003523826599, 'reward_repeat_soft_std': 0.19509495794773102, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.40990525484085083, 'reward_total_composite_std': 0.16834187507629395, 'epoch': 0.06}

 20%|█▉        | 645/3300 [1:37:29<5:43:08,  7.75s/it]INFO 04-13 08:58:27 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 646/3300 [1:37:35<5:22:22,  7.29s/it]2026-04-13 08:58:33,266 | INFO | train_grpo_train | metrics_logged mode=train step=646

                                                      
{'loss': 0.0154, 'grad_norm': 16.53147315979004, 'learning_rate': 8.045454545454546e-06, 'num_tokens': 1142892.0, 'completions/mean_length': 31.0, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8005962371826172, 'rewards/meter/std': 0.32852140069007874, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9661219716072083, 'rewards/repeat_soft/std': 0.03351852297782898, 'rewards/judge_quality/mean': 0.5225000381469727, 'rewards/judge_quality/std': 0.24598202109336853, 'rewards/total_composite/mean': 0.5804761052131653, 'rewards/total_composite/std': 0.29398468136787415, 'reward': 0.5804761052131653, 'reward_std': 0.29398468136787415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.131461963057518, 'sampling/sampling_logp_difference/max': 1.008401870727539, 'sampling/importance_sampling_ratio/min': 0.3648015260696411, 'sampling/importance_sampling_ratio/mean': 1.0192826986312866, 'sampling/importance_sampling_ratio/max': 1.921644687652588, 'entropy': 1.0241151303052902, 'clip_ratio/low_mean': 0.025862068869173527, 'clip_ratio/low_min': 0.025862068869173527, 'clip_ratio/high_mean': 0.0889372294768691, 'clip_ratio/high_max': 0.0889372294768691, 'clip_ratio/region_mean': 0.11479929834604263, 'reward_total_mean': 0.5804761052131653, 'reward_meter_mean': 0.8005962371826172, 'reward_meter_std': 0.32852140069007874, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9661219716072083, 'reward_repeat_soft_std': 0.03351852297782898, 'reward_judge_quality_mean': 0.5225000381469727, 'reward_judge_quality_std': 0.24598202109336853, 'reward_total_composite_mean': 0.5804761052131653, 'reward_total_composite_std': 0.29398468136787415, 'epoch': 0.06}

 20%|█▉        | 646/3300 [1:37:35<5:22:22,  7.29s/it]INFO 04-13 08:58:33 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 647/3300 [1:37:42<5:16:39,  7.16s/it]2026-04-13 08:58:40,132 | INFO | train_grpo_train | metrics_logged mode=train step=647

                                                      
{'loss': 0.023, 'grad_norm': 15.462921142578125, 'learning_rate': 8.042424242424243e-06, 'num_tokens': 1145026.0, 'completions/mean_length': 73.75, 'completions/min_length': 60.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.5346683263778687, 'rewards/meter/std': 0.4203801155090332, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8131506443023682, 'rewards/repeat_soft/std': 0.07363694161176682, 'rewards/judge_quality/mean': 0.6274999976158142, 'rewards/judge_quality/std': 0.18013885617256165, 'rewards/total_composite/mean': 0.517020046710968, 'rewards/total_composite/std': 0.1653960943222046, 'reward': 0.517020046710968, 'reward_std': 0.1653960943222046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17536382377147675, 'sampling/sampling_logp_difference/max': 2.880128860473633, 'sampling/importance_sampling_ratio/min': 0.056127529591321945, 'sampling/importance_sampling_ratio/mean': 1.0183418989181519, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9807343557476997, 'clip_ratio/low_mean': 0.07348845060914755, 'clip_ratio/low_min': 0.07348845060914755, 'clip_ratio/high_mean': 0.06786853261291981, 'clip_ratio/high_max': 0.06786853261291981, 'clip_ratio/region_mean': 0.14135698322206736, 'reward_total_mean': 0.517020046710968, 'reward_meter_mean': 0.5346683263778687, 'reward_meter_std': 0.4203801155090332, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8131506443023682, 'reward_repeat_soft_std': 0.07363694161176682, 'reward_judge_quality_mean': 0.6274999976158142, 'reward_judge_quality_std': 0.18013885617256165, 'reward_total_composite_mean': 0.517020046710968, 'reward_total_composite_std': 0.1653960943222046, 'epoch': 0.06}

 20%|█▉        | 647/3300 [1:37:42<5:16:39,  7.16s/it]INFO 04-13 08:58:40 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 648/3300 [1:37:48<4:59:58,  6.79s/it]2026-04-13 08:58:46,044 | INFO | train_grpo_train | metrics_logged mode=train step=648

                                                      
{'loss': 0.082, 'grad_norm': 18.118453979492188, 'learning_rate': 8.03939393939394e-06, 'num_tokens': 1146451.0, 'completions/mean_length': 35.125, 'completions/min_length': 28.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.47180289030075073, 'rewards/meter/std': 0.3639613389968872, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9971110224723816, 'rewards/repeat_soft/std': 0.002904646098613739, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5076788663864136, 'rewards/total_composite/std': 0.11163168400526047, 'reward': 0.5076788663864136, 'reward_std': 0.11163167655467987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17425420880317688, 'sampling/sampling_logp_difference/max': 1.1415696144104004, 'sampling/importance_sampling_ratio/min': 0.31931743025779724, 'sampling/importance_sampling_ratio/mean': 1.0409749746322632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1029235646128654, 'clip_ratio/low_mean': 0.1174242440611124, 'clip_ratio/low_min': 0.1174242440611124, 'clip_ratio/high_mean': 0.07134091667830944, 'clip_ratio/high_max': 0.07134091667830944, 'clip_ratio/region_mean': 0.18876516073942184, 'reward_total_mean': 0.5076788663864136, 'reward_meter_mean': 0.47180289030075073, 'reward_meter_std': 0.3639613389968872, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9971110224723816, 'reward_repeat_soft_std': 0.002904646098613739, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5076788663864136, 'reward_total_composite_std': 0.11163168400526047, 'epoch': 0.07}

 20%|█▉        | 648/3300 [1:37:48<4:59:58,  6.79s/it]INFO 04-13 08:58:46 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 649/3300 [1:37:54<4:47:44,  6.51s/it]2026-04-13 08:58:51,920 | INFO | train_grpo_train | metrics_logged mode=train step=649

                                                      
{'loss': 0.021, 'grad_norm': 22.424516677856445, 'learning_rate': 8.036363636363636e-06, 'num_tokens': 1148153.0, 'completions/mean_length': 33.75, 'completions/min_length': 29.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.623406708240509, 'rewards/meter/std': 0.38744696974754333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.971734881401062, 'rewards/repeat_soft/std': 0.04234902188181877, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6547472476959229, 'rewards/total_composite/std': 0.24986574053764343, 'reward': 0.6547472476959229, 'reward_std': 0.24986572563648224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15227848291397095, 'sampling/sampling_logp_difference/max': 0.9737920761108398, 'sampling/importance_sampling_ratio/min': 0.3776482343673706, 'sampling/importance_sampling_ratio/mean': 1.0295783281326294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1171982809901237, 'clip_ratio/low_mean': 0.10233087697997689, 'clip_ratio/low_min': 0.10233087697997689, 'clip_ratio/high_mean': 0.06856060773134232, 'clip_ratio/high_max': 0.06856060773134232, 'clip_ratio/region_mean': 0.1708914847113192, 'reward_total_mean': 0.6547472476959229, 'reward_meter_mean': 0.623406708240509, 'reward_meter_std': 0.38744696974754333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.971734881401062, 'reward_repeat_soft_std': 0.04234902188181877, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6547472476959229, 'reward_total_composite_std': 0.24986574053764343, 'epoch': 0.07}

 20%|█▉        | 649/3300 [1:37:54<4:47:44,  6.51s/it]INFO 04-13 08:58:52 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 650/3300 [1:38:00<4:50:50,  6.58s/it]2026-04-13 08:58:58,669 | INFO | train_grpo_train | metrics_logged mode=train step=650

                                                      
{'loss': 0.0298, 'grad_norm': 14.326630592346191, 'learning_rate': 8.033333333333335e-06, 'num_tokens': 1149938.0, 'completions/mean_length': 59.125, 'completions/min_length': 54.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8271393179893494, 'rewards/meter/std': 0.16652563214302063, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8895382881164551, 'rewards/repeat_soft/std': 0.06680367887020111, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.5602148175239563, 'rewards/total_composite/std': 0.2508557438850403, 'reward': 0.5602148175239563, 'reward_std': 0.2508557438850403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19041897356510162, 'sampling/sampling_logp_difference/max': 2.1767501831054688, 'sampling/importance_sampling_ratio/min': 0.11340949684381485, 'sampling/importance_sampling_ratio/mean': 1.004520058631897, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2831703871488571, 'clip_ratio/low_mean': 0.057008782401680946, 'clip_ratio/low_min': 0.057008782401680946, 'clip_ratio/high_mean': 0.1350773647427559, 'clip_ratio/high_max': 0.1350773647427559, 'clip_ratio/region_mean': 0.19208614714443684, 'reward_total_mean': 0.5602148175239563, 'reward_meter_mean': 0.8271393179893494, 'reward_meter_std': 0.16652563214302063, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8895382881164551, 'reward_repeat_soft_std': 0.06680367887020111, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.5602148175239563, 'reward_total_composite_std': 0.2508557438850403, 'epoch': 0.07}

 20%|█▉        | 650/3300 [1:38:00<4:50:50,  6.58s/it]INFO 04-13 08:58:58 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:13,  1.66s/it]

 30%|███       | 3/10 [00:06<00:17,  2.50s/it]

 40%|████      | 4/10 [00:09<00:16,  2.68s/it]

 50%|█████     | 5/10 [00:13<00:15,  3.03s/it]

 60%|██████    | 6/10 [00:17<00:13,  3.25s/it]

 70%|███████   | 7/10 [00:26<00:15,  5.03s/it]

 80%|████████  | 8/10 [00:29<00:09,  4.66s/it]

 90%|█████████ | 9/10 [00:33<00:04,  4.24s/it]

100%|██████████| 10/10 [00:36<00:00,  3.90s/it]2026-04-13 08:59:44,348 | INFO | train_grpo_train | metrics_logged mode=eval step=650

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 45.6379, 'eval_samples_per_second': 1.753, 'eval_steps_per_second': 0.219, 'eval_num_tokens': 1149938.0, 'eval_completions/mean_length': 65.9625, 'eval_completions/min_length': 28.2, 'eval_completions/max_length': 173.4, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 54.728572082519534, 'eval_completions/min_terminated_length': 28.2, 'eval_completions/max_terminated_length': 91.8, 'eval_rewards/meter/mean': 0.5899448305368423, 'eval_rewards/meter/std': 0.35944695919752123, 'eval_rewards/count_adherence/mean': 0.9735416650772095, 'eval_rewards/count_adherence/std': 0.060814641788601874, 'eval_rewards/hard_gate/mean': 0.9, 'eval_rewards/hard_gate/std': 0.23400336503982544, 'eval_rewards/repeat_soft/mean': 0.8852393925189972, 'eval_rewards/repeat_soft/std': 0.11525812894105911, 'eval_rewards/judge_quality/mean': 0.48649999797344207, 'eval_rewards/judge_quality/std': 0.1396345805376768, 'eval_rewards/total_composite/mean': 0.4717728167772293, 'eval_rewards/total_composite/std': 0.18849601075053216, 'eval_reward': 0.4717728167772293, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0842033289372921, 'eval_sampling/sampling_logp_difference/max': 0.9717438220977783, 'eval_sampling/importance_sampling_ratio/min': 0.38339222967624664, 'eval_sampling/importance_sampling_ratio/mean': 1.019832593202591, 'eval_sampling/importance_sampling_ratio/max': 1.474166202545166, 'eval_entropy': 0.9724018573760986, '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': 0.4717728167772293, 'eval_reward_meter_mean': 0.5899448305368423, 'eval_reward_meter_std': 0.35944695919752123, 'eval_reward_count_adherence_mean': 0.9735416650772095, 'eval_reward_count_adherence_std': 0.060814641788601874, 'eval_reward_hard_gate_mean': 0.9, 'eval_reward_hard_gate_std': 0.23400336503982544, 'eval_reward_repeat_soft_mean': 0.8852393925189972, 'eval_reward_repeat_soft_std': 0.11525812894105911, 'eval_reward_judge_quality_mean': 0.48649999797344207, 'eval_reward_judge_quality_std': 0.1396345805376768, 'eval_reward_total_composite_mean': 0.4717728167772293, 'eval_reward_total_composite_std': 0.18849601075053216, 'epoch': 0.07}

 20%|█▉        | 650/3300 [1:38:46<4:50:50,  6.58s/it]

100%|██████████| 10/10 [00:36<00:00,  3.90s/it]

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

 20%|█▉        | 651/3300 [1:38:56<15:35:16, 21.18s/it]2026-04-13 08:59:53,922 | INFO | train_grpo_train | metrics_logged mode=train step=651

                                                       
{'loss': 0.0216, 'grad_norm': 12.091530799865723, 'learning_rate': 8.03030303030303e-06, 'num_tokens': 1151871.0, 'completions/mean_length': 48.625, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.981431245803833, 'rewards/meter/std': 0.007213943172246218, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9497766494750977, 'rewards/repeat_soft/std': 0.03229496628046036, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.610397219657898, 'rewards/total_composite/std': 0.006398757919669151, 'reward': 0.610397219657898, 'reward_std': 0.006398768164217472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15845441818237305, 'sampling/sampling_logp_difference/max': 1.833327293395996, 'sampling/importance_sampling_ratio/min': 0.15988071262836456, 'sampling/importance_sampling_ratio/mean': 1.0219941139221191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8763070181012154, 'clip_ratio/low_mean': 0.03524390235543251, 'clip_ratio/low_min': 0.03524390235543251, 'clip_ratio/high_mean': 0.13669268880039454, 'clip_ratio/high_max': 0.13669268880039454, 'clip_ratio/region_mean': 0.17193659115582705, 'reward_total_mean': 0.610397219657898, 'reward_meter_mean': 0.981431245803833, 'reward_meter_std': 0.007213943172246218, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9497766494750977, 'reward_repeat_soft_std': 0.03229496628046036, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.610397219657898, 'reward_total_composite_std': 0.006398757919669151, 'epoch': 0.07}

 20%|█▉        | 651/3300 [1:38:56<15:35:16, 21.18s/it]INFO 04-13 08:59:54 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 652/3300 [1:39:02<12:16:15, 16.68s/it]2026-04-13 09:00:00,098 | INFO | train_grpo_train | metrics_logged mode=train step=652

                                                       
{'loss': 0.0362, 'grad_norm': 12.559391975402832, 'learning_rate': 8.027272727272728e-06, 'num_tokens': 1153579.0, 'completions/mean_length': 52.5, 'completions/min_length': 46.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5216165781021118, 'rewards/meter/std': 0.34197962284088135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9827027320861816, 'rewards/repeat_soft/std': 0.0072836256586015224, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.10091545432806015, 'rewards/total_composite/mean': 0.5056251287460327, 'rewards/total_composite/std': 0.24777263402938843, 'reward': 0.5056251287460327, 'reward_std': 0.24777261912822723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18527501821517944, 'sampling/sampling_logp_difference/max': 1.9833526611328125, 'sampling/importance_sampling_ratio/min': 0.13760711252689362, 'sampling/importance_sampling_ratio/mean': 1.0162345170974731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.240620642900467, 'clip_ratio/low_mean': 0.10041659697890282, 'clip_ratio/low_min': 0.10041659697890282, 'clip_ratio/high_mean': 0.06934077944606543, 'clip_ratio/high_max': 0.06934077944606543, 'clip_ratio/region_mean': 0.16975737642496824, 'reward_total_mean': 0.5056251287460327, 'reward_meter_mean': 0.5216165781021118, 'reward_meter_std': 0.34197962284088135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9827027320861816, 'reward_repeat_soft_std': 0.0072836256586015224, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.10091545432806015, 'reward_total_composite_mean': 0.5056251287460327, 'reward_total_composite_std': 0.24777263402938843, 'epoch': 0.07}

 20%|█▉        | 652/3300 [1:39:02<12:16:15, 16.68s/it]INFO 04-13 09:00:00 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 653/3300 [1:39:13<11:07:52, 15.14s/it]2026-04-13 09:00:11,635 | INFO | train_grpo_train | metrics_logged mode=train step=653

                                                       
{'loss': -0.0844, 'grad_norm': 4.386373043060303, 'learning_rate': 8.024242424242425e-06, 'num_tokens': 1155180.0, 'completions/mean_length': 98.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7292758822441101, 'rewards/meter/std': 0.42446523904800415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9710483551025391, 'rewards/repeat_soft/std': 0.023260634392499924, 'rewards/judge_quality/mean': 0.6862499713897705, 'rewards/judge_quality/std': 0.34221702814102173, 'rewards/total_composite/mean': 0.6550273895263672, 'rewards/total_composite/std': 0.3373952805995941, 'reward': 0.6550273895263672, 'reward_std': 0.3373952805995941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1691875159740448, 'sampling/sampling_logp_difference/max': 1.1866025924682617, 'sampling/importance_sampling_ratio/min': 0.30525660514831543, 'sampling/importance_sampling_ratio/mean': 1.006740927696228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2591385692358017, 'clip_ratio/low_mean': 0.05931802652776241, 'clip_ratio/low_min': 0.05931802652776241, 'clip_ratio/high_mean': 0.08195196930319071, 'clip_ratio/high_max': 0.08195196930319071, 'clip_ratio/region_mean': 0.14126999583095312, 'reward_total_mean': 0.6550273895263672, 'reward_meter_mean': 0.7292758822441101, 'reward_meter_std': 0.42446523904800415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9710483551025391, 'reward_repeat_soft_std': 0.023260634392499924, 'reward_judge_quality_mean': 0.6862499713897705, 'reward_judge_quality_std': 0.34221702814102173, 'reward_total_composite_mean': 0.6550273895263672, 'reward_total_composite_std': 0.3373952805995941, 'epoch': 0.07}

 20%|█▉        | 653/3300 [1:39:13<11:07:52, 15.14s/it]INFO 04-13 09:00:11 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 654/3300 [1:39:20<9:13:04, 12.54s/it] 2026-04-13 09:00:18,118 | INFO | train_grpo_train | metrics_logged mode=train step=654

                                                      
{'loss': 0.0471, 'grad_norm': 14.922459602355957, 'learning_rate': 8.021212121212122e-06, 'num_tokens': 1157122.0, 'completions/mean_length': 68.75, 'completions/min_length': 49.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.7787767052650452, 'rewards/meter/std': 0.23661600053310394, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8986475467681885, 'rewards/repeat_soft/std': 0.11464295536279678, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.4754413068294525, 'rewards/total_composite/std': 0.2196243852376938, 'reward': 0.4754413068294525, 'reward_std': 0.2196243703365326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1943458765745163, 'sampling/sampling_logp_difference/max': 2.324733257293701, 'sampling/importance_sampling_ratio/min': 0.09780952334403992, 'sampling/importance_sampling_ratio/mean': 1.0001096725463867, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5427189767360687, 'clip_ratio/low_mean': 0.062207804061472416, 'clip_ratio/low_min': 0.062207804061472416, 'clip_ratio/high_mean': 0.11934742797166109, 'clip_ratio/high_max': 0.11934742797166109, 'clip_ratio/region_mean': 0.1815552320331335, 'reward_total_mean': 0.4754413068294525, 'reward_meter_mean': 0.7787767052650452, 'reward_meter_std': 0.23661600053310394, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8986475467681885, 'reward_repeat_soft_std': 0.11464295536279678, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.4754413068294525, 'reward_total_composite_std': 0.2196243852376938, 'epoch': 0.07}

 20%|█▉        | 654/3300 [1:39:20<9:13:04, 12.54s/it]INFO 04-13 09:00:18 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 655/3300 [1:39:27<7:59:22, 10.87s/it]2026-04-13 09:00:25,102 | INFO | train_grpo_train | metrics_logged mode=train step=655

                                                      
{'loss': -0.0993, 'grad_norm': 13.134452819824219, 'learning_rate': 8.018181818181818e-06, 'num_tokens': 1158873.0, 'completions/mean_length': 51.875, 'completions/min_length': 40.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.22241005301475525, 'rewards/meter/std': 0.26595112681388855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8868178129196167, 'rewards/repeat_soft/std': 0.061087459325790405, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.14957083761692047, 'rewards/total_composite/mean': 0.3530135750770569, 'rewards/total_composite/std': 0.17210303246974945, 'reward': 0.3530135750770569, 'reward_std': 0.17210301756858826, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18812473118305206, 'sampling/sampling_logp_difference/max': 1.8689632415771484, 'sampling/importance_sampling_ratio/min': 0.1542835384607315, 'sampling/importance_sampling_ratio/mean': 1.0281426906585693, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3644793555140495, 'clip_ratio/low_mean': 0.09198097139596939, 'clip_ratio/low_min': 0.09198097139596939, 'clip_ratio/high_mean': 0.05292119272053242, 'clip_ratio/high_max': 0.05292119272053242, 'clip_ratio/region_mean': 0.1449021641165018, 'reward_total_mean': 0.3530135750770569, 'reward_meter_mean': 0.22241005301475525, 'reward_meter_std': 0.26595112681388855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8868178129196167, 'reward_repeat_soft_std': 0.061087459325790405, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.14957083761692047, 'reward_total_composite_mean': 0.3530135750770569, 'reward_total_composite_std': 0.17210303246974945, 'epoch': 0.07}

 20%|█▉        | 655/3300 [1:39:27<7:59:22, 10.87s/it]INFO 04-13 09:00:25 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 656/3300 [1:39:34<7:07:39,  9.70s/it]2026-04-13 09:00:32,076 | INFO | train_grpo_train | metrics_logged mode=train step=656

                                                      
{'loss': 0.0221, 'grad_norm': 12.217185020446777, 'learning_rate': 8.015151515151515e-06, 'num_tokens': 1160927.0, 'completions/mean_length': 68.75, 'completions/min_length': 53.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.75, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.5280009508132935, 'rewards/meter/std': 0.4282487630844116, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8394860029220581, 'rewards/repeat_soft/std': 0.11667388677597046, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.45864760875701904, 'rewards/total_composite/std': 0.13028350472450256, 'reward': 0.45864760875701904, 'reward_std': 0.13028350472450256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16322258114814758, 'sampling/sampling_logp_difference/max': 1.7329931259155273, 'sampling/importance_sampling_ratio/min': 0.17675456404685974, 'sampling/importance_sampling_ratio/mean': 1.0246920585632324, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2120042070746422, 'clip_ratio/low_mean': 0.05625782907009125, 'clip_ratio/low_min': 0.05625782907009125, 'clip_ratio/high_mean': 0.08202130533754826, 'clip_ratio/high_max': 0.08202130533754826, 'clip_ratio/region_mean': 0.1382791344076395, 'reward_total_mean': 0.45864760875701904, 'reward_meter_mean': 0.5280009508132935, 'reward_meter_std': 0.4282487630844116, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8394860029220581, 'reward_repeat_soft_std': 0.11667388677597046, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.45864760875701904, 'reward_total_composite_std': 0.13028350472450256, 'epoch': 0.07}

 20%|█▉        | 656/3300 [1:39:34<7:07:39,  9.70s/it]INFO 04-13 09:00:32 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 657/3300 [1:39:40<6:27:11,  8.79s/it]2026-04-13 09:00:38,731 | INFO | train_grpo_train | metrics_logged mode=train step=657

                                                      
{'loss': -0.0175, 'grad_norm': 14.234102249145508, 'learning_rate': 8.012121212121214e-06, 'num_tokens': 1162766.0, 'completions/mean_length': 54.875, 'completions/min_length': 43.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.5937342643737793, 'rewards/meter/std': 0.2847641110420227, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9064621925354004, 'rewards/repeat_soft/std': 0.09552107751369476, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.40473976731300354, 'rewards/total_composite/std': 0.2639545202255249, 'reward': 0.40473976731300354, 'reward_std': 0.2639545202255249, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17224234342575073, 'sampling/sampling_logp_difference/max': 2.0596227645874023, 'sampling/importance_sampling_ratio/min': 0.12750205397605896, 'sampling/importance_sampling_ratio/mean': 1.0342533588409424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2091957330703735, 'clip_ratio/low_mean': 0.05395756848156452, 'clip_ratio/low_min': 0.05395756848156452, 'clip_ratio/high_mean': 0.11370428279042244, 'clip_ratio/high_max': 0.11370428279042244, 'clip_ratio/region_mean': 0.16766185127198696, 'reward_total_mean': 0.40473976731300354, 'reward_meter_mean': 0.5937342643737793, 'reward_meter_std': 0.2847641110420227, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9064621925354004, 'reward_repeat_soft_std': 0.09552107751369476, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.40473976731300354, 'reward_total_composite_std': 0.2639545202255249, 'epoch': 0.07}

 20%|█▉        | 657/3300 [1:39:40<6:27:11,  8.79s/it]INFO 04-13 09:00:38 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 658/3300 [1:39:47<5:54:21,  8.05s/it]2026-04-13 09:00:45,046 | INFO | train_grpo_train | metrics_logged mode=train step=658

                                                      
{'loss': 0.0362, 'grad_norm': 15.805983543395996, 'learning_rate': 8.00909090909091e-06, 'num_tokens': 1164402.0, 'completions/mean_length': 33.5, 'completions/min_length': 31.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.32758235931396484, 'rewards/meter/std': 0.3652130663394928, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9861834049224854, 'rewards/repeat_soft/std': 0.030291561037302017, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.4663577079772949, 'rewards/total_composite/std': 0.11732984334230423, 'reward': 0.4663577079772949, 'reward_std': 0.11732983589172363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18540358543395996, 'sampling/sampling_logp_difference/max': 1.5263197422027588, 'sampling/importance_sampling_ratio/min': 0.21733404695987701, 'sampling/importance_sampling_ratio/mean': 1.0284390449523926, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1702303737401962, 'clip_ratio/low_mean': 0.11255082860589027, 'clip_ratio/low_min': 0.11255082860589027, 'clip_ratio/high_mean': 0.09381268778815866, 'clip_ratio/high_max': 0.09381268778815866, 'clip_ratio/region_mean': 0.20636351639404893, 'reward_total_mean': 0.4663577079772949, 'reward_meter_mean': 0.32758235931396484, 'reward_meter_std': 0.3652130663394928, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9861834049224854, 'reward_repeat_soft_std': 0.030291561037302017, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.4663577079772949, 'reward_total_composite_std': 0.11732984334230423, 'epoch': 0.07}

 20%|█▉        | 658/3300 [1:39:47<5:54:21,  8.05s/it]INFO 04-13 09:00:45 [block_pool.py:378] Successfully reset prefix cache

 20%|█▉        | 659/3300 [1:39:53<5:24:50,  7.38s/it]2026-04-13 09:00:50,871 | INFO | train_grpo_train | metrics_logged mode=train step=659

                                                      
{'loss': 0.0285, 'grad_norm': 13.327178955078125, 'learning_rate': 8.006060606060607e-06, 'num_tokens': 1165935.0, 'completions/mean_length': 34.625, 'completions/min_length': 31.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7355386018753052, 'rewards/meter/std': 0.33000099658966064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7617665529251099, 'rewards/repeat_soft/std': 0.1503032147884369, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.4926672577857971, 'rewards/total_composite/std': 0.09403003752231598, 'reward': 0.4926672577857971, 'reward_std': 0.09403003752231598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16989152133464813, 'sampling/sampling_logp_difference/max': 2.105639934539795, 'sampling/importance_sampling_ratio/min': 0.12176772207021713, 'sampling/importance_sampling_ratio/mean': 1.024625539779663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1311260685324669, 'clip_ratio/low_mean': 0.06997071579098701, 'clip_ratio/low_min': 0.06997071579098701, 'clip_ratio/high_mean': 0.06208881549537182, 'clip_ratio/high_max': 0.06208881549537182, 'clip_ratio/region_mean': 0.13205953128635883, 'reward_total_mean': 0.4926672577857971, 'reward_meter_mean': 0.7355386018753052, 'reward_meter_std': 0.33000099658966064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7617665529251099, 'reward_repeat_soft_std': 0.1503032147884369, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.4926672577857971, 'reward_total_composite_std': 0.09403003752231598, 'epoch': 0.07}

 20%|█▉        | 659/3300 [1:39:53<5:24:50,  7.38s/it]INFO 04-13 09:00:51 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 660/3300 [1:39:59<5:05:44,  6.95s/it]2026-04-13 09:00:56,813 | INFO | train_grpo_train | metrics_logged mode=train step=660

                                                      
{'loss': 0.0277, 'grad_norm': 16.149572372436523, 'learning_rate': 8.003030303030304e-06, 'num_tokens': 1167513.0, 'completions/mean_length': 30.25, 'completions/min_length': 21.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.536858856678009, 'rewards/meter/std': 0.34027794003486633, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9758846759796143, 'rewards/repeat_soft/std': 0.035142626613378525, 'rewards/judge_quality/mean': 0.7825000286102295, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6137505769729614, 'rewards/total_composite/std': 0.19050054252147675, 'reward': 0.6137505769729614, 'reward_std': 0.19050054252147675, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16642440855503082, 'sampling/sampling_logp_difference/max': 1.8366745710372925, 'sampling/importance_sampling_ratio/min': 0.15934644639492035, 'sampling/importance_sampling_ratio/mean': 0.9946097731590271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8458282127976418, 'clip_ratio/low_mean': 0.05387118645012379, 'clip_ratio/low_min': 0.05387118645012379, 'clip_ratio/high_mean': 0.08504902245476842, 'clip_ratio/high_max': 0.08504902245476842, 'clip_ratio/region_mean': 0.1389202089048922, 'reward_total_mean': 0.6137505769729614, 'reward_meter_mean': 0.536858856678009, 'reward_meter_std': 0.34027794003486633, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9758846759796143, 'reward_repeat_soft_std': 0.035142626613378525, 'reward_judge_quality_mean': 0.7825000286102295, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6137505769729614, 'reward_total_composite_std': 0.19050054252147675, 'epoch': 0.07}

 20%|██        | 660/3300 [1:39:59<5:05:44,  6.95s/it]INFO 04-13 09:00:57 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 661/3300 [1:40:04<4:49:22,  6.58s/it]2026-04-13 09:01:02,536 | INFO | train_grpo_train | metrics_logged mode=train step=661

                                                      
{'loss': 0.0367, 'grad_norm': 21.243154525756836, 'learning_rate': 8.000000000000001e-06, 'num_tokens': 1168785.0, 'completions/mean_length': 20.0, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9752902388572693, 'rewards/meter/std': 0.026702914386987686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.617825984954834, 'rewards/total_composite/std': 0.013858720660209656, 'reward': 0.617825984954834, 'reward_std': 0.01385872345417738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2052830010652542, 'sampling/sampling_logp_difference/max': 1.6482620239257812, 'sampling/importance_sampling_ratio/min': 0.19238397479057312, 'sampling/importance_sampling_ratio/mean': 1.0423879623413086, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2811303362250328, 'clip_ratio/low_mean': 0.10084818955510855, 'clip_ratio/low_min': 0.10084818955510855, 'clip_ratio/high_mean': 0.04298563348129392, 'clip_ratio/high_max': 0.04298563348129392, 'clip_ratio/region_mean': 0.14383382303640246, 'reward_total_mean': 0.617825984954834, 'reward_meter_mean': 0.9752902388572693, 'reward_meter_std': 0.026702914386987686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.617825984954834, 'reward_total_composite_std': 0.013858720660209656, 'epoch': 0.07}

 20%|██        | 661/3300 [1:40:04<4:49:22,  6.58s/it]INFO 04-13 09:01:02 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 662/3300 [1:40:10<4:44:54,  6.48s/it]2026-04-13 09:01:08,782 | INFO | train_grpo_train | metrics_logged mode=train step=662

                                                      
{'loss': 0.0195, 'grad_norm': 15.285932540893555, 'learning_rate': 7.996969696969697e-06, 'num_tokens': 1170839.0, 'completions/mean_length': 60.75, 'completions/min_length': 46.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8252530097961426, 'rewards/meter/std': 0.16105490922927856, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8301280736923218, 'rewards/repeat_soft/std': 0.1212448701262474, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5117673873901367, 'rewards/total_composite/std': 0.2360391467809677, 'reward': 0.5117673873901367, 'reward_std': 0.2360391467809677, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1757470816373825, 'sampling/sampling_logp_difference/max': 2.6512575149536133, 'sampling/importance_sampling_ratio/min': 0.07056242227554321, 'sampling/importance_sampling_ratio/mean': 1.018597960472107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1257533058524132, 'clip_ratio/low_mean': 0.05017806403338909, 'clip_ratio/low_min': 0.05017806403338909, 'clip_ratio/high_mean': 0.08945727348327637, 'clip_ratio/high_max': 0.08945727348327637, 'clip_ratio/region_mean': 0.13963533751666546, 'reward_total_mean': 0.5117673873901367, 'reward_meter_mean': 0.8252530097961426, 'reward_meter_std': 0.16105490922927856, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8301280736923218, 'reward_repeat_soft_std': 0.1212448701262474, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5117673873901367, 'reward_total_composite_std': 0.2360391467809677, 'epoch': 0.07}

 20%|██        | 662/3300 [1:40:11<4:44:54,  6.48s/it]INFO 04-13 09:01:09 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 663/3300 [1:40:23<5:59:06,  8.17s/it]2026-04-13 09:01:20,892 | INFO | train_grpo_train | metrics_logged mode=train step=663

                                                      
{'loss': -0.2039, 'grad_norm': 2.632770299911499, 'learning_rate': 7.993939393939396e-06, 'num_tokens': 1173615.0, 'completions/mean_length': 216.0, 'completions/min_length': 112.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 117.33333587646484, 'completions/min_terminated_length': 112.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.6868778467178345, 'rewards/meter/std': 0.37099072337150574, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9341126680374146, 'rewards/repeat_soft/std': 0.03347941115498543, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.22025957703590393, 'rewards/total_composite/mean': 0.42581695318222046, 'rewards/total_composite/std': 0.28781354427337646, 'reward': 0.42581695318222046, 'reward_std': 0.28781354427337646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13848777115345, 'sampling/sampling_logp_difference/max': 3.1296353340148926, 'sampling/importance_sampling_ratio/min': 0.04373374208807945, 'sampling/importance_sampling_ratio/mean': 1.0213404893875122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9202366322278976, 'clip_ratio/low_mean': 0.01508620660752058, 'clip_ratio/low_min': 0.01508620660752058, 'clip_ratio/high_mean': 0.08940434642136097, 'clip_ratio/high_max': 0.08940434642136097, 'clip_ratio/region_mean': 0.10449055302888155, 'reward_total_mean': 0.42581695318222046, 'reward_meter_mean': 0.6868778467178345, 'reward_meter_std': 0.37099072337150574, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9341126680374146, 'reward_repeat_soft_std': 0.03347941115498543, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.22025957703590393, 'reward_total_composite_mean': 0.42581695318222046, 'reward_total_composite_std': 0.28781354427337646, 'epoch': 0.07}

 20%|██        | 663/3300 [1:40:23<5:59:06,  8.17s/it]INFO 04-13 09:01:21 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 664/3300 [1:40:29<5:34:13,  7.61s/it]2026-04-13 09:01:27,267 | INFO | train_grpo_train | metrics_logged mode=train step=664

                                                      
{'loss': 0.0502, 'grad_norm': 15.893335342407227, 'learning_rate': 7.990909090909091e-06, 'num_tokens': 1175225.0, 'completions/mean_length': 35.25, 'completions/min_length': 28.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5234280824661255, 'rewards/meter/std': 0.4268744885921478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9757056832313538, 'rewards/repeat_soft/std': 0.03202519565820694, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.653202474117279, 'rewards/total_composite/std': 0.2629137635231018, 'reward': 0.653202474117279, 'reward_std': 0.2629137635231018, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1433289796113968, 'sampling/sampling_logp_difference/max': 1.4580602645874023, 'sampling/importance_sampling_ratio/min': 0.23268719017505646, 'sampling/importance_sampling_ratio/mean': 1.003405213356018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8421033546328545, 'clip_ratio/low_mean': 0.06581361405551434, 'clip_ratio/low_min': 0.06581361405551434, 'clip_ratio/high_mean': 0.05746727483347058, 'clip_ratio/high_max': 0.05746727483347058, 'clip_ratio/region_mean': 0.12328088888898492, 'reward_total_mean': 0.653202474117279, 'reward_meter_mean': 0.5234280824661255, 'reward_meter_std': 0.4268744885921478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9757056832313538, 'reward_repeat_soft_std': 0.03202519565820694, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.653202474117279, 'reward_total_composite_std': 0.2629137635231018, 'epoch': 0.07}

 20%|██        | 664/3300 [1:40:29<5:34:13,  7.61s/it]INFO 04-13 09:01:27 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 665/3300 [1:40:35<5:12:56,  7.13s/it]2026-04-13 09:01:33,188 | INFO | train_grpo_train | metrics_logged mode=train step=665

                                                      
{'loss': 0.0233, 'grad_norm': 16.9578857421875, 'learning_rate': 7.987878787878789e-06, 'num_tokens': 1176813.0, 'completions/mean_length': 39.5, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7170044183731079, 'rewards/meter/std': 0.35669296979904175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9703220725059509, 'rewards/repeat_soft/std': 0.03601430356502533, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.7743169665336609, 'rewards/total_composite/std': 0.21356315910816193, 'reward': 0.7743169665336609, 'reward_std': 0.21356317400932312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15803343057632446, 'sampling/sampling_logp_difference/max': 1.2026662826538086, 'sampling/importance_sampling_ratio/min': 0.300392210483551, 'sampling/importance_sampling_ratio/mean': 1.0351260900497437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0985603258013725, 'clip_ratio/low_mean': 0.03826013393700123, 'clip_ratio/low_min': 0.03826013393700123, 'clip_ratio/high_mean': 0.08442175667732954, 'clip_ratio/high_max': 0.08442175667732954, 'clip_ratio/region_mean': 0.12268189061433077, 'reward_total_mean': 0.7743169665336609, 'reward_meter_mean': 0.7170044183731079, 'reward_meter_std': 0.35669296979904175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9703220725059509, 'reward_repeat_soft_std': 0.03601430356502533, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.7743169665336609, 'reward_total_composite_std': 0.21356315910816193, 'epoch': 0.07}

 20%|██        | 665/3300 [1:40:35<5:12:56,  7.13s/it]INFO 04-13 09:01:33 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 666/3300 [1:40:41<4:59:04,  6.81s/it]2026-04-13 09:01:39,272 | INFO | train_grpo_train | metrics_logged mode=train step=666

                                                      
{'loss': 0.0314, 'grad_norm': 17.484010696411133, 'learning_rate': 7.984848484848486e-06, 'num_tokens': 1178291.0, 'completions/mean_length': 29.75, 'completions/min_length': 26.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.4946022927761078, 'rewards/meter/std': 0.4055790603160858, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9355142116546631, 'rewards/repeat_soft/std': 0.062372948974370956, 'rewards/judge_quality/mean': 0.731249988079071, 'rewards/judge_quality/std': 0.20131267607212067, 'rewards/total_composite/mean': 0.5652487874031067, 'rewards/total_composite/std': 0.19610926508903503, 'reward': 0.5652487874031067, 'reward_std': 0.19610926508903503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19448567926883698, 'sampling/sampling_logp_difference/max': 1.7358198165893555, 'sampling/importance_sampling_ratio/min': 0.17625565826892853, 'sampling/importance_sampling_ratio/mean': 0.9778200387954712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1346284076571465, 'clip_ratio/low_mean': 0.059890261851251125, 'clip_ratio/low_min': 0.059890261851251125, 'clip_ratio/high_mean': 0.0824202811345458, 'clip_ratio/high_max': 0.0824202811345458, 'clip_ratio/region_mean': 0.14231054298579693, 'reward_total_mean': 0.5652487874031067, 'reward_meter_mean': 0.4946022927761078, 'reward_meter_std': 0.4055790603160858, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9355142116546631, 'reward_repeat_soft_std': 0.062372948974370956, 'reward_judge_quality_mean': 0.731249988079071, 'reward_judge_quality_std': 0.20131267607212067, 'reward_total_composite_mean': 0.5652487874031067, 'reward_total_composite_std': 0.19610926508903503, 'epoch': 0.07}

 20%|██        | 666/3300 [1:40:41<4:59:04,  6.81s/it]INFO 04-13 09:01:39 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 667/3300 [1:40:47<4:45:40,  6.51s/it]2026-04-13 09:01:45,073 | INFO | train_grpo_train | metrics_logged mode=train step=667

                                                      
{'loss': 0.1176, 'grad_norm': 31.319604873657227, 'learning_rate': 7.981818181818183e-06, 'num_tokens': 1179608.0, 'completions/mean_length': 15.625, 'completions/min_length': 13.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 15.625, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.723710298538208, 'rewards/meter/std': 0.35061559081077576, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.18616428971290588, 'rewards/total_composite/mean': 0.5872097611427307, 'rewards/total_composite/std': 0.17698632180690765, 'reward': 0.5872097611427307, 'reward_std': 0.17698630690574646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17195676267147064, 'sampling/sampling_logp_difference/max': 1.2872560024261475, 'sampling/importance_sampling_ratio/min': 0.29236650466918945, 'sampling/importance_sampling_ratio/mean': 1.0469970703125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2363938093185425, 'clip_ratio/low_mean': 0.16971050202846527, 'clip_ratio/low_min': 0.16971050202846527, 'clip_ratio/high_mean': 0.05444139428436756, 'clip_ratio/high_max': 0.05444139428436756, 'clip_ratio/region_mean': 0.22415189631283283, 'reward_total_mean': 0.5872097611427307, 'reward_meter_mean': 0.723710298538208, 'reward_meter_std': 0.35061559081077576, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.18616428971290588, 'reward_total_composite_mean': 0.5872097611427307, 'reward_total_composite_std': 0.17698632180690765, 'epoch': 0.07}

 20%|██        | 667/3300 [1:40:47<4:45:40,  6.51s/it]INFO 04-13 09:01:45 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 668/3300 [1:40:52<4:34:17,  6.25s/it]2026-04-13 09:01:50,725 | INFO | train_grpo_train | metrics_logged mode=train step=668

                                                      
{'loss': 0.0449, 'grad_norm': 17.066247940063477, 'learning_rate': 7.978787878787879e-06, 'num_tokens': 1181107.0, 'completions/mean_length': 27.375, 'completions/min_length': 24.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.64689040184021, 'rewards/meter/std': 0.3717164993286133, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9223263263702393, 'rewards/repeat_soft/std': 0.11539124697446823, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.5421292781829834, 'rewards/total_composite/std': 0.13657505810260773, 'reward': 0.5421292781829834, 'reward_std': 0.13657505810260773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.163919135928154, 'sampling/sampling_logp_difference/max': 1.8664093017578125, 'sampling/importance_sampling_ratio/min': 0.15467806160449982, 'sampling/importance_sampling_ratio/mean': 1.010160207748413, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8885713741183281, 'clip_ratio/low_mean': 0.09868130926042795, 'clip_ratio/low_min': 0.09868130926042795, 'clip_ratio/high_mean': 0.07606082316488028, 'clip_ratio/high_max': 0.07606082316488028, 'clip_ratio/region_mean': 0.17474213242530823, 'reward_total_mean': 0.5421292781829834, 'reward_meter_mean': 0.64689040184021, 'reward_meter_std': 0.3717164993286133, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9223263263702393, 'reward_repeat_soft_std': 0.11539124697446823, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.5421292781829834, 'reward_total_composite_std': 0.13657505810260773, 'epoch': 0.07}

 20%|██        | 668/3300 [1:40:52<4:34:17,  6.25s/it]INFO 04-13 09:01:50 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 669/3300 [1:40:58<4:24:31,  6.03s/it]2026-04-13 09:01:56,246 | INFO | train_grpo_train | metrics_logged mode=train step=669

                                                      
{'loss': 0.0342, 'grad_norm': 25.62593650817871, 'learning_rate': 7.975757575757576e-06, 'num_tokens': 1182586.0, 'completions/mean_length': 25.875, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.46250832080841064, 'rewards/meter/std': 0.3943176865577698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9281853437423706, 'rewards/repeat_soft/std': 0.05050138756632805, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4678257703781128, 'rewards/total_composite/std': 0.10863402485847473, 'reward': 0.4678257703781128, 'reward_std': 0.10863400250673294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21301527321338654, 'sampling/sampling_logp_difference/max': 2.5426676273345947, 'sampling/importance_sampling_ratio/min': 0.07865629345178604, 'sampling/importance_sampling_ratio/mean': 1.003975510597229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2145325914025307, 'clip_ratio/low_mean': 0.08599405828863382, 'clip_ratio/low_min': 0.08599405828863382, 'clip_ratio/high_mean': 0.07259339187294245, 'clip_ratio/high_max': 0.07259339187294245, 'clip_ratio/region_mean': 0.15858745016157627, 'reward_total_mean': 0.4678257703781128, 'reward_meter_mean': 0.46250832080841064, 'reward_meter_std': 0.3943176865577698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9281853437423706, 'reward_repeat_soft_std': 0.05050138756632805, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4678257703781128, 'reward_total_composite_std': 0.10863402485847473, 'epoch': 0.07}

 20%|██        | 669/3300 [1:40:58<4:24:31,  6.03s/it]INFO 04-13 09:01:56 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 670/3300 [1:41:05<4:31:48,  6.20s/it]2026-04-13 09:02:02,838 | INFO | train_grpo_train | metrics_logged mode=train step=670

                                                      
{'loss': 0.032, 'grad_norm': 8.41564655303955, 'learning_rate': 7.972727272727273e-06, 'num_tokens': 1184823.0, 'completions/mean_length': 99.625, 'completions/min_length': 93.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.6294859647750854, 'rewards/meter/std': 0.25513356924057007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8679848909378052, 'rewards/repeat_soft/std': 0.07367976009845734, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5435631275177002, 'rewards/total_composite/std': 0.09454849362373352, 'reward': 0.5435631275177002, 'reward_std': 0.09454847127199173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16039173305034637, 'sampling/sampling_logp_difference/max': 3.5806386470794678, 'sampling/importance_sampling_ratio/min': 0.02785790152847767, 'sampling/importance_sampling_ratio/mean': 1.0072435140609741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1585305109620094, 'clip_ratio/low_mean': 0.06731393747031689, 'clip_ratio/low_min': 0.06731393747031689, 'clip_ratio/high_mean': 0.06668214779347181, 'clip_ratio/high_max': 0.06668214779347181, 'clip_ratio/region_mean': 0.1339960852637887, 'reward_total_mean': 0.5435631275177002, 'reward_meter_mean': 0.6294859647750854, 'reward_meter_std': 0.25513356924057007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8679848909378052, 'reward_repeat_soft_std': 0.07367976009845734, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5435631275177002, 'reward_total_composite_std': 0.09454849362373352, 'epoch': 0.07}

 20%|██        | 670/3300 [1:41:05<4:31:48,  6.20s/it]INFO 04-13 09:02:03 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 671/3300 [1:41:10<4:27:05,  6.10s/it]2026-04-13 09:02:08,690 | INFO | train_grpo_train | metrics_logged mode=train step=671

                                                      
{'loss': 0.0513, 'grad_norm': 21.132844924926758, 'learning_rate': 7.96969696969697e-06, 'num_tokens': 1186325.0, 'completions/mean_length': 28.75, 'completions/min_length': 26.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7009587287902832, 'rewards/meter/std': 0.40324610471725464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9373697638511658, 'rewards/repeat_soft/std': 0.0465090349316597, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/total_composite/mean': 0.7222044467926025, 'rewards/total_composite/std': 0.2330915629863739, 'reward': 0.7222044467926025, 'reward_std': 0.2330915480852127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1750107854604721, 'sampling/sampling_logp_difference/max': 1.9450829029083252, 'sampling/importance_sampling_ratio/min': 0.14297537505626678, 'sampling/importance_sampling_ratio/mean': 1.0134085416793823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.214572735130787, 'clip_ratio/low_mean': 0.07826228905469179, 'clip_ratio/low_min': 0.07826228905469179, 'clip_ratio/high_mean': 0.07405641861259937, 'clip_ratio/high_max': 0.07405641861259937, 'clip_ratio/region_mean': 0.15231870766729116, 'reward_total_mean': 0.7222044467926025, 'reward_meter_mean': 0.7009587287902832, 'reward_meter_std': 0.40324610471725464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9373697638511658, 'reward_repeat_soft_std': 0.0465090349316597, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.7222044467926025, 'reward_total_composite_std': 0.2330915629863739, 'epoch': 0.07}

 20%|██        | 671/3300 [1:41:10<4:27:05,  6.10s/it]INFO 04-13 09:02:08 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 672/3300 [1:41:17<4:38:47,  6.36s/it]2026-04-13 09:02:15,682 | INFO | train_grpo_train | metrics_logged mode=train step=672

                                                      
{'loss': 0.0669, 'grad_norm': 11.541581153869629, 'learning_rate': 7.966666666666668e-06, 'num_tokens': 1188303.0, 'completions/mean_length': 79.25, 'completions/min_length': 75.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9345486164093018, 'rewards/meter/std': 0.08608411997556686, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8871944546699524, 'rewards/repeat_soft/std': 0.08728492259979248, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6362572312355042, 'rewards/total_composite/std': 0.10502653568983078, 'reward': 0.6362572312355042, 'reward_std': 0.10502654314041138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17951758205890656, 'sampling/sampling_logp_difference/max': 1.841355800628662, 'sampling/importance_sampling_ratio/min': 0.17673450708389282, 'sampling/importance_sampling_ratio/mean': 1.0232856273651123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.519042745232582, 'clip_ratio/low_mean': 0.09671382559463382, 'clip_ratio/low_min': 0.09671382559463382, 'clip_ratio/high_mean': 0.04755411110818386, 'clip_ratio/high_max': 0.04755411110818386, 'clip_ratio/region_mean': 0.14426793670281768, 'reward_total_mean': 0.6362572312355042, 'reward_meter_mean': 0.9345486164093018, 'reward_meter_std': 0.08608411997556686, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8871944546699524, 'reward_repeat_soft_std': 0.08728492259979248, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6362572312355042, 'reward_total_composite_std': 0.10502653568983078, 'epoch': 0.07}

 20%|██        | 672/3300 [1:41:17<4:38:47,  6.36s/it]INFO 04-13 09:02:15 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 673/3300 [1:41:29<5:42:44,  7.83s/it]2026-04-13 09:02:26,925 | INFO | train_grpo_train | metrics_logged mode=train step=673

                                                      
{'loss': -0.1507, 'grad_norm': 3.268763542175293, 'learning_rate': 7.963636363636365e-06, 'num_tokens': 1190056.0, 'completions/mean_length': 123.125, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 67.5714340209961, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8443180918693542, 'rewards/meter/std': 0.2273014634847641, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8706649541854858, 'rewards/repeat_soft/std': 0.17246706783771515, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.18845234811306, 'rewards/total_composite/mean': 0.5125667452812195, 'rewards/total_composite/std': 0.23342394828796387, 'reward': 0.5125667452812195, 'reward_std': 0.23342394828796387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18086622655391693, 'sampling/sampling_logp_difference/max': 1.1409001350402832, 'sampling/importance_sampling_ratio/min': 0.31953126192092896, 'sampling/importance_sampling_ratio/mean': 1.0215522050857544, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1888233423233032, 'clip_ratio/low_mean': 0.022275641094893217, 'clip_ratio/low_min': 0.022275641094893217, 'clip_ratio/high_mean': 0.11251303926110268, 'clip_ratio/high_max': 0.11251303926110268, 'clip_ratio/region_mean': 0.1347886803559959, 'reward_total_mean': 0.5125667452812195, 'reward_meter_mean': 0.8443180918693542, 'reward_meter_std': 0.2273014634847641, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8706649541854858, 'reward_repeat_soft_std': 0.17246706783771515, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.18845234811306, 'reward_total_composite_mean': 0.5125667452812195, 'reward_total_composite_std': 0.23342394828796387, 'epoch': 0.07}

 20%|██        | 673/3300 [1:41:29<5:42:44,  7.83s/it]INFO 04-13 09:02:27 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 674/3300 [1:41:40<6:30:41,  8.93s/it]2026-04-13 09:02:38,415 | INFO | train_grpo_train | metrics_logged mode=train step=674

                                                      
{'loss': -0.1358, 'grad_norm': 5.081193923950195, 'learning_rate': 7.96060606060606e-06, 'num_tokens': 1192216.0, 'completions/mean_length': 147.0, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 94.85714721679688, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.7837413549423218, 'rewards/meter/std': 0.22885428369045258, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.7908656597137451, 'rewards/repeat_soft/std': 0.17838376760482788, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.2105392962694168, 'rewards/total_composite/mean': 0.448514461517334, 'rewards/total_composite/std': 0.2988634407520294, 'reward': 0.448514461517334, 'reward_std': 0.2988634407520294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1417771726846695, 'sampling/sampling_logp_difference/max': 1.746849536895752, 'sampling/importance_sampling_ratio/min': 0.17432227730751038, 'sampling/importance_sampling_ratio/mean': 1.0056918859481812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7560857981443405, 'clip_ratio/low_mean': 0.014044944196939468, 'clip_ratio/low_min': 0.014044944196939468, 'clip_ratio/high_mean': 0.08761232905089855, 'clip_ratio/high_max': 0.08761232905089855, 'clip_ratio/region_mean': 0.10165727324783802, 'reward_total_mean': 0.448514461517334, 'reward_meter_mean': 0.7837413549423218, 'reward_meter_std': 0.22885428369045258, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.7908656597137451, 'reward_repeat_soft_std': 0.17838376760482788, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.2105392962694168, 'reward_total_composite_mean': 0.448514461517334, 'reward_total_composite_std': 0.2988634407520294, 'epoch': 0.07}

 20%|██        | 674/3300 [1:41:40<6:30:41,  8.93s/it]INFO 04-13 09:02:38 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 675/3300 [1:41:46<5:54:54,  8.11s/it]2026-04-13 09:02:44,631 | INFO | train_grpo_train | metrics_logged mode=train step=675

                                                      
{'loss': 0.0521, 'grad_norm': 15.669044494628906, 'learning_rate': 7.957575757575758e-06, 'num_tokens': 1193848.0, 'completions/mean_length': 35.0, 'completions/min_length': 32.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.30102837085723877, 'rewards/meter/std': 0.28132569789886475, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9912567138671875, 'rewards/repeat_soft/std': 0.008987357839941978, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.44334959983825684, 'rewards/total_composite/std': 0.08618506044149399, 'reward': 0.44334959983825684, 'reward_std': 0.08618505299091339, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15274398028850555, 'sampling/sampling_logp_difference/max': 1.2224321365356445, 'sampling/importance_sampling_ratio/min': 0.2945129871368408, 'sampling/importance_sampling_ratio/mean': 1.026187777519226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2333584874868393, 'clip_ratio/low_mean': 0.06395565439015627, 'clip_ratio/low_min': 0.06395565439015627, 'clip_ratio/high_mean': 0.06867985613644123, 'clip_ratio/high_max': 0.06867985613644123, 'clip_ratio/region_mean': 0.1326355105265975, 'reward_total_mean': 0.44334959983825684, 'reward_meter_mean': 0.30102837085723877, 'reward_meter_std': 0.28132569789886475, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9912567138671875, 'reward_repeat_soft_std': 0.008987357839941978, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.44334959983825684, 'reward_total_composite_std': 0.08618506044149399, 'epoch': 0.07}

 20%|██        | 675/3300 [1:41:46<5:54:54,  8.11s/it]INFO 04-13 09:02:44 [block_pool.py:378] Successfully reset prefix cache

 20%|██        | 676/3300 [1:41:53<5:37:07,  7.71s/it]2026-04-13 09:02:51,394 | INFO | train_grpo_train | metrics_logged mode=train step=676

                                                      
{'loss': 0.1237, 'grad_norm': 9.95625114440918, 'learning_rate': 7.954545454545455e-06, 'num_tokens': 1195821.0, 'completions/mean_length': 74.625, 'completions/min_length': 56.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.7109752893447876, 'rewards/meter/std': 0.34340932965278625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6913044452667236, 'rewards/repeat_soft/std': 0.18606877326965332, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5181210041046143, 'rewards/total_composite/std': 0.12576277554035187, 'reward': 0.5181210041046143, 'reward_std': 0.12576277554035187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12566405534744263, 'sampling/sampling_logp_difference/max': 1.4823908805847168, 'sampling/importance_sampling_ratio/min': 0.22709408402442932, 'sampling/importance_sampling_ratio/mean': 1.0080106258392334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6787445619702339, 'clip_ratio/low_mean': 0.05535984970629215, 'clip_ratio/low_min': 0.05535984970629215, 'clip_ratio/high_mean': 0.06927759945392609, 'clip_ratio/high_max': 0.06927759945392609, 'clip_ratio/region_mean': 0.12463744916021824, 'reward_total_mean': 0.5181210041046143, 'reward_meter_mean': 0.7109752893447876, 'reward_meter_std': 0.34340932965278625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6913044452667236, 'reward_repeat_soft_std': 0.18606877326965332, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5181210041046143, 'reward_total_composite_std': 0.12576277554035187, 'epoch': 0.07}

 20%|██        | 676/3300 [1:41:53<5:37:07,  7.71s/it]INFO 04-13 09:02:51 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 677/3300 [1:41:59<5:15:07,  7.21s/it]2026-04-13 09:02:57,435 | INFO | train_grpo_train | metrics_logged mode=train step=677

                                                      
{'loss': 0.0415, 'grad_norm': 15.976642608642578, 'learning_rate': 7.951515151515152e-06, 'num_tokens': 1197544.0, 'completions/mean_length': 38.375, 'completions/min_length': 32.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.3758412003517151, 'rewards/meter/std': 0.3048185706138611, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9766628742218018, 'rewards/repeat_soft/std': 0.03151514008641243, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.4973105490207672, 'rewards/total_composite/std': 0.15020503103733063, 'reward': 0.4973105490207672, 'reward_std': 0.15020503103733063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16625650227069855, 'sampling/sampling_logp_difference/max': 1.1946462392807007, 'sampling/importance_sampling_ratio/min': 0.30281105637550354, 'sampling/importance_sampling_ratio/mean': 1.0364453792572021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3895529732108116, 'clip_ratio/low_mean': 0.09560188185423613, 'clip_ratio/low_min': 0.09560188185423613, 'clip_ratio/high_mean': 0.02656250074505806, 'clip_ratio/high_max': 0.02656250074505806, 'clip_ratio/region_mean': 0.12216438259929419, 'reward_total_mean': 0.4973105490207672, 'reward_meter_mean': 0.3758412003517151, 'reward_meter_std': 0.3048185706138611, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9766628742218018, 'reward_repeat_soft_std': 0.03151514008641243, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.4973105490207672, 'reward_total_composite_std': 0.15020503103733063, 'epoch': 0.07}

 21%|██        | 677/3300 [1:41:59<5:15:07,  7.21s/it]INFO 04-13 09:02:57 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 678/3300 [1:42:05<4:58:38,  6.83s/it]2026-04-13 09:03:03,396 | INFO | train_grpo_train | metrics_logged mode=train step=678

                                                      
{'loss': 0.1921, 'grad_norm': 34.70186996459961, 'learning_rate': 7.948484848484848e-06, 'num_tokens': 1199079.0, 'completions/mean_length': 37.875, 'completions/min_length': 26.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.29696419835090637, 'rewards/meter/std': 0.3232930302619934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.972385048866272, 'rewards/repeat_soft/std': 0.025973200798034668, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.24859607219696045, 'rewards/total_composite/mean': 0.46499231457710266, 'rewards/total_composite/std': 0.10828929394483566, 'reward': 0.46499231457710266, 'reward_std': 0.10828928649425507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19013623893260956, 'sampling/sampling_logp_difference/max': 2.3498706817626953, 'sampling/importance_sampling_ratio/min': 0.0953814908862114, 'sampling/importance_sampling_ratio/mean': 0.9908564686775208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4125063568353653, 'clip_ratio/low_mean': 0.0727268448099494, 'clip_ratio/low_min': 0.0727268448099494, 'clip_ratio/high_mean': 0.06334841810166836, 'clip_ratio/high_max': 0.06334841810166836, 'clip_ratio/region_mean': 0.13607526291161776, 'reward_total_mean': 0.46499231457710266, 'reward_meter_mean': 0.29696419835090637, 'reward_meter_std': 0.3232930302619934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.972385048866272, 'reward_repeat_soft_std': 0.025973200798034668, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.24859607219696045, 'reward_total_composite_mean': 0.46499231457710266, 'reward_total_composite_std': 0.10828929394483566, 'epoch': 0.07}

 21%|██        | 678/3300 [1:42:05<4:58:38,  6.83s/it]INFO 04-13 09:03:03 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 679/3300 [1:42:11<4:48:09,  6.60s/it]2026-04-13 09:03:09,440 | INFO | train_grpo_train | metrics_logged mode=train step=679

                                                      
{'loss': 0.1029, 'grad_norm': 13.571893692016602, 'learning_rate': 7.945454545454547e-06, 'num_tokens': 1200869.0, 'completions/mean_length': 40.75, 'completions/min_length': 29.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5557988882064819, 'rewards/meter/std': 0.353110671043396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.900698184967041, 'rewards/repeat_soft/std': 0.07676287740468979, 'rewards/judge_quality/mean': 0.5600000619888306, 'rewards/judge_quality/std': 0.31550416350364685, 'rewards/total_composite/mean': 0.5425010919570923, 'rewards/total_composite/std': 0.178108811378479, 'reward': 0.5425010919570923, 'reward_std': 0.178108811378479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1761329472064972, 'sampling/sampling_logp_difference/max': 1.5521812438964844, 'sampling/importance_sampling_ratio/min': 0.21178551018238068, 'sampling/importance_sampling_ratio/mean': 1.023783802986145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.722318395972252, 'clip_ratio/low_mean': 0.08291913568973541, 'clip_ratio/low_min': 0.08291913568973541, 'clip_ratio/high_mean': 0.04797446262091398, 'clip_ratio/high_max': 0.04797446262091398, 'clip_ratio/region_mean': 0.1308935983106494, 'reward_total_mean': 0.5425010919570923, 'reward_meter_mean': 0.5557988882064819, 'reward_meter_std': 0.353110671043396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.900698184967041, 'reward_repeat_soft_std': 0.07676287740468979, 'reward_judge_quality_mean': 0.5600000619888306, 'reward_judge_quality_std': 0.31550416350364685, 'reward_total_composite_mean': 0.5425010919570923, 'reward_total_composite_std': 0.178108811378479, 'epoch': 0.07}

 21%|██        | 679/3300 [1:42:11<4:48:09,  6.60s/it]INFO 04-13 09:03:09 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 680/3300 [1:42:17<4:36:49,  6.34s/it]2026-04-13 09:03:15,179 | INFO | train_grpo_train | metrics_logged mode=train step=680

                                                      
{'loss': 0.0217, 'grad_norm': 19.444847106933594, 'learning_rate': 7.942424242424242e-06, 'num_tokens': 1202320.0, 'completions/mean_length': 33.375, 'completions/min_length': 32.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9708325266838074, 'rewards/meter/std': 0.024752020835876465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9009986519813538, 'rewards/repeat_soft/std': 0.07457141578197479, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.2775370180606842, 'rewards/total_composite/mean': 0.7108228206634521, 'rewards/total_composite/std': 0.18475474417209625, 'reward': 0.7108228206634521, 'reward_std': 0.18475474417209625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17875127494335175, 'sampling/sampling_logp_difference/max': 3.197706699371338, 'sampling/importance_sampling_ratio/min': 0.04085579141974449, 'sampling/importance_sampling_ratio/mean': 1.0433377027511597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3223681598901749, 'clip_ratio/low_mean': 0.11509859748184681, 'clip_ratio/low_min': 0.11509859748184681, 'clip_ratio/high_mean': 0.043231177143752575, 'clip_ratio/high_max': 0.043231177143752575, 'clip_ratio/region_mean': 0.15832977462559938, 'reward_total_mean': 0.7108228206634521, 'reward_meter_mean': 0.9708325266838074, 'reward_meter_std': 0.024752020835876465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9009986519813538, 'reward_repeat_soft_std': 0.07457141578197479, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.2775370180606842, 'reward_total_composite_mean': 0.7108228206634521, 'reward_total_composite_std': 0.18475474417209625, 'epoch': 0.07}

 21%|██        | 680/3300 [1:42:17<4:36:49,  6.34s/it]INFO 04-13 09:03:15 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 681/3300 [1:42:23<4:30:09,  6.19s/it]2026-04-13 09:03:21,020 | INFO | train_grpo_train | metrics_logged mode=train step=681

                                                      
{'loss': -0.0315, 'grad_norm': 13.85059928894043, 'learning_rate': 7.93939393939394e-06, 'num_tokens': 1204056.0, 'completions/mean_length': 40.0, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8067896962165833, 'rewards/meter/std': 0.33366015553474426, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9807987213134766, 'rewards/repeat_soft/std': 0.023438427597284317, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.6513296365737915, 'rewards/total_composite/std': 0.18861781060695648, 'reward': 0.6513296365737915, 'reward_std': 0.1886177957057953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17290246486663818, 'sampling/sampling_logp_difference/max': 1.6026787757873535, 'sampling/importance_sampling_ratio/min': 0.20135639607906342, 'sampling/importance_sampling_ratio/mean': 1.0108685493469238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.389533892273903, 'clip_ratio/low_mean': 0.10763272503390908, 'clip_ratio/low_min': 0.10763272503390908, 'clip_ratio/high_mean': 0.02724359044805169, 'clip_ratio/high_max': 0.02724359044805169, 'clip_ratio/region_mean': 0.13487631548196077, 'reward_total_mean': 0.6513296365737915, 'reward_meter_mean': 0.8067896962165833, 'reward_meter_std': 0.33366015553474426, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9807987213134766, 'reward_repeat_soft_std': 0.023438427597284317, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.6513296365737915, 'reward_total_composite_std': 0.18861781060695648, 'epoch': 0.07}

 21%|██        | 681/3300 [1:42:23<4:30:09,  6.19s/it]INFO 04-13 09:03:21 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 682/3300 [1:42:29<4:26:42,  6.11s/it]2026-04-13 09:03:26,950 | INFO | train_grpo_train | metrics_logged mode=train step=682

                                                      
{'loss': 0.0067, 'grad_norm': 15.246498107910156, 'learning_rate': 7.936363636363637e-06, 'num_tokens': 1205804.0, 'completions/mean_length': 43.5, 'completions/min_length': 39.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.7407575845718384, 'rewards/meter/std': 0.2606121599674225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9272711873054504, 'rewards/repeat_soft/std': 0.04879491776227951, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2353682667016983, 'rewards/total_composite/mean': 0.6343569755554199, 'rewards/total_composite/std': 0.175083190202713, 'reward': 0.6343569755554199, 'reward_std': 0.1750832051038742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15224520862102509, 'sampling/sampling_logp_difference/max': 1.8238554000854492, 'sampling/importance_sampling_ratio/min': 0.16140228509902954, 'sampling/importance_sampling_ratio/mean': 1.0225683450698853, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8424210101366043, 'clip_ratio/low_mean': 0.08994126692414284, 'clip_ratio/low_min': 0.08994126692414284, 'clip_ratio/high_mean': 0.04844512231647968, 'clip_ratio/high_max': 0.04844512231647968, 'clip_ratio/region_mean': 0.13838638924062252, 'reward_total_mean': 0.6343569755554199, 'reward_meter_mean': 0.7407575845718384, 'reward_meter_std': 0.2606121599674225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9272711873054504, 'reward_repeat_soft_std': 0.04879491776227951, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2353682667016983, 'reward_total_composite_mean': 0.6343569755554199, 'reward_total_composite_std': 0.175083190202713, 'epoch': 0.07}

 21%|██        | 682/3300 [1:42:29<4:26:42,  6.11s/it]INFO 04-13 09:03:27 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 683/3300 [1:42:35<4:23:21,  6.04s/it]2026-04-13 09:03:32,817 | INFO | train_grpo_train | metrics_logged mode=train step=683

                                                      
{'loss': -0.0441, 'grad_norm': 19.87747573852539, 'learning_rate': 7.933333333333334e-06, 'num_tokens': 1207392.0, 'completions/mean_length': 38.5, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7616804838180542, 'rewards/meter/std': 0.3156481087207794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8836441040039062, 'rewards/repeat_soft/std': 0.06109395623207092, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5482738018035889, 'rewards/total_composite/std': 0.09274234622716904, 'reward': 0.5482738018035889, 'reward_std': 0.09274233877658844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1745234578847885, 'sampling/sampling_logp_difference/max': 2.012331008911133, 'sampling/importance_sampling_ratio/min': 0.1336767077445984, 'sampling/importance_sampling_ratio/mean': 1.0172439813613892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2272211387753487, 'clip_ratio/low_mean': 0.030560662038624287, 'clip_ratio/low_min': 0.030560662038624287, 'clip_ratio/high_mean': 0.14728633873164654, 'clip_ratio/high_max': 0.14728633873164654, 'clip_ratio/region_mean': 0.17784700077027082, 'reward_total_mean': 0.5482738018035889, 'reward_meter_mean': 0.7616804838180542, 'reward_meter_std': 0.3156481087207794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8836441040039062, 'reward_repeat_soft_std': 0.06109395623207092, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5482738018035889, 'reward_total_composite_std': 0.09274234622716904, 'epoch': 0.07}

 21%|██        | 683/3300 [1:42:35<4:23:21,  6.04s/it]INFO 04-13 09:03:33 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 684/3300 [1:42:40<4:21:32,  6.00s/it]2026-04-13 09:03:38,722 | INFO | train_grpo_train | metrics_logged mode=train step=684

                                                      
{'loss': 0.0117, 'grad_norm': 15.160962104797363, 'learning_rate': 7.930303030303031e-06, 'num_tokens': 1208791.0, 'completions/mean_length': 30.875, 'completions/min_length': 27.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9185324907302856, 'rewards/meter/std': 0.1610766500234604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8485807776451111, 'rewards/repeat_soft/std': 0.08786656707525253, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2822834253311157, 'rewards/total_composite/mean': 0.6363656520843506, 'rewards/total_composite/std': 0.3035244345664978, 'reward': 0.6363656520843506, 'reward_std': 0.3035244345664978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16926787793636322, 'sampling/sampling_logp_difference/max': 1.9378900527954102, 'sampling/importance_sampling_ratio/min': 0.14400747418403625, 'sampling/importance_sampling_ratio/mean': 1.0151031017303467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9692705720663071, 'clip_ratio/low_mean': 0.06495348457247019, 'clip_ratio/low_min': 0.06495348457247019, 'clip_ratio/high_mean': 0.037830933928489685, 'clip_ratio/high_max': 0.037830933928489685, 'clip_ratio/region_mean': 0.10278441850095987, 'reward_total_mean': 0.6363656520843506, 'reward_meter_mean': 0.9185324907302856, 'reward_meter_std': 0.1610766500234604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8485807776451111, 'reward_repeat_soft_std': 0.08786656707525253, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2822834253311157, 'reward_total_composite_mean': 0.6363656520843506, 'reward_total_composite_std': 0.3035244345664978, 'epoch': 0.07}

 21%|██        | 684/3300 [1:42:40<4:21:32,  6.00s/it]INFO 04-13 09:03:38 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 685/3300 [1:42:46<4:19:48,  5.96s/it]2026-04-13 09:03:44,622 | INFO | train_grpo_train | metrics_logged mode=train step=685

                                                      
{'loss': 0.0278, 'grad_norm': 15.482486724853516, 'learning_rate': 7.927272727272729e-06, 'num_tokens': 1210396.0, 'completions/mean_length': 35.625, 'completions/min_length': 33.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5370067358016968, 'rewards/meter/std': 0.42286112904548645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7139323949813843, 'rewards/repeat_soft/std': 0.21135301887989044, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.4486607313156128, 'rewards/total_composite/std': 0.10594851523637772, 'reward': 0.4486607313156128, 'reward_std': 0.10594853013753891, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16646844148635864, 'sampling/sampling_logp_difference/max': 1.4042901992797852, 'sampling/importance_sampling_ratio/min': 0.2455412894487381, 'sampling/importance_sampling_ratio/mean': 1.025395154953003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1219540163874626, 'clip_ratio/low_mean': 0.06411921512335539, 'clip_ratio/low_min': 0.06411921512335539, 'clip_ratio/high_mean': 0.0721204336732626, 'clip_ratio/high_max': 0.0721204336732626, 'clip_ratio/region_mean': 0.13623964879661798, 'reward_total_mean': 0.4486607313156128, 'reward_meter_mean': 0.5370067358016968, 'reward_meter_std': 0.42286112904548645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7139323949813843, 'reward_repeat_soft_std': 0.21135301887989044, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.4486607313156128, 'reward_total_composite_std': 0.10594851523637772, 'epoch': 0.07}

 21%|██        | 685/3300 [1:42:46<4:19:48,  5.96s/it]INFO 04-13 09:03:44 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 686/3300 [1:42:52<4:21:05,  5.99s/it]2026-04-13 09:03:50,663 | INFO | train_grpo_train | metrics_logged mode=train step=686

                                                      
{'loss': -0.0335, 'grad_norm': 15.697717666625977, 'learning_rate': 7.924242424242426e-06, 'num_tokens': 1211919.0, 'completions/mean_length': 32.375, 'completions/min_length': 26.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.3621957302093506, 'rewards/meter/std': 0.3310295343399048, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.940528154373169, 'rewards/repeat_soft/std': 0.11192068457603455, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.4836890399456024, 'rewards/total_composite/std': 0.1959802210330963, 'reward': 0.4836890399456024, 'reward_std': 0.1959802210330963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17518004775047302, 'sampling/sampling_logp_difference/max': 1.3122613430023193, 'sampling/importance_sampling_ratio/min': 0.2692105770111084, 'sampling/importance_sampling_ratio/mean': 1.0158535242080688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4639041870832443, 'clip_ratio/low_mean': 0.08152802241966128, 'clip_ratio/low_min': 0.08152802241966128, 'clip_ratio/high_mean': 0.07402259670197964, 'clip_ratio/high_max': 0.07402259670197964, 'clip_ratio/region_mean': 0.15555061912164092, 'reward_total_mean': 0.4836890399456024, 'reward_meter_mean': 0.3621957302093506, 'reward_meter_std': 0.3310295343399048, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.940528154373169, 'reward_repeat_soft_std': 0.11192068457603455, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.4836890399456024, 'reward_total_composite_std': 0.1959802210330963, 'epoch': 0.07}

 21%|██        | 686/3300 [1:42:52<4:21:05,  5.99s/it]INFO 04-13 09:03:50 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 687/3300 [1:42:58<4:19:28,  5.96s/it]2026-04-13 09:03:56,540 | INFO | train_grpo_train | metrics_logged mode=train step=687

                                                      
{'loss': -0.0076, 'grad_norm': 16.958520889282227, 'learning_rate': 7.921212121212122e-06, 'num_tokens': 1213395.0, 'completions/mean_length': 31.5, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7834794521331787, 'rewards/meter/std': 0.3169131278991699, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9552586078643799, 'rewards/repeat_soft/std': 0.044948361814022064, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6794179677963257, 'rewards/total_composite/std': 0.21409960091114044, 'reward': 0.6794179677963257, 'reward_std': 0.21409958600997925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16569404304027557, 'sampling/sampling_logp_difference/max': 1.0826793909072876, 'sampling/importance_sampling_ratio/min': 0.33868685364723206, 'sampling/importance_sampling_ratio/mean': 1.0096803903579712, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0858681723475456, 'clip_ratio/low_mean': 0.10208950191736221, 'clip_ratio/low_min': 0.10208950191736221, 'clip_ratio/high_mean': 0.0692730974406004, 'clip_ratio/high_max': 0.0692730974406004, 'clip_ratio/region_mean': 0.1713625993579626, 'reward_total_mean': 0.6794179677963257, 'reward_meter_mean': 0.7834794521331787, 'reward_meter_std': 0.3169131278991699, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9552586078643799, 'reward_repeat_soft_std': 0.044948361814022064, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6794179677963257, 'reward_total_composite_std': 0.21409960091114044, 'epoch': 0.07}

 21%|██        | 687/3300 [1:42:58<4:19:28,  5.96s/it]INFO 04-13 09:03:56 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 688/3300 [1:43:10<5:36:23,  7.73s/it]2026-04-13 09:04:08,398 | INFO | train_grpo_train | metrics_logged mode=train step=688

                                                      
{'loss': -0.053, 'grad_norm': 4.748079776763916, 'learning_rate': 7.918181818181819e-06, 'num_tokens': 1214873.0, 'completions/mean_length': 94.75, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.142860412597656, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8547170162200928, 'rewards/meter/std': 0.2940731942653656, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8981757164001465, 'rewards/repeat_soft/std': 0.10416964441537857, 'rewards/judge_quality/mean': 0.5724999904632568, 'rewards/judge_quality/std': 0.31702861189842224, 'rewards/total_composite/mean': 0.672359824180603, 'rewards/total_composite/std': 0.2378748655319214, 'reward': 0.672359824180603, 'reward_std': 0.2378748655319214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17198236286640167, 'sampling/sampling_logp_difference/max': 1.9869155883789062, 'sampling/importance_sampling_ratio/min': 0.13711769878864288, 'sampling/importance_sampling_ratio/mean': 1.0300874710083008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2422862350940704, 'clip_ratio/low_mean': 0.056661609560251236, 'clip_ratio/low_min': 0.056661609560251236, 'clip_ratio/high_mean': 0.03858352266252041, 'clip_ratio/high_max': 0.03858352266252041, 'clip_ratio/region_mean': 0.09524513222277164, 'reward_total_mean': 0.672359824180603, 'reward_meter_mean': 0.8547170162200928, 'reward_meter_std': 0.2940731942653656, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8981757164001465, 'reward_repeat_soft_std': 0.10416964441537857, 'reward_judge_quality_mean': 0.5724999904632568, 'reward_judge_quality_std': 0.31702861189842224, 'reward_total_composite_mean': 0.672359824180603, 'reward_total_composite_std': 0.2378748655319214, 'epoch': 0.07}

 21%|██        | 688/3300 [1:43:10<5:36:23,  7.73s/it]INFO 04-13 09:04:08 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 689/3300 [1:43:18<5:39:11,  7.79s/it]2026-04-13 09:04:16,349 | INFO | train_grpo_train | metrics_logged mode=train step=689

                                                      
{'loss': 0.0227, 'grad_norm': 14.012819290161133, 'learning_rate': 7.915151515151516e-06, 'num_tokens': 1216639.0, 'completions/mean_length': 53.75, 'completions/min_length': 50.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4307352900505066, 'rewards/meter/std': 0.34267833828926086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9898139238357544, 'rewards/repeat_soft/std': 0.011434405110776424, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.1527603417634964, 'rewards/total_composite/mean': 0.48308122158050537, 'rewards/total_composite/std': 0.2503104507923126, 'reward': 0.48308122158050537, 'reward_std': 0.25031042098999023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2066158801317215, 'sampling/sampling_logp_difference/max': 1.6739482879638672, 'sampling/importance_sampling_ratio/min': 0.18750527501106262, 'sampling/importance_sampling_ratio/mean': 1.0345228910446167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4963724464178085, 'clip_ratio/low_mean': 0.11144568957388401, 'clip_ratio/low_min': 0.11144568957388401, 'clip_ratio/high_mean': 0.05153766833245754, 'clip_ratio/high_max': 0.05153766833245754, 'clip_ratio/region_mean': 0.16298335790634155, 'reward_total_mean': 0.48308122158050537, 'reward_meter_mean': 0.4307352900505066, 'reward_meter_std': 0.34267833828926086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9898139238357544, 'reward_repeat_soft_std': 0.011434405110776424, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.1527603417634964, 'reward_total_composite_mean': 0.48308122158050537, 'reward_total_composite_std': 0.2503104507923126, 'epoch': 0.07}

 21%|██        | 689/3300 [1:43:18<5:39:11,  7.79s/it]INFO 04-13 09:04:16 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 690/3300 [1:43:25<5:28:03,  7.54s/it]2026-04-13 09:04:23,299 | INFO | train_grpo_train | metrics_logged mode=train step=690

                                                      
{'loss': 0.0229, 'grad_norm': 26.791959762573242, 'learning_rate': 7.912121212121213e-06, 'num_tokens': 1217928.0, 'completions/mean_length': 17.125, 'completions/min_length': 15.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.125, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.24231413006782532, 'rewards/meter/std': 0.3449556231498718, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9539605379104614, 'rewards/repeat_soft/std': 0.022088520228862762, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.47506770491600037, 'rewards/total_composite/std': 0.20589597523212433, 'reward': 0.47506770491600037, 'reward_std': 0.20589596033096313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15316243469715118, 'sampling/sampling_logp_difference/max': 1.390833854675293, 'sampling/importance_sampling_ratio/min': 0.2488677054643631, 'sampling/importance_sampling_ratio/mean': 1.047309398651123, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.260609120130539, 'clip_ratio/low_mean': 0.1275846753269434, 'clip_ratio/low_min': 0.1275846753269434, 'clip_ratio/high_mean': 0.028594771400094032, 'clip_ratio/high_max': 0.028594771400094032, 'clip_ratio/region_mean': 0.15617944672703743, 'reward_total_mean': 0.47506770491600037, 'reward_meter_mean': 0.24231413006782532, 'reward_meter_std': 0.3449556231498718, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9539605379104614, 'reward_repeat_soft_std': 0.022088520228862762, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.47506770491600037, 'reward_total_composite_std': 0.20589597523212433, 'epoch': 0.07}

 21%|██        | 690/3300 [1:43:25<5:28:03,  7.54s/it]INFO 04-13 09:04:23 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 691/3300 [1:43:31<5:12:16,  7.18s/it]2026-04-13 09:04:29,641 | INFO | train_grpo_train | metrics_logged mode=train step=691

                                                      
{'loss': -0.0811, 'grad_norm': 16.744775772094727, 'learning_rate': 7.909090909090909e-06, 'num_tokens': 1219267.0, 'completions/mean_length': 31.375, 'completions/min_length': 26.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8912780284881592, 'rewards/meter/std': 0.22956277430057526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7346838712692261, 'rewards/repeat_soft/std': 0.17986097931861877, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.6554957628250122, 'rewards/total_composite/std': 0.16776107251644135, 'reward': 0.6554957628250122, 'reward_std': 0.16776105761528015, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1292048692703247, 'sampling/sampling_logp_difference/max': 1.3336560726165771, 'sampling/importance_sampling_ratio/min': 0.2635120749473572, 'sampling/importance_sampling_ratio/mean': 1.0278366804122925, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9303603619337082, 'clip_ratio/low_mean': 0.08603221224620938, 'clip_ratio/low_min': 0.08603221224620938, 'clip_ratio/high_mean': 0.0324730109423399, 'clip_ratio/high_max': 0.0324730109423399, 'clip_ratio/region_mean': 0.11850522318854928, 'reward_total_mean': 0.6554957628250122, 'reward_meter_mean': 0.8912780284881592, 'reward_meter_std': 0.22956277430057526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7346838712692261, 'reward_repeat_soft_std': 0.17986097931861877, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.6554957628250122, 'reward_total_composite_std': 0.16776107251644135, 'epoch': 0.07}

 21%|██        | 691/3300 [1:43:31<5:12:16,  7.18s/it]INFO 04-13 09:04:29 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 692/3300 [1:43:38<5:07:15,  7.07s/it]2026-04-13 09:04:36,454 | INFO | train_grpo_train | metrics_logged mode=train step=692

                                                      
{'loss': -0.0282, 'grad_norm': 14.995078086853027, 'learning_rate': 7.906060606060608e-06, 'num_tokens': 1220877.0, 'completions/mean_length': 39.25, 'completions/min_length': 33.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6253436803817749, 'rewards/meter/std': 0.3929384648799896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8773505687713623, 'rewards/repeat_soft/std': 0.11419042199850082, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.15412774682044983, 'rewards/total_composite/mean': 0.4678286910057068, 'rewards/total_composite/std': 0.24292142689228058, 'reward': 0.4678286910057068, 'reward_std': 0.24292141199111938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1778692752122879, 'sampling/sampling_logp_difference/max': 1.5066614151000977, 'sampling/importance_sampling_ratio/min': 0.22164873778820038, 'sampling/importance_sampling_ratio/mean': 1.030326008796692, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4388520568609238, 'clip_ratio/low_mean': 0.05966448783874512, 'clip_ratio/low_min': 0.05966448783874512, 'clip_ratio/high_mean': 0.11110597476363182, 'clip_ratio/high_max': 0.11110597476363182, 'clip_ratio/region_mean': 0.17077046260237694, 'reward_total_mean': 0.4678286910057068, 'reward_meter_mean': 0.6253436803817749, 'reward_meter_std': 0.3929384648799896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8773505687713623, 'reward_repeat_soft_std': 0.11419042199850082, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.15412774682044983, 'reward_total_composite_mean': 0.4678286910057068, 'reward_total_composite_std': 0.24292142689228058, 'epoch': 0.07}

 21%|██        | 692/3300 [1:43:38<5:07:15,  7.07s/it]INFO 04-13 09:04:36 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 693/3300 [1:43:45<5:05:20,  7.03s/it]2026-04-13 09:04:43,446 | INFO | train_grpo_train | metrics_logged mode=train step=693

                                                      
{'loss': 0.006, 'grad_norm': 11.027209281921387, 'learning_rate': 7.903030303030303e-06, 'num_tokens': 1222687.0, 'completions/mean_length': 61.25, 'completions/min_length': 52.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.25, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.76621413230896, 'rewards/meter/std': 0.3152752220630646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8095709085464478, 'rewards/repeat_soft/std': 0.14905677735805511, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.5088721513748169, 'rewards/total_composite/std': 0.10416708141565323, 'reward': 0.5088721513748169, 'reward_std': 0.10416708141565323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1528916209936142, 'sampling/sampling_logp_difference/max': 1.7419815063476562, 'sampling/importance_sampling_ratio/min': 0.17517295479774475, 'sampling/importance_sampling_ratio/mean': 1.021276831626892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1214285790920258, 'clip_ratio/low_mean': 0.041261966340243816, 'clip_ratio/low_min': 0.041261966340243816, 'clip_ratio/high_mean': 0.10720415972173214, 'clip_ratio/high_max': 0.10720415972173214, 'clip_ratio/region_mean': 0.14846612606197596, 'reward_total_mean': 0.5088721513748169, 'reward_meter_mean': 0.76621413230896, 'reward_meter_std': 0.3152752220630646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8095709085464478, 'reward_repeat_soft_std': 0.14905677735805511, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.5088721513748169, 'reward_total_composite_std': 0.10416708141565323, 'epoch': 0.07}

 21%|██        | 693/3300 [1:43:45<5:05:20,  7.03s/it]INFO 04-13 09:04:43 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 694/3300 [1:43:52<5:04:10,  7.00s/it]2026-04-13 09:04:50,499 | INFO | train_grpo_train | metrics_logged mode=train step=694

                                                      
{'loss': 0.0897, 'grad_norm': 14.673514366149902, 'learning_rate': 7.9e-06, 'num_tokens': 1224270.0, 'completions/mean_length': 38.875, 'completions/min_length': 29.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.5953438878059387, 'rewards/meter/std': 0.3709369897842407, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9078002572059631, 'rewards/repeat_soft/std': 0.08058351278305054, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.5727308988571167, 'rewards/total_composite/std': 0.20136988162994385, 'reward': 0.5727308988571167, 'reward_std': 0.20136989653110504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11624979227781296, 'sampling/sampling_logp_difference/max': 1.484346866607666, 'sampling/importance_sampling_ratio/min': 0.22665032744407654, 'sampling/importance_sampling_ratio/mean': 1.0144704580307007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9147335216403008, 'clip_ratio/low_mean': 0.04443535767495632, 'clip_ratio/low_min': 0.04443535767495632, 'clip_ratio/high_mean': 0.08160406164824963, 'clip_ratio/high_max': 0.08160406164824963, 'clip_ratio/region_mean': 0.12603941932320595, 'reward_total_mean': 0.5727308988571167, 'reward_meter_mean': 0.5953438878059387, 'reward_meter_std': 0.3709369897842407, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9078002572059631, 'reward_repeat_soft_std': 0.08058351278305054, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.5727308988571167, 'reward_total_composite_std': 0.20136988162994385, 'epoch': 0.07}

 21%|██        | 694/3300 [1:43:52<5:04:10,  7.00s/it]INFO 04-13 09:04:50 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 695/3300 [1:43:59<5:04:45,  7.02s/it]2026-04-13 09:04:57,380 | INFO | train_grpo_train | metrics_logged mode=train step=695

                                                      
{'loss': -0.0441, 'grad_norm': 17.265445709228516, 'learning_rate': 7.896969696969698e-06, 'num_tokens': 1225945.0, 'completions/mean_length': 36.375, 'completions/min_length': 29.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.42902883887290955, 'rewards/meter/std': 0.45850393176078796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9710671901702881, 'rewards/repeat_soft/std': 0.03647122159600258, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.506554901599884, 'rewards/total_composite/std': 0.3096925914287567, 'reward': 0.506554901599884, 'reward_std': 0.3096925616264343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21644631028175354, 'sampling/sampling_logp_difference/max': 1.8106884956359863, 'sampling/importance_sampling_ratio/min': 0.1635415107011795, 'sampling/importance_sampling_ratio/mean': 1.0212366580963135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5601515024900436, 'clip_ratio/low_mean': 0.12057080771774054, 'clip_ratio/low_min': 0.12057080771774054, 'clip_ratio/high_mean': 0.07483388856053352, 'clip_ratio/high_max': 0.07483388856053352, 'clip_ratio/region_mean': 0.19540469627827406, 'reward_total_mean': 0.506554901599884, 'reward_meter_mean': 0.42902883887290955, 'reward_meter_std': 0.45850393176078796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9710671901702881, 'reward_repeat_soft_std': 0.03647122159600258, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.506554901599884, 'reward_total_composite_std': 0.3096925914287567, 'epoch': 0.07}

 21%|██        | 695/3300 [1:43:59<5:04:45,  7.02s/it]INFO 04-13 09:04:57 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 696/3300 [1:44:07<5:10:33,  7.16s/it]2026-04-13 09:05:04,859 | INFO | train_grpo_train | metrics_logged mode=train step=696

                                                      
{'loss': 0.0479, 'grad_norm': 11.10001277923584, 'learning_rate': 7.893939393939395e-06, 'num_tokens': 1228150.0, 'completions/mean_length': 80.625, 'completions/min_length': 77.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.3874858617782593, 'rewards/meter/std': 0.2686140239238739, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8899877071380615, 'rewards/repeat_soft/std': 0.0854894369840622, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.3540748953819275, 'rewards/total_composite/std': 0.25735554099082947, 'reward': 0.3540748953819275, 'reward_std': 0.25735554099082947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1570495069026947, 'sampling/sampling_logp_difference/max': 1.967787742614746, 'sampling/importance_sampling_ratio/min': 0.13976570963859558, 'sampling/importance_sampling_ratio/mean': 1.0180790424346924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1517724245786667, 'clip_ratio/low_mean': 0.040153549052774906, 'clip_ratio/low_min': 0.040153549052774906, 'clip_ratio/high_mean': 0.08434855472296476, 'clip_ratio/high_max': 0.08434855472296476, 'clip_ratio/region_mean': 0.12450210377573967, 'reward_total_mean': 0.3540748953819275, 'reward_meter_mean': 0.3874858617782593, 'reward_meter_std': 0.2686140239238739, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8899877071380615, 'reward_repeat_soft_std': 0.0854894369840622, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.3540748953819275, 'reward_total_composite_std': 0.25735554099082947, 'epoch': 0.07}

 21%|██        | 696/3300 [1:44:07<5:10:33,  7.16s/it]INFO 04-13 09:05:05 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 697/3300 [1:44:13<5:01:00,  6.94s/it]2026-04-13 09:05:11,323 | INFO | train_grpo_train | metrics_logged mode=train step=697

                                                      
{'loss': 0.0636, 'grad_norm': 16.595430374145508, 'learning_rate': 7.89090909090909e-06, 'num_tokens': 1229771.0, 'completions/mean_length': 38.625, 'completions/min_length': 33.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8890440464019775, 'rewards/meter/std': 0.14752990007400513, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9806209206581116, 'rewards/repeat_soft/std': 0.023214057087898254, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.6869492530822754, 'rewards/total_composite/std': 0.12026029080152512, 'reward': 0.6869492530822754, 'reward_std': 0.12026028335094452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15296423435211182, 'sampling/sampling_logp_difference/max': 1.3662142753601074, 'sampling/importance_sampling_ratio/min': 0.2687130868434906, 'sampling/importance_sampling_ratio/mean': 1.0238951444625854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1991469413042068, 'clip_ratio/low_mean': 0.09507615398615599, 'clip_ratio/low_min': 0.09507615398615599, 'clip_ratio/high_mean': 0.060897436924278736, 'clip_ratio/high_max': 0.060897436924278736, 'clip_ratio/region_mean': 0.15597359091043472, 'reward_total_mean': 0.6869492530822754, 'reward_meter_mean': 0.8890440464019775, 'reward_meter_std': 0.14752990007400513, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9806209206581116, 'reward_repeat_soft_std': 0.023214057087898254, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.6869492530822754, 'reward_total_composite_std': 0.12026029080152512, 'epoch': 0.07}

 21%|██        | 697/3300 [1:44:13<5:01:00,  6.94s/it]INFO 04-13 09:05:11 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 698/3300 [1:44:25<6:04:07,  8.40s/it]2026-04-13 09:05:23,087 | INFO | train_grpo_train | metrics_logged mode=train step=698

                                                      
{'loss': -0.1319, 'grad_norm': 2.9320316314697266, 'learning_rate': 7.88787878787879e-06, 'num_tokens': 1231756.0, 'completions/mean_length': 114.125, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.28571701049805, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.5878835916519165, 'rewards/meter/std': 0.27737826108932495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.880814790725708, 'rewards/repeat_soft/std': 0.09316648542881012, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4399494528770447, 'rewards/total_composite/std': 0.19185154139995575, 'reward': 0.4399494528770447, 'reward_std': 0.19185155630111694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15781275928020477, 'sampling/sampling_logp_difference/max': 1.5711240768432617, 'sampling/importance_sampling_ratio/min': 0.20781145989894867, 'sampling/importance_sampling_ratio/mean': 1.0406392812728882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0126184672117233, 'clip_ratio/low_mean': 0.01923076994717121, 'clip_ratio/low_min': 0.01923076994717121, 'clip_ratio/high_mean': 0.09851447865366936, 'clip_ratio/high_max': 0.09851447865366936, 'clip_ratio/region_mean': 0.11774524860084057, 'reward_total_mean': 0.4399494528770447, 'reward_meter_mean': 0.5878835916519165, 'reward_meter_std': 0.27737826108932495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.880814790725708, 'reward_repeat_soft_std': 0.09316648542881012, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4399494528770447, 'reward_total_composite_std': 0.19185154139995575, 'epoch': 0.07}

 21%|██        | 698/3300 [1:44:25<6:04:07,  8.40s/it]INFO 04-13 09:05:23 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 699/3300 [1:44:32<5:54:47,  8.18s/it]2026-04-13 09:05:30,786 | INFO | train_grpo_train | metrics_logged mode=train step=699

                                                      
{'loss': -0.0128, 'grad_norm': 9.37679386138916, 'learning_rate': 7.884848484848485e-06, 'num_tokens': 1234569.0, 'completions/mean_length': 116.625, 'completions/min_length': 108.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.625, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.43015190958976746, 'rewards/meter/std': 0.3419077694416046, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8563051819801331, 'rewards/repeat_soft/std': 0.1234939694404602, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.15638209879398346, 'rewards/total_composite/mean': 0.4412240982055664, 'rewards/total_composite/std': 0.1003093272447586, 'reward': 0.4412240982055664, 'reward_std': 0.10030930489301682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13429109752178192, 'sampling/sampling_logp_difference/max': 2.4016456604003906, 'sampling/importance_sampling_ratio/min': 0.09056878834962845, 'sampling/importance_sampling_ratio/mean': 1.0258796215057373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9660872519016266, 'clip_ratio/low_mean': 0.052333916537463665, 'clip_ratio/low_min': 0.052333916537463665, 'clip_ratio/high_mean': 0.06842698622494936, 'clip_ratio/high_max': 0.06842698622494936, 'clip_ratio/region_mean': 0.12076090276241302, 'reward_total_mean': 0.4412240982055664, 'reward_meter_mean': 0.43015190958976746, 'reward_meter_std': 0.3419077694416046, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8563051819801331, 'reward_repeat_soft_std': 0.1234939694404602, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.15638209879398346, 'reward_total_composite_mean': 0.4412240982055664, 'reward_total_composite_std': 0.1003093272447586, 'epoch': 0.07}

 21%|██        | 699/3300 [1:44:33<5:54:47,  8.18s/it]INFO 04-13 09:05:31 [block_pool.py:378] Successfully reset prefix cache

 21%|██        | 700/3300 [1:44:39<5:37:10,  7.78s/it]2026-04-13 09:05:37,614 | INFO | train_grpo_train | metrics_logged mode=train step=700

                                                      
{'loss': 0.0249, 'grad_norm': 10.956107139587402, 'learning_rate': 7.881818181818182e-06, 'num_tokens': 1236389.0, 'completions/mean_length': 56.5, 'completions/min_length': 51.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.699101448059082, 'rewards/meter/std': 0.26901111006736755, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9079020023345947, 'rewards/repeat_soft/std': 0.044114850461483, 'rewards/judge_quality/mean': 0.8075000047683716, 'rewards/judge_quality/std': 0.18077215552330017, 'rewards/total_composite/mean': 0.6893068552017212, 'rewards/total_composite/std': 0.14247862994670868, 'reward': 0.6893068552017212, 'reward_std': 0.14247861504554749, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16456753015518188, 'sampling/sampling_logp_difference/max': 1.9089515209197998, 'sampling/importance_sampling_ratio/min': 0.1482357233762741, 'sampling/importance_sampling_ratio/mean': 0.9832058548927307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0508423298597336, 'clip_ratio/low_mean': 0.0792367160320282, 'clip_ratio/low_min': 0.0792367160320282, 'clip_ratio/high_mean': 0.07942836545407772, 'clip_ratio/high_max': 0.07942836545407772, 'clip_ratio/region_mean': 0.15866508148610592, 'reward_total_mean': 0.6893068552017212, 'reward_meter_mean': 0.699101448059082, 'reward_meter_std': 0.26901111006736755, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9079020023345947, 'reward_repeat_soft_std': 0.044114850461483, 'reward_judge_quality_mean': 0.8075000047683716, 'reward_judge_quality_std': 0.18077215552330017, 'reward_total_composite_mean': 0.6893068552017212, 'reward_total_composite_std': 0.14247862994670868, 'epoch': 0.07}

 21%|██        | 700/3300 [1:44:39<5:37:10,  7.78s/it]INFO 04-13 09:05:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:12,  1.53s/it]

 30%|███       | 3/10 [00:07<00:20,  2.93s/it]

 40%|████      | 4/10 [00:11<00:18,  3.07s/it]

 50%|█████     | 5/10 [00:15<00:16,  3.39s/it]

 60%|██████    | 6/10 [00:18<00:13,  3.47s/it]

 70%|███████   | 7/10 [00:22<00:10,  3.61s/it]

 80%|████████  | 8/10 [00:26<00:07,  3.59s/it]

 90%|█████████ | 9/10 [00:35<00:05,  5.18s/it]

100%|██████████| 10/10 [00:40<00:00,  5.19s/it]2026-04-13 09:06:24,791 | INFO | train_grpo_train | metrics_logged mode=eval step=700

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 47.094, 'eval_samples_per_second': 1.699, 'eval_steps_per_second': 0.212, 'eval_num_tokens': 1236389.0, 'eval_completions/mean_length': 63.2375, 'eval_completions/min_length': 26.6, 'eval_completions/max_length': 140.4, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 57.42678604125977, 'eval_completions/min_terminated_length': 26.6, 'eval_completions/max_terminated_length': 99.0, 'eval_rewards/meter/mean': 0.562597519159317, 'eval_rewards/meter/std': 0.3473539650440216, 'eval_rewards/count_adherence/mean': 0.9895833313465119, 'eval_rewards/count_adherence/std': 0.029462784901261328, 'eval_rewards/hard_gate/mean': 0.9625, 'eval_rewards/hard_gate/std': 0.10606601536273956, 'eval_rewards/repeat_soft/mean': 0.875518399477005, 'eval_rewards/repeat_soft/std': 0.11491637155413628, 'eval_rewards/judge_quality/mean': 0.5047500103712081, 'eval_rewards/judge_quality/std': 0.20631255060434342, 'eval_rewards/total_composite/mean': 0.4972403317689896, 'eval_rewards/total_composite/std': 0.1763676755130291, 'eval_reward': 0.4972403317689896, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0910422645509243, 'eval_sampling/sampling_logp_difference/max': 1.0577627658843993, 'eval_sampling/importance_sampling_ratio/min': 0.353044393658638, 'eval_sampling/importance_sampling_ratio/mean': 1.0241777658462525, 'eval_sampling/importance_sampling_ratio/max': 1.5206967353820802, 'eval_entropy': 1.0876900553703308, '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': 0.4972403317689896, 'eval_reward_meter_mean': 0.562597519159317, 'eval_reward_meter_std': 0.3473539650440216, 'eval_reward_count_adherence_mean': 0.9895833313465119, 'eval_reward_count_adherence_std': 0.029462784901261328, 'eval_reward_hard_gate_mean': 0.9625, 'eval_reward_hard_gate_std': 0.10606601536273956, 'eval_reward_repeat_soft_mean': 0.875518399477005, 'eval_reward_repeat_soft_std': 0.11491637155413628, 'eval_reward_judge_quality_mean': 0.5047500103712081, 'eval_reward_judge_quality_std': 0.20631255060434342, 'eval_reward_total_composite_mean': 0.4972403317689896, 'eval_reward_total_composite_std': 0.1763676755130291, 'epoch': 0.07}

 21%|██        | 700/3300 [1:45:27<5:37:10,  7.78s/it]

100%|██████████| 10/10 [00:40<00:00,  5.19s/it]

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

 21%|██        | 701/3300 [1:45:35<16:06:10, 22.30s/it]2026-04-13 09:06:33,810 | INFO | train_grpo_train | metrics_logged mode=train step=701

                                                       
{'loss': 0.0305, 'grad_norm': 16.376840591430664, 'learning_rate': 7.87878787878788e-06, 'num_tokens': 1237978.0, 'completions/mean_length': 32.625, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.48679500818252563, 'rewards/meter/std': 0.3600373864173889, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8722328543663025, 'rewards/repeat_soft/std': 0.07642845809459686, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.47136837244033813, 'rewards/total_composite/std': 0.09856859594583511, 'reward': 0.47136837244033813, 'reward_std': 0.09856861084699631, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11635728180408478, 'sampling/sampling_logp_difference/max': 1.0491828918457031, 'sampling/importance_sampling_ratio/min': 0.3502238094806671, 'sampling/importance_sampling_ratio/mean': 1.0372222661972046, 'sampling/importance_sampling_ratio/max': 1.8948490619659424, 'entropy': 0.9957318902015686, 'clip_ratio/low_mean': 0.06924716010689735, 'clip_ratio/low_min': 0.06924716010689735, 'clip_ratio/high_mean': 0.05016236566007137, 'clip_ratio/high_max': 0.05016236566007137, 'clip_ratio/region_mean': 0.11940952576696873, 'reward_total_mean': 0.47136837244033813, 'reward_meter_mean': 0.48679500818252563, 'reward_meter_std': 0.3600373864173889, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8722328543663025, 'reward_repeat_soft_std': 0.07642845809459686, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.47136837244033813, 'reward_total_composite_std': 0.09856859594583511, 'epoch': 0.07}

 21%|██        | 701/3300 [1:45:36<16:06:10, 22.30s/it]INFO 04-13 09:06:34 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 702/3300 [1:45:43<12:47:26, 17.72s/it]2026-04-13 09:06:40,852 | INFO | train_grpo_train | metrics_logged mode=train step=702

                                                       
{'loss': -0.0126, 'grad_norm': 13.545258522033691, 'learning_rate': 7.875757575757577e-06, 'num_tokens': 1240039.0, 'completions/mean_length': 67.625, 'completions/min_length': 61.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.12041516602039337, 'rewards/meter/std': 0.11534888297319412, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8962299227714539, 'rewards/repeat_soft/std': 0.11184307932853699, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.38222992420196533, 'rewards/total_composite/std': 0.06179346889257431, 'reward': 0.38222992420196533, 'reward_std': 0.06179346516728401, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1679002195596695, 'sampling/sampling_logp_difference/max': 1.4634904861450195, 'sampling/importance_sampling_ratio/min': 0.23142707347869873, 'sampling/importance_sampling_ratio/mean': 1.0432548522949219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4721038043498993, 'clip_ratio/low_mean': 0.05922293849289417, 'clip_ratio/low_min': 0.05922293849289417, 'clip_ratio/high_mean': 0.06900245696306229, 'clip_ratio/high_max': 0.06900245696306229, 'clip_ratio/region_mean': 0.12822539545595646, 'reward_total_mean': 0.38222992420196533, 'reward_meter_mean': 0.12041516602039337, 'reward_meter_std': 0.11534888297319412, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8962299227714539, 'reward_repeat_soft_std': 0.11184307932853699, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.38222992420196533, 'reward_total_composite_std': 0.06179346889257431, 'epoch': 0.07}

 21%|██▏       | 702/3300 [1:45:43<12:47:26, 17.72s/it]INFO 04-13 09:06:41 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 703/3300 [1:45:54<11:27:04, 15.87s/it]2026-04-13 09:06:52,412 | INFO | train_grpo_train | metrics_logged mode=train step=703

                                                       
{'loss': -0.1726, 'grad_norm': 2.6292431354522705, 'learning_rate': 7.872727272727273e-06, 'num_tokens': 1242190.0, 'completions/mean_length': 140.875, 'completions/min_length': 74.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 87.85714721679688, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.7005671262741089, 'rewards/meter/std': 0.3341540992259979, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8083329200744629, 'rewards/repeat_soft/std': 0.10700494050979614, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.15733835101127625, 'rewards/total_composite/mean': 0.4865454435348511, 'rewards/total_composite/std': 0.20896413922309875, 'reward': 0.4865454435348511, 'reward_std': 0.20896413922309875, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1460670679807663, 'sampling/sampling_logp_difference/max': 1.56398344039917, 'sampling/importance_sampling_ratio/min': 0.2152041792869568, 'sampling/importance_sampling_ratio/mean': 1.011475920677185, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.854569286108017, 'clip_ratio/low_mean': 0.03128632344305515, 'clip_ratio/low_min': 0.03128632344305515, 'clip_ratio/high_mean': 0.07816854817792773, 'clip_ratio/high_max': 0.07816854817792773, 'clip_ratio/region_mean': 0.10945487162098289, 'reward_total_mean': 0.4865454435348511, 'reward_meter_mean': 0.7005671262741089, 'reward_meter_std': 0.3341540992259979, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8083329200744629, 'reward_repeat_soft_std': 0.10700494050979614, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.15733835101127625, 'reward_total_composite_mean': 0.4865454435348511, 'reward_total_composite_std': 0.20896413922309875, 'epoch': 0.07}

 21%|██▏       | 703/3300 [1:45:54<11:27:04, 15.87s/it]INFO 04-13 09:06:52 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 704/3300 [1:46:01<9:26:09, 13.09s/it] 2026-04-13 09:06:59,065 | INFO | train_grpo_train | metrics_logged mode=train step=704

                                                      
{'loss': 0.0239, 'grad_norm': 16.691980361938477, 'learning_rate': 7.86969696969697e-06, 'num_tokens': 1244265.0, 'completions/mean_length': 70.375, 'completions/min_length': 62.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.375, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.3741535544395447, 'rewards/meter/std': 0.345613956451416, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9200815558433533, 'rewards/repeat_soft/std': 0.08679132908582687, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.37162142992019653, 'rewards/total_composite/std': 0.16446079313755035, 'reward': 0.37162142992019653, 'reward_std': 0.16446080803871155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20231188833713531, 'sampling/sampling_logp_difference/max': 2.399839401245117, 'sampling/importance_sampling_ratio/min': 0.09073252975940704, 'sampling/importance_sampling_ratio/mean': 0.9932599067687988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9807969853281975, 'clip_ratio/low_mean': 0.014925372786819935, 'clip_ratio/low_min': 0.014925372786819935, 'clip_ratio/high_mean': 0.18753970600664616, 'clip_ratio/high_max': 0.18753970600664616, 'clip_ratio/region_mean': 0.2024650787934661, 'reward_total_mean': 0.37162142992019653, 'reward_meter_mean': 0.3741535544395447, 'reward_meter_std': 0.345613956451416, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9200815558433533, 'reward_repeat_soft_std': 0.08679132908582687, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.37162142992019653, 'reward_total_composite_std': 0.16446079313755035, 'epoch': 0.07}

 21%|██▏       | 704/3300 [1:46:01<9:26:09, 13.09s/it]INFO 04-13 09:06:59 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 705/3300 [1:46:08<8:08:28, 11.29s/it]2026-04-13 09:07:06,145 | INFO | train_grpo_train | metrics_logged mode=train step=705

                                                      
{'loss': 0.0388, 'grad_norm': 10.425191879272461, 'learning_rate': 7.866666666666667e-06, 'num_tokens': 1246456.0, 'completions/mean_length': 96.875, 'completions/min_length': 88.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.875, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.6700539588928223, 'rewards/meter/std': 0.3244590163230896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9819424152374268, 'rewards/repeat_soft/std': 0.01510180626064539, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.18516401946544647, 'rewards/total_composite/mean': 0.5936654806137085, 'rewards/total_composite/std': 0.1849101483821869, 'reward': 0.5936654806137085, 'reward_std': 0.1849101483821869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1460622400045395, 'sampling/sampling_logp_difference/max': 1.8150241374969482, 'sampling/importance_sampling_ratio/min': 0.16283397376537323, 'sampling/importance_sampling_ratio/mean': 0.9983146786689758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7747109085321426, 'clip_ratio/low_mean': 0.07223216630518436, 'clip_ratio/low_min': 0.07223216630518436, 'clip_ratio/high_mean': 0.07081507612019777, 'clip_ratio/high_max': 0.07081507612019777, 'clip_ratio/region_mean': 0.14304724242538214, 'reward_total_mean': 0.5936654806137085, 'reward_meter_mean': 0.6700539588928223, 'reward_meter_std': 0.3244590163230896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9819424152374268, 'reward_repeat_soft_std': 0.01510180626064539, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.18516401946544647, 'reward_total_composite_mean': 0.5936654806137085, 'reward_total_composite_std': 0.1849101483821869, 'epoch': 0.07}

 21%|██▏       | 705/3300 [1:46:08<8:08:28, 11.29s/it]INFO 04-13 09:07:06 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 706/3300 [1:46:19<8:12:18, 11.39s/it]2026-04-13 09:07:17,705 | INFO | train_grpo_train | metrics_logged mode=train step=706

                                                      
{'loss': -0.1275, 'grad_norm': 6.171966075897217, 'learning_rate': 7.863636363636364e-06, 'num_tokens': 1248227.0, 'completions/mean_length': 116.375, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.857147216796875, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.20292825996875763, 'rewards/meter/std': 0.3384077847003937, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9816411733627319, 'rewards/repeat_soft/std': 0.02059292607009411, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.3845466673374176, 'rewards/total_composite/std': 0.2538537383079529, 'reward': 0.3845466673374176, 'reward_std': 0.2538537383079529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22096796333789825, 'sampling/sampling_logp_difference/max': 1.3180179595947266, 'sampling/importance_sampling_ratio/min': 0.267665296792984, 'sampling/importance_sampling_ratio/mean': 1.0719823837280273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.050292491912842, 'clip_ratio/low_mean': 0.1373369675129652, 'clip_ratio/low_min': 0.1373369675129652, 'clip_ratio/high_mean': 0.01909722201526165, 'clip_ratio/high_max': 0.01909722201526165, 'clip_ratio/region_mean': 0.15643418952822685, 'reward_total_mean': 0.3845466673374176, 'reward_meter_mean': 0.20292825996875763, 'reward_meter_std': 0.3384077847003937, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9816411733627319, 'reward_repeat_soft_std': 0.02059292607009411, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.3845466673374176, 'reward_total_composite_std': 0.2538537383079529, 'epoch': 0.07}

 21%|██▏       | 706/3300 [1:46:19<8:12:18, 11.39s/it]INFO 04-13 09:07:18 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 707/3300 [1:46:27<7:21:14, 10.21s/it]2026-04-13 09:07:25,174 | INFO | train_grpo_train | metrics_logged mode=train step=707

                                                      
{'loss': -0.015, 'grad_norm': 8.014169692993164, 'learning_rate': 7.860606060606062e-06, 'num_tokens': 1250734.0, 'completions/mean_length': 112.375, 'completions/min_length': 105.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.375, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.5072179436683655, 'rewards/meter/std': 0.3544645607471466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7135148048400879, 'rewards/repeat_soft/std': 0.0966494232416153, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.423446387052536, 'rewards/total_composite/std': 0.22093810141086578, 'reward': 0.423446387052536, 'reward_std': 0.2209380865097046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11801961064338684, 'sampling/sampling_logp_difference/max': 1.608119249343872, 'sampling/importance_sampling_ratio/min': 0.20026391744613647, 'sampling/importance_sampling_ratio/mean': 1.0145412683486938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8730439320206642, 'clip_ratio/low_mean': 0.03968938160687685, 'clip_ratio/low_min': 0.03968938160687685, 'clip_ratio/high_mean': 0.07308758050203323, 'clip_ratio/high_max': 0.07308758050203323, 'clip_ratio/region_mean': 0.11277696210891008, 'reward_total_mean': 0.423446387052536, 'reward_meter_mean': 0.5072179436683655, 'reward_meter_std': 0.3544645607471466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7135148048400879, 'reward_repeat_soft_std': 0.0966494232416153, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.423446387052536, 'reward_total_composite_std': 0.22093810141086578, 'epoch': 0.07}

 21%|██▏       | 707/3300 [1:46:27<7:21:14, 10.21s/it]INFO 04-13 09:07:25 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 708/3300 [1:46:39<7:50:47, 10.90s/it]2026-04-13 09:07:37,672 | INFO | train_grpo_train | metrics_logged mode=train step=708

                                                      
{'loss': -0.1553, 'grad_norm': 1.827104926109314, 'learning_rate': 7.857575757575759e-06, 'num_tokens': 1252520.0, 'completions/mean_length': 247.25, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 88.4000015258789, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.584979772567749, 'rewards/meter/std': 0.2787756323814392, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.31160587072372437, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.7171851396560669, 'rewards/repeat_soft/std': 0.09273076057434082, 'rewards/judge_quality/mean': 0.23874999582767487, 'rewards/judge_quality/std': 0.17141741514205933, 'rewards/total_composite/mean': 0.2731322646141052, 'rewards/total_composite/std': 0.22805926203727722, 'reward': 0.2731322646141052, 'reward_std': 0.22805923223495483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13770842552185059, 'sampling/sampling_logp_difference/max': 2.203799247741699, 'sampling/importance_sampling_ratio/min': 0.11038298159837723, 'sampling/importance_sampling_ratio/mean': 1.0084809064865112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6015427559614182, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08695725258439779, 'clip_ratio/high_max': 0.08695725258439779, 'clip_ratio/region_mean': 0.08695725258439779, 'reward_total_mean': 0.2731322646141052, 'reward_meter_mean': 0.584979772567749, 'reward_meter_std': 0.2787756323814392, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.31160587072372437, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.7171851396560669, 'reward_repeat_soft_std': 0.09273076057434082, 'reward_judge_quality_mean': 0.23874999582767487, 'reward_judge_quality_std': 0.17141741514205933, 'reward_total_composite_mean': 0.2731322646141052, 'reward_total_composite_std': 0.22805926203727722, 'epoch': 0.07}

 21%|██▏       | 708/3300 [1:46:39<7:50:47, 10.90s/it]INFO 04-13 09:07:37 [block_pool.py:378] Successfully reset prefix cache

 21%|██▏       | 709/3300 [1:46:46<6:59:10,  9.71s/it]2026-04-13 09:07:44,601 | INFO | train_grpo_train | metrics_logged mode=train step=709

                                                      
{'loss': 0.0795, 'grad_norm': 15.316978454589844, 'learning_rate': 7.854545454545454e-06, 'num_tokens': 1253990.0, 'completions/mean_length': 33.75, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4370405077934265, 'rewards/meter/std': 0.31697505712509155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9854952096939087, 'rewards/repeat_soft/std': 0.012407306581735611, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.22258225083351135, 'rewards/total_composite/mean': 0.5258222222328186, 'rewards/total_composite/std': 0.16406190395355225, 'reward': 0.5258222222328186, 'reward_std': 0.16406190395355225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1508551388978958, 'sampling/sampling_logp_difference/max': 1.278902530670166, 'sampling/importance_sampling_ratio/min': 0.278342604637146, 'sampling/importance_sampling_ratio/mean': 1.0059592723846436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0151208117604256, 'clip_ratio/low_mean': 0.06558240670710802, 'clip_ratio/low_min': 0.06558240670710802, 'clip_ratio/high_mean': 0.06477042380720377, 'clip_ratio/high_max': 0.06477042380720377, 'clip_ratio/region_mean': 0.1303528305143118, 'reward_total_mean': 0.5258222222328186, 'reward_meter_mean': 0.4370405077934265, 'reward_meter_std': 0.31697505712509155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9854952096939087, 'reward_repeat_soft_std': 0.012407306581735611, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.22258225083351135, 'reward_total_composite_mean': 0.5258222222328186, 'reward_total_composite_std': 0.16406190395355225, 'epoch': 0.07}

 21%|██▏       | 709/3300 [1:46:46<6:59:10,  9.71s/it]INFO 04-13 09:07:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 710/3300 [1:46:53<6:19:06,  8.78s/it]2026-04-13 09:07:51,224 | INFO | train_grpo_train | metrics_logged mode=train step=710

                                                      
{'loss': 0.0538, 'grad_norm': 13.484710693359375, 'learning_rate': 7.851515151515152e-06, 'num_tokens': 1255632.0, 'completions/mean_length': 42.25, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8817436099052429, 'rewards/meter/std': 0.178593248128891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.799583375453949, 'rewards/repeat_soft/std': 0.14987260103225708, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5841915011405945, 'rewards/total_composite/std': 0.09952611476182938, 'reward': 0.5841915011405945, 'reward_std': 0.09952611476182938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1483382135629654, 'sampling/sampling_logp_difference/max': 1.8885650634765625, 'sampling/importance_sampling_ratio/min': 0.15128874778747559, 'sampling/importance_sampling_ratio/mean': 1.0168503522872925, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.999920554459095, 'clip_ratio/low_mean': 0.08082265174016356, 'clip_ratio/low_min': 0.08082265174016356, 'clip_ratio/high_mean': 0.04457815829664469, 'clip_ratio/high_max': 0.04457815829664469, 'clip_ratio/region_mean': 0.12540081003680825, 'reward_total_mean': 0.5841915011405945, 'reward_meter_mean': 0.8817436099052429, 'reward_meter_std': 0.178593248128891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.799583375453949, 'reward_repeat_soft_std': 0.14987260103225708, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5841915011405945, 'reward_total_composite_std': 0.09952611476182938, 'epoch': 0.07}

 22%|██▏       | 710/3300 [1:46:53<6:19:06,  8.78s/it]INFO 04-13 09:07:51 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 711/3300 [1:46:59<5:47:21,  8.05s/it]2026-04-13 09:07:57,560 | INFO | train_grpo_train | metrics_logged mode=train step=711

                                                      
{'loss': 0.0095, 'grad_norm': 16.092296600341797, 'learning_rate': 7.848484848484849e-06, 'num_tokens': 1257169.0, 'completions/mean_length': 31.125, 'completions/min_length': 30.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6596676111221313, 'rewards/meter/std': 0.38747116923332214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9813710451126099, 'rewards/repeat_soft/std': 0.017808331176638603, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6435909271240234, 'rewards/total_composite/std': 0.24125070869922638, 'reward': 0.6435909271240234, 'reward_std': 0.24125069379806519, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12443126738071442, 'sampling/sampling_logp_difference/max': 2.83040452003479, 'sampling/importance_sampling_ratio/min': 0.058988988399505615, 'sampling/importance_sampling_ratio/mean': 1.0007023811340332, 'sampling/importance_sampling_ratio/max': 1.543837308883667, 'entropy': 0.8394577726721764, 'clip_ratio/low_mean': 0.08185483934357762, 'clip_ratio/low_min': 0.08185483934357762, 'clip_ratio/high_mean': 0.06256670970469713, 'clip_ratio/high_max': 0.06256670970469713, 'clip_ratio/region_mean': 0.14442154904827476, 'reward_total_mean': 0.6435909271240234, 'reward_meter_mean': 0.6596676111221313, 'reward_meter_std': 0.38747116923332214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9813710451126099, 'reward_repeat_soft_std': 0.017808331176638603, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6435909271240234, 'reward_total_composite_std': 0.24125070869922638, 'epoch': 0.07}

 22%|██▏       | 711/3300 [1:46:59<5:47:21,  8.05s/it]INFO 04-13 09:07:57 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 712/3300 [1:47:06<5:28:59,  7.63s/it]2026-04-13 09:08:04,207 | INFO | train_grpo_train | metrics_logged mode=train step=712

                                                      
{'loss': 0.0357, 'grad_norm': 18.334585189819336, 'learning_rate': 7.845454545454546e-06, 'num_tokens': 1258632.0, 'completions/mean_length': 34.875, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.14928486943244934, 'rewards/meter/std': 0.14895722270011902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9914079308509827, 'rewards/repeat_soft/std': 0.010679380968213081, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.39553284645080566, 'rewards/total_composite/std': 0.044622622430324554, 'reward': 0.39553284645080566, 'reward_std': 0.04462261125445366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16986387968063354, 'sampling/sampling_logp_difference/max': 1.7969157695770264, 'sampling/importance_sampling_ratio/min': 0.1658094972372055, 'sampling/importance_sampling_ratio/mean': 1.0382651090621948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4820304661989212, 'clip_ratio/low_mean': 0.08604302210733294, 'clip_ratio/low_min': 0.08604302210733294, 'clip_ratio/high_mean': 0.06095559988170862, 'clip_ratio/high_max': 0.06095559988170862, 'clip_ratio/region_mean': 0.14699862198904157, 'reward_total_mean': 0.39553284645080566, 'reward_meter_mean': 0.14928486943244934, 'reward_meter_std': 0.14895722270011902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9914079308509827, 'reward_repeat_soft_std': 0.010679380968213081, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.39553284645080566, 'reward_total_composite_std': 0.044622622430324554, 'epoch': 0.07}

 22%|██▏       | 712/3300 [1:47:06<5:28:59,  7.63s/it]INFO 04-13 09:08:04 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 713/3300 [1:47:13<5:22:00,  7.47s/it]2026-04-13 09:08:11,301 | INFO | train_grpo_train | metrics_logged mode=train step=713

                                                      
{'loss': 0.0261, 'grad_norm': 17.9517879486084, 'learning_rate': 7.842424242424243e-06, 'num_tokens': 1260195.0, 'completions/mean_length': 36.375, 'completions/min_length': 29.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8564504981040955, 'rewards/meter/std': 0.21740178763866425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9412529468536377, 'rewards/repeat_soft/std': 0.07152710855007172, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.194054514169693, 'rewards/total_composite/mean': 0.6544015407562256, 'rewards/total_composite/std': 0.1400221437215805, 'reward': 0.6544015407562256, 'reward_std': 0.1400221586227417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18278828263282776, 'sampling/sampling_logp_difference/max': 1.9740679264068604, 'sampling/importance_sampling_ratio/min': 0.13889069855213165, 'sampling/importance_sampling_ratio/mean': 0.9994227886199951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0823270753026009, 'clip_ratio/low_mean': 0.07796026673167944, 'clip_ratio/low_min': 0.07796026673167944, 'clip_ratio/high_mean': 0.0432076808065176, 'clip_ratio/high_max': 0.0432076808065176, 'clip_ratio/region_mean': 0.12116794753819704, 'reward_total_mean': 0.6544015407562256, 'reward_meter_mean': 0.8564504981040955, 'reward_meter_std': 0.21740178763866425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9412529468536377, 'reward_repeat_soft_std': 0.07152710855007172, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.194054514169693, 'reward_total_composite_mean': 0.6544015407562256, 'reward_total_composite_std': 0.1400221437215805, 'epoch': 0.07}

 22%|██▏       | 713/3300 [1:47:13<5:22:00,  7.47s/it]INFO 04-13 09:08:11 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 714/3300 [1:47:20<5:19:28,  7.41s/it]2026-04-13 09:08:18,577 | INFO | train_grpo_train | metrics_logged mode=train step=714

                                                      
{'loss': 0.0862, 'grad_norm': 20.394691467285156, 'learning_rate': 7.83939393939394e-06, 'num_tokens': 1261660.0, 'completions/mean_length': 23.125, 'completions/min_length': 17.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.5563315749168396, 'rewards/meter/std': 0.46619367599487305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.926868736743927, 'rewards/repeat_soft/std': 0.08202483505010605, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5017377138137817, 'rewards/total_composite/std': 0.14172938466072083, 'reward': 0.5017377138137817, 'reward_std': 0.14172936975955963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1598486304283142, 'sampling/sampling_logp_difference/max': 1.5706138610839844, 'sampling/importance_sampling_ratio/min': 0.2079174965620041, 'sampling/importance_sampling_ratio/mean': 1.0144904851913452, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1529867574572563, 'clip_ratio/low_mean': 0.030833333730697632, 'clip_ratio/low_min': 0.030833333730697632, 'clip_ratio/high_mean': 0.10804452188313007, 'clip_ratio/high_max': 0.10804452188313007, 'clip_ratio/region_mean': 0.1388778556138277, 'reward_total_mean': 0.5017377138137817, 'reward_meter_mean': 0.5563315749168396, 'reward_meter_std': 0.46619367599487305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.926868736743927, 'reward_repeat_soft_std': 0.08202483505010605, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5017377138137817, 'reward_total_composite_std': 0.14172938466072083, 'epoch': 0.07}

 22%|██▏       | 714/3300 [1:47:20<5:19:28,  7.41s/it]INFO 04-13 09:08:18 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 715/3300 [1:47:27<5:16:04,  7.34s/it]2026-04-13 09:08:25,738 | INFO | train_grpo_train | metrics_logged mode=train step=715

                                                      
{'loss': -0.0115, 'grad_norm': 12.384317398071289, 'learning_rate': 7.836363636363638e-06, 'num_tokens': 1263534.0, 'completions/mean_length': 70.25, 'completions/min_length': 61.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.37978553771972656, 'rewards/meter/std': 0.3818439245223999, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9859769344329834, 'rewards/repeat_soft/std': 0.007877059280872345, 'rewards/judge_quality/mean': 0.7487500309944153, 'rewards/judge_quality/std': 0.0699872374534607, 'rewards/total_composite/mean': 0.5291460156440735, 'rewards/total_composite/std': 0.18106192350387573, 'reward': 0.5291460156440735, 'reward_std': 0.18106190860271454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1295756697654724, 'sampling/sampling_logp_difference/max': 1.5767947435379028, 'sampling/importance_sampling_ratio/min': 0.20663635432720184, 'sampling/importance_sampling_ratio/mean': 1.0221647024154663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7575226053595543, 'clip_ratio/low_mean': 0.07443731464445591, 'clip_ratio/low_min': 0.07443731464445591, 'clip_ratio/high_mean': 0.046102119609713554, 'clip_ratio/high_max': 0.046102119609713554, 'clip_ratio/region_mean': 0.12053943425416946, 'reward_total_mean': 0.5291460156440735, 'reward_meter_mean': 0.37978553771972656, 'reward_meter_std': 0.3818439245223999, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9859769344329834, 'reward_repeat_soft_std': 0.007877059280872345, 'reward_judge_quality_mean': 0.7487500309944153, 'reward_judge_quality_std': 0.0699872374534607, 'reward_total_composite_mean': 0.5291460156440735, 'reward_total_composite_std': 0.18106192350387573, 'epoch': 0.07}

 22%|██▏       | 715/3300 [1:47:27<5:16:04,  7.34s/it]INFO 04-13 09:08:25 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 716/3300 [1:47:35<5:12:35,  7.26s/it]2026-04-13 09:08:32,812 | INFO | train_grpo_train | metrics_logged mode=train step=716

                                                      
{'loss': 0.0699, 'grad_norm': 12.60446834564209, 'learning_rate': 7.833333333333333e-06, 'num_tokens': 1265443.0, 'completions/mean_length': 62.625, 'completions/min_length': 56.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.2881770730018616, 'rewards/meter/std': 0.13510683178901672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9140480160713196, 'rewards/repeat_soft/std': 0.09184758365154266, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.4229186177253723, 'rewards/total_composite/std': 0.05308286473155022, 'reward': 0.4229186177253723, 'reward_std': 0.05308286473155022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15849167108535767, 'sampling/sampling_logp_difference/max': 1.5524301528930664, 'sampling/importance_sampling_ratio/min': 0.21173278987407684, 'sampling/importance_sampling_ratio/mean': 1.0272974967956543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2141590863466263, 'clip_ratio/low_mean': 0.09201069548726082, 'clip_ratio/low_min': 0.09201069548726082, 'clip_ratio/high_mean': 0.05910267308354378, 'clip_ratio/high_max': 0.05910267308354378, 'clip_ratio/region_mean': 0.1511133685708046, 'reward_total_mean': 0.4229186177253723, 'reward_meter_mean': 0.2881770730018616, 'reward_meter_std': 0.13510683178901672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9140480160713196, 'reward_repeat_soft_std': 0.09184758365154266, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.4229186177253723, 'reward_total_composite_std': 0.05308286473155022, 'epoch': 0.07}

 22%|██▏       | 716/3300 [1:47:35<5:12:35,  7.26s/it]INFO 04-13 09:08:33 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 717/3300 [1:47:47<6:13:59,  8.69s/it]2026-04-13 09:08:44,839 | INFO | train_grpo_train | metrics_logged mode=train step=717

                                                      
{'loss': -0.1462, 'grad_norm': 3.343062400817871, 'learning_rate': 7.83030303030303e-06, 'num_tokens': 1267277.0, 'completions/mean_length': 120.25, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.28572082519531, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.31134355068206787, 'rewards/meter/std': 0.24840696156024933, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8691983222961426, 'rewards/repeat_soft/std': 0.09458259493112564, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.3813011646270752, 'rewards/total_composite/std': 0.173007532954216, 'reward': 0.3813011646270752, 'reward_std': 0.1730075180530548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13522842526435852, 'sampling/sampling_logp_difference/max': 1.5793147087097168, 'sampling/importance_sampling_ratio/min': 0.20611628890037537, 'sampling/importance_sampling_ratio/mean': 1.0342615842819214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9020535796880722, 'clip_ratio/low_mean': 0.05329301115125418, 'clip_ratio/low_min': 0.05329301115125418, 'clip_ratio/high_mean': 0.07759564556181431, 'clip_ratio/high_max': 0.07759564556181431, 'clip_ratio/region_mean': 0.13088865671306849, 'reward_total_mean': 0.3813011646270752, 'reward_meter_mean': 0.31134355068206787, 'reward_meter_std': 0.24840696156024933, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8691983222961426, 'reward_repeat_soft_std': 0.09458259493112564, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.3813011646270752, 'reward_total_composite_std': 0.173007532954216, 'epoch': 0.07}

 22%|██▏       | 717/3300 [1:47:47<6:13:59,  8.69s/it]INFO 04-13 09:08:45 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 718/3300 [1:47:53<5:47:37,  8.08s/it]2026-04-13 09:08:51,490 | INFO | train_grpo_train | metrics_logged mode=train step=718

                                                      
{'loss': 0.0157, 'grad_norm': 18.708446502685547, 'learning_rate': 7.827272727272728e-06, 'num_tokens': 1268898.0, 'completions/mean_length': 41.625, 'completions/min_length': 40.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7582693099975586, 'rewards/meter/std': 0.3461999297142029, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9715173244476318, 'rewards/repeat_soft/std': 0.04452996328473091, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5789026021957397, 'rewards/total_composite/std': 0.09932930022478104, 'reward': 0.5789026021957397, 'reward_std': 0.09932930022478104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12216933816671371, 'sampling/sampling_logp_difference/max': 1.7323646545410156, 'sampling/importance_sampling_ratio/min': 0.17686569690704346, 'sampling/importance_sampling_ratio/mean': 1.0042716264724731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8733915016055107, 'clip_ratio/low_mean': 0.03689024318009615, 'clip_ratio/low_min': 0.03689024318009615, 'clip_ratio/high_mean': 0.10406699776649475, 'clip_ratio/high_max': 0.10406699776649475, 'clip_ratio/region_mean': 0.1409572409465909, 'reward_total_mean': 0.5789026021957397, 'reward_meter_mean': 0.7582693099975586, 'reward_meter_std': 0.3461999297142029, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9715173244476318, 'reward_repeat_soft_std': 0.04452996328473091, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5789026021957397, 'reward_total_composite_std': 0.09932930022478104, 'epoch': 0.07}

 22%|██▏       | 718/3300 [1:47:53<5:47:37,  8.08s/it]INFO 04-13 09:08:51 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 719/3300 [1:48:00<5:25:01,  7.56s/it]2026-04-13 09:08:57,829 | INFO | train_grpo_train | metrics_logged mode=train step=719

                                                      
{'loss': 0.0427, 'grad_norm': 18.660600662231445, 'learning_rate': 7.824242424242425e-06, 'num_tokens': 1270425.0, 'completions/mean_length': 32.875, 'completions/min_length': 28.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7567814588546753, 'rewards/meter/std': 0.3158141076564789, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9840539693832397, 'rewards/repeat_soft/std': 0.007520779967308044, 'rewards/judge_quality/mean': 0.48874998092651367, 'rewards/judge_quality/std': 0.28048110008239746, 'rewards/total_composite/mean': 0.5574514269828796, 'rewards/total_composite/std': 0.13825681805610657, 'reward': 0.5574514269828796, 'reward_std': 0.13825680315494537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15731342136859894, 'sampling/sampling_logp_difference/max': 1.4657902717590332, 'sampling/importance_sampling_ratio/min': 0.23089544475078583, 'sampling/importance_sampling_ratio/mean': 1.0008772611618042, 'sampling/importance_sampling_ratio/max': 1.9630663394927979, 'entropy': 0.9993829131126404, 'clip_ratio/low_mean': 0.06591235660016537, 'clip_ratio/low_min': 0.06591235660016537, 'clip_ratio/high_mean': 0.05210291314870119, 'clip_ratio/high_max': 0.05210291314870119, 'clip_ratio/region_mean': 0.11801526974886656, 'reward_total_mean': 0.5574514269828796, 'reward_meter_mean': 0.7567814588546753, 'reward_meter_std': 0.3158141076564789, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9840539693832397, 'reward_repeat_soft_std': 0.007520779967308044, 'reward_judge_quality_mean': 0.48874998092651367, 'reward_judge_quality_std': 0.28048110008239746, 'reward_total_composite_mean': 0.5574514269828796, 'reward_total_composite_std': 0.13825681805610657, 'epoch': 0.07}

 22%|██▏       | 719/3300 [1:48:00<5:25:01,  7.56s/it]INFO 04-13 09:08:58 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 720/3300 [1:48:05<5:04:26,  7.08s/it]2026-04-13 09:09:03,802 | INFO | train_grpo_train | metrics_logged mode=train step=720

                                                      
{'loss': 0.0734, 'grad_norm': 23.673458099365234, 'learning_rate': 7.821212121212122e-06, 'num_tokens': 1271911.0, 'completions/mean_length': 19.75, 'completions/min_length': 14.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.75, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8487032651901245, 'rewards/meter/std': 0.29530757665634155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9546234011650085, 'rewards/repeat_soft/std': 0.022278277203440666, 'rewards/judge_quality/mean': 0.3087500035762787, 'rewards/judge_quality/std': 0.1470119059085846, 'rewards/total_composite/mean': 0.5278379321098328, 'rewards/total_composite/std': 0.11549441516399384, 'reward': 0.5278379321098328, 'reward_std': 0.11549440771341324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17154358327388763, 'sampling/sampling_logp_difference/max': 1.6976003646850586, 'sampling/importance_sampling_ratio/min': 0.1831224262714386, 'sampling/importance_sampling_ratio/mean': 0.9992610216140747, 'sampling/importance_sampling_ratio/max': 1.7448310852050781, 'entropy': 1.1536924913525581, 'clip_ratio/low_mean': 0.09748476650565863, 'clip_ratio/low_min': 0.09748476650565863, 'clip_ratio/high_mean': 0.09762183204293251, 'clip_ratio/high_max': 0.09762183204293251, 'clip_ratio/region_mean': 0.19510659854859114, 'reward_total_mean': 0.5278379321098328, 'reward_meter_mean': 0.8487032651901245, 'reward_meter_std': 0.29530757665634155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9546234011650085, 'reward_repeat_soft_std': 0.022278277203440666, 'reward_judge_quality_mean': 0.3087500035762787, 'reward_judge_quality_std': 0.1470119059085846, 'reward_total_composite_mean': 0.5278379321098328, 'reward_total_composite_std': 0.11549441516399384, 'epoch': 0.07}

 22%|██▏       | 720/3300 [1:48:06<5:04:26,  7.08s/it]INFO 04-13 09:09:04 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 721/3300 [1:48:12<5:02:09,  7.03s/it]2026-04-13 09:09:10,720 | INFO | train_grpo_train | metrics_logged mode=train step=721

                                                      
{'loss': 0.0194, 'grad_norm': 7.824896812438965, 'learning_rate': 7.81818181818182e-06, 'num_tokens': 1274152.0, 'completions/mean_length': 104.125, 'completions/min_length': 87.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.125, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.4792739450931549, 'rewards/meter/std': 0.3051510751247406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6932368874549866, 'rewards/repeat_soft/std': 0.16597597301006317, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.4246557354927063, 'rewards/total_composite/std': 0.08762908726930618, 'reward': 0.4246557354927063, 'reward_std': 0.08762907981872559, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11447266489267349, 'sampling/sampling_logp_difference/max': 2.2866077423095703, 'sampling/importance_sampling_ratio/min': 0.10161057114601135, 'sampling/importance_sampling_ratio/mean': 0.9976920485496521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8533372655510902, 'clip_ratio/low_mean': 0.04566267365589738, 'clip_ratio/low_min': 0.04566267365589738, 'clip_ratio/high_mean': 0.06220975797623396, 'clip_ratio/high_max': 0.06220975797623396, 'clip_ratio/region_mean': 0.10787243163213134, 'reward_total_mean': 0.4246557354927063, 'reward_meter_mean': 0.4792739450931549, 'reward_meter_std': 0.3051510751247406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6932368874549866, 'reward_repeat_soft_std': 0.16597597301006317, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.4246557354927063, 'reward_total_composite_std': 0.08762908726930618, 'epoch': 0.07}

 22%|██▏       | 721/3300 [1:48:12<5:02:09,  7.03s/it]INFO 04-13 09:09:10 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 722/3300 [1:48:19<4:52:25,  6.81s/it]2026-04-13 09:09:16,995 | INFO | train_grpo_train | metrics_logged mode=train step=722

                                                      
{'loss': 0.0906, 'grad_norm': 15.575541496276855, 'learning_rate': 7.815151515151515e-06, 'num_tokens': 1276024.0, 'completions/mean_length': 57.0, 'completions/min_length': 47.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.655722439289093, 'rewards/meter/std': 0.35337594151496887, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9352512359619141, 'rewards/repeat_soft/std': 0.04997103288769722, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.20057062804698944, 'rewards/total_composite/mean': 0.6082093715667725, 'rewards/total_composite/std': 0.20887351036071777, 'reward': 0.6082093715667725, 'reward_std': 0.20887349545955658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1639634221792221, 'sampling/sampling_logp_difference/max': 1.4926071166992188, 'sampling/importance_sampling_ratio/min': 0.22478584945201874, 'sampling/importance_sampling_ratio/mean': 1.0217629671096802, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2073974460363388, 'clip_ratio/low_mean': 0.06932212691754103, 'clip_ratio/low_min': 0.06932212691754103, 'clip_ratio/high_mean': 0.07310424745082855, 'clip_ratio/high_max': 0.07310424745082855, 'clip_ratio/region_mean': 0.14242637436836958, 'reward_total_mean': 0.6082093715667725, 'reward_meter_mean': 0.655722439289093, 'reward_meter_std': 0.35337594151496887, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9352512359619141, 'reward_repeat_soft_std': 0.04997103288769722, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.20057062804698944, 'reward_total_composite_mean': 0.6082093715667725, 'reward_total_composite_std': 0.20887351036071777, 'epoch': 0.07}

 22%|██▏       | 722/3300 [1:48:19<4:52:25,  6.81s/it]INFO 04-13 09:09:17 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 723/3300 [1:48:30<5:52:33,  8.21s/it]2026-04-13 09:09:28,477 | INFO | train_grpo_train | metrics_logged mode=train step=723

                                                      
{'loss': -0.0893, 'grad_norm': 3.653560161590576, 'learning_rate': 7.812121212121213e-06, 'num_tokens': 1277560.0, 'completions/mean_length': 104.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.71428680419922, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8652969598770142, 'rewards/meter/std': 0.2713530361652374, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9274253845214844, 'rewards/repeat_soft/std': 0.04185902699828148, 'rewards/judge_quality/mean': 0.518750011920929, 'rewards/judge_quality/std': 0.2695730924606323, 'rewards/total_composite/mean': 0.5890629291534424, 'rewards/total_composite/std': 0.2866470217704773, 'reward': 0.5890629291534424, 'reward_std': 0.2866469919681549, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15090833604335785, 'sampling/sampling_logp_difference/max': 1.5242571830749512, 'sampling/importance_sampling_ratio/min': 0.21778278052806854, 'sampling/importance_sampling_ratio/mean': 1.0158202648162842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8286675438284874, 'clip_ratio/low_mean': 0.03413120657205582, 'clip_ratio/low_min': 0.03413120657205582, 'clip_ratio/high_mean': 0.08307747496291995, 'clip_ratio/high_max': 0.08307747496291995, 'clip_ratio/region_mean': 0.11720868153497577, 'reward_total_mean': 0.5890629291534424, 'reward_meter_mean': 0.8652969598770142, 'reward_meter_std': 0.2713530361652374, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9274253845214844, 'reward_repeat_soft_std': 0.04185902699828148, 'reward_judge_quality_mean': 0.518750011920929, 'reward_judge_quality_std': 0.2695730924606323, 'reward_total_composite_mean': 0.5890629291534424, 'reward_total_composite_std': 0.2866470217704773, 'epoch': 0.07}

 22%|██▏       | 723/3300 [1:48:30<5:52:33,  8.21s/it]INFO 04-13 09:09:28 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 724/3300 [1:48:36<5:19:19,  7.44s/it]2026-04-13 09:09:34,116 | INFO | train_grpo_train | metrics_logged mode=train step=724

                                                      
{'loss': -0.0251, 'grad_norm': 24.232315063476562, 'learning_rate': 7.80909090909091e-06, 'num_tokens': 1279079.0, 'completions/mean_length': 18.875, 'completions/min_length': 15.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7440844774246216, 'rewards/meter/std': 0.4326940178871155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9621384143829346, 'rewards/repeat_soft/std': 0.0010226722806692123, 'rewards/judge_quality/mean': 0.2587500214576721, 'rewards/judge_quality/std': 0.15037453174591064, 'rewards/total_composite/mean': 0.4419988989830017, 'rewards/total_composite/std': 0.08182511478662491, 'reward': 0.4419988989830017, 'reward_std': 0.08182510733604431, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13539238274097443, 'sampling/sampling_logp_difference/max': 1.2402918338775635, 'sampling/importance_sampling_ratio/min': 0.28929978609085083, 'sampling/importance_sampling_ratio/mean': 1.0327802896499634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8540509678423405, 'clip_ratio/low_mean': 0.07351715723052621, 'clip_ratio/low_min': 0.07351715723052621, 'clip_ratio/high_mean': 0.01923076994717121, 'clip_ratio/high_max': 0.01923076994717121, 'clip_ratio/region_mean': 0.09274792717769742, 'reward_total_mean': 0.4419988989830017, 'reward_meter_mean': 0.7440844774246216, 'reward_meter_std': 0.4326940178871155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9621384143829346, 'reward_repeat_soft_std': 0.0010226722806692123, 'reward_judge_quality_mean': 0.2587500214576721, 'reward_judge_quality_std': 0.15037453174591064, 'reward_total_composite_mean': 0.4419988989830017, 'reward_total_composite_std': 0.08182511478662491, 'epoch': 0.07}

 22%|██▏       | 724/3300 [1:48:36<5:19:19,  7.44s/it]INFO 04-13 09:09:34 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 725/3300 [1:48:42<5:05:07,  7.11s/it]2026-04-13 09:09:40,460 | INFO | train_grpo_train | metrics_logged mode=train step=725

                                                      
{'loss': 0.0342, 'grad_norm': 20.793373107910156, 'learning_rate': 7.806060606060607e-06, 'num_tokens': 1281054.0, 'completions/mean_length': 64.875, 'completions/min_length': 52.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.5379571914672852, 'rewards/meter/std': 0.27355244755744934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9909393787384033, 'rewards/repeat_soft/std': 0.010423214174807072, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.5543895959854126, 'rewards/total_composite/std': 0.08682841807603836, 'reward': 0.5543895959854126, 'reward_std': 0.08682840317487717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1904158592224121, 'sampling/sampling_logp_difference/max': 2.63891339302063, 'sampling/importance_sampling_ratio/min': 0.07143885642290115, 'sampling/importance_sampling_ratio/mean': 1.0042673349380493, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7586786299943924, 'clip_ratio/low_mean': 0.09301826171576977, 'clip_ratio/low_min': 0.09301826171576977, 'clip_ratio/high_mean': 0.07457197550684214, 'clip_ratio/high_max': 0.07457197550684214, 'clip_ratio/region_mean': 0.1675902372226119, 'reward_total_mean': 0.5543895959854126, 'reward_meter_mean': 0.5379571914672852, 'reward_meter_std': 0.27355244755744934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9909393787384033, 'reward_repeat_soft_std': 0.010423214174807072, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.5543895959854126, 'reward_total_composite_std': 0.08682841807603836, 'epoch': 0.07}

 22%|██▏       | 725/3300 [1:48:42<5:05:07,  7.11s/it]INFO 04-13 09:09:40 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 726/3300 [1:48:50<5:12:57,  7.29s/it]2026-04-13 09:09:48,188 | INFO | train_grpo_train | metrics_logged mode=train step=726

                                                      
{'loss': 0.0355, 'grad_norm': 8.718574523925781, 'learning_rate': 7.803030303030303e-06, 'num_tokens': 1283328.0, 'completions/mean_length': 113.25, 'completions/min_length': 97.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.25, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.5364246368408203, 'rewards/meter/std': 0.3864142596721649, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6865354776382446, 'rewards/repeat_soft/std': 0.20342059433460236, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.4773806929588318, 'rewards/total_composite/std': 0.11725659668445587, 'reward': 0.4773806929588318, 'reward_std': 0.11725659668445587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13357757031917572, 'sampling/sampling_logp_difference/max': 2.550213575363159, 'sampling/importance_sampling_ratio/min': 0.07806499302387238, 'sampling/importance_sampling_ratio/mean': 1.0096253156661987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8650649413466454, 'clip_ratio/low_mean': 0.06561764236539602, 'clip_ratio/low_min': 0.06561764236539602, 'clip_ratio/high_mean': 0.059514397755265236, 'clip_ratio/high_max': 0.059514397755265236, 'clip_ratio/region_mean': 0.12513204012066126, 'reward_total_mean': 0.4773806929588318, 'reward_meter_mean': 0.5364246368408203, 'reward_meter_std': 0.3864142596721649, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6865354776382446, 'reward_repeat_soft_std': 0.20342059433460236, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.4773806929588318, 'reward_total_composite_std': 0.11725659668445587, 'epoch': 0.07}

 22%|██▏       | 726/3300 [1:48:50<5:12:57,  7.29s/it]INFO 04-13 09:09:48 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 727/3300 [1:48:56<4:54:28,  6.87s/it]2026-04-13 09:09:54,057 | INFO | train_grpo_train | metrics_logged mode=train step=727

                                                      
{'loss': -0.0132, 'grad_norm': 15.35572624206543, 'learning_rate': 7.800000000000002e-06, 'num_tokens': 1285105.0, 'completions/mean_length': 42.125, 'completions/min_length': 36.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.835135817527771, 'rewards/meter/std': 0.3140829801559448, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9693187475204468, 'rewards/repeat_soft/std': 0.043177492916584015, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.8076326847076416, 'rewards/total_composite/std': 0.2045147866010666, 'reward': 0.8076326847076416, 'reward_std': 0.2045147866010666, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18121851980686188, 'sampling/sampling_logp_difference/max': 3.835233211517334, 'sampling/importance_sampling_ratio/min': 0.021596301347017288, 'sampling/importance_sampling_ratio/mean': 1.021759271621704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3519465550780296, 'clip_ratio/low_mean': 0.04869186133146286, 'clip_ratio/low_min': 0.04869186133146286, 'clip_ratio/high_mean': 0.13332146871834993, 'clip_ratio/high_max': 0.13332146871834993, 'clip_ratio/region_mean': 0.1820133300498128, 'reward_total_mean': 0.8076326847076416, 'reward_meter_mean': 0.835135817527771, 'reward_meter_std': 0.3140829801559448, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9693187475204468, 'reward_repeat_soft_std': 0.043177492916584015, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.8076326847076416, 'reward_total_composite_std': 0.2045147866010666, 'epoch': 0.07}

 22%|██▏       | 727/3300 [1:48:56<4:54:28,  6.87s/it]INFO 04-13 09:09:54 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 728/3300 [1:49:02<4:43:38,  6.62s/it]2026-04-13 09:10:00,090 | INFO | train_grpo_train | metrics_logged mode=train step=728

                                                      
{'loss': 0.0498, 'grad_norm': 15.05860710144043, 'learning_rate': 7.796969696969697e-06, 'num_tokens': 1286799.0, 'completions/mean_length': 49.75, 'completions/min_length': 44.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.75, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.4592335820198059, 'rewards/meter/std': 0.3384995758533478, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9714906215667725, 'rewards/repeat_soft/std': 0.022578958421945572, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.1968638300895691, 'rewards/total_composite/mean': 0.5114455819129944, 'rewards/total_composite/std': 0.1362476795911789, 'reward': 0.5114455819129944, 'reward_std': 0.1362476795911789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1806032508611679, 'sampling/sampling_logp_difference/max': 2.0891857147216797, 'sampling/importance_sampling_ratio/min': 0.2237635999917984, 'sampling/importance_sampling_ratio/mean': 1.0257023572921753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0947838053107262, 'clip_ratio/low_mean': 0.09659948013722897, 'clip_ratio/low_min': 0.09659948013722897, 'clip_ratio/high_mean': 0.073857381939888, 'clip_ratio/high_max': 0.073857381939888, 'clip_ratio/region_mean': 0.17045686207711697, 'reward_total_mean': 0.5114455819129944, 'reward_meter_mean': 0.4592335820198059, 'reward_meter_std': 0.3384995758533478, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9714906215667725, 'reward_repeat_soft_std': 0.022578958421945572, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.1968638300895691, 'reward_total_composite_mean': 0.5114455819129944, 'reward_total_composite_std': 0.1362476795911789, 'epoch': 0.07}

 22%|██▏       | 728/3300 [1:49:02<4:43:38,  6.62s/it]INFO 04-13 09:10:00 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 729/3300 [1:49:08<4:38:36,  6.50s/it]2026-04-13 09:10:06,324 | INFO | train_grpo_train | metrics_logged mode=train step=729

                                                      
{'loss': 0.0938, 'grad_norm': 13.141268730163574, 'learning_rate': 7.793939393939394e-06, 'num_tokens': 1288418.0, 'completions/mean_length': 58.375, 'completions/min_length': 46.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6245149374008179, 'rewards/meter/std': 0.3908325433731079, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8847557306289673, 'rewards/repeat_soft/std': 0.12024150043725967, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.1970088928937912, 'rewards/total_composite/mean': 0.4973753094673157, 'rewards/total_composite/std': 0.25831887125968933, 'reward': 0.4973753094673157, 'reward_std': 0.2583189010620117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1637883484363556, 'sampling/sampling_logp_difference/max': 1.5596323013305664, 'sampling/importance_sampling_ratio/min': 0.21021334826946259, 'sampling/importance_sampling_ratio/mean': 1.0219950675964355, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8918347582221031, 'clip_ratio/low_mean': 0.07465672492980957, 'clip_ratio/low_min': 0.07465672492980957, 'clip_ratio/high_mean': 0.07172060199081898, 'clip_ratio/high_max': 0.07172060199081898, 'clip_ratio/region_mean': 0.14637732692062855, 'reward_total_mean': 0.4973753094673157, 'reward_meter_mean': 0.6245149374008179, 'reward_meter_std': 0.3908325433731079, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8847557306289673, 'reward_repeat_soft_std': 0.12024150043725967, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.1970088928937912, 'reward_total_composite_mean': 0.4973753094673157, 'reward_total_composite_std': 0.25831887125968933, 'epoch': 0.07}

 22%|██▏       | 729/3300 [1:49:08<4:38:36,  6.50s/it]INFO 04-13 09:10:06 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 730/3300 [1:49:15<4:43:09,  6.61s/it]2026-04-13 09:10:13,189 | INFO | train_grpo_train | metrics_logged mode=train step=730

                                                      
{'loss': -0.039, 'grad_norm': 14.46064567565918, 'learning_rate': 7.790909090909092e-06, 'num_tokens': 1290611.0, 'completions/mean_length': 75.125, 'completions/min_length': 59.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.568236231803894, 'rewards/meter/std': 0.300004780292511, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9620645642280579, 'rewards/repeat_soft/std': 0.029723232612013817, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.2030482292175293, 'rewards/total_composite/mean': 0.5131289958953857, 'rewards/total_composite/std': 0.10526996105909348, 'reward': 0.5131289958953857, 'reward_std': 0.10526996105909348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18893152475357056, 'sampling/sampling_logp_difference/max': 2.1716699600219727, 'sampling/importance_sampling_ratio/min': 0.14820317924022675, 'sampling/importance_sampling_ratio/mean': 1.0112147331237793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1518063470721245, 'clip_ratio/low_mean': 0.1104865912348032, 'clip_ratio/low_min': 0.1104865912348032, 'clip_ratio/high_mean': 0.04992319457232952, 'clip_ratio/high_max': 0.04992319457232952, 'clip_ratio/region_mean': 0.16040978580713272, 'reward_total_mean': 0.5131289958953857, 'reward_meter_mean': 0.568236231803894, 'reward_meter_std': 0.300004780292511, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9620645642280579, 'reward_repeat_soft_std': 0.029723232612013817, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.2030482292175293, 'reward_total_composite_mean': 0.5131289958953857, 'reward_total_composite_std': 0.10526996105909348, 'epoch': 0.07}

 22%|██▏       | 730/3300 [1:49:15<4:43:09,  6.61s/it]INFO 04-13 09:10:13 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 731/3300 [1:49:21<4:43:14,  6.62s/it]2026-04-13 09:10:19,820 | INFO | train_grpo_train | metrics_logged mode=train step=731

                                                      
{'loss': -0.0022, 'grad_norm': 11.444204330444336, 'learning_rate': 7.787878787878789e-06, 'num_tokens': 1292653.0, 'completions/mean_length': 86.25, 'completions/min_length': 66.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.6808890104293823, 'rewards/meter/std': 0.3985695242881775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9651210308074951, 'rewards/repeat_soft/std': 0.024766085669398308, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.1968638300895691, 'rewards/total_composite/mean': 0.6168433427810669, 'rewards/total_composite/std': 0.1998293399810791, 'reward': 0.6168433427810669, 'reward_std': 0.1998293399810791, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1904446929693222, 'sampling/sampling_logp_difference/max': 2.048414707183838, 'sampling/importance_sampling_ratio/min': 0.12893915176391602, 'sampling/importance_sampling_ratio/mean': 1.0512608289718628, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6167262345552444, 'clip_ratio/low_mean': 0.08131587691605091, 'clip_ratio/low_min': 0.08131587691605091, 'clip_ratio/high_mean': 0.1005734745413065, 'clip_ratio/high_max': 0.1005734745413065, 'clip_ratio/region_mean': 0.1818893514573574, 'reward_total_mean': 0.6168433427810669, 'reward_meter_mean': 0.6808890104293823, 'reward_meter_std': 0.3985695242881775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9651210308074951, 'reward_repeat_soft_std': 0.024766085669398308, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.1968638300895691, 'reward_total_composite_mean': 0.6168433427810669, 'reward_total_composite_std': 0.1998293399810791, 'epoch': 0.07}

 22%|██▏       | 731/3300 [1:49:22<4:43:14,  6.62s/it]INFO 04-13 09:10:20 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 732/3300 [1:49:28<4:47:00,  6.71s/it]2026-04-13 09:10:26,735 | INFO | train_grpo_train | metrics_logged mode=train step=732

                                                      
{'loss': 0.036, 'grad_norm': 9.532846450805664, 'learning_rate': 7.784848484848484e-06, 'num_tokens': 1294857.0, 'completions/mean_length': 96.5, 'completions/min_length': 88.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.6556195020675659, 'rewards/meter/std': 0.27973002195358276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.88929283618927, 'rewards/repeat_soft/std': 0.10339207947254181, 'rewards/judge_quality/mean': 0.6699999570846558, 'rewards/judge_quality/std': 0.22038927674293518, 'rewards/total_composite/mean': 0.6277689933776855, 'rewards/total_composite/std': 0.16709183156490326, 'reward': 0.6277689933776855, 'reward_std': 0.16709184646606445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15185225009918213, 'sampling/sampling_logp_difference/max': 1.9622669219970703, 'sampling/importance_sampling_ratio/min': 0.14053946733474731, 'sampling/importance_sampling_ratio/mean': 1.037898063659668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2167585268616676, 'clip_ratio/low_mean': 0.10452262032777071, 'clip_ratio/low_min': 0.10452262032777071, 'clip_ratio/high_mean': 0.04544956237077713, 'clip_ratio/high_max': 0.04544956237077713, 'clip_ratio/region_mean': 0.14997218269854784, 'reward_total_mean': 0.6277689933776855, 'reward_meter_mean': 0.6556195020675659, 'reward_meter_std': 0.27973002195358276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.88929283618927, 'reward_repeat_soft_std': 0.10339207947254181, 'reward_judge_quality_mean': 0.6699999570846558, 'reward_judge_quality_std': 0.22038927674293518, 'reward_total_composite_mean': 0.6277689933776855, 'reward_total_composite_std': 0.16709183156490326, 'epoch': 0.07}

 22%|██▏       | 732/3300 [1:49:28<4:47:00,  6.71s/it]INFO 04-13 09:10:26 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 733/3300 [1:49:40<5:43:05,  8.02s/it]2026-04-13 09:10:37,823 | INFO | train_grpo_train | metrics_logged mode=train step=733

                                                      
{'loss': -0.0943, 'grad_norm': 4.793450355529785, 'learning_rate': 7.781818181818183e-06, 'num_tokens': 1296295.0, 'completions/mean_length': 101.75, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7315326929092407, 'rewards/meter/std': 0.41042831540107727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8619771003723145, 'rewards/repeat_soft/std': 0.1800392121076584, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.254611074924469, 'rewards/total_composite/mean': 0.516670823097229, 'rewards/total_composite/std': 0.2958741784095764, 'reward': 0.516670823097229, 'reward_std': 0.29587414860725403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13379375636577606, 'sampling/sampling_logp_difference/max': 1.787715196609497, 'sampling/importance_sampling_ratio/min': 0.16734206676483154, 'sampling/importance_sampling_ratio/mean': 1.0093183517456055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7671684697270393, 'clip_ratio/low_mean': 0.023473837645724416, 'clip_ratio/low_min': 0.023473837645724416, 'clip_ratio/high_mean': 0.09144049324095249, 'clip_ratio/high_max': 0.09144049324095249, 'clip_ratio/region_mean': 0.11491433088667691, 'reward_total_mean': 0.516670823097229, 'reward_meter_mean': 0.7315326929092407, 'reward_meter_std': 0.41042831540107727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8619771003723145, 'reward_repeat_soft_std': 0.1800392121076584, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.254611074924469, 'reward_total_composite_mean': 0.516670823097229, 'reward_total_composite_std': 0.2958741784095764, 'epoch': 0.07}

 22%|██▏       | 733/3300 [1:49:40<5:43:05,  8.02s/it]INFO 04-13 09:10:38 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 734/3300 [1:49:45<5:16:56,  7.41s/it]2026-04-13 09:10:43,811 | INFO | train_grpo_train | metrics_logged mode=train step=734

                                                      
{'loss': 0.0212, 'grad_norm': 14.931814193725586, 'learning_rate': 7.778787878787879e-06, 'num_tokens': 1297938.0, 'completions/mean_length': 39.375, 'completions/min_length': 30.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7002102136611938, 'rewards/meter/std': 0.29298287630081177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9946794509887695, 'rewards/repeat_soft/std': 0.009923518635332584, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.6478596925735474, 'rewards/total_composite/std': 0.19373920559883118, 'reward': 0.6478596925735474, 'reward_std': 0.19373920559883118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15438002347946167, 'sampling/sampling_logp_difference/max': 1.7901697158813477, 'sampling/importance_sampling_ratio/min': 0.16693183779716492, 'sampling/importance_sampling_ratio/mean': 1.0184907913208008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9812079221010208, 'clip_ratio/low_mean': 0.08412531949579716, 'clip_ratio/low_min': 0.08412531949579716, 'clip_ratio/high_mean': 0.04582786979153752, 'clip_ratio/high_max': 0.04582786979153752, 'clip_ratio/region_mean': 0.12995318928733468, 'reward_total_mean': 0.6478596925735474, 'reward_meter_mean': 0.7002102136611938, 'reward_meter_std': 0.29298287630081177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9946794509887695, 'reward_repeat_soft_std': 0.009923518635332584, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.6478596925735474, 'reward_total_composite_std': 0.19373920559883118, 'epoch': 0.07}

 22%|██▏       | 734/3300 [1:49:46<5:16:56,  7.41s/it]INFO 04-13 09:10:44 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 735/3300 [1:49:51<4:54:56,  6.90s/it]2026-04-13 09:10:49,514 | INFO | train_grpo_train | metrics_logged mode=train step=735

                                                      
{'loss': -0.0292, 'grad_norm': 17.531740188598633, 'learning_rate': 7.775757575757576e-06, 'num_tokens': 1299416.0, 'completions/mean_length': 22.75, 'completions/min_length': 18.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6155325174331665, 'rewards/meter/std': 0.34199750423431396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9615821838378906, 'rewards/repeat_soft/std': 0.002596014179289341, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5228148698806763, 'rewards/total_composite/std': 0.0878102108836174, 'reward': 0.5228148698806763, 'reward_std': 0.087810218334198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19599971175193787, 'sampling/sampling_logp_difference/max': 1.1137094497680664, 'sampling/importance_sampling_ratio/min': 0.32833874225616455, 'sampling/importance_sampling_ratio/mean': 1.085088849067688, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.71340611577034, 'clip_ratio/low_mean': 0.08386752288788557, 'clip_ratio/low_min': 0.08386752288788557, 'clip_ratio/high_mean': 0.07213768130168319, 'clip_ratio/high_max': 0.07213768130168319, 'clip_ratio/region_mean': 0.15600520418956876, 'reward_total_mean': 0.5228148698806763, 'reward_meter_mean': 0.6155325174331665, 'reward_meter_std': 0.34199750423431396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9615821838378906, 'reward_repeat_soft_std': 0.002596014179289341, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5228148698806763, 'reward_total_composite_std': 0.0878102108836174, 'epoch': 0.07}

 22%|██▏       | 735/3300 [1:49:51<4:54:56,  6.90s/it]INFO 04-13 09:10:49 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 736/3300 [1:50:02<5:47:06,  8.12s/it]2026-04-13 09:11:00,496 | INFO | train_grpo_train | metrics_logged mode=train step=736

                                                      
{'loss': -0.1665, 'grad_norm': 2.189173936843872, 'learning_rate': 7.772727272727273e-06, 'num_tokens': 1301239.0, 'completions/mean_length': 191.875, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 85.16667175292969, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7461813688278198, 'rewards/meter/std': 0.4131945073604584, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8395068049430847, 'rewards/repeat_soft/std': 0.11717653274536133, 'rewards/judge_quality/mean': 0.5225000381469727, 'rewards/judge_quality/std': 0.3446219563484192, 'rewards/total_composite/mean': 0.5662407875061035, 'rewards/total_composite/std': 0.3655031621456146, 'reward': 0.5662407875061035, 'reward_std': 0.36550313234329224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1478935331106186, 'sampling/sampling_logp_difference/max': 1.9637460708618164, 'sampling/importance_sampling_ratio/min': 0.14033174514770508, 'sampling/importance_sampling_ratio/mean': 1.0013198852539062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7034620717167854, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11650931276381016, 'clip_ratio/high_max': 0.11650931276381016, 'clip_ratio/region_mean': 0.11650931276381016, 'reward_total_mean': 0.5662407875061035, 'reward_meter_mean': 0.7461813688278198, 'reward_meter_std': 0.4131945073604584, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8395068049430847, 'reward_repeat_soft_std': 0.11717653274536133, 'reward_judge_quality_mean': 0.5225000381469727, 'reward_judge_quality_std': 0.3446219563484192, 'reward_total_composite_mean': 0.5662407875061035, 'reward_total_composite_std': 0.3655031621456146, 'epoch': 0.07}

 22%|██▏       | 736/3300 [1:50:02<5:47:06,  8.12s/it]INFO 04-13 09:11:00 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 737/3300 [1:50:08<5:19:16,  7.47s/it]2026-04-13 09:11:06,455 | INFO | train_grpo_train | metrics_logged mode=train step=737

                                                      
{'loss': 0.1024, 'grad_norm': 21.48760414123535, 'learning_rate': 7.76969696969697e-06, 'num_tokens': 1302578.0, 'completions/mean_length': 25.375, 'completions/min_length': 21.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.31592369079589844, 'rewards/meter/std': 0.3158826231956482, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9594065546989441, 'rewards/repeat_soft/std': 0.008749538101255894, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/total_composite/mean': 0.45355793833732605, 'rewards/total_composite/std': 0.1119605228304863, 'reward': 0.45355793833732605, 'reward_std': 0.1119605302810669, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.21762022376060486, 'sampling/sampling_logp_difference/max': 1.8392524719238281, 'sampling/importance_sampling_ratio/min': 0.15893620252609253, 'sampling/importance_sampling_ratio/mean': 1.0408340692520142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0567884147167206, 'clip_ratio/low_mean': 0.08554933313280344, 'clip_ratio/low_min': 0.08554933313280344, 'clip_ratio/high_mean': 0.07183441706001759, 'clip_ratio/high_max': 0.07183441706001759, 'clip_ratio/region_mean': 0.15738375019282103, 'reward_total_mean': 0.45355793833732605, 'reward_meter_mean': 0.31592369079589844, 'reward_meter_std': 0.3158826231956482, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9594065546989441, 'reward_repeat_soft_std': 0.008749538101255894, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.45355793833732605, 'reward_total_composite_std': 0.1119605228304863, 'epoch': 0.07}

 22%|██▏       | 737/3300 [1:50:08<5:19:16,  7.47s/it]INFO 04-13 09:11:06 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 738/3300 [1:50:14<5:00:18,  7.03s/it]2026-04-13 09:11:12,459 | INFO | train_grpo_train | metrics_logged mode=train step=738

                                                      
{'loss': 0.0724, 'grad_norm': 13.87582015991211, 'learning_rate': 7.766666666666666e-06, 'num_tokens': 1304338.0, 'completions/mean_length': 42.0, 'completions/min_length': 29.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6752654314041138, 'rewards/meter/std': 0.32563450932502747, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9729601144790649, 'rewards/repeat_soft/std': 0.03770548850297928, 'rewards/judge_quality/mean': 0.8612500429153442, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/total_composite/mean': 0.7150299549102783, 'rewards/total_composite/std': 0.18782474100589752, 'reward': 0.7150299549102783, 'reward_std': 0.18782474100589752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16547410190105438, 'sampling/sampling_logp_difference/max': 1.6455574035644531, 'sampling/importance_sampling_ratio/min': 0.1929050087928772, 'sampling/importance_sampling_ratio/mean': 0.9987207055091858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8475840240716934, 'clip_ratio/low_mean': 0.0568147050216794, 'clip_ratio/low_min': 0.0568147050216794, 'clip_ratio/high_mean': 0.08551109861582518, 'clip_ratio/high_max': 0.08551109861582518, 'clip_ratio/region_mean': 0.14232580363750458, 'reward_total_mean': 0.7150299549102783, 'reward_meter_mean': 0.6752654314041138, 'reward_meter_std': 0.32563450932502747, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9729601144790649, 'reward_repeat_soft_std': 0.03770548850297928, 'reward_judge_quality_mean': 0.8612500429153442, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.7150299549102783, 'reward_total_composite_std': 0.18782474100589752, 'epoch': 0.07}

 22%|██▏       | 738/3300 [1:50:14<5:00:18,  7.03s/it]INFO 04-13 09:11:12 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 739/3300 [1:50:21<4:54:16,  6.89s/it]2026-04-13 09:11:19,037 | INFO | train_grpo_train | metrics_logged mode=train step=739

                                                      
{'loss': 0.0354, 'grad_norm': 11.87075138092041, 'learning_rate': 7.763636363636364e-06, 'num_tokens': 1306516.0, 'completions/mean_length': 95.25, 'completions/min_length': 72.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.25, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.6119871139526367, 'rewards/meter/std': 0.3081716001033783, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9510897397994995, 'rewards/repeat_soft/std': 0.040236297994852066, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.20469054579734802, 'rewards/total_composite/mean': 0.5248219966888428, 'rewards/total_composite/std': 0.1047632247209549, 'reward': 0.5248219966888428, 'reward_std': 0.10476323962211609, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16723228991031647, 'sampling/sampling_logp_difference/max': 1.7635916471481323, 'sampling/importance_sampling_ratio/min': 0.17142803966999054, 'sampling/importance_sampling_ratio/mean': 1.0197486877441406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0619538724422455, 'clip_ratio/low_mean': 0.11025430075824261, 'clip_ratio/low_min': 0.11025430075824261, 'clip_ratio/high_mean': 0.06122303195297718, 'clip_ratio/high_max': 0.06122303195297718, 'clip_ratio/region_mean': 0.1714773327112198, 'reward_total_mean': 0.5248219966888428, 'reward_meter_mean': 0.6119871139526367, 'reward_meter_std': 0.3081716001033783, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9510897397994995, 'reward_repeat_soft_std': 0.040236297994852066, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.20469054579734802, 'reward_total_composite_mean': 0.5248219966888428, 'reward_total_composite_std': 0.1047632247209549, 'epoch': 0.07}

 22%|██▏       | 739/3300 [1:50:21<4:54:16,  6.89s/it]INFO 04-13 09:11:19 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 740/3300 [1:50:32<5:47:09,  8.14s/it]2026-04-13 09:11:30,059 | INFO | train_grpo_train | metrics_logged mode=train step=740

                                                      
{'loss': -0.0624, 'grad_norm': 4.005544185638428, 'learning_rate': 7.76060606060606e-06, 'num_tokens': 1307801.0, 'completions/mean_length': 80.625, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.39341825246810913, 'rewards/meter/std': 0.41285014152526855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.627500057220459, 'rewards/judge_quality/std': 0.3366537392139435, 'rewards/total_composite/mean': 0.4925876557826996, 'rewards/total_composite/std': 0.3017560839653015, 'reward': 0.4925876557826996, 'reward_std': 0.3017560541629791, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17517143487930298, 'sampling/sampling_logp_difference/max': 1.2181692123413086, 'sampling/importance_sampling_ratio/min': 0.29577115178108215, 'sampling/importance_sampling_ratio/mean': 1.0153645277023315, 'sampling/importance_sampling_ratio/max': 1.9104666709899902, 'entropy': 1.3799055814743042, 'clip_ratio/low_mean': 0.07118055690079927, 'clip_ratio/low_min': 0.07118055690079927, 'clip_ratio/high_mean': 0.06757382769137621, 'clip_ratio/high_max': 0.06757382769137621, 'clip_ratio/region_mean': 0.13875438459217548, 'reward_total_mean': 0.4925876557826996, 'reward_meter_mean': 0.39341825246810913, 'reward_meter_std': 0.41285014152526855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.627500057220459, 'reward_judge_quality_std': 0.3366537392139435, 'reward_total_composite_mean': 0.4925876557826996, 'reward_total_composite_std': 0.3017560839653015, 'epoch': 0.07}

 22%|██▏       | 740/3300 [1:50:32<5:47:09,  8.14s/it]INFO 04-13 09:11:30 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 741/3300 [1:50:43<6:25:33,  9.04s/it]2026-04-13 09:11:41,213 | INFO | train_grpo_train | metrics_logged mode=train step=741

                                                      
{'loss': -0.1253, 'grad_norm': 3.169534683227539, 'learning_rate': 7.757575757575758e-06, 'num_tokens': 1309498.0, 'completions/mean_length': 107.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.28571701049805, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9664974808692932, 'rewards/meter/std': 0.01657053641974926, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9849109053611755, 'rewards/repeat_soft/std': 0.03397379815578461, 'rewards/judge_quality/mean': 0.48124998807907104, 'rewards/judge_quality/std': 0.2582046389579773, 'rewards/total_composite/mean': 0.6017608642578125, 'rewards/total_composite/std': 0.27139541506767273, 'reward': 0.6017608642578125, 'reward_std': 0.27139541506767273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1672983467578888, 'sampling/sampling_logp_difference/max': 1.6825370788574219, 'sampling/importance_sampling_ratio/min': 0.1859017312526703, 'sampling/importance_sampling_ratio/mean': 1.0501664876937866, 'sampling/importance_sampling_ratio/max': 1.9643293619155884, 'entropy': 1.2438460141420364, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.159818759188056, 'clip_ratio/high_max': 0.159818759188056, 'clip_ratio/region_mean': 0.159818759188056, 'reward_total_mean': 0.6017608642578125, 'reward_meter_mean': 0.9664974808692932, 'reward_meter_std': 0.01657053641974926, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9849109053611755, 'reward_repeat_soft_std': 0.03397379815578461, 'reward_judge_quality_mean': 0.48124998807907104, 'reward_judge_quality_std': 0.2582046389579773, 'reward_total_composite_mean': 0.6017608642578125, 'reward_total_composite_std': 0.27139541506767273, 'epoch': 0.07}

 22%|██▏       | 741/3300 [1:50:43<6:25:33,  9.04s/it]INFO 04-13 09:11:41 [block_pool.py:378] Successfully reset prefix cache

 22%|██▏       | 742/3300 [1:50:54<6:54:27,  9.72s/it]2026-04-13 09:11:52,525 | INFO | train_grpo_train | metrics_logged mode=train step=742

                                                      
{'loss': -0.0849, 'grad_norm': 3.7173469066619873, 'learning_rate': 7.754545454545455e-06, 'num_tokens': 1311239.0, 'completions/mean_length': 105.625, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.57143020629883, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.46908485889434814, 'rewards/meter/std': 0.44270479679107666, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9901416301727295, 'rewards/repeat_soft/std': 0.013197098858654499, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.2563444972038269, 'rewards/total_composite/mean': 0.4840984344482422, 'rewards/total_composite/std': 0.26524755358695984, 'reward': 0.4840984344482422, 'reward_std': 0.26524755358695984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20756895840168, 'sampling/sampling_logp_difference/max': 1.9614200592041016, 'sampling/importance_sampling_ratio/min': 0.14065852761268616, 'sampling/importance_sampling_ratio/mean': 1.032393217086792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4433472752571106, 'clip_ratio/low_mean': 0.06576655060052872, 'clip_ratio/low_min': 0.06576655060052872, 'clip_ratio/high_mean': 0.09691833518445492, 'clip_ratio/high_max': 0.09691833518445492, 'clip_ratio/region_mean': 0.16268488578498363, 'reward_total_mean': 0.4840984344482422, 'reward_meter_mean': 0.46908485889434814, 'reward_meter_std': 0.44270479679107666, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9901416301727295, 'reward_repeat_soft_std': 0.013197098858654499, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.2563444972038269, 'reward_total_composite_mean': 0.4840984344482422, 'reward_total_composite_std': 0.26524755358695984, 'epoch': 0.07}

 22%|██▏       | 742/3300 [1:50:54<6:54:27,  9.72s/it]INFO 04-13 09:11:52 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 743/3300 [1:51:01<6:12:48,  8.75s/it]2026-04-13 09:11:58,997 | INFO | train_grpo_train | metrics_logged mode=train step=743

                                                      
{'loss': -0.0189, 'grad_norm': 12.860284805297852, 'learning_rate': 7.751515151515153e-06, 'num_tokens': 1313094.0, 'completions/mean_length': 58.875, 'completions/min_length': 40.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.8150298595428467, 'rewards/meter/std': 0.29089590907096863, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9740594625473022, 'rewards/repeat_soft/std': 0.016727657988667488, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5241194367408752, 'rewards/total_composite/std': 0.0947740450501442, 'reward': 0.5241194367408752, 'reward_std': 0.09477405250072479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20211203396320343, 'sampling/sampling_logp_difference/max': 3.094081401824951, 'sampling/importance_sampling_ratio/min': 0.144920215010643, 'sampling/importance_sampling_ratio/mean': 1.024543046951294, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.351658508181572, 'clip_ratio/low_mean': 0.07901785522699356, 'clip_ratio/low_min': 0.07901785522699356, 'clip_ratio/high_mean': 0.11662971321493387, 'clip_ratio/high_max': 0.11662971321493387, 'clip_ratio/region_mean': 0.19564756844192743, 'reward_total_mean': 0.5241194367408752, 'reward_meter_mean': 0.8150298595428467, 'reward_meter_std': 0.29089590907096863, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9740594625473022, 'reward_repeat_soft_std': 0.016727657988667488, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5241194367408752, 'reward_total_composite_std': 0.0947740450501442, 'epoch': 0.07}

 23%|██▎       | 743/3300 [1:51:01<6:12:48,  8.75s/it]INFO 04-13 09:11:59 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 744/3300 [1:51:08<5:49:07,  8.20s/it]2026-04-13 09:12:05,912 | INFO | train_grpo_train | metrics_logged mode=train step=744

                                                      
{'loss': 0.0216, 'grad_norm': 9.15574836730957, 'learning_rate': 7.74848484848485e-06, 'num_tokens': 1315335.0, 'completions/mean_length': 113.125, 'completions/min_length': 94.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.125, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.7801750898361206, 'rewards/meter/std': 0.27828752994537354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9237654805183411, 'rewards/repeat_soft/std': 0.04357816278934479, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5798206329345703, 'rewards/total_composite/std': 0.2641700506210327, 'reward': 0.5798206329345703, 'reward_std': 0.2641700506210327, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1706664115190506, 'sampling/sampling_logp_difference/max': 2.086062431335449, 'sampling/importance_sampling_ratio/min': 0.12417511641979218, 'sampling/importance_sampling_ratio/mean': 1.0122138261795044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2584090828895569, 'clip_ratio/low_mean': 0.0580312255769968, 'clip_ratio/low_min': 0.0580312255769968, 'clip_ratio/high_mean': 0.1189239863306284, 'clip_ratio/high_max': 0.1189239863306284, 'clip_ratio/region_mean': 0.1769552119076252, 'reward_total_mean': 0.5798206329345703, 'reward_meter_mean': 0.7801750898361206, 'reward_meter_std': 0.27828752994537354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9237654805183411, 'reward_repeat_soft_std': 0.04357816278934479, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5798206329345703, 'reward_total_composite_std': 0.2641700506210327, 'epoch': 0.07}

 23%|██▎       | 744/3300 [1:51:08<5:49:07,  8.20s/it]INFO 04-13 09:12:06 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 745/3300 [1:51:14<5:21:58,  7.56s/it]2026-04-13 09:12:11,987 | INFO | train_grpo_train | metrics_logged mode=train step=745

                                                      
{'loss': 0.0347, 'grad_norm': 17.920703887939453, 'learning_rate': 7.745454545454545e-06, 'num_tokens': 1316895.0, 'completions/mean_length': 43.0, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9510095119476318, 'rewards/meter/std': 0.10056521743535995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.968070387840271, 'rewards/repeat_soft/std': 0.03921138867735863, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465452551841736, 'rewards/total_composite/mean': 0.8477826714515686, 'rewards/total_composite/std': 0.16634705662727356, 'reward': 0.8477826714515686, 'reward_std': 0.16634705662727356, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18730603158473969, 'sampling/sampling_logp_difference/max': 1.5259227752685547, 'sampling/importance_sampling_ratio/min': 0.21742035448551178, 'sampling/importance_sampling_ratio/mean': 1.0188100337982178, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.546481341123581, 'clip_ratio/low_mean': 0.029415836557745934, 'clip_ratio/low_min': 0.029415836557745934, 'clip_ratio/high_mean': 0.14870378375053406, 'clip_ratio/high_max': 0.14870378375053406, 'clip_ratio/region_mean': 0.17811962030828, 'reward_total_mean': 0.8477826714515686, 'reward_meter_mean': 0.9510095119476318, 'reward_meter_std': 0.10056521743535995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.968070387840271, 'reward_repeat_soft_std': 0.03921138867735863, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465452551841736, 'reward_total_composite_mean': 0.8477826714515686, 'reward_total_composite_std': 0.16634705662727356, 'epoch': 0.07}

 23%|██▎       | 745/3300 [1:51:14<5:21:58,  7.56s/it]INFO 04-13 09:12:12 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 746/3300 [1:51:20<5:02:21,  7.10s/it]2026-04-13 09:12:18,027 | INFO | train_grpo_train | metrics_logged mode=train step=746

                                                      
{'loss': -0.0258, 'grad_norm': 16.017648696899414, 'learning_rate': 7.742424242424244e-06, 'num_tokens': 1318476.0, 'completions/mean_length': 43.625, 'completions/min_length': 38.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7205237746238708, 'rewards/meter/std': 0.39658796787261963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9746197462081909, 'rewards/repeat_soft/std': 0.03359459713101387, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5865886211395264, 'rewards/total_composite/std': 0.17276552319526672, 'reward': 0.5865886211395264, 'reward_std': 0.17276552319526672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1927284151315689, 'sampling/sampling_logp_difference/max': 1.848165512084961, 'sampling/importance_sampling_ratio/min': 0.15752588212490082, 'sampling/importance_sampling_ratio/mean': 0.9987056851387024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4613386616110802, 'clip_ratio/low_mean': 0.04135736916214228, 'clip_ratio/low_min': 0.04135736916214228, 'clip_ratio/high_mean': 0.09174805972725153, 'clip_ratio/high_max': 0.09174805972725153, 'clip_ratio/region_mean': 0.1331054288893938, 'reward_total_mean': 0.5865886211395264, 'reward_meter_mean': 0.7205237746238708, 'reward_meter_std': 0.39658796787261963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9746197462081909, 'reward_repeat_soft_std': 0.03359459713101387, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5865886211395264, 'reward_total_composite_std': 0.17276552319526672, 'epoch': 0.07}

 23%|██▎       | 746/3300 [1:51:20<5:02:21,  7.10s/it]INFO 04-13 09:12:18 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 747/3300 [1:51:26<4:46:50,  6.74s/it]2026-04-13 09:12:23,922 | INFO | train_grpo_train | metrics_logged mode=train step=747

                                                      
{'loss': 0.0372, 'grad_norm': 21.14970588684082, 'learning_rate': 7.73939393939394e-06, 'num_tokens': 1319893.0, 'completions/mean_length': 23.125, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.6219501495361328, 'rewards/meter/std': 0.5007457733154297, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8457368612289429, 'rewards/repeat_soft/std': 0.09808874130249023, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.31717222929000854, 'rewards/total_composite/mean': 0.5735586881637573, 'rewards/total_composite/std': 0.3366280794143677, 'reward': 0.5735586881637573, 'reward_std': 0.3366280794143677, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19304467737674713, 'sampling/sampling_logp_difference/max': 1.6328563690185547, 'sampling/importance_sampling_ratio/min': 0.19537071883678436, 'sampling/importance_sampling_ratio/mean': 1.0243732929229736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4444321170449257, 'clip_ratio/low_mean': 0.07059548143297434, 'clip_ratio/low_min': 0.07059548143297434, 'clip_ratio/high_mean': 0.10715579986572266, 'clip_ratio/high_max': 0.10715579986572266, 'clip_ratio/region_mean': 0.177751281298697, 'reward_total_mean': 0.5735586881637573, 'reward_meter_mean': 0.6219501495361328, 'reward_meter_std': 0.5007457733154297, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8457368612289429, 'reward_repeat_soft_std': 0.09808874130249023, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.31717222929000854, 'reward_total_composite_mean': 0.5735586881637573, 'reward_total_composite_std': 0.3366280794143677, 'epoch': 0.08}

 23%|██▎       | 747/3300 [1:51:26<4:46:50,  6.74s/it]INFO 04-13 09:12:24 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 748/3300 [1:51:31<4:34:01,  6.44s/it]2026-04-13 09:12:29,666 | INFO | train_grpo_train | metrics_logged mode=train step=748

                                                      
{'loss': 0.0543, 'grad_norm': 20.271127700805664, 'learning_rate': 7.736363636363637e-06, 'num_tokens': 1321439.0, 'completions/mean_length': 37.25, 'completions/min_length': 32.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9286379814147949, 'rewards/meter/std': 0.13010568916797638, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9717549085617065, 'rewards/repeat_soft/std': 0.019136307761073112, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.655158519744873, 'rewards/total_composite/std': 0.3113307058811188, 'reward': 0.655158519744873, 'reward_std': 0.3113307058811188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17222309112548828, 'sampling/sampling_logp_difference/max': 2.0556488037109375, 'sampling/importance_sampling_ratio/min': 0.12800975143909454, 'sampling/importance_sampling_ratio/mean': 1.0102615356445312, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3393578231334686, 'clip_ratio/low_mean': 0.09474449045956135, 'clip_ratio/low_min': 0.09474449045956135, 'clip_ratio/high_mean': 0.07014034502208233, 'clip_ratio/high_max': 0.07014034502208233, 'clip_ratio/region_mean': 0.16488483548164368, 'reward_total_mean': 0.655158519744873, 'reward_meter_mean': 0.9286379814147949, 'reward_meter_std': 0.13010568916797638, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9717549085617065, 'reward_repeat_soft_std': 0.019136307761073112, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.655158519744873, 'reward_total_composite_std': 0.3113307058811188, 'epoch': 0.08}

 23%|██▎       | 748/3300 [1:51:31<4:34:01,  6.44s/it]INFO 04-13 09:12:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 749/3300 [1:51:43<5:37:49,  7.95s/it]2026-04-13 09:12:41,120 | INFO | train_grpo_train | metrics_logged mode=train step=749

                                                      
{'loss': -0.1137, 'grad_norm': 4.881791114807129, 'learning_rate': 7.733333333333334e-06, 'num_tokens': 1323294.0, 'completions/mean_length': 120.875, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.41662102937698364, 'rewards/meter/std': 0.33569326996803284, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671957492828369, 'rewards/repeat_soft/std': 0.020684966817498207, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.317265123128891, 'rewards/total_composite/mean': 0.4998413622379303, 'rewards/total_composite/std': 0.27943938970565796, 'reward': 0.4998413622379303, 'reward_std': 0.27943938970565796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20088011026382446, 'sampling/sampling_logp_difference/max': 3.438444137573242, 'sampling/importance_sampling_ratio/min': 0.03211461380124092, 'sampling/importance_sampling_ratio/mean': 0.993660569190979, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.191350631415844, 'clip_ratio/low_mean': 0.06448888964951038, 'clip_ratio/low_min': 0.06448888964951038, 'clip_ratio/high_mean': 0.08905814588069916, 'clip_ratio/high_max': 0.08905814588069916, 'clip_ratio/region_mean': 0.15354703553020954, 'reward_total_mean': 0.4998413622379303, 'reward_meter_mean': 0.41662102937698364, 'reward_meter_std': 0.33569326996803284, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671957492828369, 'reward_repeat_soft_std': 0.020684966817498207, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.317265123128891, 'reward_total_composite_mean': 0.4998413622379303, 'reward_total_composite_std': 0.27943938970565796, 'epoch': 0.08}

 23%|██▎       | 749/3300 [1:51:43<5:37:49,  7.95s/it]INFO 04-13 09:12:41 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 750/3300 [1:51:54<6:20:03,  8.94s/it]2026-04-13 09:12:52,390 | INFO | train_grpo_train | metrics_logged mode=train step=750

                                                      
{'loss': -0.0773, 'grad_norm': 4.638439178466797, 'learning_rate': 7.730303030303032e-06, 'num_tokens': 1325106.0, 'completions/mean_length': 95.5, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6533169746398926, 'rewards/meter/std': 0.4125302731990814, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.992682933807373, 'rewards/repeat_soft/std': 0.010398227721452713, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.33417007327079773, 'rewards/total_composite/mean': 0.622667133808136, 'rewards/total_composite/std': 0.32134515047073364, 'reward': 0.622667133808136, 'reward_std': 0.32134518027305603, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18471704423427582, 'sampling/sampling_logp_difference/max': 1.7912163734436035, 'sampling/importance_sampling_ratio/min': 0.16675721108913422, 'sampling/importance_sampling_ratio/mean': 1.0475273132324219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9543810561299324, 'clip_ratio/low_mean': 0.05844298284500837, 'clip_ratio/low_min': 0.05844298284500837, 'clip_ratio/high_mean': 0.07930672587826848, 'clip_ratio/high_max': 0.07930672587826848, 'clip_ratio/region_mean': 0.13774970872327685, 'reward_total_mean': 0.622667133808136, 'reward_meter_mean': 0.6533169746398926, 'reward_meter_std': 0.4125302731990814, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.992682933807373, 'reward_repeat_soft_std': 0.010398227721452713, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.33417007327079773, 'reward_total_composite_mean': 0.622667133808136, 'reward_total_composite_std': 0.32134515047073364, 'epoch': 0.08}

 23%|██▎       | 750/3300 [1:51:54<6:20:03,  8.94s/it]INFO 04-13 09:12:52 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.85s/it]

 30%|███       | 3/10 [00:13<00:34,  4.99s/it]

 40%|████      | 4/10 [00:16<00:26,  4.48s/it]

 50%|█████     | 5/10 [00:21<00:22,  4.55s/it]

 60%|██████    | 6/10 [00:25<00:17,  4.37s/it]

 70%|███████   | 7/10 [00:34<00:17,  5.83s/it]

 80%|████████  | 8/10 [00:38<00:10,  5.28s/it]

 90%|█████████ | 9/10 [00:46<00:06,  6.28s/it]

100%|██████████| 10/10 [00:55<00:00,  7.06s/it]2026-04-13 09:13:52,285 | INFO | train_grpo_train | metrics_logged mode=eval step=750

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 59.8464, 'eval_samples_per_second': 1.337, 'eval_steps_per_second': 0.167, 'eval_num_tokens': 1325106.0, 'eval_completions/mean_length': 100.5, 'eval_completions/min_length': 37.2, 'eval_completions/max_length': 281.9, 'eval_completions/clipped_ratio': 0.0625, 'eval_completions/mean_terminated_length': 72.86488265991211, 'eval_completions/min_terminated_length': 37.2, 'eval_completions/max_terminated_length': 125.0, 'eval_rewards/meter/mean': 0.6636147797107697, 'eval_rewards/meter/std': 0.37332658767700194, 'eval_rewards/count_adherence/mean': 0.9685416698455811, 'eval_rewards/count_adherence/std': 0.08160217814147472, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.2230676978826523, 'eval_rewards/repeat_soft/mean': 0.9539894819259643, 'eval_rewards/repeat_soft/std': 0.05015313681215048, 'eval_rewards/judge_quality/mean': 0.5278750061988831, 'eval_rewards/judge_quality/std': 0.22390185445547103, 'eval_rewards/total_composite/mean': 0.5406864643096924, 'eval_rewards/total_composite/std': 0.22603418827056884, 'eval_reward': 0.5406864643096924, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.09669260457158088, 'eval_sampling/sampling_logp_difference/max': 1.0985046863555907, 'eval_sampling/importance_sampling_ratio/min': 0.34071766436100004, 'eval_sampling/importance_sampling_ratio/mean': 1.030054771900177, 'eval_sampling/importance_sampling_ratio/max': 1.5093917369842529, 'eval_entropy': 1.175079160928726, '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': 0.5406864643096924, 'eval_reward_meter_mean': 0.6636147797107697, 'eval_reward_meter_std': 0.37332658767700194, 'eval_reward_count_adherence_mean': 0.9685416698455811, 'eval_reward_count_adherence_std': 0.08160217814147472, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.2230676978826523, 'eval_reward_repeat_soft_mean': 0.9539894819259643, 'eval_reward_repeat_soft_std': 0.05015313681215048, 'eval_reward_judge_quality_mean': 0.5278750061988831, 'eval_reward_judge_quality_std': 0.22390185445547103, 'eval_reward_total_composite_mean': 0.5406864643096924, 'eval_reward_total_composite_std': 0.22603418827056884, 'epoch': 0.08}

 23%|██▎       | 750/3300 [1:52:54<6:20:03,  8.94s/it]

100%|██████████| 10/10 [00:55<00:00,  7.06s/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(

 23%|██▎       | 751/3300 [1:53:08<20:11:21, 28.51s/it]2026-04-13 09:14:06,568 | INFO | train_grpo_train | metrics_logged mode=train step=751

                                                       
{'loss': -0.1672, 'grad_norm': 3.318563938140869, 'learning_rate': 7.727272727272727e-06, 'num_tokens': 1326802.0, 'completions/mean_length': 121.0, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.14286041259766, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8778524398803711, 'rewards/meter/std': 0.2851432263851166, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9820973873138428, 'rewards/repeat_soft/std': 0.038947753608226776, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.25616681575775146, 'rewards/total_composite/mean': 0.6581735014915466, 'rewards/total_composite/std': 0.2878771722316742, 'reward': 0.6581735014915466, 'reward_std': 0.2878771722316742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1561141312122345, 'sampling/sampling_logp_difference/max': 1.4459128379821777, 'sampling/importance_sampling_ratio/min': 0.23553098738193512, 'sampling/importance_sampling_ratio/mean': 1.0225884914398193, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7705533057451248, 'clip_ratio/low_mean': 0.032418000511825085, 'clip_ratio/low_min': 0.032418000511825085, 'clip_ratio/high_mean': 0.07891025021672249, 'clip_ratio/high_max': 0.07891025021672249, 'clip_ratio/region_mean': 0.11132825072854757, 'reward_total_mean': 0.6581735014915466, 'reward_meter_mean': 0.8778524398803711, 'reward_meter_std': 0.2851432263851166, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9820973873138428, 'reward_repeat_soft_std': 0.038947753608226776, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.25616681575775146, 'reward_total_composite_mean': 0.6581735014915466, 'reward_total_composite_std': 0.2878771722316742, 'epoch': 0.08}

 23%|██▎       | 751/3300 [1:53:08<20:11:21, 28.51s/it]INFO 04-13 09:14:06 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 752/3300 [1:53:19<16:30:53, 23.33s/it]2026-04-13 09:14:17,816 | INFO | train_grpo_train | metrics_logged mode=train step=752

                                                       
{'loss': -0.1935, 'grad_norm': 2.901785373687744, 'learning_rate': 7.724242424242424e-06, 'num_tokens': 1329034.0, 'completions/mean_length': 210.0, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 109.33333587646484, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.6406359076499939, 'rewards/meter/std': 0.41214004158973694, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9574401378631592, 'rewards/repeat_soft/std': 0.0687209814786911, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.3117433786392212, 'rewards/total_composite/mean': 0.5163625478744507, 'rewards/total_composite/std': 0.3412361741065979, 'reward': 0.5163625478744507, 'reward_std': 0.3412361741065979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1524454653263092, 'sampling/sampling_logp_difference/max': 1.9500782489776611, 'sampling/importance_sampling_ratio/min': 0.14226293563842773, 'sampling/importance_sampling_ratio/mean': 1.0099399089813232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0000339895486832, 'clip_ratio/low_mean': 0.012376237660646439, 'clip_ratio/low_min': 0.012376237660646439, 'clip_ratio/high_mean': 0.10809026099741459, 'clip_ratio/high_max': 0.10809026099741459, 'clip_ratio/region_mean': 0.12046649865806103, 'reward_total_mean': 0.5163625478744507, 'reward_meter_mean': 0.6406359076499939, 'reward_meter_std': 0.41214004158973694, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9574401378631592, 'reward_repeat_soft_std': 0.0687209814786911, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.3117433786392212, 'reward_total_composite_mean': 0.5163625478744507, 'reward_total_composite_std': 0.3412361741065979, 'epoch': 0.08}

 23%|██▎       | 752/3300 [1:53:20<16:30:53, 23.33s/it]INFO 04-13 09:14:18 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 753/3300 [1:53:26<12:55:11, 18.26s/it]2026-04-13 09:14:24,235 | INFO | train_grpo_train | metrics_logged mode=train step=753

                                                       
{'loss': 0.0651, 'grad_norm': 13.571142196655273, 'learning_rate': 7.721212121212122e-06, 'num_tokens': 1330752.0, 'completions/mean_length': 44.75, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7906677722930908, 'rewards/meter/std': 0.3044937551021576, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9969684481620789, 'rewards/repeat_soft/std': 0.00582959083840251, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.24628673493862152, 'rewards/total_composite/mean': 0.6356469392776489, 'rewards/total_composite/std': 0.1697525978088379, 'reward': 0.6356469392776489, 'reward_std': 0.1697525978088379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1771460473537445, 'sampling/sampling_logp_difference/max': 1.5940003395080566, 'sampling/importance_sampling_ratio/min': 0.20311148464679718, 'sampling/importance_sampling_ratio/mean': 1.0426604747772217, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4844245836138725, 'clip_ratio/low_mean': 0.10905131604522467, 'clip_ratio/low_min': 0.10905131604522467, 'clip_ratio/high_mean': 0.06427770107984543, 'clip_ratio/high_max': 0.06427770107984543, 'clip_ratio/region_mean': 0.1733290171250701, 'reward_total_mean': 0.6356469392776489, 'reward_meter_mean': 0.7906677722930908, 'reward_meter_std': 0.3044937551021576, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9969684481620789, 'reward_repeat_soft_std': 0.00582959083840251, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.24628673493862152, 'reward_total_composite_mean': 0.6356469392776489, 'reward_total_composite_std': 0.1697525978088379, 'epoch': 0.08}

 23%|██▎       | 753/3300 [1:53:26<12:55:11, 18.26s/it]INFO 04-13 09:14:24 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 754/3300 [1:53:32<10:20:05, 14.61s/it]2026-04-13 09:14:30,336 | INFO | train_grpo_train | metrics_logged mode=train step=754

                                                       
{'loss': 0.0222, 'grad_norm': 11.563011169433594, 'learning_rate': 7.718181818181819e-06, 'num_tokens': 1332278.0, 'completions/mean_length': 45.75, 'completions/min_length': 40.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.31257134675979614, 'rewards/meter/std': 0.32782265543937683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9939672946929932, 'rewards/repeat_soft/std': 0.006959215737879276, 'rewards/judge_quality/mean': 0.5350000262260437, 'rewards/judge_quality/std': 0.18431341648101807, 'rewards/total_composite/mean': 0.40538159012794495, 'rewards/total_composite/std': 0.18614186346530914, 'reward': 0.40538159012794495, 'reward_std': 0.18614186346530914, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14600770175457, 'sampling/sampling_logp_difference/max': 1.466710090637207, 'sampling/importance_sampling_ratio/min': 0.23068316280841827, 'sampling/importance_sampling_ratio/mean': 1.0287079811096191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9013760164380074, 'clip_ratio/low_mean': 0.04290668945759535, 'clip_ratio/low_min': 0.04290668945759535, 'clip_ratio/high_mean': 0.07478864770382643, 'clip_ratio/high_max': 0.07478864770382643, 'clip_ratio/region_mean': 0.11769533716142178, 'reward_total_mean': 0.40538159012794495, 'reward_meter_mean': 0.31257134675979614, 'reward_meter_std': 0.32782265543937683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9939672946929932, 'reward_repeat_soft_std': 0.006959215737879276, 'reward_judge_quality_mean': 0.5350000262260437, 'reward_judge_quality_std': 0.18431341648101807, 'reward_total_composite_mean': 0.40538159012794495, 'reward_total_composite_std': 0.18614186346530914, 'epoch': 0.08}

 23%|██▎       | 754/3300 [1:53:32<10:20:05, 14.61s/it]INFO 04-13 09:14:30 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 755/3300 [1:53:43<9:34:45, 13.55s/it] 2026-04-13 09:14:41,406 | INFO | train_grpo_train | metrics_logged mode=train step=755

                                                      
{'loss': -0.0203, 'grad_norm': 6.581203937530518, 'learning_rate': 7.715151515151516e-06, 'num_tokens': 1333897.0, 'completions/mean_length': 90.375, 'completions/min_length': 24.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 30.142858505249023, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.3430681824684143, 'rewards/meter/std': 0.3828832507133484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9945133328437805, 'rewards/repeat_soft/std': 0.004590123891830444, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.3381884694099426, 'rewards/total_composite/mean': 0.3937174677848816, 'rewards/total_composite/std': 0.28141385316848755, 'reward': 0.3937174677848816, 'reward_std': 0.28141382336616516, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17463988065719604, 'sampling/sampling_logp_difference/max': 1.883415937423706, 'sampling/importance_sampling_ratio/min': 0.15206974744796753, 'sampling/importance_sampling_ratio/mean': 0.9966902732849121, 'sampling/importance_sampling_ratio/max': 1.9669967889785767, 'entropy': 0.890459232032299, 'clip_ratio/low_mean': 0.044477523770183325, 'clip_ratio/low_min': 0.044477523770183325, 'clip_ratio/high_mean': 0.07569331582635641, 'clip_ratio/high_max': 0.07569331582635641, 'clip_ratio/region_mean': 0.12017083959653974, 'reward_total_mean': 0.3937174677848816, 'reward_meter_mean': 0.3430681824684143, 'reward_meter_std': 0.3828832507133484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9945133328437805, 'reward_repeat_soft_std': 0.004590123891830444, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.3381884694099426, 'reward_total_composite_mean': 0.3937174677848816, 'reward_total_composite_std': 0.28141385316848755, 'epoch': 0.08}

 23%|██▎       | 755/3300 [1:53:43<9:34:45, 13.55s/it]INFO 04-13 09:14:41 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 756/3300 [1:53:49<7:59:13, 11.30s/it]2026-04-13 09:14:47,463 | INFO | train_grpo_train | metrics_logged mode=train step=756

                                                      
{'loss': 0.0408, 'grad_norm': 14.017119407653809, 'learning_rate': 7.712121212121213e-06, 'num_tokens': 1335650.0, 'completions/mean_length': 52.125, 'completions/min_length': 47.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4777580499649048, 'rewards/meter/std': 0.36191046237945557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877919554710388, 'rewards/repeat_soft/std': 0.027078699320554733, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.5244680047035217, 'rewards/total_composite/std': 0.16301079094409943, 'reward': 0.5244680047035217, 'reward_std': 0.16301079094409943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.175713449716568, 'sampling/sampling_logp_difference/max': 2.6853513717651367, 'sampling/importance_sampling_ratio/min': 0.06819722801446915, 'sampling/importance_sampling_ratio/mean': 1.02373206615448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.249506413936615, 'clip_ratio/low_mean': 0.08428190555423498, 'clip_ratio/low_min': 0.08428190555423498, 'clip_ratio/high_mean': 0.05872160289436579, 'clip_ratio/high_max': 0.05872160289436579, 'clip_ratio/region_mean': 0.14300350844860077, 'reward_total_mean': 0.5244680047035217, 'reward_meter_mean': 0.4777580499649048, 'reward_meter_std': 0.36191046237945557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877919554710388, 'reward_repeat_soft_std': 0.027078699320554733, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.5244680047035217, 'reward_total_composite_std': 0.16301079094409943, 'epoch': 0.08}

 23%|██▎       | 756/3300 [1:53:49<7:59:13, 11.30s/it]INFO 04-13 09:14:47 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 757/3300 [1:53:55<6:52:47,  9.74s/it]2026-04-13 09:14:53,555 | INFO | train_grpo_train | metrics_logged mode=train step=757

                                                      
{'loss': 0.0962, 'grad_norm': 18.8282527923584, 'learning_rate': 7.709090909090909e-06, 'num_tokens': 1337288.0, 'completions/mean_length': 42.75, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.3488878607749939, 'rewards/meter/std': 0.3507755696773529, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9945725798606873, 'rewards/repeat_soft/std': 0.005440371111035347, 'rewards/judge_quality/mean': 0.5187499523162842, 'rewards/judge_quality/std': 0.1799553632736206, 'rewards/total_composite/mean': 0.45065516233444214, 'rewards/total_composite/std': 0.100885771214962, 'reward': 0.45065516233444214, 'reward_std': 0.100885771214962, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20386451482772827, 'sampling/sampling_logp_difference/max': 1.976332426071167, 'sampling/importance_sampling_ratio/min': 0.13857655227184296, 'sampling/importance_sampling_ratio/mean': 0.9969514608383179, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8613913357257843, 'clip_ratio/low_mean': 0.11712655983865261, 'clip_ratio/low_min': 0.11712655983865261, 'clip_ratio/high_mean': 0.07038371358066797, 'clip_ratio/high_max': 0.07038371358066797, 'clip_ratio/region_mean': 0.18751027341932058, 'reward_total_mean': 0.45065516233444214, 'reward_meter_mean': 0.3488878607749939, 'reward_meter_std': 0.3507755696773529, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9945725798606873, 'reward_repeat_soft_std': 0.005440371111035347, 'reward_judge_quality_mean': 0.5187499523162842, 'reward_judge_quality_std': 0.1799553632736206, 'reward_total_composite_mean': 0.45065516233444214, 'reward_total_composite_std': 0.100885771214962, 'epoch': 0.08}

 23%|██▎       | 757/3300 [1:53:55<6:52:47,  9.74s/it]INFO 04-13 09:14:53 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 758/3300 [1:54:01<6:05:32,  8.63s/it]2026-04-13 09:14:59,590 | INFO | train_grpo_train | metrics_logged mode=train step=758

                                                      
{'loss': 0.0404, 'grad_norm': 17.19386100769043, 'learning_rate': 7.706060606060606e-06, 'num_tokens': 1338817.0, 'completions/mean_length': 46.125, 'completions/min_length': 38.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.5067713260650635, 'rewards/meter/std': 0.3364039361476898, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9228519201278687, 'rewards/repeat_soft/std': 0.13501907885074615, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4796302914619446, 'rewards/total_composite/std': 0.08091279119253159, 'reward': 0.4796302914619446, 'reward_std': 0.08091279119253159, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19849666953086853, 'sampling/sampling_logp_difference/max': 1.488032341003418, 'sampling/importance_sampling_ratio/min': 0.225816547870636, 'sampling/importance_sampling_ratio/mean': 1.0271035432815552, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.723602756857872, 'clip_ratio/low_mean': 0.07782778888940811, 'clip_ratio/low_min': 0.07782778888940811, 'clip_ratio/high_mean': 0.0803694874048233, 'clip_ratio/high_max': 0.0803694874048233, 'clip_ratio/region_mean': 0.15819727629423141, 'reward_total_mean': 0.4796302914619446, 'reward_meter_mean': 0.5067713260650635, 'reward_meter_std': 0.3364039361476898, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9228519201278687, 'reward_repeat_soft_std': 0.13501907885074615, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4796302914619446, 'reward_total_composite_std': 0.08091279119253159, 'epoch': 0.08}

 23%|██▎       | 758/3300 [1:54:01<6:05:32,  8.63s/it]INFO 04-13 09:14:59 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 759/3300 [1:54:13<6:42:16,  9.50s/it]2026-04-13 09:15:11,124 | INFO | train_grpo_train | metrics_logged mode=train step=759

                                                      
{'loss': -0.184, 'grad_norm': 3.580249309539795, 'learning_rate': 7.703030303030304e-06, 'num_tokens': 1341202.0, 'completions/mean_length': 168.125, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 119.00000762939453, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.5677632093429565, 'rewards/meter/std': 0.3104856312274933, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.14880475401878357, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9539953470230103, 'rewards/repeat_soft/std': 0.027996771037578583, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.23046152293682098, 'rewards/total_composite/mean': 0.49184703826904297, 'rewards/total_composite/std': 0.22149476408958435, 'reward': 0.49184703826904297, 'reward_std': 0.22149474918842316, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17595018446445465, 'sampling/sampling_logp_difference/max': 1.7975578308105469, 'sampling/importance_sampling_ratio/min': 0.16570307314395905, 'sampling/importance_sampling_ratio/mean': 1.0275319814682007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1057077646255493, 'clip_ratio/low_mean': 0.041847825050354004, 'clip_ratio/low_min': 0.041847825050354004, 'clip_ratio/high_mean': 0.09729828126728535, 'clip_ratio/high_max': 0.09729828126728535, 'clip_ratio/region_mean': 0.13914610631763935, 'reward_total_mean': 0.49184703826904297, 'reward_meter_mean': 0.5677632093429565, 'reward_meter_std': 0.3104856312274933, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.14880475401878357, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9539953470230103, 'reward_repeat_soft_std': 0.027996771037578583, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.23046152293682098, 'reward_total_composite_mean': 0.49184703826904297, 'reward_total_composite_std': 0.22149476408958435, 'epoch': 0.08}

 23%|██▎       | 759/3300 [1:54:13<6:42:16,  9.50s/it]INFO 04-13 09:15:11 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 760/3300 [1:54:19<6:01:58,  8.55s/it]2026-04-13 09:15:17,465 | INFO | train_grpo_train | metrics_logged mode=train step=760

                                                      
{'loss': 0.0051, 'grad_norm': 14.544023513793945, 'learning_rate': 7.7e-06, 'num_tokens': 1342872.0, 'completions/mean_length': 45.75, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8717202544212341, 'rewards/meter/std': 0.3210260570049286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9277708530426025, 'rewards/repeat_soft/std': 0.043750494718551636, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.601246178150177, 'rewards/total_composite/std': 0.11626143753528595, 'reward': 0.601246178150177, 'reward_std': 0.11626144498586655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14291509985923767, 'sampling/sampling_logp_difference/max': 1.8930234909057617, 'sampling/importance_sampling_ratio/min': 0.15061573684215546, 'sampling/importance_sampling_ratio/mean': 1.024094820022583, 'sampling/importance_sampling_ratio/max': 1.982306957244873, 'entropy': 0.9761926680803299, 'clip_ratio/low_mean': 0.052287583239376545, 'clip_ratio/low_min': 0.052287583239376545, 'clip_ratio/high_mean': 0.09623707085847855, 'clip_ratio/high_max': 0.09623707085847855, 'clip_ratio/region_mean': 0.1485246540978551, 'reward_total_mean': 0.601246178150177, 'reward_meter_mean': 0.8717202544212341, 'reward_meter_std': 0.3210260570049286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9277708530426025, 'reward_repeat_soft_std': 0.043750494718551636, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.601246178150177, 'reward_total_composite_std': 0.11626143753528595, 'epoch': 0.08}

 23%|██▎       | 760/3300 [1:54:19<6:01:58,  8.55s/it]INFO 04-13 09:15:17 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 761/3300 [1:54:31<6:44:09,  9.55s/it]2026-04-13 09:15:29,346 | INFO | train_grpo_train | metrics_logged mode=train step=761

                                                      
{'loss': -0.2033, 'grad_norm': 2.647010087966919, 'learning_rate': 7.696969696969696e-06, 'num_tokens': 1345400.0, 'completions/mean_length': 188.0, 'completions/min_length': 124.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 141.71429443359375, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.8462285399436951, 'rewards/meter/std': 0.2651536762714386, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.12400396168231964, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9577072262763977, 'rewards/repeat_soft/std': 0.024801496416330338, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.5141828060150146, 'rewards/total_composite/std': 0.21469427645206451, 'reward': 0.5141828060150146, 'reward_std': 0.21469426155090332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1781354397535324, 'sampling/sampling_logp_difference/max': 1.9977011680603027, 'sampling/importance_sampling_ratio/min': 0.1356467604637146, 'sampling/importance_sampling_ratio/mean': 1.0326799154281616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3835592418909073, 'clip_ratio/low_mean': 0.018786126747727394, 'clip_ratio/low_min': 0.018786126747727394, 'clip_ratio/high_mean': 0.11646628007292747, 'clip_ratio/high_max': 0.11646628007292747, 'clip_ratio/region_mean': 0.13525240682065487, 'reward_total_mean': 0.5141828060150146, 'reward_meter_mean': 0.8462285399436951, 'reward_meter_std': 0.2651536762714386, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.12400396168231964, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9577072262763977, 'reward_repeat_soft_std': 0.024801496416330338, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.5141828060150146, 'reward_total_composite_std': 0.21469427645206451, 'epoch': 0.08}

 23%|██▎       | 761/3300 [1:54:31<6:44:09,  9.55s/it]INFO 04-13 09:15:29 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 762/3300 [1:54:39<6:17:59,  8.94s/it]2026-04-13 09:15:36,850 | INFO | train_grpo_train | metrics_logged mode=train step=762

                                                      
{'loss': 0.1127, 'grad_norm': 12.662678718566895, 'learning_rate': 7.693939393939395e-06, 'num_tokens': 1347943.0, 'completions/mean_length': 107.875, 'completions/min_length': 84.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.875, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.842235267162323, 'rewards/meter/std': 0.20652878284454346, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9619722366333008, 'rewards/repeat_soft/std': 0.032441169023513794, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.44063663482666016, 'rewards/total_composite/std': 0.2853558659553528, 'reward': 0.44063663482666016, 'reward_std': 0.28535589575767517, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17923206090927124, 'sampling/sampling_logp_difference/max': 2.0224971771240234, 'sampling/importance_sampling_ratio/min': 0.13232462108135223, 'sampling/importance_sampling_ratio/mean': 0.9922921657562256, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0965612456202507, 'clip_ratio/low_mean': 0.02926973532885313, 'clip_ratio/low_min': 0.02926973532885313, 'clip_ratio/high_mean': 0.1326370146125555, 'clip_ratio/high_max': 0.1326370146125555, 'clip_ratio/region_mean': 0.16190674994140863, 'reward_total_mean': 0.44063663482666016, 'reward_meter_mean': 0.842235267162323, 'reward_meter_std': 0.20652878284454346, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9619722366333008, 'reward_repeat_soft_std': 0.032441169023513794, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.44063663482666016, 'reward_total_composite_std': 0.2853558659553528, 'epoch': 0.08}

 23%|██▎       | 762/3300 [1:54:39<6:17:59,  8.94s/it]INFO 04-13 09:15:37 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 763/3300 [1:54:50<6:44:30,  9.57s/it]2026-04-13 09:15:47,887 | INFO | train_grpo_train | metrics_logged mode=train step=763

                                                      
{'loss': -0.1103, 'grad_norm': 3.1770293712615967, 'learning_rate': 7.690909090909091e-06, 'num_tokens': 1349696.0, 'completions/mean_length': 123.125, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 67.5714340209961, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.3883039951324463, 'rewards/meter/std': 0.3838312327861786, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9750188589096069, 'rewards/repeat_soft/std': 0.023989897221326828, 'rewards/judge_quality/mean': 0.5987499952316284, 'rewards/judge_quality/std': 0.28316769003868103, 'rewards/total_composite/mean': 0.4377855360507965, 'rewards/total_composite/std': 0.2552489638328552, 'reward': 0.4377855360507965, 'reward_std': 0.2552489638328552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13883692026138306, 'sampling/sampling_logp_difference/max': 3.3449044227600098, 'sampling/importance_sampling_ratio/min': 0.03526358678936958, 'sampling/importance_sampling_ratio/mean': 1.0103102922439575, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7316030263900757, 'clip_ratio/low_mean': 0.057411215268075466, 'clip_ratio/low_min': 0.057411215268075466, 'clip_ratio/high_mean': 0.03219697065651417, 'clip_ratio/high_max': 0.03219697065651417, 'clip_ratio/region_mean': 0.08960818592458963, 'reward_total_mean': 0.4377855360507965, 'reward_meter_mean': 0.3883039951324463, 'reward_meter_std': 0.3838312327861786, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9750188589096069, 'reward_repeat_soft_std': 0.023989897221326828, 'reward_judge_quality_mean': 0.5987499952316284, 'reward_judge_quality_std': 0.28316769003868103, 'reward_total_composite_mean': 0.4377855360507965, 'reward_total_composite_std': 0.2552489638328552, 'epoch': 0.08}

 23%|██▎       | 763/3300 [1:54:50<6:44:30,  9.57s/it]INFO 04-13 09:15:48 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 764/3300 [1:54:56<6:03:11,  8.59s/it]2026-04-13 09:15:54,202 | INFO | train_grpo_train | metrics_logged mode=train step=764

                                                      
{'loss': 0.028, 'grad_norm': 14.50843334197998, 'learning_rate': 7.687878787878788e-06, 'num_tokens': 1351209.0, 'completions/mean_length': 44.125, 'completions/min_length': 42.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9126839637756348, 'rewards/meter/std': 0.20989079773426056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9592670202255249, 'rewards/repeat_soft/std': 0.04412023723125458, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.6975982189178467, 'rewards/total_composite/std': 0.1767217367887497, 'reward': 0.6975982189178467, 'reward_std': 0.1767217367887497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15735089778900146, 'sampling/sampling_logp_difference/max': 1.154123306274414, 'sampling/importance_sampling_ratio/min': 0.31533387303352356, 'sampling/importance_sampling_ratio/mean': 1.0255959033966064, 'sampling/importance_sampling_ratio/max': 1.9996552467346191, 'entropy': 1.3363846838474274, 'clip_ratio/low_mean': 0.08857548795640469, 'clip_ratio/low_min': 0.08857548795640469, 'clip_ratio/high_mean': 0.06425290182232857, 'clip_ratio/high_max': 0.06425290182232857, 'clip_ratio/region_mean': 0.15282838977873325, 'reward_total_mean': 0.6975982189178467, 'reward_meter_mean': 0.9126839637756348, 'reward_meter_std': 0.20989079773426056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9592670202255249, 'reward_repeat_soft_std': 0.04412023723125458, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.6975982189178467, 'reward_total_composite_std': 0.1767217367887497, 'epoch': 0.08}

 23%|██▎       | 764/3300 [1:54:56<6:03:11,  8.59s/it]INFO 04-13 09:15:54 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 765/3300 [1:55:07<6:39:41,  9.46s/it]2026-04-13 09:16:05,689 | INFO | train_grpo_train | metrics_logged mode=train step=765

                                                      
{'loss': -0.072, 'grad_norm': 4.216747283935547, 'learning_rate': 7.684848484848485e-06, 'num_tokens': 1352574.0, 'completions/mean_length': 85.625, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.71428680419922, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6552296876907349, 'rewards/meter/std': 0.433675616979599, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9599957466125488, 'rewards/repeat_soft/std': 0.0070829675532877445, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.14029940962791443, 'rewards/total_composite/mean': 0.45586252212524414, 'rewards/total_composite/std': 0.2198062390089035, 'reward': 0.45586252212524414, 'reward_std': 0.2198062390089035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19094108045101166, 'sampling/sampling_logp_difference/max': 1.128258228302002, 'sampling/importance_sampling_ratio/min': 0.32359638810157776, 'sampling/importance_sampling_ratio/mean': 1.0259119272232056, 'sampling/importance_sampling_ratio/max': 1.927560567855835, 'entropy': 1.506519839167595, 'clip_ratio/low_mean': 0.06678511761128902, 'clip_ratio/low_min': 0.06678511761128902, 'clip_ratio/high_mean': 0.10454833135008812, 'clip_ratio/high_max': 0.10454833135008812, 'clip_ratio/region_mean': 0.17133344896137714, 'reward_total_mean': 0.45586252212524414, 'reward_meter_mean': 0.6552296876907349, 'reward_meter_std': 0.433675616979599, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9599957466125488, 'reward_repeat_soft_std': 0.0070829675532877445, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.14029940962791443, 'reward_total_composite_mean': 0.45586252212524414, 'reward_total_composite_std': 0.2198062390089035, 'epoch': 0.08}

 23%|██▎       | 765/3300 [1:55:07<6:39:41,  9.46s/it]INFO 04-13 09:16:05 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 766/3300 [1:55:14<5:59:11,  8.50s/it]2026-04-13 09:16:11,964 | INFO | train_grpo_train | metrics_logged mode=train step=766

                                                      
{'loss': 0.0643, 'grad_norm': 13.96207332611084, 'learning_rate': 7.681818181818183e-06, 'num_tokens': 1354414.0, 'completions/mean_length': 52.0, 'completions/min_length': 46.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9453797340393066, 'rewards/meter/std': 0.10096082091331482, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9896356463432312, 'rewards/repeat_soft/std': 0.018915308639407158, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.7106248140335083, 'rewards/total_composite/std': 0.158878356218338, 'reward': 0.7106248140335083, 'reward_std': 0.15887834131717682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15723644196987152, 'sampling/sampling_logp_difference/max': 1.598418951034546, 'sampling/importance_sampling_ratio/min': 0.2022159844636917, 'sampling/importance_sampling_ratio/mean': 1.0136470794677734, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1226835399866104, 'clip_ratio/low_mean': 0.09415237419307232, 'clip_ratio/low_min': 0.09415237419307232, 'clip_ratio/high_mean': 0.06614341773092747, 'clip_ratio/high_max': 0.06614341773092747, 'clip_ratio/region_mean': 0.1602957919239998, 'reward_total_mean': 0.7106248140335083, 'reward_meter_mean': 0.9453797340393066, 'reward_meter_std': 0.10096082091331482, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9896356463432312, 'reward_repeat_soft_std': 0.018915308639407158, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.7106248140335083, 'reward_total_composite_std': 0.158878356218338, 'epoch': 0.08}

 23%|██▎       | 766/3300 [1:55:14<5:59:11,  8.50s/it]INFO 04-13 09:16:12 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 767/3300 [1:55:20<5:33:58,  7.91s/it]2026-04-13 09:16:18,490 | INFO | train_grpo_train | metrics_logged mode=train step=767

                                                      
{'loss': 0.0326, 'grad_norm': 15.807914733886719, 'learning_rate': 7.678787878787878e-06, 'num_tokens': 1356408.0, 'completions/mean_length': 80.25, 'completions/min_length': 65.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.5978748798370361, 'rewards/meter/std': 0.2955169975757599, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9931373596191406, 'rewards/repeat_soft/std': 0.005969149060547352, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5544592142105103, 'rewards/total_composite/std': 0.12938223779201508, 'reward': 0.5544592142105103, 'reward_std': 0.12938222289085388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19241324067115784, 'sampling/sampling_logp_difference/max': 2.6200027465820312, 'sampling/importance_sampling_ratio/min': 0.07280266284942627, 'sampling/importance_sampling_ratio/mean': 0.9966608285903931, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7311543300747871, 'clip_ratio/low_mean': 0.11057438887655735, 'clip_ratio/low_min': 0.11057438887655735, 'clip_ratio/high_mean': 0.05676700547337532, 'clip_ratio/high_max': 0.05676700547337532, 'clip_ratio/region_mean': 0.16734139434993267, 'reward_total_mean': 0.5544592142105103, 'reward_meter_mean': 0.5978748798370361, 'reward_meter_std': 0.2955169975757599, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9931373596191406, 'reward_repeat_soft_std': 0.005969149060547352, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5544592142105103, 'reward_total_composite_std': 0.12938223779201508, 'epoch': 0.08}

 23%|██▎       | 767/3300 [1:55:20<5:33:58,  7.91s/it]INFO 04-13 09:16:18 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 768/3300 [1:55:26<5:09:36,  7.34s/it]2026-04-13 09:16:24,489 | INFO | train_grpo_train | metrics_logged mode=train step=768

                                                      
{'loss': 0.126, 'grad_norm': 17.3958683013916, 'learning_rate': 7.675757575757577e-06, 'num_tokens': 1357981.0, 'completions/mean_length': 27.625, 'completions/min_length': 22.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7603672742843628, 'rewards/meter/std': 0.3146916329860687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9957367777824402, 'rewards/repeat_soft/std': 0.0060963560827076435, 'rewards/judge_quality/mean': 0.7737500667572021, 'rewards/judge_quality/std': 0.27458739280700684, 'rewards/total_composite/mean': 0.728003203868866, 'rewards/total_composite/std': 0.3442491888999939, 'reward': 0.728003203868866, 'reward_std': 0.3442491888999939, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1403573751449585, 'sampling/sampling_logp_difference/max': 1.4121198654174805, 'sampling/importance_sampling_ratio/min': 0.24362626671791077, 'sampling/importance_sampling_ratio/mean': 1.0180941820144653, 'sampling/importance_sampling_ratio/max': 1.9689271450042725, 'entropy': 0.9881499111652374, 'clip_ratio/low_mean': 0.03348214365541935, 'clip_ratio/low_min': 0.03348214365541935, 'clip_ratio/high_mean': 0.10506078135222197, 'clip_ratio/high_max': 0.10506078135222197, 'clip_ratio/region_mean': 0.13854292500764132, 'reward_total_mean': 0.728003203868866, 'reward_meter_mean': 0.7603672742843628, 'reward_meter_std': 0.3146916329860687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9957367777824402, 'reward_repeat_soft_std': 0.0060963560827076435, 'reward_judge_quality_mean': 0.7737500667572021, 'reward_judge_quality_std': 0.27458739280700684, 'reward_total_composite_mean': 0.728003203868866, 'reward_total_composite_std': 0.3442491888999939, 'epoch': 0.08}

 23%|██▎       | 768/3300 [1:55:26<5:09:36,  7.34s/it]INFO 04-13 09:16:24 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 769/3300 [1:55:32<4:49:51,  6.87s/it]2026-04-13 09:16:30,275 | INFO | train_grpo_train | metrics_logged mode=train step=769

                                                      
{'loss': 0.0226, 'grad_norm': 18.644556045532227, 'learning_rate': 7.672727272727273e-06, 'num_tokens': 1359397.0, 'completions/mean_length': 25.0, 'completions/min_length': 19.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8413082957267761, 'rewards/meter/std': 0.3381594717502594, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8590782880783081, 'rewards/repeat_soft/std': 0.1899154782295227, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.21185827255249023, 'rewards/total_composite/mean': 0.5824275016784668, 'rewards/total_composite/std': 0.1867597997188568, 'reward': 0.5824275016784668, 'reward_std': 0.18675978481769562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14683429896831512, 'sampling/sampling_logp_difference/max': 2.0840606689453125, 'sampling/importance_sampling_ratio/min': 0.12442393600940704, 'sampling/importance_sampling_ratio/mean': 1.0149449110031128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7047759927809238, 'clip_ratio/low_mean': 0.04124658228829503, 'clip_ratio/low_min': 0.04124658228829503, 'clip_ratio/high_mean': 0.07379701640456915, 'clip_ratio/high_max': 0.07379701640456915, 'clip_ratio/region_mean': 0.11504359869286418, 'reward_total_mean': 0.5824275016784668, 'reward_meter_mean': 0.8413082957267761, 'reward_meter_std': 0.3381594717502594, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8590782880783081, 'reward_repeat_soft_std': 0.1899154782295227, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.21185827255249023, 'reward_total_composite_mean': 0.5824275016784668, 'reward_total_composite_std': 0.1867597997188568, 'epoch': 0.08}

 23%|██▎       | 769/3300 [1:55:32<4:49:51,  6.87s/it]INFO 04-13 09:16:30 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 770/3300 [1:55:43<5:47:26,  8.24s/it]2026-04-13 09:16:41,705 | INFO | train_grpo_train | metrics_logged mode=train step=770

                                                      
{'loss': -0.1785, 'grad_norm': 3.450512170791626, 'learning_rate': 7.66969696969697e-06, 'num_tokens': 1361206.0, 'completions/mean_length': 129.125, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.42857360839844, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9283363819122314, 'rewards/meter/std': 0.13040593266487122, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9898082613945007, 'rewards/repeat_soft/std': 0.010055318474769592, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.2517049014568329, 'rewards/total_composite/mean': 0.587196946144104, 'rewards/total_composite/std': 0.2623291611671448, 'reward': 0.587196946144104, 'reward_std': 0.2623291611671448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16380071640014648, 'sampling/sampling_logp_difference/max': 1.6352181434631348, 'sampling/importance_sampling_ratio/min': 0.2790594696998596, 'sampling/importance_sampling_ratio/mean': 1.0115994215011597, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0749797523021698, 'clip_ratio/low_mean': 0.06256871297955513, 'clip_ratio/low_min': 0.06256871297955513, 'clip_ratio/high_mean': 0.08013118803501129, 'clip_ratio/high_max': 0.08013118803501129, 'clip_ratio/region_mean': 0.14269990101456642, 'reward_total_mean': 0.587196946144104, 'reward_meter_mean': 0.9283363819122314, 'reward_meter_std': 0.13040593266487122, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9898082613945007, 'reward_repeat_soft_std': 0.010055318474769592, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.2517049014568329, 'reward_total_composite_mean': 0.587196946144104, 'reward_total_composite_std': 0.2623291611671448, 'epoch': 0.08}

 23%|██▎       | 770/3300 [1:55:43<5:47:26,  8.24s/it]INFO 04-13 09:16:41 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 771/3300 [1:55:50<5:23:10,  7.67s/it]2026-04-13 09:16:48,038 | INFO | train_grpo_train | metrics_logged mode=train step=771

                                                      
{'loss': 0.0829, 'grad_norm': 12.254379272460938, 'learning_rate': 7.666666666666667e-06, 'num_tokens': 1362865.0, 'completions/mean_length': 57.375, 'completions/min_length': 45.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9686471223831177, 'rewards/meter/std': 0.045269280672073364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9945254325866699, 'rewards/repeat_soft/std': 0.008712784387171268, 'rewards/judge_quality/mean': 0.48000001907348633, 'rewards/judge_quality/std': 0.19071295857429504, 'rewards/total_composite/mean': 0.6532793045043945, 'rewards/total_composite/std': 0.12339124828577042, 'reward': 0.6532793045043945, 'reward_std': 0.12339124828577042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1452702134847641, 'sampling/sampling_logp_difference/max': 1.4859485626220703, 'sampling/importance_sampling_ratio/min': 0.2262876033782959, 'sampling/importance_sampling_ratio/mean': 1.0236133337020874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0772370174527168, 'clip_ratio/low_mean': 0.13615667633712292, 'clip_ratio/low_min': 0.13615667633712292, 'clip_ratio/high_mean': 0.01715686358511448, 'clip_ratio/high_max': 0.01715686358511448, 'clip_ratio/region_mean': 0.1533135399222374, 'reward_total_mean': 0.6532793045043945, 'reward_meter_mean': 0.9686471223831177, 'reward_meter_std': 0.045269280672073364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9945254325866699, 'reward_repeat_soft_std': 0.008712784387171268, 'reward_judge_quality_mean': 0.48000001907348633, 'reward_judge_quality_std': 0.19071295857429504, 'reward_total_composite_mean': 0.6532793045043945, 'reward_total_composite_std': 0.12339124828577042, 'epoch': 0.08}

 23%|██▎       | 771/3300 [1:55:50<5:23:10,  7.67s/it]INFO 04-13 09:16:48 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 772/3300 [1:56:01<6:06:48,  8.71s/it]2026-04-13 09:16:59,166 | INFO | train_grpo_train | metrics_logged mode=train step=772

                                                      
{'loss': -0.0452, 'grad_norm': 4.1646294593811035, 'learning_rate': 7.663636363636364e-06, 'num_tokens': 1364220.0, 'completions/mean_length': 84.375, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.285715103149414, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.45469433069229126, 'rewards/meter/std': 0.47846660017967224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.14029940962791443, 'rewards/total_composite/mean': 0.4319379925727844, 'rewards/total_composite/std': 0.2155667245388031, 'reward': 0.4319379925727844, 'reward_std': 0.2155667245388031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22159840166568756, 'sampling/sampling_logp_difference/max': 1.7078592777252197, 'sampling/importance_sampling_ratio/min': 0.18125338852405548, 'sampling/importance_sampling_ratio/mean': 1.0563842058181763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5208516269922256, 'clip_ratio/low_mean': 0.08533653989434242, 'clip_ratio/low_min': 0.08533653989434242, 'clip_ratio/high_mean': 0.08314394112676382, 'clip_ratio/high_max': 0.08314394112676382, 'clip_ratio/region_mean': 0.16848048102110624, 'reward_total_mean': 0.4319379925727844, 'reward_meter_mean': 0.45469433069229126, 'reward_meter_std': 0.47846660017967224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.14029940962791443, 'reward_total_composite_mean': 0.4319379925727844, 'reward_total_composite_std': 0.2155667245388031, 'epoch': 0.08}

 23%|██▎       | 772/3300 [1:56:01<6:06:48,  8.71s/it]INFO 04-13 09:16:59 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 773/3300 [1:56:07<5:31:34,  7.87s/it]2026-04-13 09:17:05,097 | INFO | train_grpo_train | metrics_logged mode=train step=773

                                                      
{'loss': 0.0204, 'grad_norm': 16.329193115234375, 'learning_rate': 7.660606060606062e-06, 'num_tokens': 1365963.0, 'completions/mean_length': 46.875, 'completions/min_length': 40.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9151001572608948, 'rewards/meter/std': 0.09497155994176865, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9848058223724365, 'rewards/repeat_soft/std': 0.01939382031559944, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.7250627279281616, 'rewards/total_composite/std': 0.17324359714984894, 'reward': 0.7250627279281616, 'reward_std': 0.17324359714984894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19107262790203094, 'sampling/sampling_logp_difference/max': 2.556926727294922, 'sampling/importance_sampling_ratio/min': 0.07754268497228622, 'sampling/importance_sampling_ratio/mean': 1.00608491897583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1131526306271553, 'clip_ratio/low_mean': 0.07445781538262963, 'clip_ratio/low_min': 0.07445781538262963, 'clip_ratio/high_mean': 0.05299872066825628, 'clip_ratio/high_max': 0.05299872066825628, 'clip_ratio/region_mean': 0.12745653605088592, 'reward_total_mean': 0.7250627279281616, 'reward_meter_mean': 0.9151001572608948, 'reward_meter_std': 0.09497155994176865, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9848058223724365, 'reward_repeat_soft_std': 0.01939382031559944, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.7250627279281616, 'reward_total_composite_std': 0.17324359714984894, 'epoch': 0.08}

 23%|██▎       | 773/3300 [1:56:07<5:31:34,  7.87s/it]INFO 04-13 09:17:05 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 774/3300 [1:56:14<5:18:31,  7.57s/it]2026-04-13 09:17:11,942 | INFO | train_grpo_train | metrics_logged mode=train step=774

                                                      
{'loss': 0.0682, 'grad_norm': 11.950896263122559, 'learning_rate': 7.657575757575757e-06, 'num_tokens': 1367683.0, 'completions/mean_length': 51.0, 'completions/min_length': 38.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9761290550231934, 'rewards/meter/std': 0.009273960255086422, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9809232950210571, 'rewards/repeat_soft/std': 0.02996324747800827, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.214609295129776, 'rewards/total_composite/mean': 0.8164946436882019, 'rewards/total_composite/std': 0.1343652456998825, 'reward': 0.8164946436882019, 'reward_std': 0.1343652606010437, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1684437394142151, 'sampling/sampling_logp_difference/max': 1.9768190383911133, 'sampling/importance_sampling_ratio/min': 0.1385091245174408, 'sampling/importance_sampling_ratio/mean': 1.0028111934661865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1285023465752602, 'clip_ratio/low_mean': 0.09706372953951359, 'clip_ratio/low_min': 0.09706372953951359, 'clip_ratio/high_mean': 0.0847039483487606, 'clip_ratio/high_max': 0.0847039483487606, 'clip_ratio/region_mean': 0.1817676778882742, 'reward_total_mean': 0.8164946436882019, 'reward_meter_mean': 0.9761290550231934, 'reward_meter_std': 0.009273960255086422, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9809232950210571, 'reward_repeat_soft_std': 0.02996324747800827, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.214609295129776, 'reward_total_composite_mean': 0.8164946436882019, 'reward_total_composite_std': 0.1343652456998825, 'epoch': 0.08}

 23%|██▎       | 774/3300 [1:56:14<5:18:31,  7.57s/it]INFO 04-13 09:17:12 [block_pool.py:378] Successfully reset prefix cache

 23%|██▎       | 775/3300 [1:56:26<6:16:29,  8.95s/it]2026-04-13 09:17:24,112 | INFO | train_grpo_train | metrics_logged mode=train step=775

                                                      
{'loss': -0.077, 'grad_norm': 5.5654168128967285, 'learning_rate': 7.654545454545456e-06, 'num_tokens': 1369386.0, 'completions/mean_length': 110.875, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.57143020629883, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7354999780654907, 'rewards/meter/std': 0.42168018221855164, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9402517080307007, 'rewards/repeat_soft/std': 0.06503720581531525, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.5412236452102661, 'rewards/total_composite/std': 0.11183484643697739, 'reward': 0.5412236452102661, 'reward_std': 0.11183483898639679, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.187318354845047, 'sampling/sampling_logp_difference/max': 1.4507150650024414, 'sampling/importance_sampling_ratio/min': 0.2344026267528534, 'sampling/importance_sampling_ratio/mean': 1.0134236812591553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1164140701293945, 'clip_ratio/low_mean': 0.02500000037252903, 'clip_ratio/low_min': 0.02500000037252903, 'clip_ratio/high_mean': 0.12107518129050732, 'clip_ratio/high_max': 0.12107518129050732, 'clip_ratio/region_mean': 0.14607518166303635, 'reward_total_mean': 0.5412236452102661, 'reward_meter_mean': 0.7354999780654907, 'reward_meter_std': 0.42168018221855164, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9402517080307007, 'reward_repeat_soft_std': 0.06503720581531525, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.5412236452102661, 'reward_total_composite_std': 0.11183484643697739, 'epoch': 0.08}

 23%|██▎       | 775/3300 [1:56:26<6:16:29,  8.95s/it]INFO 04-13 09:17:24 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 776/3300 [1:56:38<6:52:20,  9.80s/it]2026-04-13 09:17:35,911 | INFO | train_grpo_train | metrics_logged mode=train step=776

                                                      
{'loss': -0.1016, 'grad_norm': 3.9890058040618896, 'learning_rate': 7.651515151515152e-06, 'num_tokens': 1370871.0, 'completions/mean_length': 105.625, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.57143020629883, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6314957141876221, 'rewards/meter/std': 0.37059882283210754, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9793222546577454, 'rewards/repeat_soft/std': 0.03290608525276184, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.25147777795791626, 'rewards/total_composite/mean': 0.5274893045425415, 'rewards/total_composite/std': 0.2628191411495209, 'reward': 0.5274893045425415, 'reward_std': 0.2628191411495209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15183840692043304, 'sampling/sampling_logp_difference/max': 2.443131446838379, 'sampling/importance_sampling_ratio/min': 0.08688833564519882, 'sampling/importance_sampling_ratio/mean': 1.0251280069351196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8372940421104431, 'clip_ratio/low_mean': 0.04993386194109917, 'clip_ratio/low_min': 0.04993386194109917, 'clip_ratio/high_mean': 0.1101582683622837, 'clip_ratio/high_max': 0.1101582683622837, 'clip_ratio/region_mean': 0.16009213030338287, 'reward_total_mean': 0.5274893045425415, 'reward_meter_mean': 0.6314957141876221, 'reward_meter_std': 0.37059882283210754, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9793222546577454, 'reward_repeat_soft_std': 0.03290608525276184, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.25147777795791626, 'reward_total_composite_mean': 0.5274893045425415, 'reward_total_composite_std': 0.2628191411495209, 'epoch': 0.08}

 24%|██▎       | 776/3300 [1:56:38<6:52:20,  9.80s/it]INFO 04-13 09:17:36 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 777/3300 [1:56:44<6:11:40,  8.84s/it]2026-04-13 09:17:42,503 | INFO | train_grpo_train | metrics_logged mode=train step=777

                                                      
{'loss': -0.0499, 'grad_norm': 19.725584030151367, 'learning_rate': 7.648484848484849e-06, 'num_tokens': 1372369.0, 'completions/mean_length': 24.25, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8873765468597412, 'rewards/meter/std': 0.20461690425872803, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9608622789382935, 'rewards/repeat_soft/std': 0.0046321069821715355, 'rewards/judge_quality/mean': 0.46000000834465027, 'rewards/judge_quality/std': 0.21138995885849, 'rewards/total_composite/mean': 0.6119965314865112, 'rewards/total_composite/std': 0.15210434794425964, 'reward': 0.6119965314865112, 'reward_std': 0.15210434794425964, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15863265097141266, 'sampling/sampling_logp_difference/max': 1.2769782543182373, 'sampling/importance_sampling_ratio/min': 0.3066890239715576, 'sampling/importance_sampling_ratio/mean': 1.0200525522232056, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.434073582291603, 'clip_ratio/low_mean': 0.048363096080720425, 'clip_ratio/low_min': 0.048363096080720425, 'clip_ratio/high_mean': 0.10358843859285116, 'clip_ratio/high_max': 0.10358843859285116, 'clip_ratio/region_mean': 0.1519515346735716, 'reward_total_mean': 0.6119965314865112, 'reward_meter_mean': 0.8873765468597412, 'reward_meter_std': 0.20461690425872803, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9608622789382935, 'reward_repeat_soft_std': 0.0046321069821715355, 'reward_judge_quality_mean': 0.46000000834465027, 'reward_judge_quality_std': 0.21138995885849, 'reward_total_composite_mean': 0.6119965314865112, 'reward_total_composite_std': 0.15210434794425964, 'epoch': 0.08}

 24%|██▎       | 777/3300 [1:56:44<6:11:40,  8.84s/it]INFO 04-13 09:17:42 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 778/3300 [1:56:51<5:43:07,  8.16s/it]2026-04-13 09:17:49,087 | INFO | train_grpo_train | metrics_logged mode=train step=778

                                                      
{'loss': -0.0521, 'grad_norm': 15.71272087097168, 'learning_rate': 7.645454545454546e-06, 'num_tokens': 1374322.0, 'completions/mean_length': 64.125, 'completions/min_length': 42.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.5612162351608276, 'rewards/meter/std': 0.3150908052921295, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9959571957588196, 'rewards/repeat_soft/std': 0.007008410524576902, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.513142466545105, 'rewards/total_composite/std': 0.11905184388160706, 'reward': 0.513142466545105, 'reward_std': 0.11905186623334885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20146779716014862, 'sampling/sampling_logp_difference/max': 1.9037284851074219, 'sampling/importance_sampling_ratio/min': 0.1490119993686676, 'sampling/importance_sampling_ratio/mean': 1.0142766237258911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6660637855529785, 'clip_ratio/low_mean': 0.08029035665094852, 'clip_ratio/low_min': 0.08029035665094852, 'clip_ratio/high_mean': 0.09320173598825932, 'clip_ratio/high_max': 0.09320173598825932, 'clip_ratio/region_mean': 0.17349209263920784, 'reward_total_mean': 0.513142466545105, 'reward_meter_mean': 0.5612162351608276, 'reward_meter_std': 0.3150908052921295, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9959571957588196, 'reward_repeat_soft_std': 0.007008410524576902, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.513142466545105, 'reward_total_composite_std': 0.11905184388160706, 'epoch': 0.08}

 24%|██▎       | 778/3300 [1:56:51<5:43:07,  8.16s/it]INFO 04-13 09:17:49 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 779/3300 [1:56:57<5:15:49,  7.52s/it]2026-04-13 09:17:55,096 | INFO | train_grpo_train | metrics_logged mode=train step=779

                                                      
{'loss': 0.0016, 'grad_norm': 13.117396354675293, 'learning_rate': 7.642424242424244e-06, 'num_tokens': 1376027.0, 'completions/mean_length': 45.125, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.3465504050254822, 'rewards/meter/std': 0.26037847995758057, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9035850763320923, 'rewards/repeat_soft/std': 0.11488734185695648, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4339674711227417, 'rewards/total_composite/std': 0.08193355053663254, 'reward': 0.4339674711227417, 'reward_std': 0.08193355798721313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15073300898075104, 'sampling/sampling_logp_difference/max': 1.0328598022460938, 'sampling/importance_sampling_ratio/min': 0.35598745942115784, 'sampling/importance_sampling_ratio/mean': 0.9965004920959473, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9764309599995613, 'clip_ratio/low_mean': 0.08635712508112192, 'clip_ratio/low_min': 0.08635712508112192, 'clip_ratio/high_mean': 0.04932301864027977, 'clip_ratio/high_max': 0.04932301864027977, 'clip_ratio/region_mean': 0.1356801437214017, 'reward_total_mean': 0.4339674711227417, 'reward_meter_mean': 0.3465504050254822, 'reward_meter_std': 0.26037847995758057, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9035850763320923, 'reward_repeat_soft_std': 0.11488734185695648, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4339674711227417, 'reward_total_composite_std': 0.08193355053663254, 'epoch': 0.08}

 24%|██▎       | 779/3300 [1:56:57<5:15:49,  7.52s/it]INFO 04-13 09:17:55 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 780/3300 [1:57:08<6:04:36,  8.68s/it]2026-04-13 09:18:06,490 | INFO | train_grpo_train | metrics_logged mode=train step=780

                                                      
{'loss': -0.1464, 'grad_norm': 2.2677340507507324, 'learning_rate': 7.639393939393939e-06, 'num_tokens': 1377876.0, 'completions/mean_length': 249.125, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 91.4000015258789, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.6196017265319824, 'rewards/meter/std': 0.33285483717918396, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3471825420856476, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.7700977325439453, 'rewards/repeat_soft/std': 0.1705545336008072, 'rewards/judge_quality/mean': 0.3999999761581421, 'rewards/judge_quality/std': 0.36781206727027893, 'rewards/total_composite/mean': 0.3964031934738159, 'rewards/total_composite/std': 0.3589183986186981, 'reward': 0.3964031934738159, 'reward_std': 0.35891836881637573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12576289474964142, 'sampling/sampling_logp_difference/max': 1.7001616954803467, 'sampling/importance_sampling_ratio/min': 0.1826539784669876, 'sampling/importance_sampling_ratio/mean': 1.0203227996826172, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5451840795576572, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06860982486978173, 'clip_ratio/high_max': 0.06860982486978173, 'clip_ratio/region_mean': 0.06860982486978173, 'reward_total_mean': 0.3964031934738159, 'reward_meter_mean': 0.6196017265319824, 'reward_meter_std': 0.33285483717918396, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3471825420856476, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.7700977325439453, 'reward_repeat_soft_std': 0.1705545336008072, 'reward_judge_quality_mean': 0.3999999761581421, 'reward_judge_quality_std': 0.36781206727027893, 'reward_total_composite_mean': 0.3964031934738159, 'reward_total_composite_std': 0.3589183986186981, 'epoch': 0.08}

 24%|██▎       | 780/3300 [1:57:08<6:04:36,  8.68s/it]INFO 04-13 09:18:06 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 781/3300 [1:57:20<6:42:37,  9.59s/it]2026-04-13 09:18:18,205 | INFO | train_grpo_train | metrics_logged mode=train step=781

                                                      
{'loss': -0.1202, 'grad_norm': 3.2003910541534424, 'learning_rate': 7.636363636363638e-06, 'num_tokens': 1379420.0, 'completions/mean_length': 107.0, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.142860412597656, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.3380328416824341, 'rewards/meter/std': 0.3630189299583435, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9931943416595459, 'rewards/repeat_soft/std': 0.01353566162288189, 'rewards/judge_quality/mean': 0.5212500095367432, 'rewards/judge_quality/std': 0.2591986358165741, 'rewards/total_composite/mean': 0.37957361340522766, 'rewards/total_composite/std': 0.17293740808963776, 'reward': 0.37957361340522766, 'reward_std': 0.17293739318847656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15182960033416748, 'sampling/sampling_logp_difference/max': 1.8792412281036377, 'sampling/importance_sampling_ratio/min': 0.15270593762397766, 'sampling/importance_sampling_ratio/mean': 1.0260132551193237, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9700796380639076, 'clip_ratio/low_mean': 0.060232844203710556, 'clip_ratio/low_min': 0.060232844203710556, 'clip_ratio/high_mean': 0.0648200111463666, 'clip_ratio/high_max': 0.0648200111463666, 'clip_ratio/region_mean': 0.12505285535007715, 'reward_total_mean': 0.37957361340522766, 'reward_meter_mean': 0.3380328416824341, 'reward_meter_std': 0.3630189299583435, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9931943416595459, 'reward_repeat_soft_std': 0.01353566162288189, 'reward_judge_quality_mean': 0.5212500095367432, 'reward_judge_quality_std': 0.2591986358165741, 'reward_total_composite_mean': 0.37957361340522766, 'reward_total_composite_std': 0.17293740808963776, 'epoch': 0.08}

 24%|██▎       | 781/3300 [1:57:20<6:42:37,  9.59s/it]INFO 04-13 09:18:18 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 782/3300 [1:57:31<7:01:05, 10.03s/it]2026-04-13 09:18:29,277 | INFO | train_grpo_train | metrics_logged mode=train step=782

                                                      
{'loss': -0.0638, 'grad_norm': 3.63277530670166, 'learning_rate': 7.633333333333334e-06, 'num_tokens': 1380808.0, 'completions/mean_length': 87.5, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.85714340209961, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.5601080656051636, 'rewards/meter/std': 0.4184094965457916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.33161941170692444, 'rewards/total_composite/mean': 0.5623338222503662, 'rewards/total_composite/std': 0.2842698395252228, 'reward': 0.5623338222503662, 'reward_std': 0.2842698395252228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17743803560733795, 'sampling/sampling_logp_difference/max': 1.9913415908813477, 'sampling/importance_sampling_ratio/min': 0.1365121603012085, 'sampling/importance_sampling_ratio/mean': 1.0241336822509766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8500548712909222, 'clip_ratio/low_mean': 0.04614695254713297, 'clip_ratio/low_min': 0.04614695254713297, 'clip_ratio/high_mean': 0.08065619971603155, 'clip_ratio/high_max': 0.08065619971603155, 'clip_ratio/region_mean': 0.12680315226316452, 'reward_total_mean': 0.5623338222503662, 'reward_meter_mean': 0.5601080656051636, 'reward_meter_std': 0.4184094965457916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.33161941170692444, 'reward_total_composite_mean': 0.5623338222503662, 'reward_total_composite_std': 0.2842698395252228, 'epoch': 0.08}

 24%|██▎       | 782/3300 [1:57:31<7:01:05, 10.03s/it]INFO 04-13 09:18:29 [block_pool.py:378] Successfully reset prefix cache

 24%|██▎       | 783/3300 [1:57:42<7:15:40, 10.39s/it]2026-04-13 09:18:40,484 | INFO | train_grpo_train | metrics_logged mode=train step=783

                                                      
{'loss': -0.1072, 'grad_norm': 1.7008535861968994, 'learning_rate': 7.630303030303031e-06, 'num_tokens': 1382436.0, 'completions/mean_length': 164.5, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 48.66666793823242, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9663210511207581, 'rewards/meter/std': 0.04540238901972771, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9975972771644592, 'rewards/repeat_soft/std': 0.004779578186571598, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.2602711617946625, 'rewards/total_composite/mean': 0.5138510465621948, 'rewards/total_composite/std': 0.3300139904022217, 'reward': 0.5138510465621948, 'reward_std': 0.3300139904022217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16095556318759918, 'sampling/sampling_logp_difference/max': 2.0277771949768066, 'sampling/importance_sampling_ratio/min': 0.13162778317928314, 'sampling/importance_sampling_ratio/mean': 0.9975730180740356, 'sampling/importance_sampling_ratio/max': 1.8377701044082642, 'entropy': 1.0407484769821167, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.12459279596805573, 'clip_ratio/high_max': 0.12459279596805573, 'clip_ratio/region_mean': 0.12459279596805573, 'reward_total_mean': 0.5138510465621948, 'reward_meter_mean': 0.9663210511207581, 'reward_meter_std': 0.04540238901972771, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9975972771644592, 'reward_repeat_soft_std': 0.004779578186571598, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.2602711617946625, 'reward_total_composite_mean': 0.5138510465621948, 'reward_total_composite_std': 0.3300139904022217, 'epoch': 0.08}

 24%|██▎       | 783/3300 [1:57:42<7:15:40, 10.39s/it]INFO 04-13 09:18:40 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 784/3300 [1:57:54<7:27:50, 10.68s/it]2026-04-13 09:18:51,856 | INFO | train_grpo_train | metrics_logged mode=train step=784

                                                      
{'loss': -0.157, 'grad_norm': 4.411523342132568, 'learning_rate': 7.627272727272727e-06, 'num_tokens': 1384468.0, 'completions/mean_length': 134.0, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.0, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.8589165806770325, 'rewards/meter/std': 0.26121780276298523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9155592918395996, 'rewards/repeat_soft/std': 0.04644685238599777, 'rewards/judge_quality/mean': 0.6862499713897705, 'rewards/judge_quality/std': 0.34221702814102173, 'rewards/total_composite/mean': 0.7048117518424988, 'rewards/total_composite/std': 0.34871798753738403, 'reward': 0.7048117518424988, 'reward_std': 0.34871798753738403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17850680649280548, 'sampling/sampling_logp_difference/max': 1.966983675956726, 'sampling/importance_sampling_ratio/min': 0.13987812399864197, 'sampling/importance_sampling_ratio/mean': 1.0244739055633545, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.102909229695797, 'clip_ratio/low_mean': 0.032152908854186535, 'clip_ratio/low_min': 0.032152908854186535, 'clip_ratio/high_mean': 0.09117072448134422, 'clip_ratio/high_max': 0.09117072448134422, 'clip_ratio/region_mean': 0.12332363333553076, 'reward_total_mean': 0.7048117518424988, 'reward_meter_mean': 0.8589165806770325, 'reward_meter_std': 0.26121780276298523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9155592918395996, 'reward_repeat_soft_std': 0.04644685238599777, 'reward_judge_quality_mean': 0.6862499713897705, 'reward_judge_quality_std': 0.34221702814102173, 'reward_total_composite_mean': 0.7048117518424988, 'reward_total_composite_std': 0.34871798753738403, 'epoch': 0.08}

 24%|██▍       | 784/3300 [1:57:54<7:27:50, 10.68s/it]INFO 04-13 09:18:52 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 785/3300 [1:58:01<6:41:19,  9.57s/it]2026-04-13 09:18:58,843 | INFO | train_grpo_train | metrics_logged mode=train step=785

                                                      
{'loss': 0.0508, 'grad_norm': 12.999509811401367, 'learning_rate': 7.6242424242424254e-06, 'num_tokens': 1386131.0, 'completions/mean_length': 50.875, 'completions/min_length': 43.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8049522042274475, 'rewards/meter/std': 0.3317568898200989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9843548536300659, 'rewards/repeat_soft/std': 0.029881445690989494, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.6261520385742188, 'rewards/total_composite/std': 0.14769762754440308, 'reward': 0.6261520385742188, 'reward_std': 0.14769762754440308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17569702863693237, 'sampling/sampling_logp_difference/max': 2.9056591987609863, 'sampling/importance_sampling_ratio/min': 0.05471271649003029, 'sampling/importance_sampling_ratio/mean': 1.011540174484253, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1616344675421715, 'clip_ratio/low_mean': 0.0884865066036582, 'clip_ratio/low_min': 0.0884865066036582, 'clip_ratio/high_mean': 0.09933181293308735, 'clip_ratio/high_max': 0.09933181293308735, 'clip_ratio/region_mean': 0.18781831953674555, 'reward_total_mean': 0.6261520385742188, 'reward_meter_mean': 0.8049522042274475, 'reward_meter_std': 0.3317568898200989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9843548536300659, 'reward_repeat_soft_std': 0.029881445690989494, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.6261520385742188, 'reward_total_composite_std': 0.14769762754440308, 'epoch': 0.08}

 24%|██▍       | 785/3300 [1:58:01<6:41:19,  9.57s/it]INFO 04-13 09:18:59 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 786/3300 [1:58:06<5:52:47,  8.42s/it]2026-04-13 09:19:04,572 | INFO | train_grpo_train | metrics_logged mode=train step=786

                                                      
{'loss': 0.0085, 'grad_norm': 21.665050506591797, 'learning_rate': 7.621212121212122e-06, 'num_tokens': 1387476.0, 'completions/mean_length': 25.125, 'completions/min_length': 22.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6657093167304993, 'rewards/meter/std': 0.3802475929260254, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9570496082305908, 'rewards/repeat_soft/std': 0.015415864065289497, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5710169076919556, 'rewards/total_composite/std': 0.1791563630104065, 'reward': 0.5710169076919556, 'reward_std': 0.1791563332080841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14274726808071136, 'sampling/sampling_logp_difference/max': 0.998407244682312, 'sampling/importance_sampling_ratio/min': 0.3684658706188202, 'sampling/importance_sampling_ratio/mean': 1.031922459602356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0329799726605415, 'clip_ratio/low_mean': 0.04865320026874542, 'clip_ratio/low_min': 0.04865320026874542, 'clip_ratio/high_mean': 0.05201863497495651, 'clip_ratio/high_max': 0.05201863497495651, 'clip_ratio/region_mean': 0.10067183524370193, 'reward_total_mean': 0.5710169076919556, 'reward_meter_mean': 0.6657093167304993, 'reward_meter_std': 0.3802475929260254, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9570496082305908, 'reward_repeat_soft_std': 0.015415864065289497, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5710169076919556, 'reward_total_composite_std': 0.1791563630104065, 'epoch': 0.08}

 24%|██▍       | 786/3300 [1:58:06<5:52:47,  8.42s/it]INFO 04-13 09:19:04 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 787/3300 [1:58:13<5:25:48,  7.78s/it]2026-04-13 09:19:10,856 | INFO | train_grpo_train | metrics_logged mode=train step=787

                                                      
{'loss': 0.0898, 'grad_norm': 15.552066802978516, 'learning_rate': 7.618181818181819e-06, 'num_tokens': 1389106.0, 'completions/mean_length': 40.75, 'completions/min_length': 33.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.680222749710083, 'rewards/meter/std': 0.3175264000892639, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9770693778991699, 'rewards/repeat_soft/std': 0.020964395254850388, 'rewards/judge_quality/mean': 0.6062500476837158, 'rewards/judge_quality/std': 0.19390259683132172, 'rewards/total_composite/mean': 0.6077351570129395, 'rewards/total_composite/std': 0.13894174993038177, 'reward': 0.6077351570129395, 'reward_std': 0.1389417052268982, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1768689900636673, 'sampling/sampling_logp_difference/max': 1.7020316123962402, 'sampling/importance_sampling_ratio/min': 0.1823127567768097, 'sampling/importance_sampling_ratio/mean': 1.0149352550506592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.059500440955162, 'clip_ratio/low_mean': 0.060787507332861423, 'clip_ratio/low_min': 0.060787507332861423, 'clip_ratio/high_mean': 0.08163997158408165, 'clip_ratio/high_max': 0.08163997158408165, 'clip_ratio/region_mean': 0.14242747891694307, 'reward_total_mean': 0.6077351570129395, 'reward_meter_mean': 0.680222749710083, 'reward_meter_std': 0.3175264000892639, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9770693778991699, 'reward_repeat_soft_std': 0.020964395254850388, 'reward_judge_quality_mean': 0.6062500476837158, 'reward_judge_quality_std': 0.19390259683132172, 'reward_total_composite_mean': 0.6077351570129395, 'reward_total_composite_std': 0.13894174993038177, 'epoch': 0.08}

 24%|██▍       | 787/3300 [1:58:13<5:25:48,  7.78s/it]INFO 04-13 09:19:11 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 788/3300 [1:58:19<5:11:41,  7.44s/it]2026-04-13 09:19:17,519 | INFO | train_grpo_train | metrics_logged mode=train step=788

                                                      
{'loss': -0.0161, 'grad_norm': 14.044755935668945, 'learning_rate': 7.6151515151515155e-06, 'num_tokens': 1390843.0, 'completions/mean_length': 47.125, 'completions/min_length': 40.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.7587764263153076, 'rewards/meter/std': 0.327594518661499, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9857375621795654, 'rewards/repeat_soft/std': 0.01699240505695343, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.215406596660614, 'rewards/total_composite/mean': 0.6128135919570923, 'rewards/total_composite/std': 0.1883668303489685, 'reward': 0.6128135919570923, 'reward_std': 0.1883668452501297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14258180558681488, 'sampling/sampling_logp_difference/max': 1.9234857559204102, 'sampling/importance_sampling_ratio/min': 0.14609681069850922, 'sampling/importance_sampling_ratio/mean': 1.0270355939865112, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.011030912399292, 'clip_ratio/low_mean': 0.1028314083814621, 'clip_ratio/low_min': 0.1028314083814621, 'clip_ratio/high_mean': 0.07616372779011726, 'clip_ratio/high_max': 0.07616372779011726, 'clip_ratio/region_mean': 0.17899513617157936, 'reward_total_mean': 0.6128135919570923, 'reward_meter_mean': 0.7587764263153076, 'reward_meter_std': 0.327594518661499, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9857375621795654, 'reward_repeat_soft_std': 0.01699240505695343, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.215406596660614, 'reward_total_composite_mean': 0.6128135919570923, 'reward_total_composite_std': 0.1883668303489685, 'epoch': 0.08}

 24%|██▍       | 788/3300 [1:58:19<5:11:41,  7.44s/it]INFO 04-13 09:19:17 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 789/3300 [1:58:25<4:54:26,  7.04s/it]2026-04-13 09:19:23,598 | INFO | train_grpo_train | metrics_logged mode=train step=789

                                                      
{'loss': 0.0352, 'grad_norm': 16.33643913269043, 'learning_rate': 7.612121212121213e-06, 'num_tokens': 1392585.0, 'completions/mean_length': 44.75, 'completions/min_length': 36.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9510666728019714, 'rewards/meter/std': 0.05670829862356186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9683701992034912, 'rewards/repeat_soft/std': 0.0318145677447319, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5726941823959351, 'rewards/total_composite/std': 0.250593900680542, 'reward': 0.5726941823959351, 'reward_std': 0.250593900680542, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1539141684770584, 'sampling/sampling_logp_difference/max': 2.2820472717285156, 'sampling/importance_sampling_ratio/min': 0.1020750179886818, 'sampling/importance_sampling_ratio/mean': 1.0005478858947754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8115236088633537, 'clip_ratio/low_mean': 0.025510204955935478, 'clip_ratio/low_min': 0.025510204955935478, 'clip_ratio/high_mean': 0.1278712498024106, 'clip_ratio/high_max': 0.1278712498024106, 'clip_ratio/region_mean': 0.15338145475834608, 'reward_total_mean': 0.5726941823959351, 'reward_meter_mean': 0.9510666728019714, 'reward_meter_std': 0.05670829862356186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9683701992034912, 'reward_repeat_soft_std': 0.0318145677447319, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5726941823959351, 'reward_total_composite_std': 0.250593900680542, 'epoch': 0.08}

 24%|██▍       | 789/3300 [1:58:25<4:54:26,  7.04s/it]INFO 04-13 09:19:23 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 790/3300 [1:58:33<5:02:01,  7.22s/it]2026-04-13 09:19:31,248 | INFO | train_grpo_train | metrics_logged mode=train step=790

                                                      
{'loss': 0.0481, 'grad_norm': 9.999359130859375, 'learning_rate': 7.609090909090909e-06, 'num_tokens': 1395365.0, 'completions/mean_length': 127.5, 'completions/min_length': 116.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.5, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.6306518316268921, 'rewards/meter/std': 0.3782559037208557, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9702376127243042, 'rewards/repeat_soft/std': 0.031808041036129, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.2133910059928894, 'rewards/total_composite/mean': 0.6025485396385193, 'rewards/total_composite/std': 0.2148790806531906, 'reward': 0.6025485396385193, 'reward_std': 0.2148790806531906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16975581645965576, 'sampling/sampling_logp_difference/max': 2.3685126304626465, 'sampling/importance_sampling_ratio/min': 0.09361986815929413, 'sampling/importance_sampling_ratio/mean': 1.0210152864456177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0747182965278625, 'clip_ratio/low_mean': 0.05339346919208765, 'clip_ratio/low_min': 0.05339346919208765, 'clip_ratio/high_mean': 0.10562057327479124, 'clip_ratio/high_max': 0.10562057327479124, 'clip_ratio/region_mean': 0.1590140424668789, 'reward_total_mean': 0.6025485396385193, 'reward_meter_mean': 0.6306518316268921, 'reward_meter_std': 0.3782559037208557, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9702376127243042, 'reward_repeat_soft_std': 0.031808041036129, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.2133910059928894, 'reward_total_composite_mean': 0.6025485396385193, 'reward_total_composite_std': 0.2148790806531906, 'epoch': 0.08}

 24%|██▍       | 790/3300 [1:58:33<5:02:01,  7.22s/it]INFO 04-13 09:19:31 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 791/3300 [1:58:45<5:56:56,  8.54s/it]2026-04-13 09:19:42,858 | INFO | train_grpo_train | metrics_logged mode=train step=791

                                                      
{'loss': -0.141, 'grad_norm': 3.7419774532318115, 'learning_rate': 7.606060606060606e-06, 'num_tokens': 1397425.0, 'completions/mean_length': 155.5, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 104.5714340209961, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.6867519617080688, 'rewards/meter/std': 0.4089537560939789, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.35197150707244873, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9493618011474609, 'rewards/repeat_soft/std': 0.05099398270249367, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.21022097766399384, 'rewards/total_composite/mean': 0.5299545526504517, 'rewards/total_composite/std': 0.26412343978881836, 'reward': 0.5299545526504517, 'reward_std': 0.26412343978881836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16901877522468567, 'sampling/sampling_logp_difference/max': 1.9053401947021484, 'sampling/importance_sampling_ratio/min': 0.17543934285640717, 'sampling/importance_sampling_ratio/mean': 1.016891360282898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9964922145009041, 'clip_ratio/low_mean': 0.03585714288055897, 'clip_ratio/low_min': 0.03585714288055897, 'clip_ratio/high_mean': 0.08133437763899565, 'clip_ratio/high_max': 0.08133437763899565, 'clip_ratio/region_mean': 0.11719152051955462, 'reward_total_mean': 0.5299545526504517, 'reward_meter_mean': 0.6867519617080688, 'reward_meter_std': 0.4089537560939789, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.35197150707244873, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9493618011474609, 'reward_repeat_soft_std': 0.05099398270249367, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.21022097766399384, 'reward_total_composite_mean': 0.5299545526504517, 'reward_total_composite_std': 0.26412343978881836, 'epoch': 0.08}

 24%|██▍       | 791/3300 [1:58:45<5:56:56,  8.54s/it]INFO 04-13 09:19:43 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 792/3300 [1:58:50<5:24:20,  7.76s/it]2026-04-13 09:19:48,805 | INFO | train_grpo_train | metrics_logged mode=train step=792

                                                      
{'loss': 0.0258, 'grad_norm': 15.24664306640625, 'learning_rate': 7.603030303030303e-06, 'num_tokens': 1398939.0, 'completions/mean_length': 21.25, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9517276287078857, 'rewards/meter/std': 0.0785837322473526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.7555526494979858, 'rewards/total_composite/std': 0.16212528944015503, 'reward': 0.7555526494979858, 'reward_std': 0.16212528944015503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10954467952251434, 'sampling/sampling_logp_difference/max': 0.9737614393234253, 'sampling/importance_sampling_ratio/min': 0.3776598274707794, 'sampling/importance_sampling_ratio/mean': 1.0057436227798462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5773142501711845, 'clip_ratio/low_mean': 0.08900091052055359, 'clip_ratio/low_min': 0.08900091052055359, 'clip_ratio/high_mean': 0.07013889029622078, 'clip_ratio/high_max': 0.07013889029622078, 'clip_ratio/region_mean': 0.15913980081677437, 'reward_total_mean': 0.7555526494979858, 'reward_meter_mean': 0.9517276287078857, 'reward_meter_std': 0.0785837322473526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.7555526494979858, 'reward_total_composite_std': 0.16212528944015503, 'epoch': 0.08}

 24%|██▍       | 792/3300 [1:58:51<5:24:20,  7.76s/it]INFO 04-13 09:19:49 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 793/3300 [1:59:02<6:10:14,  8.86s/it]2026-04-13 09:20:00,232 | INFO | train_grpo_train | metrics_logged mode=train step=793

                                                      
{'loss': -0.0863, 'grad_norm': 3.825547695159912, 'learning_rate': 7.600000000000001e-06, 'num_tokens': 1400431.0, 'completions/mean_length': 98.5, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.42857360839844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.37778598070144653, 'rewards/meter/std': 0.36534878611564636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9773523807525635, 'rewards/repeat_soft/std': 0.03622420132160187, 'rewards/judge_quality/mean': 0.6062500476837158, 'rewards/judge_quality/std': 0.32053250074386597, 'rewards/total_composite/mean': 0.4723506569862366, 'rewards/total_composite/std': 0.23831723630428314, 'reward': 0.4723506569862366, 'reward_std': 0.23831720650196075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1502223163843155, 'sampling/sampling_logp_difference/max': 2.3693289756774902, 'sampling/importance_sampling_ratio/min': 0.09354346990585327, 'sampling/importance_sampling_ratio/mean': 0.9848490357398987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9237863123416901, 'clip_ratio/low_mean': 0.038352273404598236, 'clip_ratio/low_min': 0.038352273404598236, 'clip_ratio/high_mean': 0.08816816285252571, 'clip_ratio/high_max': 0.08816816285252571, 'clip_ratio/region_mean': 0.12652043625712395, 'reward_total_mean': 0.4723506569862366, 'reward_meter_mean': 0.37778598070144653, 'reward_meter_std': 0.36534878611564636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9773523807525635, 'reward_repeat_soft_std': 0.03622420132160187, 'reward_judge_quality_mean': 0.6062500476837158, 'reward_judge_quality_std': 0.32053250074386597, 'reward_total_composite_mean': 0.4723506569862366, 'reward_total_composite_std': 0.23831723630428314, 'epoch': 0.08}

 24%|██▍       | 793/3300 [1:59:02<6:10:14,  8.86s/it]INFO 04-13 09:20:00 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 794/3300 [1:59:15<7:03:16, 10.13s/it]2026-04-13 09:20:13,336 | INFO | train_grpo_train | metrics_logged mode=train step=794

                                                      
{'loss': 0.0329, 'grad_norm': 9.547015190124512, 'learning_rate': 7.596969696969697e-06, 'num_tokens': 1402650.0, 'completions/mean_length': 99.375, 'completions/min_length': 91.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9521958827972412, 'rewards/meter/std': 0.07844063639640808, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9547011256217957, 'rewards/repeat_soft/std': 0.025254275649785995, 'rewards/judge_quality/mean': 0.7074999809265137, 'rewards/judge_quality/std': 0.2474873960018158, 'rewards/total_composite/mean': 0.7787088751792908, 'rewards/total_composite/std': 0.15237843990325928, 'reward': 0.7787088751792908, 'reward_std': 0.15237843990325928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1505635529756546, 'sampling/sampling_logp_difference/max': 1.9828691482543945, 'sampling/importance_sampling_ratio/min': 0.13767366111278534, 'sampling/importance_sampling_ratio/mean': 1.0115715265274048, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0196355730295181, 'clip_ratio/low_mean': 0.048223040997982025, 'clip_ratio/low_min': 0.048223040997982025, 'clip_ratio/high_mean': 0.0992923779413104, 'clip_ratio/high_max': 0.0992923779413104, 'clip_ratio/region_mean': 0.14751541893929243, 'reward_total_mean': 0.7787088751792908, 'reward_meter_mean': 0.9521958827972412, 'reward_meter_std': 0.07844063639640808, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9547011256217957, 'reward_repeat_soft_std': 0.025254275649785995, 'reward_judge_quality_mean': 0.7074999809265137, 'reward_judge_quality_std': 0.2474873960018158, 'reward_total_composite_mean': 0.7787088751792908, 'reward_total_composite_std': 0.15237843990325928, 'epoch': 0.08}

 24%|██▍       | 794/3300 [1:59:15<7:03:16, 10.13s/it]INFO 04-13 09:20:13 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 795/3300 [1:59:22<6:25:56,  9.24s/it]2026-04-13 09:20:20,505 | INFO | train_grpo_train | metrics_logged mode=train step=795

                                                      
{'loss': 0.0164, 'grad_norm': 7.827341079711914, 'learning_rate': 7.593939393939395e-06, 'num_tokens': 1405076.0, 'completions/mean_length': 114.25, 'completions/min_length': 96.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.25, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9653751850128174, 'rewards/meter/std': 0.04670365899801254, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9599523544311523, 'rewards/repeat_soft/std': 0.03862236067652702, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.2232191562652588, 'rewards/total_composite/mean': 0.738872766494751, 'rewards/total_composite/std': 0.15536025166511536, 'reward': 0.738872766494751, 'reward_std': 0.15536023676395416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16115161776542664, 'sampling/sampling_logp_difference/max': 2.0109481811523438, 'sampling/importance_sampling_ratio/min': 0.1338616907596588, 'sampling/importance_sampling_ratio/mean': 1.0239852666854858, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1828294694423676, 'clip_ratio/low_mean': 0.0740177072584629, 'clip_ratio/low_min': 0.0740177072584629, 'clip_ratio/high_mean': 0.07718428783118725, 'clip_ratio/high_max': 0.07718428783118725, 'clip_ratio/region_mean': 0.15120199508965015, 'reward_total_mean': 0.738872766494751, 'reward_meter_mean': 0.9653751850128174, 'reward_meter_std': 0.04670365899801254, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9599523544311523, 'reward_repeat_soft_std': 0.03862236067652702, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.2232191562652588, 'reward_total_composite_mean': 0.738872766494751, 'reward_total_composite_std': 0.15536025166511536, 'epoch': 0.08}

 24%|██▍       | 795/3300 [1:59:22<6:25:56,  9.24s/it]INFO 04-13 09:20:20 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 796/3300 [1:59:33<6:48:45,  9.79s/it]2026-04-13 09:20:31,586 | INFO | train_grpo_train | metrics_logged mode=train step=796

                                                      
{'loss': -0.076, 'grad_norm': 3.3293709754943848, 'learning_rate': 7.590909090909091e-06, 'num_tokens': 1406630.0, 'completions/mean_length': 102.25, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.71428680419922, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.664996862411499, 'rewards/meter/std': 0.3527509570121765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.993262529373169, 'rewards/repeat_soft/std': 0.00963639747351408, 'rewards/judge_quality/mean': 0.668749988079071, 'rewards/judge_quality/std': 0.3276948928833008, 'rewards/total_composite/mean': 0.6403868198394775, 'rewards/total_composite/std': 0.30043327808380127, 'reward': 0.6403868198394775, 'reward_std': 0.30043327808380127, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1478281021118164, 'sampling/sampling_logp_difference/max': 1.687039852142334, 'sampling/importance_sampling_ratio/min': 0.18506652116775513, 'sampling/importance_sampling_ratio/mean': 0.9968125224113464, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7571103498339653, 'clip_ratio/low_mean': 0.03758741356432438, 'clip_ratio/low_min': 0.03758741356432438, 'clip_ratio/high_mean': 0.09591932641342282, 'clip_ratio/high_max': 0.09591932641342282, 'clip_ratio/region_mean': 0.1335067399777472, 'reward_total_mean': 0.6403868198394775, 'reward_meter_mean': 0.664996862411499, 'reward_meter_std': 0.3527509570121765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.993262529373169, 'reward_repeat_soft_std': 0.00963639747351408, 'reward_judge_quality_mean': 0.668749988079071, 'reward_judge_quality_std': 0.3276948928833008, 'reward_total_composite_mean': 0.6403868198394775, 'reward_total_composite_std': 0.30043327808380127, 'epoch': 0.08}

 24%|██▍       | 796/3300 [1:59:33<6:48:45,  9.79s/it]INFO 04-13 09:20:31 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 797/3300 [1:59:45<7:11:03, 10.33s/it]2026-04-13 09:20:43,172 | INFO | train_grpo_train | metrics_logged mode=train step=797

                                                      
{'loss': -0.1188, 'grad_norm': 5.404216289520264, 'learning_rate': 7.587878787878788e-06, 'num_tokens': 1408726.0, 'completions/mean_length': 154.0, 'completions/min_length': 94.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 102.85714721679688, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.7925493717193604, 'rewards/meter/std': 0.23417937755584717, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8938276767730713, 'rewards/repeat_soft/std': 0.07476700097322464, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.22222496569156647, 'rewards/total_composite/mean': 0.4308021366596222, 'rewards/total_composite/std': 0.2820538878440857, 'reward': 0.4308021366596222, 'reward_std': 0.2820538580417633, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1384115219116211, 'sampling/sampling_logp_difference/max': 2.7043967247009277, 'sampling/importance_sampling_ratio/min': 0.06691067665815353, 'sampling/importance_sampling_ratio/mean': 0.9951090812683105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7167848758399487, 'clip_ratio/low_mean': 0.014627659693360329, 'clip_ratio/low_min': 0.014627659693360329, 'clip_ratio/high_mean': 0.11153530143201351, 'clip_ratio/high_max': 0.11153530143201351, 'clip_ratio/region_mean': 0.12616296112537384, 'reward_total_mean': 0.4308021366596222, 'reward_meter_mean': 0.7925493717193604, 'reward_meter_std': 0.23417937755584717, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8938276767730713, 'reward_repeat_soft_std': 0.07476700097322464, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.22222496569156647, 'reward_total_composite_mean': 0.4308021366596222, 'reward_total_composite_std': 0.2820538878440857, 'epoch': 0.08}

 24%|██▍       | 797/3300 [1:59:45<7:11:03, 10.33s/it]INFO 04-13 09:20:43 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 798/3300 [1:59:52<6:24:53,  9.23s/it]2026-04-13 09:20:49,837 | INFO | train_grpo_train | metrics_logged mode=train step=798

                                                      
{'loss': 0.0659, 'grad_norm': 13.294465065002441, 'learning_rate': 7.584848484848486e-06, 'num_tokens': 1410885.0, 'completions/mean_length': 88.875, 'completions/min_length': 66.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.6034026145935059, 'rewards/meter/std': 0.3710324764251709, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9907156229019165, 'rewards/repeat_soft/std': 0.008669731207191944, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5133362412452698, 'rewards/total_composite/std': 0.10131111741065979, 'reward': 0.5133362412452698, 'reward_std': 0.1013111099600792, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16643524169921875, 'sampling/sampling_logp_difference/max': 8.26811408996582, 'sampling/importance_sampling_ratio/min': 0.0002565686882007867, 'sampling/importance_sampling_ratio/mean': 0.9939804077148438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6501155868172646, 'clip_ratio/low_mean': 0.07310330867767334, 'clip_ratio/low_min': 0.07310330867767334, 'clip_ratio/high_mean': 0.07407133094966412, 'clip_ratio/high_max': 0.07407133094966412, 'clip_ratio/region_mean': 0.14717463962733746, 'reward_total_mean': 0.5133362412452698, 'reward_meter_mean': 0.6034026145935059, 'reward_meter_std': 0.3710324764251709, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9907156229019165, 'reward_repeat_soft_std': 0.008669731207191944, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5133362412452698, 'reward_total_composite_std': 0.10131111741065979, 'epoch': 0.08}

 24%|██▍       | 798/3300 [1:59:52<6:24:53,  9.23s/it]INFO 04-13 09:20:50 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 799/3300 [1:59:58<5:51:45,  8.44s/it]2026-04-13 09:20:56,430 | INFO | train_grpo_train | metrics_logged mode=train step=799

                                                      
{'loss': 0.0763, 'grad_norm': 17.824810028076172, 'learning_rate': 7.581818181818183e-06, 'num_tokens': 1412684.0, 'completions/mean_length': 57.875, 'completions/min_length': 52.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.875, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4198606312274933, 'rewards/meter/std': 0.23079407215118408, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9929182529449463, 'rewards/repeat_soft/std': 0.008453218266367912, 'rewards/judge_quality/mean': 0.6612499952316284, 'rewards/judge_quality/std': 0.2654612064361572, 'rewards/total_composite/mean': 0.5252758264541626, 'rewards/total_composite/std': 0.1283113658428192, 'reward': 0.5252758264541626, 'reward_std': 0.1283113658428192, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18238553404808044, 'sampling/sampling_logp_difference/max': 2.830293655395508, 'sampling/importance_sampling_ratio/min': 0.058995530009269714, 'sampling/importance_sampling_ratio/mean': 1.0064098834991455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.822184830904007, 'clip_ratio/low_mean': 0.09742045029997826, 'clip_ratio/low_min': 0.09742045029997826, 'clip_ratio/high_mean': 0.05481120944023132, 'clip_ratio/high_max': 0.05481120944023132, 'clip_ratio/region_mean': 0.15223165974020958, 'reward_total_mean': 0.5252758264541626, 'reward_meter_mean': 0.4198606312274933, 'reward_meter_std': 0.23079407215118408, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9929182529449463, 'reward_repeat_soft_std': 0.008453218266367912, 'reward_judge_quality_mean': 0.6612499952316284, 'reward_judge_quality_std': 0.2654612064361572, 'reward_total_composite_mean': 0.5252758264541626, 'reward_total_composite_std': 0.1283113658428192, 'epoch': 0.08}

 24%|██▍       | 799/3300 [1:59:58<5:51:45,  8.44s/it]INFO 04-13 09:20:56 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 800/3300 [2:00:04<5:23:33,  7.77s/it]2026-04-13 09:21:02,618 | INFO | train_grpo_train | metrics_logged mode=train step=800

                                                      
{'loss': 0.0126, 'grad_norm': 11.47986125946045, 'learning_rate': 7.57878787878788e-06, 'num_tokens': 1414466.0, 'completions/mean_length': 50.75, 'completions/min_length': 43.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.3981095552444458, 'rewards/meter/std': 0.4597620666027069, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975108504295349, 'rewards/repeat_soft/std': 0.004614746663719416, 'rewards/judge_quality/mean': 0.78125, 'rewards/judge_quality/std': 0.2126995027065277, 'rewards/total_composite/mean': 0.5501787662506104, 'rewards/total_composite/std': 0.25034084916114807, 'reward': 0.5501787662506104, 'reward_std': 0.2503408193588257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1491750329732895, 'sampling/sampling_logp_difference/max': 1.6211563348770142, 'sampling/importance_sampling_ratio/min': 0.19766999781131744, 'sampling/importance_sampling_ratio/mean': 1.0138750076293945, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9692709296941757, 'clip_ratio/low_mean': 0.09596501477062702, 'clip_ratio/low_min': 0.09596501477062702, 'clip_ratio/high_mean': 0.061666665598750114, 'clip_ratio/high_max': 0.061666665598750114, 'clip_ratio/region_mean': 0.15763168036937714, 'reward_total_mean': 0.5501787662506104, 'reward_meter_mean': 0.3981095552444458, 'reward_meter_std': 0.4597620666027069, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975108504295349, 'reward_repeat_soft_std': 0.004614746663719416, 'reward_judge_quality_mean': 0.78125, 'reward_judge_quality_std': 0.2126995027065277, 'reward_total_composite_mean': 0.5501787662506104, 'reward_total_composite_std': 0.25034084916114807, 'epoch': 0.08}

 24%|██▍       | 800/3300 [2:00:04<5:23:33,  7.77s/it]INFO 04-13 09:21:02 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:08<00:34,  4.37s/it]

 30%|███       | 3/10 [00:12<00:30,  4.30s/it]

 40%|████      | 4/10 [00:16<00:24,  4.03s/it]

 50%|█████     | 5/10 [00:20<00:20,  4.02s/it]

 60%|██████    | 6/10 [00:29<00:22,  5.60s/it]

 70%|███████   | 7/10 [00:38<00:19,  6.61s/it]

 80%|████████  | 8/10 [00:41<00:11,  5.64s/it]

 90%|█████████ | 9/10 [00:45<00:05,  5.17s/it]

100%|██████████| 10/10 [00:48<00:00,  4.54s/it]2026-04-13 09:21:55,610 | INFO | train_grpo_train | metrics_logged mode=eval step=800

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 52.9423, 'eval_samples_per_second': 1.511, 'eval_steps_per_second': 0.189, 'eval_num_tokens': 1414466.0, 'eval_completions/mean_length': 96.1125, 'eval_completions/min_length': 39.4, 'eval_completions/max_length': 237.7, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 74.84107208251953, 'eval_completions/min_terminated_length': 39.4, 'eval_completions/max_terminated_length': 128.6, 'eval_rewards/meter/mean': 0.5917634904384613, 'eval_rewards/meter/std': 0.34982125759124755, 'eval_rewards/count_adherence/mean': 0.9554166615009307, 'eval_rewards/count_adherence/std': 0.09762779586017131, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.15235702097415924, 'eval_rewards/repeat_soft/mean': 0.9491789221763611, 'eval_rewards/repeat_soft/std': 0.06461945287883282, 'eval_rewards/judge_quality/mean': 0.5437499970197678, 'eval_rewards/judge_quality/std': 0.21584835201501845, 'eval_rewards/total_composite/mean': 0.5118943601846695, 'eval_rewards/total_composite/std': 0.19045581221580504, 'eval_reward': 0.5118943601846695, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07544206082820892, 'eval_sampling/sampling_logp_difference/max': 1.1279006004333496, 'eval_sampling/importance_sampling_ratio/min': 0.33538752645254133, 'eval_sampling/importance_sampling_ratio/mean': 1.02023948431015, 'eval_sampling/importance_sampling_ratio/max': 1.4546257734298706, 'eval_entropy': 0.865834218263626, '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': 0.5118943601846695, 'eval_reward_meter_mean': 0.5917634904384613, 'eval_reward_meter_std': 0.34982125759124755, 'eval_reward_count_adherence_mean': 0.9554166615009307, 'eval_reward_count_adherence_std': 0.09762779586017131, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.15235702097415924, 'eval_reward_repeat_soft_mean': 0.9491789221763611, 'eval_reward_repeat_soft_std': 0.06461945287883282, 'eval_reward_judge_quality_mean': 0.5437499970197678, 'eval_reward_judge_quality_std': 0.21584835201501845, 'eval_reward_total_composite_mean': 0.5118943601846695, 'eval_reward_total_composite_std': 0.19045581221580504, 'epoch': 0.08}

 24%|██▍       | 800/3300 [2:00:57<5:23:33,  7.77s/it]

100%|██████████| 10/10 [00:48<00:00,  4.54s/it]

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

 24%|██▍       | 801/3300 [2:01:06<16:40:52, 24.03s/it]2026-04-13 09:22:04,602 | INFO | train_grpo_train | metrics_logged mode=train step=801

                                                       
{'loss': 0.0383, 'grad_norm': 14.906476974487305, 'learning_rate': 7.5757575757575764e-06, 'num_tokens': 1416082.0, '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': 0.29853975772857666, 'rewards/meter/std': 0.3955845832824707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9885627627372742, 'rewards/repeat_soft/std': 0.010870145633816719, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.2133910059928894, 'rewards/total_composite/mean': 0.44581830501556396, 'rewards/total_composite/std': 0.10586800426244736, 'reward': 0.44581830501556396, 'reward_std': 0.10586800426244736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1568552553653717, 'sampling/sampling_logp_difference/max': 2.663341999053955, 'sampling/importance_sampling_ratio/min': 0.06971484422683716, 'sampling/importance_sampling_ratio/mean': 0.9949712753295898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1192193403840065, 'clip_ratio/low_mean': 0.08543961867690086, 'clip_ratio/low_min': 0.08543961867690086, 'clip_ratio/high_mean': 0.06054006889462471, 'clip_ratio/high_max': 0.06054006889462471, 'clip_ratio/region_mean': 0.14597968757152557, 'reward_total_mean': 0.44581830501556396, 'reward_meter_mean': 0.29853975772857666, 'reward_meter_std': 0.3955845832824707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9885627627372742, 'reward_repeat_soft_std': 0.010870145633816719, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.2133910059928894, 'reward_total_composite_mean': 0.44581830501556396, 'reward_total_composite_std': 0.10586800426244736, 'epoch': 0.08}

 24%|██▍       | 801/3300 [2:01:06<16:40:52, 24.03s/it]INFO 04-13 09:22:04 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 802/3300 [2:01:13<13:00:29, 18.75s/it]2026-04-13 09:22:11,017 | INFO | train_grpo_train | metrics_logged mode=train step=802

                                                       
{'loss': 0.0174, 'grad_norm': 10.958527565002441, 'learning_rate': 7.572727272727274e-06, 'num_tokens': 1418006.0, 'completions/mean_length': 56.5, 'completions/min_length': 45.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9016284942626953, 'rewards/meter/std': 0.1466209888458252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9821110963821411, 'rewards/repeat_soft/std': 0.021698538213968277, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.6615656614303589, 'rewards/total_composite/std': 0.11730705201625824, 'reward': 0.6615656614303589, 'reward_std': 0.11730704456567764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15427520871162415, 'sampling/sampling_logp_difference/max': 2.4017622470855713, 'sampling/importance_sampling_ratio/min': 0.09055822342634201, 'sampling/importance_sampling_ratio/mean': 1.010503888130188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8828255087137222, 'clip_ratio/low_mean': 0.1184957604855299, 'clip_ratio/low_min': 0.1184957604855299, 'clip_ratio/high_mean': 0.036312622018158436, 'clip_ratio/high_max': 0.036312622018158436, 'clip_ratio/region_mean': 0.15480838250368834, 'reward_total_mean': 0.6615656614303589, 'reward_meter_mean': 0.9016284942626953, 'reward_meter_std': 0.1466209888458252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9821110963821411, 'reward_repeat_soft_std': 0.021698538213968277, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.6615656614303589, 'reward_total_composite_std': 0.11730705201625824, 'epoch': 0.08}

 24%|██▍       | 802/3300 [2:01:13<13:00:29, 18.75s/it]INFO 04-13 09:22:11 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 803/3300 [2:01:24<11:27:27, 16.52s/it]2026-04-13 09:22:22,351 | INFO | train_grpo_train | metrics_logged mode=train step=803

                                                       
{'loss': -0.1125, 'grad_norm': 3.0111196041107178, 'learning_rate': 7.56969696969697e-06, 'num_tokens': 1419800.0, 'completions/mean_length': 183.25, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 73.66667175292969, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.4995984733104706, 'rewards/meter/std': 0.3973415493965149, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9811067581176758, 'rewards/repeat_soft/std': 0.029663169756531715, 'rewards/judge_quality/mean': 0.7024999856948853, 'rewards/judge_quality/std': 0.4027317464351654, 'rewards/total_composite/mean': 0.4504998028278351, 'rewards/total_composite/std': 0.3490696847438812, 'reward': 0.4504998028278351, 'reward_std': 0.3490696847438812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13643351197242737, 'sampling/sampling_logp_difference/max': 1.9621081352233887, 'sampling/importance_sampling_ratio/min': 0.1405617892742157, 'sampling/importance_sampling_ratio/mean': 0.9995192289352417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5766572505235672, 'clip_ratio/low_mean': 0.029759816825389862, 'clip_ratio/low_min': 0.029759816825389862, 'clip_ratio/high_mean': 0.06271313689649105, 'clip_ratio/high_max': 0.06271313689649105, 'clip_ratio/region_mean': 0.09247295372188091, 'reward_total_mean': 0.4504998028278351, 'reward_meter_mean': 0.4995984733104706, 'reward_meter_std': 0.3973415493965149, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.24800792336463928, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9811067581176758, 'reward_repeat_soft_std': 0.029663169756531715, 'reward_judge_quality_mean': 0.7024999856948853, 'reward_judge_quality_std': 0.4027317464351654, 'reward_total_composite_mean': 0.4504998028278351, 'reward_total_composite_std': 0.3490696847438812, 'epoch': 0.08}

 24%|██▍       | 803/3300 [2:01:24<11:27:27, 16.52s/it]INFO 04-13 09:22:22 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 804/3300 [2:01:31<9:28:27, 13.66s/it] 2026-04-13 09:22:29,351 | INFO | train_grpo_train | metrics_logged mode=train step=804

                                                      
{'loss': 0.0414, 'grad_norm': 11.315879821777344, 'learning_rate': 7.566666666666667e-06, 'num_tokens': 1422235.0, 'completions/mean_length': 112.375, 'completions/min_length': 103.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.375, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.7090272903442383, 'rewards/meter/std': 0.2736465334892273, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9467700123786926, 'rewards/repeat_soft/std': 0.030022691935300827, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5997272729873657, 'rewards/total_composite/std': 0.13130873441696167, 'reward': 0.5997272729873657, 'reward_std': 0.13130873441696167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15653181076049805, 'sampling/sampling_logp_difference/max': 2.0967979431152344, 'sampling/importance_sampling_ratio/min': 0.12284916639328003, 'sampling/importance_sampling_ratio/mean': 1.0141373872756958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9335663095116615, 'clip_ratio/low_mean': 0.06188603490591049, 'clip_ratio/low_min': 0.06188603490591049, 'clip_ratio/high_mean': 0.07365716155618429, 'clip_ratio/high_max': 0.07365716155618429, 'clip_ratio/region_mean': 0.13554319646209478, 'reward_total_mean': 0.5997272729873657, 'reward_meter_mean': 0.7090272903442383, 'reward_meter_std': 0.2736465334892273, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9467700123786926, 'reward_repeat_soft_std': 0.030022691935300827, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5997272729873657, 'reward_total_composite_std': 0.13130873441696167, 'epoch': 0.08}

 24%|██▍       | 804/3300 [2:01:31<9:28:27, 13.66s/it]INFO 04-13 09:22:29 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 805/3300 [2:01:38<8:02:20, 11.60s/it]2026-04-13 09:22:36,133 | INFO | train_grpo_train | metrics_logged mode=train step=805

                                                      
{'loss': -0.0006, 'grad_norm': 15.6214599609375, 'learning_rate': 7.563636363636364e-06, 'num_tokens': 1423947.0, 'completions/mean_length': 45.0, 'completions/min_length': 34.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.5403231978416443, 'rewards/meter/std': 0.3556986153125763, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9451793432235718, 'rewards/repeat_soft/std': 0.09756789356470108, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.22398583590984344, 'rewards/total_composite/mean': 0.5799914598464966, 'rewards/total_composite/std': 0.21093840897083282, 'reward': 0.5799914598464966, 'reward_std': 0.21093840897083282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1565788835287094, 'sampling/sampling_logp_difference/max': 1.6490745544433594, 'sampling/importance_sampling_ratio/min': 0.19222772121429443, 'sampling/importance_sampling_ratio/mean': 1.0035032033920288, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9338272884488106, 'clip_ratio/low_mean': 0.08352437056601048, 'clip_ratio/low_min': 0.08352437056601048, 'clip_ratio/high_mean': 0.07671568915247917, 'clip_ratio/high_max': 0.07671568915247917, 'clip_ratio/region_mean': 0.16024005971848965, 'reward_total_mean': 0.5799914598464966, 'reward_meter_mean': 0.5403231978416443, 'reward_meter_std': 0.3556986153125763, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9451793432235718, 'reward_repeat_soft_std': 0.09756789356470108, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.22398583590984344, 'reward_total_composite_mean': 0.5799914598464966, 'reward_total_composite_std': 0.21093840897083282, 'epoch': 0.08}

 24%|██▍       | 805/3300 [2:01:38<8:02:20, 11.60s/it]INFO 04-13 09:22:36 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 806/3300 [2:01:44<6:57:10, 10.04s/it]2026-04-13 09:22:42,519 | INFO | train_grpo_train | metrics_logged mode=train step=806

                                                      
{'loss': -0.0222, 'grad_norm': 13.966983795166016, 'learning_rate': 7.560606060606062e-06, 'num_tokens': 1425584.0, 'completions/mean_length': 43.625, 'completions/min_length': 41.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7136768102645874, 'rewards/meter/std': 0.3417411148548126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9802080392837524, 'rewards/repeat_soft/std': 0.03066585585474968, 'rewards/judge_quality/mean': 0.6399999856948853, 'rewards/judge_quality/std': 0.12224100530147552, 'rewards/total_composite/mean': 0.6391867399215698, 'rewards/total_composite/std': 0.15036608278751373, 'reward': 0.6391867399215698, 'reward_std': 0.15036608278751373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16272421181201935, 'sampling/sampling_logp_difference/max': 2.713101387023926, 'sampling/importance_sampling_ratio/min': 0.20161476731300354, 'sampling/importance_sampling_ratio/mean': 1.0206215381622314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.962073840200901, 'clip_ratio/low_mean': 0.08706512209028006, 'clip_ratio/low_min': 0.08706512209028006, 'clip_ratio/high_mean': 0.06618053466081619, 'clip_ratio/high_max': 0.06618053466081619, 'clip_ratio/region_mean': 0.15324565675109625, 'reward_total_mean': 0.6391867399215698, 'reward_meter_mean': 0.7136768102645874, 'reward_meter_std': 0.3417411148548126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9802080392837524, 'reward_repeat_soft_std': 0.03066585585474968, 'reward_judge_quality_mean': 0.6399999856948853, 'reward_judge_quality_std': 0.12224100530147552, 'reward_total_composite_mean': 0.6391867399215698, 'reward_total_composite_std': 0.15036608278751373, 'epoch': 0.08}

 24%|██▍       | 806/3300 [2:01:44<6:57:10, 10.04s/it]INFO 04-13 09:22:42 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 807/3300 [2:01:56<7:22:24, 10.65s/it]2026-04-13 09:22:54,593 | INFO | train_grpo_train | metrics_logged mode=train step=807

                                                      
{'loss': -0.0503, 'grad_norm': 4.909672737121582, 'learning_rate': 7.557575757575758e-06, 'num_tokens': 1427142.0, 'completions/mean_length': 103.75, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.42857360839844, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7983464002609253, 'rewards/meter/std': 0.28263235092163086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.985405683517456, 'rewards/repeat_soft/std': 0.010004878975450993, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.2865559160709381, 'rewards/total_composite/mean': 0.5196365118026733, 'rewards/total_composite/std': 0.36020785570144653, 'reward': 0.5196365118026733, 'reward_std': 0.36020782589912415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14581044018268585, 'sampling/sampling_logp_difference/max': 1.498145580291748, 'sampling/importance_sampling_ratio/min': 0.22354432940483093, 'sampling/importance_sampling_ratio/mean': 1.0206215381622314, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.772854670882225, 'clip_ratio/low_mean': 0.02757973689585924, 'clip_ratio/low_min': 0.02757973689585924, 'clip_ratio/high_mean': 0.09412868414074183, 'clip_ratio/high_max': 0.09412868414074183, 'clip_ratio/region_mean': 0.12170842103660107, 'reward_total_mean': 0.5196365118026733, 'reward_meter_mean': 0.7983464002609253, 'reward_meter_std': 0.28263235092163086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.985405683517456, 'reward_repeat_soft_std': 0.010004878975450993, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.2865559160709381, 'reward_total_composite_mean': 0.5196365118026733, 'reward_total_composite_std': 0.36020785570144653, 'epoch': 0.08}

 24%|██▍       | 807/3300 [2:01:56<7:22:24, 10.65s/it]INFO 04-13 09:22:54 [block_pool.py:378] Successfully reset prefix cache

 24%|██▍       | 808/3300 [2:02:03<6:39:19,  9.61s/it]2026-04-13 09:23:01,796 | INFO | train_grpo_train | metrics_logged mode=train step=808

                                                      
{'loss': -0.0245, 'grad_norm': 13.79069995880127, 'learning_rate': 7.5545454545454555e-06, 'num_tokens': 1428709.0, 'completions/mean_length': 49.875, 'completions/min_length': 45.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8329415917396545, 'rewards/meter/std': 0.34195801615715027, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9917269945144653, 'rewards/repeat_soft/std': 0.009106655605137348, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.665428876876831, 'rewards/total_composite/std': 0.19449499249458313, 'reward': 0.665428876876831, 'reward_std': 0.19449497759342194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15418878197669983, 'sampling/sampling_logp_difference/max': 5.072943687438965, 'sampling/importance_sampling_ratio/min': 0.006263953633606434, 'sampling/importance_sampling_ratio/mean': 1.0029226541519165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9787449389696121, 'clip_ratio/low_mean': 0.0927612129598856, 'clip_ratio/low_min': 0.0927612129598856, 'clip_ratio/high_mean': 0.03973354212939739, 'clip_ratio/high_max': 0.03973354212939739, 'clip_ratio/region_mean': 0.132494755089283, 'reward_total_mean': 0.665428876876831, 'reward_meter_mean': 0.8329415917396545, 'reward_meter_std': 0.34195801615715027, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9917269945144653, 'reward_repeat_soft_std': 0.009106655605137348, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.665428876876831, 'reward_total_composite_std': 0.19449499249458313, 'epoch': 0.08}

 24%|██▍       | 808/3300 [2:02:04<6:39:19,  9.61s/it]INFO 04-13 09:23:02 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 809/3300 [2:02:10<6:05:13,  8.80s/it]2026-04-13 09:23:08,686 | INFO | train_grpo_train | metrics_logged mode=train step=809

                                                      
{'loss': -0.006, 'grad_norm': 15.120176315307617, 'learning_rate': 7.551515151515152e-06, 'num_tokens': 1430435.0, 'completions/mean_length': 53.75, 'completions/min_length': 47.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7933717966079712, 'rewards/meter/std': 0.2962089776992798, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.91013103723526, 'rewards/repeat_soft/std': 0.10067904740571976, 'rewards/judge_quality/mean': 0.6200000047683716, 'rewards/judge_quality/std': 0.22677870094776154, 'rewards/total_composite/mean': 0.655008852481842, 'rewards/total_composite/std': 0.1684381663799286, 'reward': 0.655008852481842, 'reward_std': 0.1684381514787674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15119916200637817, 'sampling/sampling_logp_difference/max': 1.5658235549926758, 'sampling/importance_sampling_ratio/min': 0.20891588926315308, 'sampling/importance_sampling_ratio/mean': 1.0017062425613403, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.051535002887249, 'clip_ratio/low_mean': 0.06343050673604012, 'clip_ratio/low_min': 0.06343050673604012, 'clip_ratio/high_mean': 0.049440870992839336, 'clip_ratio/high_max': 0.049440870992839336, 'clip_ratio/region_mean': 0.11287137772887945, 'reward_total_mean': 0.655008852481842, 'reward_meter_mean': 0.7933717966079712, 'reward_meter_std': 0.2962089776992798, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.91013103723526, 'reward_repeat_soft_std': 0.10067904740571976, 'reward_judge_quality_mean': 0.6200000047683716, 'reward_judge_quality_std': 0.22677870094776154, 'reward_total_composite_mean': 0.655008852481842, 'reward_total_composite_std': 0.1684381663799286, 'epoch': 0.08}

 25%|██▍       | 809/3300 [2:02:10<6:05:13,  8.80s/it]INFO 04-13 09:23:08 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 810/3300 [2:02:18<5:47:56,  8.38s/it]2026-04-13 09:23:16,113 | INFO | train_grpo_train | metrics_logged mode=train step=810

                                                      
{'loss': -0.0396, 'grad_norm': 25.197595596313477, 'learning_rate': 7.548484848484849e-06, 'num_tokens': 1431944.0, 'completions/mean_length': 26.625, 'completions/min_length': 20.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7921427488327026, 'rewards/meter/std': 0.35519370436668396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9307898283004761, 'rewards/repeat_soft/std': 0.08015943318605423, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5676900744438171, 'rewards/total_composite/std': 0.10124175250530243, 'reward': 0.5676900744438171, 'reward_std': 0.10124175250530243, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1733849048614502, 'sampling/sampling_logp_difference/max': 1.4316306114196777, 'sampling/importance_sampling_ratio/min': 0.23891901969909668, 'sampling/importance_sampling_ratio/mean': 0.9951972961425781, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1499949097633362, 'clip_ratio/low_mean': 0.046875, 'clip_ratio/low_min': 0.046875, 'clip_ratio/high_mean': 0.1721633835695684, 'clip_ratio/high_max': 0.1721633835695684, 'clip_ratio/region_mean': 0.2190383835695684, 'reward_total_mean': 0.5676900744438171, 'reward_meter_mean': 0.7921427488327026, 'reward_meter_std': 0.35519370436668396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9307898283004761, 'reward_repeat_soft_std': 0.08015943318605423, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5676900744438171, 'reward_total_composite_std': 0.10124175250530243, 'epoch': 0.08}

 25%|██▍       | 810/3300 [2:02:18<5:47:56,  8.38s/it]INFO 04-13 09:23:16 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 811/3300 [2:02:25<5:28:46,  7.93s/it]2026-04-13 09:23:22,972 | INFO | train_grpo_train | metrics_logged mode=train step=811

                                                      
{'loss': 0.0569, 'grad_norm': 14.960481643676758, 'learning_rate': 7.545454545454546e-06, 'num_tokens': 1433724.0, 'completions/mean_length': 47.5, 'completions/min_length': 41.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.3600141406059265, 'rewards/meter/std': 0.3861323893070221, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9460911750793457, 'rewards/repeat_soft/std': 0.06058942526578903, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4455781877040863, 'rewards/total_composite/std': 0.11271742731332779, 'reward': 0.4455781877040863, 'reward_std': 0.11271742731332779, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14823615550994873, 'sampling/sampling_logp_difference/max': 1.5308938026428223, 'sampling/importance_sampling_ratio/min': 0.22016087174415588, 'sampling/importance_sampling_ratio/mean': 1.015677571296692, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8647165969014168, 'clip_ratio/low_mean': 0.10145791340619326, 'clip_ratio/low_min': 0.10145791340619326, 'clip_ratio/high_mean': 0.0511738546192646, 'clip_ratio/high_max': 0.0511738546192646, 'clip_ratio/region_mean': 0.15263176802545786, 'reward_total_mean': 0.4455781877040863, 'reward_meter_mean': 0.3600141406059265, 'reward_meter_std': 0.3861323893070221, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9460911750793457, 'reward_repeat_soft_std': 0.06058942526578903, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4455781877040863, 'reward_total_composite_std': 0.11271742731332779, 'epoch': 0.08}

 25%|██▍       | 811/3300 [2:02:25<5:28:46,  7.93s/it]INFO 04-13 09:23:23 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 812/3300 [2:02:37<6:17:43,  9.11s/it]2026-04-13 09:23:34,827 | INFO | train_grpo_train | metrics_logged mode=train step=812

                                                      
{'loss': -0.0049, 'grad_norm': 5.505806922912598, 'learning_rate': 7.542424242424244e-06, 'num_tokens': 1435525.0, 'completions/mean_length': 108.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.42857360839844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.15675875544548035, 'rewards/meter/std': 0.2787606120109558, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9964594841003418, 'rewards/repeat_soft/std': 0.004860998596996069, 'rewards/judge_quality/mean': 0.4987499713897705, 'rewards/judge_quality/std': 0.28965190052986145, 'rewards/total_composite/mean': 0.39452141523361206, 'rewards/total_composite/std': 0.07557890564203262, 'reward': 0.39452141523361206, 'reward_std': 0.07557890564203262, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1547120362520218, 'sampling/sampling_logp_difference/max': 1.3801994323730469, 'sampling/importance_sampling_ratio/min': 0.25152841210365295, 'sampling/importance_sampling_ratio/mean': 1.0177757740020752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7508349493145943, 'clip_ratio/low_mean': 0.08179171429947019, 'clip_ratio/low_min': 0.08179171429947019, 'clip_ratio/high_mean': 0.051739130169153214, 'clip_ratio/high_max': 0.051739130169153214, 'clip_ratio/region_mean': 0.1335308444686234, 'reward_total_mean': 0.39452141523361206, 'reward_meter_mean': 0.15675875544548035, 'reward_meter_std': 0.2787606120109558, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9964594841003418, 'reward_repeat_soft_std': 0.004860998596996069, 'reward_judge_quality_mean': 0.4987499713897705, 'reward_judge_quality_std': 0.28965190052986145, 'reward_total_composite_mean': 0.39452141523361206, 'reward_total_composite_std': 0.07557890564203262, 'epoch': 0.08}

 25%|██▍       | 812/3300 [2:02:37<6:17:43,  9.11s/it]INFO 04-13 09:23:35 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 813/3300 [2:02:48<6:44:18,  9.75s/it]2026-04-13 09:23:46,093 | INFO | train_grpo_train | metrics_logged mode=train step=813

                                                      
{'loss': -0.1233, 'grad_norm': 4.601381778717041, 'learning_rate': 7.53939393939394e-06, 'num_tokens': 1437221.0, 'completions/mean_length': 117.0, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.57143020629883, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.6576850414276123, 'rewards/meter/std': 0.28868722915649414, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9237185716629028, 'rewards/repeat_soft/std': 0.05879615247249603, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.23439893126487732, 'rewards/total_composite/mean': 0.5077955722808838, 'rewards/total_composite/std': 0.2563643157482147, 'reward': 0.5077955722808838, 'reward_std': 0.2563643157482147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1729813665151596, 'sampling/sampling_logp_difference/max': 3.193268060684204, 'sampling/importance_sampling_ratio/min': 0.04103753715753555, 'sampling/importance_sampling_ratio/mean': 0.9843983054161072, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6876777485013008, 'clip_ratio/low_mean': 0.04267744906246662, 'clip_ratio/low_min': 0.04267744906246662, 'clip_ratio/high_mean': 0.10265432484447956, 'clip_ratio/high_max': 0.10265432484447956, 'clip_ratio/region_mean': 0.14533177390694618, 'reward_total_mean': 0.5077955722808838, 'reward_meter_mean': 0.6576850414276123, 'reward_meter_std': 0.28868722915649414, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9237185716629028, 'reward_repeat_soft_std': 0.05879615247249603, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.23439893126487732, 'reward_total_composite_mean': 0.5077955722808838, 'reward_total_composite_std': 0.2563643157482147, 'epoch': 0.08}

 25%|██▍       | 813/3300 [2:02:48<6:44:18,  9.75s/it]INFO 04-13 09:23:46 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 814/3300 [2:02:59<7:01:19, 10.17s/it]2026-04-13 09:23:57,230 | INFO | train_grpo_train | metrics_logged mode=train step=814

                                                      
{'loss': -0.13, 'grad_norm': 3.343351364135742, 'learning_rate': 7.536363636363637e-06, 'num_tokens': 1438797.0, 'completions/mean_length': 111.0, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.71428680419922, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6222841739654541, 'rewards/meter/std': 0.3936542272567749, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9862548112869263, 'rewards/repeat_soft/std': 0.014779746532440186, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.18074746429920197, 'rewards/total_composite/mean': 0.4458679258823395, 'rewards/total_composite/std': 0.20553505420684814, 'reward': 0.4458679258823395, 'reward_std': 0.20553503930568695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18630266189575195, 'sampling/sampling_logp_difference/max': 1.6068196296691895, 'sampling/importance_sampling_ratio/min': 0.20052434504032135, 'sampling/importance_sampling_ratio/mean': 1.0044987201690674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1804257184267044, 'clip_ratio/low_mean': 0.061155762523412704, 'clip_ratio/low_min': 0.061155762523412704, 'clip_ratio/high_mean': 0.1013546446338296, 'clip_ratio/high_max': 0.1013546446338296, 'clip_ratio/region_mean': 0.1625104071572423, 'reward_total_mean': 0.4458679258823395, 'reward_meter_mean': 0.6222841739654541, 'reward_meter_std': 0.3936542272567749, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9862548112869263, 'reward_repeat_soft_std': 0.014779746532440186, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.18074746429920197, 'reward_total_composite_mean': 0.4458679258823395, 'reward_total_composite_std': 0.20553505420684814, 'epoch': 0.08}

 25%|██▍       | 814/3300 [2:02:59<7:01:19, 10.17s/it]INFO 04-13 09:23:57 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 815/3300 [2:03:05<6:11:48,  8.98s/it]2026-04-13 09:24:03,425 | INFO | train_grpo_train | metrics_logged mode=train step=815

                                                      
{'loss': 0.0498, 'grad_norm': 19.04191780090332, 'learning_rate': 7.533333333333334e-06, 'num_tokens': 1440461.0, 'completions/mean_length': 25.0, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7894392013549805, 'rewards/meter/std': 0.3342284560203552, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8165859580039978, 'rewards/repeat_soft/std': 0.12885448336601257, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/total_composite/mean': 0.7297252416610718, 'rewards/total_composite/std': 0.2254355251789093, 'reward': 0.7297252416610718, 'reward_std': 0.22543549537658691, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11707991361618042, 'sampling/sampling_logp_difference/max': 1.4313504695892334, 'sampling/importance_sampling_ratio/min': 0.23898597061634064, 'sampling/importance_sampling_ratio/mean': 0.9947558641433716, 'sampling/importance_sampling_ratio/max': 1.6051779985427856, 'entropy': 0.6842283755540848, 'clip_ratio/low_mean': 0.033491379115730524, 'clip_ratio/low_min': 0.033491379115730524, 'clip_ratio/high_mean': 0.08551282249391079, 'clip_ratio/high_max': 0.08551282249391079, 'clip_ratio/region_mean': 0.11900420160964131, 'reward_total_mean': 0.7297252416610718, 'reward_meter_mean': 0.7894392013549805, 'reward_meter_std': 0.3342284560203552, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8165859580039978, 'reward_repeat_soft_std': 0.12885448336601257, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.7297252416610718, 'reward_total_composite_std': 0.2254355251789093, 'epoch': 0.08}

 25%|██▍       | 815/3300 [2:03:05<6:11:48,  8.98s/it]INFO 04-13 09:24:03 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 816/3300 [2:03:11<5:38:10,  8.17s/it]2026-04-13 09:24:09,707 | INFO | train_grpo_train | metrics_logged mode=train step=816

                                                      
{'loss': 0.0396, 'grad_norm': 14.968425750732422, 'learning_rate': 7.530303030303031e-06, 'num_tokens': 1442114.0, 'completions/mean_length': 50.625, 'completions/min_length': 33.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.49092987179756165, 'rewards/meter/std': 0.28902894258499146, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9823921918869019, 'rewards/repeat_soft/std': 0.028371434658765793, 'rewards/judge_quality/mean': 0.5087499618530273, 'rewards/judge_quality/std': 0.16617010533809662, 'rewards/total_composite/mean': 0.5222827792167664, 'rewards/total_composite/std': 0.15004116296768188, 'reward': 0.5222827792167664, 'reward_std': 0.15004117786884308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20552361011505127, 'sampling/sampling_logp_difference/max': 1.8502120971679688, 'sampling/importance_sampling_ratio/min': 0.1572038233280182, 'sampling/importance_sampling_ratio/mean': 1.0396384000778198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.494948334991932, 'clip_ratio/low_mean': 0.10349871683865786, 'clip_ratio/low_min': 0.10349871683865786, 'clip_ratio/high_mean': 0.0664410050958395, 'clip_ratio/high_max': 0.0664410050958395, 'clip_ratio/region_mean': 0.16993972193449736, 'reward_total_mean': 0.5222827792167664, 'reward_meter_mean': 0.49092987179756165, 'reward_meter_std': 0.28902894258499146, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9823921918869019, 'reward_repeat_soft_std': 0.028371434658765793, 'reward_judge_quality_mean': 0.5087499618530273, 'reward_judge_quality_std': 0.16617010533809662, 'reward_total_composite_mean': 0.5222827792167664, 'reward_total_composite_std': 0.15004116296768188, 'epoch': 0.08}

 25%|██▍       | 816/3300 [2:03:11<5:38:10,  8.17s/it]INFO 04-13 09:24:09 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 817/3300 [2:03:18<5:22:01,  7.78s/it]2026-04-13 09:24:16,588 | INFO | train_grpo_train | metrics_logged mode=train step=817

                                                      
{'loss': 0.0807, 'grad_norm': 12.185218811035156, 'learning_rate': 7.5272727272727274e-06, 'num_tokens': 1444393.0, 'completions/mean_length': 95.875, 'completions/min_length': 76.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.875, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.7116522789001465, 'rewards/meter/std': 0.2779863178730011, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9482159614562988, 'rewards/repeat_soft/std': 0.034769054502248764, 'rewards/judge_quality/mean': 0.6450000405311584, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.6138795614242554, 'rewards/total_composite/std': 0.1565939038991928, 'reward': 0.6138795614242554, 'reward_std': 0.1565939038991928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1762741208076477, 'sampling/sampling_logp_difference/max': 2.9817864894866943, 'sampling/importance_sampling_ratio/min': 0.05070217698812485, 'sampling/importance_sampling_ratio/mean': 1.0143365859985352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0195010900497437, 'clip_ratio/low_mean': 0.103886135853827, 'clip_ratio/low_min': 0.103886135853827, 'clip_ratio/high_mean': 0.049659901298582554, 'clip_ratio/high_max': 0.049659901298582554, 'clip_ratio/region_mean': 0.15354603715240955, 'reward_total_mean': 0.6138795614242554, 'reward_meter_mean': 0.7116522789001465, 'reward_meter_std': 0.2779863178730011, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9482159614562988, 'reward_repeat_soft_std': 0.034769054502248764, 'reward_judge_quality_mean': 0.6450000405311584, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.6138795614242554, 'reward_total_composite_std': 0.1565939038991928, 'epoch': 0.08}

 25%|██▍       | 817/3300 [2:03:18<5:22:01,  7.78s/it]INFO 04-13 09:24:16 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 818/3300 [2:03:24<5:02:08,  7.30s/it]2026-04-13 09:24:22,778 | INFO | train_grpo_train | metrics_logged mode=train step=818

                                                      
{'loss': 0.1043, 'grad_norm': 16.24044418334961, 'learning_rate': 7.524242424242425e-06, 'num_tokens': 1446016.0, 'completions/mean_length': 44.875, 'completions/min_length': 40.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.42738354206085205, 'rewards/meter/std': 0.2925856113433838, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9781241416931152, 'rewards/repeat_soft/std': 0.01897178590297699, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/total_composite/mean': 0.5775432586669922, 'rewards/total_composite/std': 0.1707666516304016, 'reward': 0.5775432586669922, 'reward_std': 0.17076663672924042, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20095308125019073, 'sampling/sampling_logp_difference/max': 2.370021343231201, 'sampling/importance_sampling_ratio/min': 0.0934787318110466, 'sampling/importance_sampling_ratio/mean': 1.0148979425430298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1817000210285187, 'clip_ratio/low_mean': 0.13013580441474915, 'clip_ratio/low_min': 0.13013580441474915, 'clip_ratio/high_mean': 0.07315891608595848, 'clip_ratio/high_max': 0.07315891608595848, 'clip_ratio/region_mean': 0.20329472050070763, 'reward_total_mean': 0.5775432586669922, 'reward_meter_mean': 0.42738354206085205, 'reward_meter_std': 0.2925856113433838, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9781241416931152, 'reward_repeat_soft_std': 0.01897178590297699, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.5775432586669922, 'reward_total_composite_std': 0.1707666516304016, 'epoch': 0.08}

 25%|██▍       | 818/3300 [2:03:25<5:02:08,  7.30s/it]INFO 04-13 09:24:23 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 819/3300 [2:03:36<5:48:52,  8.44s/it]2026-04-13 09:24:33,857 | INFO | train_grpo_train | metrics_logged mode=train step=819

                                                      
{'loss': -0.1303, 'grad_norm': 3.4475998878479004, 'learning_rate': 7.521212121212121e-06, 'num_tokens': 1447870.0, 'completions/mean_length': 181.75, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 71.66667175292969, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.5192000865936279, 'rewards/meter/std': 0.3829488754272461, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9459807872772217, 'rewards/repeat_soft/std': 0.02395819127559662, 'rewards/judge_quality/mean': 0.45250001549720764, 'rewards/judge_quality/std': 0.33065953850746155, 'rewards/total_composite/mean': 0.4163227677345276, 'rewards/total_composite/std': 0.2916165888309479, 'reward': 0.4163227677345276, 'reward_std': 0.2916165888309479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1499626338481903, 'sampling/sampling_logp_difference/max': 2.4193878173828125, 'sampling/importance_sampling_ratio/min': 0.08897607028484344, 'sampling/importance_sampling_ratio/mean': 1.01922607421875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6223423704504967, 'clip_ratio/low_mean': 0.03234265837818384, 'clip_ratio/low_min': 0.03234265837818384, 'clip_ratio/high_mean': 0.06764488946646452, 'clip_ratio/high_max': 0.06764488946646452, 'clip_ratio/region_mean': 0.09998754784464836, 'reward_total_mean': 0.4163227677345276, 'reward_meter_mean': 0.5192000865936279, 'reward_meter_std': 0.3829488754272461, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9459807872772217, 'reward_repeat_soft_std': 0.02395819127559662, 'reward_judge_quality_mean': 0.45250001549720764, 'reward_judge_quality_std': 0.33065953850746155, 'reward_total_composite_mean': 0.4163227677345276, 'reward_total_composite_std': 0.2916165888309479, 'epoch': 0.08}

 25%|██▍       | 819/3300 [2:03:36<5:48:52,  8.44s/it]INFO 04-13 09:24:34 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 820/3300 [2:03:47<6:22:40,  9.26s/it]2026-04-13 09:24:45,031 | INFO | train_grpo_train | metrics_logged mode=train step=820

                                                      
{'loss': -0.1359, 'grad_norm': 3.6564669609069824, 'learning_rate': 7.518181818181819e-06, 'num_tokens': 1450045.0, 'completions/mean_length': 149.875, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 98.14286041259766, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.6863880157470703, 'rewards/meter/std': 0.37522250413894653, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8512330651283264, 'rewards/repeat_soft/std': 0.11977135390043259, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.254611074924469, 'rewards/total_composite/mean': 0.5125121474266052, 'rewards/total_composite/std': 0.2600172162055969, 'reward': 0.5125121474266052, 'reward_std': 0.2600172162055969, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14205388724803925, 'sampling/sampling_logp_difference/max': 3.338271141052246, 'sampling/importance_sampling_ratio/min': 0.035498276352882385, 'sampling/importance_sampling_ratio/mean': 0.9941490888595581, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7078659161925316, 'clip_ratio/low_mean': 0.03297199681401253, 'clip_ratio/low_min': 0.03297199681401253, 'clip_ratio/high_mean': 0.09131707530468702, 'clip_ratio/high_max': 0.09131707530468702, 'clip_ratio/region_mean': 0.12428907211869955, 'reward_total_mean': 0.5125121474266052, 'reward_meter_mean': 0.6863880157470703, 'reward_meter_std': 0.37522250413894653, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8512330651283264, 'reward_repeat_soft_std': 0.11977135390043259, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.254611074924469, 'reward_total_composite_mean': 0.5125121474266052, 'reward_total_composite_std': 0.2600172162055969, 'epoch': 0.08}

 25%|██▍       | 820/3300 [2:03:47<6:22:40,  9.26s/it]INFO 04-13 09:24:45 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 821/3300 [2:03:53<5:50:18,  8.48s/it]2026-04-13 09:24:51,693 | INFO | train_grpo_train | metrics_logged mode=train step=821

                                                      
{'loss': -0.0201, 'grad_norm': 9.674942970275879, 'learning_rate': 7.515151515151516e-06, 'num_tokens': 1452235.0, 'completions/mean_length': 90.75, 'completions/min_length': 83.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.75, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.4520382881164551, 'rewards/meter/std': 0.19169962406158447, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9560103416442871, 'rewards/repeat_soft/std': 0.05141996592283249, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.47455012798309326, 'rewards/total_composite/std': 0.06853678822517395, 'reward': 0.47455012798309326, 'reward_std': 0.06853678077459335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15331177413463593, 'sampling/sampling_logp_difference/max': 1.9285740852355957, 'sampling/importance_sampling_ratio/min': 0.14535531401634216, 'sampling/importance_sampling_ratio/mean': 1.0141502618789673, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8684991300106049, 'clip_ratio/low_mean': 0.09285540506243706, 'clip_ratio/low_min': 0.09285540506243706, 'clip_ratio/high_mean': 0.07472315896302462, 'clip_ratio/high_max': 0.07472315896302462, 'clip_ratio/region_mean': 0.16757856402546167, 'reward_total_mean': 0.47455012798309326, 'reward_meter_mean': 0.4520382881164551, 'reward_meter_std': 0.19169962406158447, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9560103416442871, 'reward_repeat_soft_std': 0.05141996592283249, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.47455012798309326, 'reward_total_composite_std': 0.06853678822517395, 'epoch': 0.08}

 25%|██▍       | 821/3300 [2:03:53<5:50:18,  8.48s/it]INFO 04-13 09:24:51 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 822/3300 [2:04:00<5:24:36,  7.86s/it]2026-04-13 09:24:58,113 | INFO | train_grpo_train | metrics_logged mode=train step=822

                                                      
{'loss': 0.0216, 'grad_norm': 19.24358558654785, 'learning_rate': 7.512121212121213e-06, 'num_tokens': 1453858.0, 'completions/mean_length': 28.875, 'completions/min_length': 23.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7364631295204163, 'rewards/meter/std': 0.40704989433288574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9546753168106079, 'rewards/repeat_soft/std': 0.013788825832307339, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5358984470367432, 'rewards/total_composite/std': 0.11425299942493439, 'reward': 0.5358984470367432, 'reward_std': 0.11425299197435379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20006316900253296, 'sampling/sampling_logp_difference/max': 1.5987379550933838, 'sampling/importance_sampling_ratio/min': 0.20215147733688354, 'sampling/importance_sampling_ratio/mean': 1.0060007572174072, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3630374073982239, 'clip_ratio/low_mean': 0.05968630267307162, 'clip_ratio/low_min': 0.05968630267307162, 'clip_ratio/high_mean': 0.15672273561358452, 'clip_ratio/high_max': 0.15672273561358452, 'clip_ratio/region_mean': 0.21640903828665614, 'reward_total_mean': 0.5358984470367432, 'reward_meter_mean': 0.7364631295204163, 'reward_meter_std': 0.40704989433288574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9546753168106079, 'reward_repeat_soft_std': 0.013788825832307339, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5358984470367432, 'reward_total_composite_std': 0.11425299942493439, 'epoch': 0.08}

 25%|██▍       | 822/3300 [2:04:00<5:24:36,  7.86s/it]INFO 04-13 09:24:58 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 823/3300 [2:04:07<5:10:33,  7.52s/it]2026-04-13 09:25:04,837 | INFO | train_grpo_train | metrics_logged mode=train step=823

                                                      
{'loss': 0.0574, 'grad_norm': 9.428529739379883, 'learning_rate': 7.509090909090909e-06, 'num_tokens': 1456342.0, 'completions/mean_length': 98.5, 'completions/min_length': 85.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.5, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.791350245475769, 'rewards/meter/std': 0.2780003249645233, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9728424549102783, 'rewards/repeat_soft/std': 0.011104121804237366, 'rewards/judge_quality/mean': 0.7237499952316284, 'rewards/judge_quality/std': 0.1418185532093048, 'rewards/total_composite/mean': 0.6681636571884155, 'rewards/total_composite/std': 0.12867885828018188, 'reward': 0.6681636571884155, 'reward_std': 0.12867885828018188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12085013091564178, 'sampling/sampling_logp_difference/max': 1.7185745239257812, 'sampling/importance_sampling_ratio/min': 0.1881616860628128, 'sampling/importance_sampling_ratio/mean': 1.0084378719329834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5888826176524162, 'clip_ratio/low_mean': 0.044817473739385605, 'clip_ratio/low_min': 0.044817473739385605, 'clip_ratio/high_mean': 0.06585449632257223, 'clip_ratio/high_max': 0.06585449632257223, 'clip_ratio/region_mean': 0.11067197006195784, 'reward_total_mean': 0.6681636571884155, 'reward_meter_mean': 0.791350245475769, 'reward_meter_std': 0.2780003249645233, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9728424549102783, 'reward_repeat_soft_std': 0.011104121804237366, 'reward_judge_quality_mean': 0.7237499952316284, 'reward_judge_quality_std': 0.1418185532093048, 'reward_total_composite_mean': 0.6681636571884155, 'reward_total_composite_std': 0.12867885828018188, 'epoch': 0.08}

 25%|██▍       | 823/3300 [2:04:07<5:10:33,  7.52s/it]INFO 04-13 09:25:05 [block_pool.py:378] Successfully reset prefix cache

 25%|██▍       | 824/3300 [2:04:13<4:59:09,  7.25s/it]2026-04-13 09:25:11,448 | INFO | train_grpo_train | metrics_logged mode=train step=824

                                                      
{'loss': 0.0782, 'grad_norm': 18.671138763427734, 'learning_rate': 7.5060606060606065e-06, 'num_tokens': 1458085.0, 'completions/mean_length': 37.875, 'completions/min_length': 28.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6317110061645508, 'rewards/meter/std': 0.4083610773086548, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9670195579528809, 'rewards/repeat_soft/std': 0.04772603139281273, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.645653486251831, 'rewards/total_composite/std': 0.2172938585281372, 'reward': 0.645653486251831, 'reward_std': 0.2172938585281372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16183391213417053, 'sampling/sampling_logp_difference/max': 1.4298913478851318, 'sampling/importance_sampling_ratio/min': 0.23933492600917816, 'sampling/importance_sampling_ratio/mean': 1.0175118446350098, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7881267480552197, 'clip_ratio/low_mean': 0.05909091094508767, 'clip_ratio/low_min': 0.05909091094508767, 'clip_ratio/high_mean': 0.08884947653859854, 'clip_ratio/high_max': 0.08884947653859854, 'clip_ratio/region_mean': 0.1479403874836862, 'reward_total_mean': 0.645653486251831, 'reward_meter_mean': 0.6317110061645508, 'reward_meter_std': 0.4083610773086548, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9670195579528809, 'reward_repeat_soft_std': 0.04772603139281273, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.645653486251831, 'reward_total_composite_std': 0.2172938585281372, 'epoch': 0.08}

 25%|██▍       | 824/3300 [2:04:13<4:59:09,  7.25s/it]INFO 04-13 09:25:11 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 825/3300 [2:04:25<5:53:15,  8.56s/it]2026-04-13 09:25:23,081 | INFO | train_grpo_train | metrics_logged mode=train step=825

                                                      
{'loss': -0.1015, 'grad_norm': 3.5623085498809814, 'learning_rate': 7.503030303030303e-06, 'num_tokens': 1459880.0, 'completions/mean_length': 107.375, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.57143020629883, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5495851635932922, 'rewards/meter/std': 0.44300636649131775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9228882193565369, 'rewards/repeat_soft/std': 0.07558905333280563, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.22947145998477936, 'rewards/total_composite/mean': 0.49460169672966003, 'rewards/total_composite/std': 0.2527218461036682, 'reward': 0.49460169672966003, 'reward_std': 0.2527218461036682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14410503208637238, 'sampling/sampling_logp_difference/max': 1.8627885580062866, 'sampling/importance_sampling_ratio/min': 0.15523913502693176, 'sampling/importance_sampling_ratio/mean': 1.0030133724212646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6821320280432701, 'clip_ratio/low_mean': 0.03609848441556096, 'clip_ratio/low_min': 0.03609848441556096, 'clip_ratio/high_mean': 0.07457344699651003, 'clip_ratio/high_max': 0.07457344699651003, 'clip_ratio/region_mean': 0.11067193141207099, 'reward_total_mean': 0.49460169672966003, 'reward_meter_mean': 0.5495851635932922, 'reward_meter_std': 0.44300636649131775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9228882193565369, 'reward_repeat_soft_std': 0.07558905333280563, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.22947145998477936, 'reward_total_composite_mean': 0.49460169672966003, 'reward_total_composite_std': 0.2527218461036682, 'epoch': 0.08}

 25%|██▌       | 825/3300 [2:04:25<5:53:15,  8.56s/it]INFO 04-13 09:25:23 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 826/3300 [2:04:36<6:27:52,  9.41s/it]2026-04-13 09:25:34,458 | INFO | train_grpo_train | metrics_logged mode=train step=826

                                                      
{'loss': -0.0448, 'grad_norm': 126.64268493652344, 'learning_rate': 7.500000000000001e-06, 'num_tokens': 1461737.0, 'completions/mean_length': 127.125, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 72.14286041259766, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.8361838459968567, 'rewards/meter/std': 0.20359478890895844, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.24800792336463928, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9886918663978577, 'rewards/repeat_soft/std': 0.00916048139333725, 'rewards/judge_quality/mean': 0.4687500298023224, 'rewards/judge_quality/std': 0.19037088751792908, 'rewards/total_composite/mean': 0.5030023455619812, 'rewards/total_composite/std': 0.25660642981529236, 'reward': 0.5030023455619812, 'reward_std': 0.25660642981529236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12969906628131866, 'sampling/sampling_logp_difference/max': 1.141408920288086, 'sampling/importance_sampling_ratio/min': 0.31936874985694885, 'sampling/importance_sampling_ratio/mean': 1.0196263790130615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7928005829453468, 'clip_ratio/low_mean': 0.03452381119132042, 'clip_ratio/low_min': 0.03452381119132042, 'clip_ratio/high_mean': 0.08575281407684088, 'clip_ratio/high_max': 0.08575281407684088, 'clip_ratio/region_mean': 0.1202766252681613, 'reward_total_mean': 0.5030023455619812, 'reward_meter_mean': 0.8361838459968567, 'reward_meter_std': 0.20359478890895844, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.24800792336463928, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9886918663978577, 'reward_repeat_soft_std': 0.00916048139333725, 'reward_judge_quality_mean': 0.4687500298023224, 'reward_judge_quality_std': 0.19037088751792908, 'reward_total_composite_mean': 0.5030023455619812, 'reward_total_composite_std': 0.25660642981529236, 'epoch': 0.08}

 25%|██▌       | 826/3300 [2:04:36<6:27:52,  9.41s/it]INFO 04-13 09:25:34 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 827/3300 [2:04:48<6:54:55, 10.07s/it]2026-04-13 09:25:46,070 | INFO | train_grpo_train | metrics_logged mode=train step=827

                                                      
{'loss': -0.1233, 'grad_norm': 3.053903818130493, 'learning_rate': 7.496969696969698e-06, 'num_tokens': 1463335.0, 'completions/mean_length': 113.75, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.857147216796875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7329822182655334, 'rewards/meter/std': 0.3762177526950836, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8214617967605591, 'rewards/repeat_soft/std': 0.09373597055673599, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.14327171444892883, 'rewards/total_composite/mean': 0.426153302192688, 'rewards/total_composite/std': 0.19938506186008453, 'reward': 0.426153302192688, 'reward_std': 0.19938506186008453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16028787195682526, 'sampling/sampling_logp_difference/max': 1.5074095726013184, 'sampling/importance_sampling_ratio/min': 0.22148297727108002, 'sampling/importance_sampling_ratio/mean': 1.0096970796585083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1295492947101593, 'clip_ratio/low_mean': 0.021753935143351555, 'clip_ratio/low_min': 0.021753935143351555, 'clip_ratio/high_mean': 0.0876676170155406, 'clip_ratio/high_max': 0.0876676170155406, 'clip_ratio/region_mean': 0.10942155215889215, 'reward_total_mean': 0.426153302192688, 'reward_meter_mean': 0.7329822182655334, 'reward_meter_std': 0.3762177526950836, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8214617967605591, 'reward_repeat_soft_std': 0.09373597055673599, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.14327171444892883, 'reward_total_composite_mean': 0.426153302192688, 'reward_total_composite_std': 0.19938506186008453, 'epoch': 0.08}

 25%|██▌       | 827/3300 [2:04:48<6:54:55, 10.07s/it]INFO 04-13 09:25:46 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 828/3300 [2:04:55<6:18:36,  9.19s/it]2026-04-13 09:25:53,212 | INFO | train_grpo_train | metrics_logged mode=train step=828

                                                      
{'loss': 0.096, 'grad_norm': 12.640822410583496, 'learning_rate': 7.493939393939395e-06, 'num_tokens': 1465031.0, 'completions/mean_length': 54.0, 'completions/min_length': 37.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5518558025360107, 'rewards/meter/std': 0.38216036558151245, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9843226671218872, 'rewards/repeat_soft/std': 0.0100281722843647, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6428855657577515, 'rewards/total_composite/std': 0.24778501689434052, 'reward': 0.6428855657577515, 'reward_std': 0.24778501689434052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1674165427684784, 'sampling/sampling_logp_difference/max': 2.5901987552642822, 'sampling/importance_sampling_ratio/min': 0.07500512897968292, 'sampling/importance_sampling_ratio/mean': 1.0255234241485596, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2542723193764687, 'clip_ratio/low_mean': 0.09152952954173088, 'clip_ratio/low_min': 0.09152952954173088, 'clip_ratio/high_mean': 0.07294967584311962, 'clip_ratio/high_max': 0.07294967584311962, 'clip_ratio/region_mean': 0.1644792053848505, 'reward_total_mean': 0.6428855657577515, 'reward_meter_mean': 0.5518558025360107, 'reward_meter_std': 0.38216036558151245, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9843226671218872, 'reward_repeat_soft_std': 0.0100281722843647, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6428855657577515, 'reward_total_composite_std': 0.24778501689434052, 'epoch': 0.08}

 25%|██▌       | 828/3300 [2:04:55<6:18:36,  9.19s/it]INFO 04-13 09:25:53 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 829/3300 [2:05:06<6:46:25,  9.87s/it]2026-04-13 09:26:04,670 | INFO | train_grpo_train | metrics_logged mode=train step=829

                                                      
{'loss': -0.0929, 'grad_norm': 3.071845054626465, 'learning_rate': 7.490909090909092e-06, 'num_tokens': 1466437.0, 'completions/mean_length': 91.75, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.71428680419922, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8275455236434937, 'rewards/meter/std': 0.33318546414375305, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9912638664245605, 'rewards/repeat_soft/std': 0.01396881602704525, 'rewards/judge_quality/mean': 0.5437500476837158, 'rewards/judge_quality/std': 0.29861289262771606, 'rewards/total_composite/mean': 0.6351829767227173, 'rewards/total_composite/std': 0.2909258306026459, 'reward': 0.6351829767227173, 'reward_std': 0.2909258008003235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1065325066447258, 'sampling/sampling_logp_difference/max': 1.6565746068954468, 'sampling/importance_sampling_ratio/min': 0.19079139828681946, 'sampling/importance_sampling_ratio/mean': 1.0092597007751465, 'sampling/importance_sampling_ratio/max': 1.9726223945617676, 'entropy': 0.39859044924378395, 'clip_ratio/low_mean': 0.04791666753590107, 'clip_ratio/low_min': 0.04791666753590107, 'clip_ratio/high_mean': 0.034446023404598236, 'clip_ratio/high_max': 0.034446023404598236, 'clip_ratio/region_mean': 0.0823626909404993, 'reward_total_mean': 0.6351829767227173, 'reward_meter_mean': 0.8275455236434937, 'reward_meter_std': 0.33318546414375305, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9912638664245605, 'reward_repeat_soft_std': 0.01396881602704525, 'reward_judge_quality_mean': 0.5437500476837158, 'reward_judge_quality_std': 0.29861289262771606, 'reward_total_composite_mean': 0.6351829767227173, 'reward_total_composite_std': 0.2909258306026459, 'epoch': 0.08}

 25%|██▌       | 829/3300 [2:05:06<6:46:25,  9.87s/it]INFO 04-13 09:26:04 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 830/3300 [2:05:14<6:20:19,  9.24s/it]2026-04-13 09:26:12,433 | INFO | train_grpo_train | metrics_logged mode=train step=830

                                                      
{'loss': 0.063, 'grad_norm': 9.040604591369629, 'learning_rate': 7.487878787878788e-06, 'num_tokens': 1468904.0, 'completions/mean_length': 107.375, 'completions/min_length': 102.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.375, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9646186232566833, 'rewards/meter/std': 0.043388731777668, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8325355052947998, 'rewards/repeat_soft/std': 0.14555595815181732, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.548221230506897, 'rewards/total_composite/std': 0.020647374913096428, 'reward': 0.548221230506897, 'reward_std': 0.020647387951612473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13437898457050323, 'sampling/sampling_logp_difference/max': 3.2449071407318115, 'sampling/importance_sampling_ratio/min': 0.03897218033671379, 'sampling/importance_sampling_ratio/mean': 0.9994519352912903, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7556704059243202, 'clip_ratio/low_mean': 0.04775570100173354, 'clip_ratio/low_min': 0.04775570100173354, 'clip_ratio/high_mean': 0.08244301937520504, 'clip_ratio/high_max': 0.08244301937520504, 'clip_ratio/region_mean': 0.13019872037693858, 'reward_total_mean': 0.548221230506897, 'reward_meter_mean': 0.9646186232566833, 'reward_meter_std': 0.043388731777668, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8325355052947998, 'reward_repeat_soft_std': 0.14555595815181732, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.548221230506897, 'reward_total_composite_std': 0.020647374913096428, 'epoch': 0.08}

 25%|██▌       | 830/3300 [2:05:14<6:20:19,  9.24s/it]INFO 04-13 09:26:12 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 831/3300 [2:05:21<5:49:23,  8.49s/it]2026-04-13 09:26:19,169 | INFO | train_grpo_train | metrics_logged mode=train step=831

                                                      
{'loss': 0.0236, 'grad_norm': 10.653721809387207, 'learning_rate': 7.484848484848486e-06, 'num_tokens': 1470919.0, 'completions/mean_length': 71.875, 'completions/min_length': 54.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.5295940637588501, 'rewards/meter/std': 0.3171357214450836, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9773417115211487, 'rewards/repeat_soft/std': 0.015895428135991096, 'rewards/judge_quality/mean': 0.6737499833106995, 'rewards/judge_quality/std': 0.21573381125926971, 'rewards/total_composite/mean': 0.5672909021377563, 'rewards/total_composite/std': 0.13962553441524506, 'reward': 0.5672909021377563, 'reward_std': 0.13962554931640625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15307047963142395, 'sampling/sampling_logp_difference/max': 1.9560296535491943, 'sampling/importance_sampling_ratio/min': 0.15023307502269745, 'sampling/importance_sampling_ratio/mean': 1.0319684743881226, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8835284784436226, 'clip_ratio/low_mean': 0.06878079008311033, 'clip_ratio/low_min': 0.06878079008311033, 'clip_ratio/high_mean': 0.07038584724068642, 'clip_ratio/high_max': 0.07038584724068642, 'clip_ratio/region_mean': 0.13916663732379675, 'reward_total_mean': 0.5672909021377563, 'reward_meter_mean': 0.5295940637588501, 'reward_meter_std': 0.3171357214450836, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9773417115211487, 'reward_repeat_soft_std': 0.015895428135991096, 'reward_judge_quality_mean': 0.6737499833106995, 'reward_judge_quality_std': 0.21573381125926971, 'reward_total_composite_mean': 0.5672909021377563, 'reward_total_composite_std': 0.13962553441524506, 'epoch': 0.08}

 25%|██▌       | 831/3300 [2:05:21<5:49:23,  8.49s/it]INFO 04-13 09:26:19 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 832/3300 [2:05:27<5:15:19,  7.67s/it]2026-04-13 09:26:24,914 | INFO | train_grpo_train | metrics_logged mode=train step=832

                                                      
{'loss': 0.0711, 'grad_norm': 17.475072860717773, 'learning_rate': 7.481818181818182e-06, 'num_tokens': 1472318.0, 'completions/mean_length': 24.875, 'completions/min_length': 17.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.643103837966919, 'rewards/meter/std': 0.37102043628692627, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.919399619102478, 'rewards/repeat_soft/std': 0.05687469244003296, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5491907596588135, 'rewards/total_composite/std': 0.12342143803834915, 'reward': 0.5491907596588135, 'reward_std': 0.12342143803834915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18710698187351227, 'sampling/sampling_logp_difference/max': 2.4872217178344727, 'sampling/importance_sampling_ratio/min': 0.08314063400030136, 'sampling/importance_sampling_ratio/mean': 1.0076881647109985, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0633917525410652, 'clip_ratio/low_mean': 0.06078296899795532, 'clip_ratio/low_min': 0.06078296899795532, 'clip_ratio/high_mean': 0.12785907462239265, 'clip_ratio/high_max': 0.12785907462239265, 'clip_ratio/region_mean': 0.18864204362034798, 'reward_total_mean': 0.5491907596588135, 'reward_meter_mean': 0.643103837966919, 'reward_meter_std': 0.37102043628692627, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.919399619102478, 'reward_repeat_soft_std': 0.05687469244003296, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5491907596588135, 'reward_total_composite_std': 0.12342143803834915, 'epoch': 0.08}

 25%|██▌       | 832/3300 [2:05:27<5:15:19,  7.67s/it]INFO 04-13 09:26:25 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 833/3300 [2:05:38<6:01:24,  8.79s/it]2026-04-13 09:26:36,325 | INFO | train_grpo_train | metrics_logged mode=train step=833

                                                      
{'loss': -0.1675, 'grad_norm': 2.430192470550537, 'learning_rate': 7.47878787878788e-06, 'num_tokens': 1474137.0, 'completions/mean_length': 189.375, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 81.83333587646484, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.8023124933242798, 'rewards/meter/std': 0.31727367639541626, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.24775780737400055, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8893700242042542, 'rewards/repeat_soft/std': 0.09991069883108139, 'rewards/judge_quality/mean': 0.3137499988079071, 'rewards/judge_quality/std': 0.17492344975471497, 'rewards/total_composite/mean': 0.402348130941391, 'rewards/total_composite/std': 0.25113871693611145, 'reward': 0.402348130941391, 'reward_std': 0.25113871693611145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13309930264949799, 'sampling/sampling_logp_difference/max': 3.3386030197143555, 'sampling/importance_sampling_ratio/min': 0.03548649698495865, 'sampling/importance_sampling_ratio/mean': 1.0186488628387451, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5958703383803368, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09196088742464781, 'clip_ratio/high_max': 0.09196088742464781, 'clip_ratio/region_mean': 0.09196088742464781, 'reward_total_mean': 0.402348130941391, 'reward_meter_mean': 0.8023124933242798, 'reward_meter_std': 0.31727367639541626, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.24775780737400055, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8893700242042542, 'reward_repeat_soft_std': 0.09991069883108139, 'reward_judge_quality_mean': 0.3137499988079071, 'reward_judge_quality_std': 0.17492344975471497, 'reward_total_composite_mean': 0.402348130941391, 'reward_total_composite_std': 0.25113871693611145, 'epoch': 0.08}

 25%|██▌       | 833/3300 [2:05:38<6:01:24,  8.79s/it]INFO 04-13 09:26:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 834/3300 [2:05:45<5:40:19,  8.28s/it]2026-04-13 09:26:43,420 | INFO | train_grpo_train | metrics_logged mode=train step=834

                                                      
{'loss': 0.0629, 'grad_norm': 11.07868480682373, 'learning_rate': 7.4757575757575765e-06, 'num_tokens': 1476039.0, 'completions/mean_length': 61.75, 'completions/min_length': 54.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.8380767107009888, 'rewards/meter/std': 0.2552626430988312, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9660685658454895, 'rewards/repeat_soft/std': 0.020650943741202354, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.5503005981445312, 'rewards/total_composite/std': 0.1013234481215477, 'reward': 0.5503005981445312, 'reward_std': 0.1013234481215477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15240879356861115, 'sampling/sampling_logp_difference/max': 2.377470016479492, 'sampling/importance_sampling_ratio/min': 0.09278502315282822, 'sampling/importance_sampling_ratio/mean': 0.9974220991134644, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8167963847517967, 'clip_ratio/low_mean': 0.0525423726066947, 'clip_ratio/low_min': 0.0525423726066947, 'clip_ratio/high_mean': 0.08835308533161879, 'clip_ratio/high_max': 0.08835308533161879, 'clip_ratio/region_mean': 0.14089545793831348, 'reward_total_mean': 0.5503005981445312, 'reward_meter_mean': 0.8380767107009888, 'reward_meter_std': 0.2552626430988312, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9660685658454895, 'reward_repeat_soft_std': 0.020650943741202354, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.5503005981445312, 'reward_total_composite_std': 0.1013234481215477, 'epoch': 0.08}

 25%|██▌       | 834/3300 [2:05:45<5:40:19,  8.28s/it]INFO 04-13 09:26:43 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 835/3300 [2:05:52<5:21:54,  7.84s/it]2026-04-13 09:26:50,217 | INFO | train_grpo_train | metrics_logged mode=train step=835

                                                      
{'loss': 0.0691, 'grad_norm': 9.015896797180176, 'learning_rate': 7.472727272727274e-06, 'num_tokens': 1478109.0, 'completions/mean_length': 90.75, 'completions/min_length': 79.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.75, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.4803101718425751, 'rewards/meter/std': 0.31699150800704956, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7889866232872009, 'rewards/repeat_soft/std': 0.13812139630317688, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.42656323313713074, 'rewards/total_composite/std': 0.10647182166576385, 'reward': 0.42656323313713074, 'reward_std': 0.10647180676460266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13574816286563873, 'sampling/sampling_logp_difference/max': 1.7294883728027344, 'sampling/importance_sampling_ratio/min': 0.17737513780593872, 'sampling/importance_sampling_ratio/mean': 0.9962683916091919, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6161966919898987, 'clip_ratio/low_mean': 0.06651968089863658, 'clip_ratio/low_min': 0.06651968089863658, 'clip_ratio/high_mean': 0.04569953680038452, 'clip_ratio/high_max': 0.04569953680038452, 'clip_ratio/region_mean': 0.1122192176990211, 'reward_total_mean': 0.42656323313713074, 'reward_meter_mean': 0.4803101718425751, 'reward_meter_std': 0.31699150800704956, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7889866232872009, 'reward_repeat_soft_std': 0.13812139630317688, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.42656323313713074, 'reward_total_composite_std': 0.10647182166576385, 'epoch': 0.08}

 25%|██▌       | 835/3300 [2:05:52<5:21:54,  7.84s/it]INFO 04-13 09:26:50 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 836/3300 [2:05:58<4:59:06,  7.28s/it]2026-04-13 09:26:56,210 | INFO | train_grpo_train | metrics_logged mode=train step=836

                                                      
{'loss': 0.0204, 'grad_norm': 15.014250755310059, 'learning_rate': 7.46969696969697e-06, 'num_tokens': 1479620.0, 'completions/mean_length': 40.875, 'completions/min_length': 36.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.24300700426101685, 'rewards/meter/std': 0.28373533487319946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9927163124084473, 'rewards/repeat_soft/std': 0.01412123627960682, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.4226791262626648, 'rewards/total_composite/std': 0.07671697437763214, 'reward': 0.4226791262626648, 'reward_std': 0.07671696692705154, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15632516145706177, 'sampling/sampling_logp_difference/max': 2.2316935062408447, 'sampling/importance_sampling_ratio/min': 0.10734648257493973, 'sampling/importance_sampling_ratio/mean': 0.9939863681793213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5575226359069347, 'clip_ratio/low_mean': 0.09041997790336609, 'clip_ratio/low_min': 0.09041997790336609, 'clip_ratio/high_mean': 0.05626385845243931, 'clip_ratio/high_max': 0.05626385845243931, 'clip_ratio/region_mean': 0.1466838363558054, 'reward_total_mean': 0.4226791262626648, 'reward_meter_mean': 0.24300700426101685, 'reward_meter_std': 0.28373533487319946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9927163124084473, 'reward_repeat_soft_std': 0.01412123627960682, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.4226791262626648, 'reward_total_composite_std': 0.07671697437763214, 'epoch': 0.08}

 25%|██▌       | 836/3300 [2:05:58<4:59:06,  7.28s/it]INFO 04-13 09:26:56 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 837/3300 [2:06:09<5:46:26,  8.44s/it]2026-04-13 09:27:07,345 | INFO | train_grpo_train | metrics_logged mode=train step=837

                                                      
{'loss': -0.0744, 'grad_norm': 2.8593173027038574, 'learning_rate': 7.4666666666666675e-06, 'num_tokens': 1481119.0, 'completions/mean_length': 86.375, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.571430206298828, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.43489348888397217, 'rewards/meter/std': 0.44809314608573914, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9363674521446228, 'rewards/repeat_soft/std': 0.042148202657699585, 'rewards/judge_quality/mean': 0.3712499737739563, 'rewards/judge_quality/std': 0.1470119059085846, 'rewards/total_composite/mean': 0.4052889943122864, 'rewards/total_composite/std': 0.19342483580112457, 'reward': 0.4052889943122864, 'reward_std': 0.19342480599880219, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18456131219863892, 'sampling/sampling_logp_difference/max': 1.0080249309539795, 'sampling/importance_sampling_ratio/min': 0.36493903398513794, 'sampling/importance_sampling_ratio/mean': 1.015906572341919, 'sampling/importance_sampling_ratio/max': 1.7779897451400757, 'entropy': 1.6088326200842857, 'clip_ratio/low_mean': 0.03570930939167738, 'clip_ratio/low_min': 0.03570930939167738, 'clip_ratio/high_mean': 0.09555131569504738, 'clip_ratio/high_max': 0.09555131569504738, 'clip_ratio/region_mean': 0.13126062508672476, 'reward_total_mean': 0.4052889943122864, 'reward_meter_mean': 0.43489348888397217, 'reward_meter_std': 0.44809314608573914, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9363674521446228, 'reward_repeat_soft_std': 0.042148202657699585, 'reward_judge_quality_mean': 0.3712499737739563, 'reward_judge_quality_std': 0.1470119059085846, 'reward_total_composite_mean': 0.4052889943122864, 'reward_total_composite_std': 0.19342483580112457, 'epoch': 0.08}

 25%|██▌       | 837/3300 [2:06:09<5:46:26,  8.44s/it]INFO 04-13 09:27:07 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 838/3300 [2:06:20<6:16:53,  9.19s/it]2026-04-13 09:27:18,273 | INFO | train_grpo_train | metrics_logged mode=train step=838

                                                      
{'loss': -0.0852, 'grad_norm': 2.8423261642456055, 'learning_rate': 7.463636363636364e-06, 'num_tokens': 1482522.0, 'completions/mean_length': 87.375, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.71428680419922, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.796506404876709, 'rewards/meter/std': 0.35017162561416626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9451595544815063, 'rewards/repeat_soft/std': 0.03314792737364769, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.13452960550785065, 'rewards/total_composite/mean': 0.49256080389022827, 'rewards/total_composite/std': 0.22193673253059387, 'reward': 0.49256080389022827, 'reward_std': 0.22193673253059387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16713367402553558, 'sampling/sampling_logp_difference/max': 2.263957977294922, 'sampling/importance_sampling_ratio/min': 0.1039382815361023, 'sampling/importance_sampling_ratio/mean': 1.02517831325531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0848853811621666, 'clip_ratio/low_mean': 0.02500000037252903, 'clip_ratio/low_min': 0.02500000037252903, 'clip_ratio/high_mean': 0.11378003749996424, 'clip_ratio/high_max': 0.11378003749996424, 'clip_ratio/region_mean': 0.13878003787249327, 'reward_total_mean': 0.49256080389022827, 'reward_meter_mean': 0.796506404876709, 'reward_meter_std': 0.35017162561416626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9451595544815063, 'reward_repeat_soft_std': 0.03314792737364769, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.13452960550785065, 'reward_total_composite_mean': 0.49256080389022827, 'reward_total_composite_std': 0.22193673253059387, 'epoch': 0.08}

 25%|██▌       | 838/3300 [2:06:20<6:16:53,  9.19s/it]INFO 04-13 09:27:18 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 839/3300 [2:06:32<6:47:08,  9.93s/it]2026-04-13 09:27:29,927 | INFO | train_grpo_train | metrics_logged mode=train step=839

                                                      
{'loss': -0.1898, 'grad_norm': 3.640242338180542, 'learning_rate': 7.460606060606061e-06, 'num_tokens': 1484565.0, 'completions/mean_length': 137.375, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 83.85714721679688, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.32945507764816284, 'rewards/meter/std': 0.2728482186794281, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9720566272735596, 'rewards/repeat_soft/std': 0.04077162221074104, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.22320717573165894, 'rewards/total_composite/mean': 0.3884682059288025, 'rewards/total_composite/std': 0.17261742055416107, 'reward': 0.3884682059288025, 'reward_std': 0.17261742055416107, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1431867629289627, 'sampling/sampling_logp_difference/max': 1.8204381465911865, 'sampling/importance_sampling_ratio/min': 0.16195477545261383, 'sampling/importance_sampling_ratio/mean': 1.0115183591842651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8769795000553131, 'clip_ratio/low_mean': 0.035330988466739655, 'clip_ratio/low_min': 0.035330988466739655, 'clip_ratio/high_mean': 0.09252869058400393, 'clip_ratio/high_max': 0.09252869058400393, 'clip_ratio/region_mean': 0.12785967905074358, 'reward_total_mean': 0.3884682059288025, 'reward_meter_mean': 0.32945507764816284, 'reward_meter_std': 0.2728482186794281, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9720566272735596, 'reward_repeat_soft_std': 0.04077162221074104, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.22320717573165894, 'reward_total_composite_mean': 0.3884682059288025, 'reward_total_composite_std': 0.17261742055416107, 'epoch': 0.08}

 25%|██▌       | 839/3300 [2:06:32<6:47:08,  9.93s/it]INFO 04-13 09:27:30 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 840/3300 [2:06:38<5:58:12,  8.74s/it]2026-04-13 09:27:35,919 | INFO | train_grpo_train | metrics_logged mode=train step=840

                                                      
{'loss': 0.0086, 'grad_norm': 28.868541717529297, 'learning_rate': 7.4575757575757575e-06, 'num_tokens': 1485934.0, 'completions/mean_length': 18.125, 'completions/min_length': 13.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.125, 'completions/min_terminated_length': 13.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.31399306654930115, 'rewards/meter/std': 0.42830389738082886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.4486703872680664, 'rewards/total_composite/std': 0.13358429074287415, 'reward': 0.4486703872680664, 'reward_std': 0.13358429074287415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19885778427124023, 'sampling/sampling_logp_difference/max': 1.3104972839355469, 'sampling/importance_sampling_ratio/min': 0.26968589425086975, 'sampling/importance_sampling_ratio/mean': 1.0157369375228882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3750434964895248, 'clip_ratio/low_mean': 0.1131312781944871, 'clip_ratio/low_min': 0.1131312781944871, 'clip_ratio/high_mean': 0.05902777798473835, 'clip_ratio/high_max': 0.05902777798473835, 'clip_ratio/region_mean': 0.17215905617922544, 'reward_total_mean': 0.4486703872680664, 'reward_meter_mean': 0.31399306654930115, 'reward_meter_std': 0.42830389738082886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.4486703872680664, 'reward_total_composite_std': 0.13358429074287415, 'epoch': 0.08}

 25%|██▌       | 840/3300 [2:06:38<5:58:12,  8.74s/it]INFO 04-13 09:27:36 [block_pool.py:378] Successfully reset prefix cache

 25%|██▌       | 841/3300 [2:06:45<5:46:31,  8.46s/it]2026-04-13 09:27:43,688 | INFO | train_grpo_train | metrics_logged mode=train step=841

                                                      
{'loss': -0.0214, 'grad_norm': 10.265496253967285, 'learning_rate': 7.454545454545456e-06, 'num_tokens': 1487932.0, 'completions/mean_length': 71.75, 'completions/min_length': 64.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.27685803174972534, 'rewards/meter/std': 0.23641550540924072, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7995047569274902, 'rewards/repeat_soft/std': 0.11472798138856888, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.3935416340827942, 'rewards/total_composite/std': 0.06690897047519684, 'reward': 0.3935416340827942, 'reward_std': 0.06690897047519684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12471328675746918, 'sampling/sampling_logp_difference/max': 2.387075901031494, 'sampling/importance_sampling_ratio/min': 0.09189800918102264, 'sampling/importance_sampling_ratio/mean': 1.0074000358581543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5988735593855381, 'clip_ratio/low_mean': 0.0575826425338164, 'clip_ratio/low_min': 0.0575826425338164, 'clip_ratio/high_mean': 0.01458030566573143, 'clip_ratio/high_max': 0.01458030566573143, 'clip_ratio/region_mean': 0.07216294819954783, 'reward_total_mean': 0.3935416340827942, 'reward_meter_mean': 0.27685803174972534, 'reward_meter_std': 0.23641550540924072, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7995047569274902, 'reward_repeat_soft_std': 0.11472798138856888, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.3935416340827942, 'reward_total_composite_std': 0.06690897047519684, 'epoch': 0.08}

 25%|██▌       | 841/3300 [2:06:45<5:46:31,  8.46s/it]INFO 04-13 09:27:44 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 842/3300 [2:06:57<6:23:32,  9.36s/it]2026-04-13 09:27:55,165 | INFO | train_grpo_train | metrics_logged mode=train step=842

                                                      
{'loss': -0.1792, 'grad_norm': 2.87349271774292, 'learning_rate': 7.451515151515152e-06, 'num_tokens': 1490164.0, 'completions/mean_length': 153.0, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 101.71428680419922, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.4708000421524048, 'rewards/meter/std': 0.2996063232421875, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8525799512863159, 'rewards/repeat_soft/std': 0.11474509537220001, 'rewards/judge_quality/mean': 0.7075000405311584, 'rewards/judge_quality/std': 0.29489707946777344, 'rewards/total_composite/mean': 0.46092456579208374, 'rewards/total_composite/std': 0.20109796524047852, 'reward': 0.46092456579208374, 'reward_std': 0.20109793543815613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1308763027191162, 'sampling/sampling_logp_difference/max': 1.7270450592041016, 'sampling/importance_sampling_ratio/min': 0.17780904471874237, 'sampling/importance_sampling_ratio/mean': 1.0077468156814575, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48516032099723816, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.10736322961747646, 'clip_ratio/high_max': 0.10736322961747646, 'clip_ratio/region_mean': 0.12298822961747646, 'reward_total_mean': 0.46092456579208374, 'reward_meter_mean': 0.4708000421524048, 'reward_meter_std': 0.2996063232421875, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8525799512863159, 'reward_repeat_soft_std': 0.11474509537220001, 'reward_judge_quality_mean': 0.7075000405311584, 'reward_judge_quality_std': 0.29489707946777344, 'reward_total_composite_mean': 0.46092456579208374, 'reward_total_composite_std': 0.20109796524047852, 'epoch': 0.08}

 26%|██▌       | 842/3300 [2:06:57<6:23:32,  9.36s/it]INFO 04-13 09:27:55 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 843/3300 [2:07:04<5:53:29,  8.63s/it]2026-04-13 09:28:02,095 | INFO | train_grpo_train | metrics_logged mode=train step=843

                                                      
{'loss': 0.004, 'grad_norm': 9.423258781433105, 'learning_rate': 7.448484848484849e-06, 'num_tokens': 1492334.0, 'completions/mean_length': 79.25, 'completions/min_length': 59.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.25, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.7220500111579895, 'rewards/meter/std': 0.34564322233200073, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9802860021591187, 'rewards/repeat_soft/std': 0.0205127764493227, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.22937417030334473, 'rewards/total_composite/mean': 0.5785923600196838, 'rewards/total_composite/std': 0.19729620218276978, 'reward': 0.5785923600196838, 'reward_std': 0.19729618728160858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13742201030254364, 'sampling/sampling_logp_difference/max': 1.9483990669250488, 'sampling/importance_sampling_ratio/min': 0.14250202476978302, 'sampling/importance_sampling_ratio/mean': 1.0099302530288696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7041908577084541, 'clip_ratio/low_mean': 0.08119471371173859, 'clip_ratio/low_min': 0.08119471371173859, 'clip_ratio/high_mean': 0.03453873749822378, 'clip_ratio/high_max': 0.03453873749822378, 'clip_ratio/region_mean': 0.11573345120996237, 'reward_total_mean': 0.5785923600196838, 'reward_meter_mean': 0.7220500111579895, 'reward_meter_std': 0.34564322233200073, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9802860021591187, 'reward_repeat_soft_std': 0.0205127764493227, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.22937417030334473, 'reward_total_composite_mean': 0.5785923600196838, 'reward_total_composite_std': 0.19729620218276978, 'epoch': 0.08}

 26%|██▌       | 843/3300 [2:07:04<5:53:29,  8.63s/it]INFO 04-13 09:28:02 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 844/3300 [2:07:10<5:25:59,  7.96s/it]2026-04-13 09:28:08,503 | INFO | train_grpo_train | metrics_logged mode=train step=844

                                                      
{'loss': 0.0033, 'grad_norm': 13.6819486618042, 'learning_rate': 7.445454545454546e-06, 'num_tokens': 1494315.0, 'completions/mean_length': 73.625, 'completions/min_length': 57.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.19047191739082336, 'rewards/meter/std': 0.254278302192688, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9960660934448242, 'rewards/repeat_soft/std': 0.0026364249642938375, 'rewards/judge_quality/mean': 0.5987499952316284, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.40190425515174866, 'rewards/total_composite/std': 0.15848101675510406, 'reward': 0.40190425515174866, 'reward_std': 0.15848100185394287, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1798246055841446, 'sampling/sampling_logp_difference/max': 3.6892249584198, 'sampling/importance_sampling_ratio/min': 0.024991365149617195, 'sampling/importance_sampling_ratio/mean': 1.0255519151687622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0761237666010857, 'clip_ratio/low_mean': 0.1708741346374154, 'clip_ratio/low_min': 0.1708741346374154, 'clip_ratio/high_mean': 0.0234375, 'clip_ratio/high_max': 0.0234375, 'clip_ratio/region_mean': 0.1943116346374154, 'reward_total_mean': 0.40190425515174866, 'reward_meter_mean': 0.19047191739082336, 'reward_meter_std': 0.254278302192688, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9960660934448242, 'reward_repeat_soft_std': 0.0026364249642938375, 'reward_judge_quality_mean': 0.5987499952316284, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.40190425515174866, 'reward_total_composite_std': 0.15848101675510406, 'epoch': 0.08}

 26%|██▌       | 844/3300 [2:07:10<5:25:59,  7.96s/it]INFO 04-13 09:28:08 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 845/3300 [2:07:16<5:00:57,  7.36s/it]2026-04-13 09:28:14,435 | INFO | train_grpo_train | metrics_logged mode=train step=845

                                                      
{'loss': 0.0103, 'grad_norm': 12.717124938964844, 'learning_rate': 7.442424242424243e-06, 'num_tokens': 1495911.0, 'completions/mean_length': 39.5, 'completions/min_length': 34.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9568942785263062, 'rewards/meter/std': 0.049314092844724655, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9216736555099487, 'rewards/repeat_soft/std': 0.11439024657011032, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6089969873428345, 'rewards/total_composite/std': 0.019405582919716835, 'reward': 0.6089969873428345, 'reward_std': 0.019405582919716835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17396031320095062, 'sampling/sampling_logp_difference/max': 1.5292034149169922, 'sampling/importance_sampling_ratio/min': 0.216708242893219, 'sampling/importance_sampling_ratio/mean': 0.9948079586029053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.194935329258442, 'clip_ratio/low_mean': 0.03553921636193991, 'clip_ratio/low_min': 0.03553921636193991, 'clip_ratio/high_mean': 0.14455649629235268, 'clip_ratio/high_max': 0.14455649629235268, 'clip_ratio/region_mean': 0.18009571265429258, 'reward_total_mean': 0.6089969873428345, 'reward_meter_mean': 0.9568942785263062, 'reward_meter_std': 0.049314092844724655, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9216736555099487, 'reward_repeat_soft_std': 0.11439024657011032, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6089969873428345, 'reward_total_composite_std': 0.019405582919716835, 'epoch': 0.08}

 26%|██▌       | 845/3300 [2:07:16<5:00:57,  7.36s/it]INFO 04-13 09:28:14 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 846/3300 [2:07:28<5:53:02,  8.63s/it]2026-04-13 09:28:26,047 | INFO | train_grpo_train | metrics_logged mode=train step=846

                                                      
{'loss': -0.1199, 'grad_norm': 3.850022554397583, 'learning_rate': 7.439393939393939e-06, 'num_tokens': 1497555.0, 'completions/mean_length': 101.5, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.85714340209961, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.33093252778053284, 'rewards/meter/std': 0.307478129863739, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9527453184127808, 'rewards/repeat_soft/std': 0.09837198257446289, 'rewards/judge_quality/mean': 0.5437500476837158, 'rewards/judge_quality/std': 0.29861289262771606, 'rewards/total_composite/mean': 0.4027683138847351, 'rewards/total_composite/std': 0.18549330532550812, 'reward': 0.4027683138847351, 'reward_std': 0.18549330532550812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17301549017429352, 'sampling/sampling_logp_difference/max': 1.2793481349945068, 'sampling/importance_sampling_ratio/min': 0.32019028067588806, 'sampling/importance_sampling_ratio/mean': 1.0440881252288818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0814722403883934, 'clip_ratio/low_mean': 0.030868902802467346, 'clip_ratio/low_min': 0.030868902802467346, 'clip_ratio/high_mean': 0.1328445840626955, 'clip_ratio/high_max': 0.1328445840626955, 'clip_ratio/region_mean': 0.16371348686516285, 'reward_total_mean': 0.4027683138847351, 'reward_meter_mean': 0.33093252778053284, 'reward_meter_std': 0.307478129863739, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9527453184127808, 'reward_repeat_soft_std': 0.09837198257446289, 'reward_judge_quality_mean': 0.5437500476837158, 'reward_judge_quality_std': 0.29861289262771606, 'reward_total_composite_mean': 0.4027683138847351, 'reward_total_composite_std': 0.18549330532550812, 'epoch': 0.08}

 26%|██▌       | 846/3300 [2:07:28<5:53:02,  8.63s/it]INFO 04-13 09:28:26 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 847/3300 [2:07:34<5:21:57,  7.87s/it]2026-04-13 09:28:32,155 | INFO | train_grpo_train | metrics_logged mode=train step=847

                                                      
{'loss': 0.0692, 'grad_norm': 14.30763053894043, 'learning_rate': 7.4363636363636375e-06, 'num_tokens': 1499209.0, 'completions/mean_length': 40.75, 'completions/min_length': 36.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.2561221420764923, 'rewards/meter/std': 0.38831233978271484, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9554833769798279, 'rewards/repeat_soft/std': 0.04576227441430092, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.11686378717422485, 'rewards/total_composite/mean': 0.4316032826900482, 'rewards/total_composite/std': 0.13052450120449066, 'reward': 0.4316032826900482, 'reward_std': 0.13052450120449066, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15778307616710663, 'sampling/sampling_logp_difference/max': 2.067727565765381, 'sampling/importance_sampling_ratio/min': 0.1264728456735611, 'sampling/importance_sampling_ratio/mean': 1.0026819705963135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0495265498757362, 'clip_ratio/low_mean': 0.11036532931029797, 'clip_ratio/low_min': 0.11036532931029797, 'clip_ratio/high_mean': 0.04537786729633808, 'clip_ratio/high_max': 0.04537786729633808, 'clip_ratio/region_mean': 0.15574319660663605, 'reward_total_mean': 0.4316032826900482, 'reward_meter_mean': 0.2561221420764923, 'reward_meter_std': 0.38831233978271484, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9554833769798279, 'reward_repeat_soft_std': 0.04576227441430092, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.11686378717422485, 'reward_total_composite_mean': 0.4316032826900482, 'reward_total_composite_std': 0.13052450120449066, 'epoch': 0.09}

 26%|██▌       | 847/3300 [2:07:34<5:21:57,  7.87s/it]INFO 04-13 09:28:32 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 848/3300 [2:07:45<6:08:01,  9.01s/it]2026-04-13 09:28:43,797 | INFO | train_grpo_train | metrics_logged mode=train step=848

                                                      
{'loss': -0.0575, 'grad_norm': 1.3515770435333252, 'learning_rate': 7.433333333333334e-06, 'num_tokens': 1500625.0, 'completions/mean_length': 274.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.5, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.1757933795452118, 'rewards/meter/std': 0.11196483671665192, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.9779536724090576, 'rewards/repeat_soft/std': 0.036843180656433105, 'rewards/judge_quality/mean': 0.2175000011920929, 'rewards/judge_quality/std': 0.18873639404773712, 'rewards/total_composite/mean': 0.20365145802497864, 'rewards/total_composite/std': 0.21790151298046112, 'reward': 0.20365145802497864, 'reward_std': 0.21790151298046112, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18480101227760315, 'sampling/sampling_logp_difference/max': 1.077570915222168, 'sampling/importance_sampling_ratio/min': 0.3404214084148407, 'sampling/importance_sampling_ratio/mean': 0.9984260201454163, 'sampling/importance_sampling_ratio/max': 1.741991639137268, 'entropy': 0.7906372845172882, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08901694789528847, 'clip_ratio/high_max': 0.08901694789528847, 'clip_ratio/region_mean': 0.08901694789528847, 'reward_total_mean': 0.20365145802497864, 'reward_meter_mean': 0.1757933795452118, 'reward_meter_std': 0.11196483671665192, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.9779536724090576, 'reward_repeat_soft_std': 0.036843180656433105, 'reward_judge_quality_mean': 0.2175000011920929, 'reward_judge_quality_std': 0.18873639404773712, 'reward_total_composite_mean': 0.20365145802497864, 'reward_total_composite_std': 0.21790151298046112, 'epoch': 0.09}

 26%|██▌       | 848/3300 [2:07:46<6:08:01,  9.01s/it]INFO 04-13 09:28:44 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 849/3300 [2:07:51<5:29:47,  8.07s/it]2026-04-13 09:28:49,697 | INFO | train_grpo_train | metrics_logged mode=train step=849

                                                      
{'loss': 0.017, 'grad_norm': 17.583232879638672, 'learning_rate': 7.430303030303031e-06, 'num_tokens': 1502151.0, 'completions/mean_length': 36.75, 'completions/min_length': 30.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.34428954124450684, 'rewards/meter/std': 0.35059359669685364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9794764518737793, 'rewards/repeat_soft/std': 0.013408654369413853, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.4079633355140686, 'rewards/total_composite/std': 0.18753710389137268, 'reward': 0.4079633355140686, 'reward_std': 0.18753710389137268, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1670643538236618, 'sampling/sampling_logp_difference/max': 2.1720149517059326, 'sampling/importance_sampling_ratio/min': 0.1139477863907814, 'sampling/importance_sampling_ratio/mean': 1.011920690536499, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8922168835997581, 'clip_ratio/low_mean': 0.026709401980042458, 'clip_ratio/low_min': 0.026709401980042458, 'clip_ratio/high_mean': 0.11197670269757509, 'clip_ratio/high_max': 0.11197670269757509, 'clip_ratio/region_mean': 0.13868610467761755, 'reward_total_mean': 0.4079633355140686, 'reward_meter_mean': 0.34428954124450684, 'reward_meter_std': 0.35059359669685364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9794764518737793, 'reward_repeat_soft_std': 0.013408654369413853, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.4079633355140686, 'reward_total_composite_std': 0.18753710389137268, 'epoch': 0.09}

 26%|██▌       | 849/3300 [2:07:51<5:29:47,  8.07s/it]INFO 04-13 09:28:49 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 850/3300 [2:07:57<5:02:55,  7.42s/it]2026-04-13 09:28:55,586 | INFO | train_grpo_train | metrics_logged mode=train step=850

                                                      
{'loss': 0.0287, 'grad_norm': 13.789236068725586, 'learning_rate': 7.4272727272727275e-06, 'num_tokens': 1503640.0, 'completions/mean_length': 38.125, 'completions/min_length': 34.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.22956950962543488, 'rewards/meter/std': 0.20842336118221283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9453689455986023, 'rewards/repeat_soft/std': 0.03891696780920029, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.4220011532306671, 'rewards/total_composite/std': 0.06589780002832413, 'reward': 0.4220011532306671, 'reward_std': 0.06589780747890472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13908246159553528, 'sampling/sampling_logp_difference/max': 1.6850318908691406, 'sampling/importance_sampling_ratio/min': 0.18543851375579834, 'sampling/importance_sampling_ratio/mean': 1.0063995122909546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9740190431475639, 'clip_ratio/low_mean': 0.065509426407516, 'clip_ratio/low_min': 0.065509426407516, 'clip_ratio/high_mean': 0.062609649496153, 'clip_ratio/high_max': 0.062609649496153, 'clip_ratio/region_mean': 0.128119075903669, 'reward_total_mean': 0.4220011532306671, 'reward_meter_mean': 0.22956950962543488, 'reward_meter_std': 0.20842336118221283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9453689455986023, 'reward_repeat_soft_std': 0.03891696780920029, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.4220011532306671, 'reward_total_composite_std': 0.06589780002832413, 'epoch': 0.09}

 26%|██▌       | 850/3300 [2:07:57<5:02:55,  7.42s/it]INFO 04-13 09:28:55 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.75s/it]

 30%|███       | 3/10 [00:12<00:33,  4.72s/it]

 40%|████      | 4/10 [00:15<00:24,  4.11s/it]

 50%|█████     | 5/10 [00:24<00:29,  5.83s/it]

 60%|██████    | 6/10 [00:27<00:20,  5.02s/it]

 70%|███████   | 7/10 [00:36<00:18,  6.19s/it]

 80%|████████  | 8/10 [00:45<00:14,  7.03s/it]

 90%|█████████ | 9/10 [00:54<00:07,  7.59s/it]

100%|██████████| 10/10 [01:02<00:00,  7.87s/it]2026-04-13 09:30:07,451 | INFO | train_grpo_train | metrics_logged mode=eval step=850

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 71.8131, 'eval_samples_per_second': 1.114, 'eval_steps_per_second': 0.139, 'eval_num_tokens': 1503640.0, 'eval_completions/mean_length': 124.15, 'eval_completions/min_length': 32.8, 'eval_completions/max_length': 385.3, 'eval_completions/clipped_ratio': 0.1375, 'eval_completions/mean_terminated_length': 62.55821685791015, 'eval_completions/min_terminated_length': 32.8, 'eval_completions/max_terminated_length': 97.5, 'eval_rewards/meter/mean': 0.5041633427143097, 'eval_rewards/meter/std': 0.3839849755167961, 'eval_rewards/count_adherence/mean': 0.8604166507720947, 'eval_rewards/count_adherence/std': 0.1260624572634697, 'eval_rewards/hard_gate/mean': 0.8625, 'eval_rewards/hard_gate/std': 0.28575828671455383, 'eval_rewards/repeat_soft/mean': 0.9330996811389923, 'eval_rewards/repeat_soft/std': 0.10413262154906988, 'eval_rewards/judge_quality/mean': 0.47837499976158143, 'eval_rewards/judge_quality/std': 0.23521116748452187, 'eval_rewards/total_composite/mean': 0.43147261142730714, 'eval_rewards/total_composite/std': 0.2192233495414257, 'eval_reward': 0.43147261142730714, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08082602322101592, 'eval_sampling/sampling_logp_difference/max': 0.9368991374969482, 'eval_sampling/importance_sampling_ratio/min': 0.3969268500804901, 'eval_sampling/importance_sampling_ratio/mean': 1.0231210112571716, 'eval_sampling/importance_sampling_ratio/max': 1.523264479637146, 'eval_entropy': 0.988732784986496, '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': 0.43147261142730714, 'eval_reward_meter_mean': 0.5041633427143097, 'eval_reward_meter_std': 0.3839849755167961, 'eval_reward_count_adherence_mean': 0.8604166507720947, 'eval_reward_count_adherence_std': 0.1260624572634697, 'eval_reward_hard_gate_mean': 0.8625, 'eval_reward_hard_gate_std': 0.28575828671455383, 'eval_reward_repeat_soft_mean': 0.9330996811389923, 'eval_reward_repeat_soft_std': 0.10413262154906988, 'eval_reward_judge_quality_mean': 0.47837499976158143, 'eval_reward_judge_quality_std': 0.23521116748452187, 'eval_reward_total_composite_mean': 0.43147261142730714, 'eval_reward_total_composite_std': 0.2192233495414257, 'epoch': 0.09}

 26%|██▌       | 850/3300 [2:09:09<5:02:55,  7.42s/it]

100%|██████████| 10/10 [01:02<00:00,  7.87s/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(

 26%|██▌       | 851/3300 [2:09:19<20:10:26, 29.66s/it]2026-04-13 09:30:17,131 | INFO | train_grpo_train | metrics_logged mode=train step=851

                                                       
{'loss': 0.1141, 'grad_norm': 14.212329864501953, 'learning_rate': 7.424242424242425e-06, 'num_tokens': 1505291.0, 'completions/mean_length': 44.375, 'completions/min_length': 37.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5698230266571045, 'rewards/meter/std': 0.3540171980857849, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.973199725151062, 'rewards/repeat_soft/std': 0.015333795920014381, 'rewards/judge_quality/mean': 0.8524999618530273, 'rewards/judge_quality/std': 0.17531605064868927, 'rewards/total_composite/mean': 0.6624869108200073, 'rewards/total_composite/std': 0.2213582694530487, 'reward': 0.6624869108200073, 'reward_std': 0.2213582694530487, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14528997242450714, 'sampling/sampling_logp_difference/max': 1.5352954864501953, 'sampling/importance_sampling_ratio/min': 0.21539203822612762, 'sampling/importance_sampling_ratio/mean': 1.0009996891021729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.897915206849575, 'clip_ratio/low_mean': 0.07398198451846838, 'clip_ratio/low_min': 0.07398198451846838, 'clip_ratio/high_mean': 0.04296270105987787, 'clip_ratio/high_max': 0.04296270105987787, 'clip_ratio/region_mean': 0.11694468557834625, 'reward_total_mean': 0.6624869108200073, 'reward_meter_mean': 0.5698230266571045, 'reward_meter_std': 0.3540171980857849, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.973199725151062, 'reward_repeat_soft_std': 0.015333795920014381, 'reward_judge_quality_mean': 0.8524999618530273, 'reward_judge_quality_std': 0.17531605064868927, 'reward_total_composite_mean': 0.6624869108200073, 'reward_total_composite_std': 0.2213582694530487, 'epoch': 0.09}

 26%|██▌       | 851/3300 [2:09:19<20:10:26, 29.66s/it]INFO 04-13 09:30:17 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 852/3300 [2:09:31<16:33:07, 24.34s/it]2026-04-13 09:30:29,071 | INFO | train_grpo_train | metrics_logged mode=train step=852

                                                       
{'loss': -0.1035, 'grad_norm': 3.831021547317505, 'learning_rate': 7.421212121212121e-06, 'num_tokens': 1507089.0, 'completions/mean_length': 109.75, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.28571701049805, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.758591890335083, 'rewards/meter/std': 0.20175181329250336, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9904145002365112, 'rewards/repeat_soft/std': 0.0077845207415521145, 'rewards/judge_quality/mean': 0.6362500190734863, 'rewards/judge_quality/std': 0.29635587334632874, 'rewards/total_composite/mean': 0.6015548706054688, 'rewards/total_composite/std': 0.27526721358299255, 'reward': 0.6015548706054688, 'reward_std': 0.27526721358299255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15900366008281708, 'sampling/sampling_logp_difference/max': 2.1218101978302, 'sampling/importance_sampling_ratio/min': 0.11981455236673355, 'sampling/importance_sampling_ratio/mean': 1.0003243684768677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6033063903450966, 'clip_ratio/low_mean': 0.06447595916688442, 'clip_ratio/low_min': 0.06447595916688442, 'clip_ratio/high_mean': 0.07329545635730028, 'clip_ratio/high_max': 0.07329545635730028, 'clip_ratio/region_mean': 0.1377714155241847, 'reward_total_mean': 0.6015548706054688, 'reward_meter_mean': 0.758591890335083, 'reward_meter_std': 0.20175181329250336, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9904145002365112, 'reward_repeat_soft_std': 0.0077845207415521145, 'reward_judge_quality_mean': 0.6362500190734863, 'reward_judge_quality_std': 0.29635587334632874, 'reward_total_composite_mean': 0.6015548706054688, 'reward_total_composite_std': 0.27526721358299255, 'epoch': 0.09}

 26%|██▌       | 852/3300 [2:09:31<16:33:07, 24.34s/it]INFO 04-13 09:30:29 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 853/3300 [2:09:42<13:53:58, 20.45s/it]2026-04-13 09:30:40,507 | INFO | train_grpo_train | metrics_logged mode=train step=853

                                                       
{'loss': -0.0834, 'grad_norm': 1.662527322769165, 'learning_rate': 7.4181818181818185e-06, 'num_tokens': 1508481.0, 'completions/mean_length': 155.0, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.2141370326280594, 'rewards/meter/std': 0.3055233061313629, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.3720119297504425, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9991523027420044, 'rewards/repeat_soft/std': 0.0015711163869127631, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.3919456899166107, 'rewards/total_composite/mean': 0.3827788233757019, 'rewards/total_composite/std': 0.29095616936683655, 'reward': 0.3827788233757019, 'reward_std': 0.29095613956451416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1022883877158165, 'sampling/sampling_logp_difference/max': 1.447120189666748, 'sampling/importance_sampling_ratio/min': 0.23524677753448486, 'sampling/importance_sampling_ratio/mean': 0.9997597336769104, 'sampling/importance_sampling_ratio/max': 1.6579594612121582, 'entropy': 0.4650096446275711, 'clip_ratio/low_mean': 0.01993534481152892, 'clip_ratio/low_min': 0.01993534481152892, 'clip_ratio/high_mean': 0.06936437916010618, 'clip_ratio/high_max': 0.06936437916010618, 'clip_ratio/region_mean': 0.0892997239716351, 'reward_total_mean': 0.3827788233757019, 'reward_meter_mean': 0.2141370326280594, 'reward_meter_std': 0.3055233061313629, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.3720119297504425, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9991523027420044, 'reward_repeat_soft_std': 0.0015711163869127631, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.3919456899166107, 'reward_total_composite_mean': 0.3827788233757019, 'reward_total_composite_std': 0.29095616936683655, 'epoch': 0.09}

 26%|██▌       | 853/3300 [2:09:42<13:53:58, 20.45s/it]INFO 04-13 09:30:40 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 854/3300 [2:09:49<11:03:02, 16.26s/it]2026-04-13 09:30:46,964 | INFO | train_grpo_train | metrics_logged mode=train step=854

                                                       
{'loss': 0.1215, 'grad_norm': 20.572551727294922, 'learning_rate': 7.415151515151515e-06, 'num_tokens': 1509884.0, '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': 0.6981173753738403, 'rewards/meter/std': 0.4420422911643982, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8873992562294006, 'rewards/repeat_soft/std': 0.11063570529222488, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.6501049995422363, 'rewards/total_composite/std': 0.25837576389312744, 'reward': 0.6501049995422363, 'reward_std': 0.25837576389312744, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18870551884174347, 'sampling/sampling_logp_difference/max': 1.3879594802856445, 'sampling/importance_sampling_ratio/min': 0.24958406388759613, 'sampling/importance_sampling_ratio/mean': 1.016324758529663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5916697159409523, 'clip_ratio/low_mean': 0.07532964693382382, 'clip_ratio/low_min': 0.07532964693382382, 'clip_ratio/high_mean': 0.07359601557254791, 'clip_ratio/high_max': 0.07359601557254791, 'clip_ratio/region_mean': 0.14892566250637174, 'reward_total_mean': 0.6501049995422363, 'reward_meter_mean': 0.6981173753738403, 'reward_meter_std': 0.4420422911643982, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8873992562294006, 'reward_repeat_soft_std': 0.11063570529222488, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.6501049995422363, 'reward_total_composite_std': 0.25837576389312744, 'epoch': 0.09}

 26%|██▌       | 854/3300 [2:09:49<11:03:02, 16.26s/it]INFO 04-13 09:30:47 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 855/3300 [2:10:00<10:04:49, 14.84s/it]2026-04-13 09:30:58,487 | INFO | train_grpo_train | metrics_logged mode=train step=855

                                                       
{'loss': -0.09, 'grad_norm': 2.11004376411438, 'learning_rate': 7.412121212121213e-06, 'num_tokens': 1511385.0, 'completions/mean_length': 160.625, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4706278443336487, 'rewards/meter/std': 0.3859804570674896, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9950848817825317, 'rewards/repeat_soft/std': 0.007068851962685585, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.32407838106155396, 'rewards/total_composite/mean': 0.4136121869087219, 'rewards/total_composite/std': 0.29443883895874023, 'reward': 0.4136121869087219, 'reward_std': 0.29443883895874023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17346595227718353, 'sampling/sampling_logp_difference/max': 1.8882384300231934, 'sampling/importance_sampling_ratio/min': 0.15133817493915558, 'sampling/importance_sampling_ratio/mean': 1.0436005592346191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1971502006053925, 'clip_ratio/low_mean': 0.05434782616794109, 'clip_ratio/low_min': 0.05434782616794109, 'clip_ratio/high_mean': 0.0963926874101162, 'clip_ratio/high_max': 0.0963926874101162, 'clip_ratio/region_mean': 0.1507405135780573, 'reward_total_mean': 0.4136121869087219, 'reward_meter_mean': 0.4706278443336487, 'reward_meter_std': 0.3859804570674896, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9950848817825317, 'reward_repeat_soft_std': 0.007068851962685585, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.32407838106155396, 'reward_total_composite_mean': 0.4136121869087219, 'reward_total_composite_std': 0.29443883895874023, 'epoch': 0.09}

 26%|██▌       | 855/3300 [2:10:00<10:04:49, 14.84s/it]INFO 04-13 09:30:58 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 856/3300 [2:10:11<9:21:11, 13.78s/it] 2026-04-13 09:31:09,751 | INFO | train_grpo_train | metrics_logged mode=train step=856

                                                      
{'loss': -0.0471, 'grad_norm': 5.2753214836120605, 'learning_rate': 7.40909090909091e-06, 'num_tokens': 1512967.0, 'completions/mean_length': 96.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.42857360839844, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4234778583049774, 'rewards/meter/std': 0.45958077907562256, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8645415306091309, 'rewards/repeat_soft/std': 0.08556653559207916, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.3520044684410095, 'rewards/total_composite/mean': 0.4821651577949524, 'rewards/total_composite/std': 0.318382203578949, 'reward': 0.4821651577949524, 'reward_std': 0.318382203578949, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1610652208328247, 'sampling/sampling_logp_difference/max': 1.7582321166992188, 'sampling/importance_sampling_ratio/min': 0.17234928905963898, 'sampling/importance_sampling_ratio/mean': 0.969700276851654, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7111929133534431, 'clip_ratio/low_mean': 0.048448143526911736, 'clip_ratio/low_min': 0.048448143526911736, 'clip_ratio/high_mean': 0.05237358249723911, 'clip_ratio/high_max': 0.05237358249723911, 'clip_ratio/region_mean': 0.10082172602415085, 'reward_total_mean': 0.4821651577949524, 'reward_meter_mean': 0.4234778583049774, 'reward_meter_std': 0.45958077907562256, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8645415306091309, 'reward_repeat_soft_std': 0.08556653559207916, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.3520044684410095, 'reward_total_composite_mean': 0.4821651577949524, 'reward_total_composite_std': 0.318382203578949, 'epoch': 0.09}

 26%|██▌       | 856/3300 [2:10:11<9:21:11, 13.78s/it]INFO 04-13 09:31:09 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 857/3300 [2:10:23<8:51:52, 13.06s/it]2026-04-13 09:31:21,149 | INFO | train_grpo_train | metrics_logged mode=train step=857

                                                      
{'loss': -0.1533, 'grad_norm': 2.2963130474090576, 'learning_rate': 7.406060606060607e-06, 'num_tokens': 1514965.0, 'completions/mean_length': 250.75, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.7481331825256348, 'rewards/meter/std': 0.23359724879264832, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.8709614872932434, 'rewards/repeat_soft/std': 0.10680233687162399, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.24686609208583832, 'rewards/total_composite/mean': 0.3642788231372833, 'rewards/total_composite/std': 0.32161790132522583, 'reward': 0.3642788231372833, 'reward_std': 0.32161790132522583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14991313219070435, 'sampling/sampling_logp_difference/max': 1.635301113128662, 'sampling/importance_sampling_ratio/min': 0.19489367306232452, 'sampling/importance_sampling_ratio/mean': 1.0165810585021973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6166017353534698, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07782910205423832, 'clip_ratio/high_max': 0.07782910205423832, 'clip_ratio/region_mean': 0.07782910205423832, 'reward_total_mean': 0.3642788231372833, 'reward_meter_mean': 0.7481331825256348, 'reward_meter_std': 0.23359724879264832, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.8709614872932434, 'reward_repeat_soft_std': 0.10680233687162399, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.24686609208583832, 'reward_total_composite_mean': 0.3642788231372833, 'reward_total_composite_std': 0.32161790132522583, 'epoch': 0.09}

 26%|██▌       | 857/3300 [2:10:23<8:51:52, 13.06s/it]INFO 04-13 09:31:21 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 858/3300 [2:10:30<7:34:46, 11.17s/it]2026-04-13 09:31:27,915 | INFO | train_grpo_train | metrics_logged mode=train step=858

                                                      
{'loss': 0.0507, 'grad_norm': 10.526352882385254, 'learning_rate': 7.403030303030304e-06, 'num_tokens': 1516484.0, 'completions/mean_length': 44.875, 'completions/min_length': 38.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5815331935882568, 'rewards/meter/std': 0.3391035795211792, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9906216263771057, 'rewards/repeat_soft/std': 0.008486546576023102, 'rewards/judge_quality/mean': 0.6412500143051147, 'rewards/judge_quality/std': 0.3280869722366333, 'rewards/total_composite/mean': 0.5933663249015808, 'rewards/total_composite/std': 0.22715814411640167, 'reward': 0.5933663249015808, 'reward_std': 0.22715814411640167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14228667318820953, 'sampling/sampling_logp_difference/max': 2.4088973999023438, 'sampling/importance_sampling_ratio/min': 0.08991438150405884, 'sampling/importance_sampling_ratio/mean': 1.0182994604110718, 'sampling/importance_sampling_ratio/max': 1.898579716682434, 'entropy': 0.9067439511418343, 'clip_ratio/low_mean': 0.07189829740673304, 'clip_ratio/low_min': 0.07189829740673304, 'clip_ratio/high_mean': 0.039669995196163654, 'clip_ratio/high_max': 0.039669995196163654, 'clip_ratio/region_mean': 0.11156829260289669, 'reward_total_mean': 0.5933663249015808, 'reward_meter_mean': 0.5815331935882568, 'reward_meter_std': 0.3391035795211792, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9906216263771057, 'reward_repeat_soft_std': 0.008486546576023102, 'reward_judge_quality_mean': 0.6412500143051147, 'reward_judge_quality_std': 0.3280869722366333, 'reward_total_composite_mean': 0.5933663249015808, 'reward_total_composite_std': 0.22715814411640167, 'epoch': 0.09}

 26%|██▌       | 858/3300 [2:10:30<7:34:46, 11.17s/it]INFO 04-13 09:31:28 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 859/3300 [2:10:41<7:35:32, 11.20s/it]2026-04-13 09:31:39,199 | INFO | train_grpo_train | metrics_logged mode=train step=859

                                                      
{'loss': -0.0944, 'grad_norm': 1.9822921752929688, 'learning_rate': 7.4e-06, 'num_tokens': 1517939.0, 'completions/mean_length': 156.875, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5484725832939148, 'rewards/meter/std': 0.34517568349838257, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9916176199913025, 'rewards/repeat_soft/std': 0.014756565913558006, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.27166154980659485, 'rewards/total_composite/mean': 0.3975730538368225, 'rewards/total_composite/std': 0.25572142004966736, 'reward': 0.3975730538368225, 'reward_std': 0.25572139024734497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15622539818286896, 'sampling/sampling_logp_difference/max': 1.3402645587921143, 'sampling/importance_sampling_ratio/min': 0.2617764174938202, 'sampling/importance_sampling_ratio/mean': 1.0125941038131714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7355927750468254, 'clip_ratio/low_mean': 0.0234375, 'clip_ratio/low_min': 0.0234375, 'clip_ratio/high_mean': 0.08745039720088243, 'clip_ratio/high_max': 0.08745039720088243, 'clip_ratio/region_mean': 0.11088789720088243, 'reward_total_mean': 0.3975730538368225, 'reward_meter_mean': 0.5484725832939148, 'reward_meter_std': 0.34517568349838257, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9916176199913025, 'reward_repeat_soft_std': 0.014756565913558006, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.27166154980659485, 'reward_total_composite_mean': 0.3975730538368225, 'reward_total_composite_std': 0.25572142004966736, 'epoch': 0.09}

 26%|██▌       | 859/3300 [2:10:41<7:35:32, 11.20s/it]INFO 04-13 09:31:39 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 860/3300 [2:10:48<6:44:40,  9.95s/it]2026-04-13 09:31:46,210 | INFO | train_grpo_train | metrics_logged mode=train step=860

                                                      
{'loss': 0.0341, 'grad_norm': 13.615326881408691, 'learning_rate': 7.396969696969698e-06, 'num_tokens': 1519874.0, 'completions/mean_length': 66.875, 'completions/min_length': 56.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8078511953353882, 'rewards/meter/std': 0.31829094886779785, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9427466988563538, 'rewards/repeat_soft/std': 0.03384443372488022, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6276594400405884, 'rewards/total_composite/std': 0.2029723823070526, 'reward': 0.6276594400405884, 'reward_std': 0.20297236740589142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15863348543643951, 'sampling/sampling_logp_difference/max': 1.6779999732971191, 'sampling/importance_sampling_ratio/min': 0.18674710392951965, 'sampling/importance_sampling_ratio/mean': 1.0145034790039062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2692635208368301, 'clip_ratio/low_mean': 0.11109228897839785, 'clip_ratio/low_min': 0.11109228897839785, 'clip_ratio/high_mean': 0.07190890610218048, 'clip_ratio/high_max': 0.07190890610218048, 'clip_ratio/region_mean': 0.18300119508057833, 'reward_total_mean': 0.6276594400405884, 'reward_meter_mean': 0.8078511953353882, 'reward_meter_std': 0.31829094886779785, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9427466988563538, 'reward_repeat_soft_std': 0.03384443372488022, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6276594400405884, 'reward_total_composite_std': 0.2029723823070526, 'epoch': 0.09}

 26%|██▌       | 860/3300 [2:10:48<6:44:40,  9.95s/it]INFO 04-13 09:31:46 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 861/3300 [2:11:00<7:05:59, 10.48s/it]2026-04-13 09:31:57,924 | INFO | train_grpo_train | metrics_logged mode=train step=861

                                                      
{'loss': -0.1475, 'grad_norm': 3.81089448928833, 'learning_rate': 7.393939393939395e-06, 'num_tokens': 1522021.0, 'completions/mean_length': 155.375, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 104.42857360839844, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.4980456233024597, 'rewards/meter/std': 0.23626483976840973, 'rewards/count_adherence/mean': 0.7916666269302368, 'rewards/count_adherence/std': 0.07715165615081787, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9310250282287598, 'rewards/repeat_soft/std': 0.04550262540578842, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.2386719286441803, 'rewards/total_composite/mean': 0.4168010950088501, 'rewards/total_composite/std': 0.19585895538330078, 'reward': 0.4168010950088501, 'reward_std': 0.1958589404821396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14212839305400848, 'sampling/sampling_logp_difference/max': 2.3954758644104004, 'sampling/importance_sampling_ratio/min': 0.09112930297851562, 'sampling/importance_sampling_ratio/mean': 1.0250738859176636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9451244100928307, 'clip_ratio/low_mean': 0.026294946670532227, 'clip_ratio/low_min': 0.026294946670532227, 'clip_ratio/high_mean': 0.07693687546998262, 'clip_ratio/high_max': 0.07693687546998262, 'clip_ratio/region_mean': 0.10323182214051485, 'reward_total_mean': 0.4168010950088501, 'reward_meter_mean': 0.4980456233024597, 'reward_meter_std': 0.23626483976840973, 'reward_count_adherence_mean': 0.7916666269302368, 'reward_count_adherence_std': 0.07715165615081787, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9310250282287598, 'reward_repeat_soft_std': 0.04550262540578842, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.2386719286441803, 'reward_total_composite_mean': 0.4168010950088501, 'reward_total_composite_std': 0.19585895538330078, 'epoch': 0.09}

 26%|██▌       | 861/3300 [2:11:00<7:05:59, 10.48s/it]INFO 04-13 09:31:58 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 862/3300 [2:11:05<6:09:28,  9.09s/it]2026-04-13 09:32:03,782 | INFO | train_grpo_train | metrics_logged mode=train step=862

                                                      
{'loss': 0.1107, 'grad_norm': 18.969633102416992, 'learning_rate': 7.390909090909092e-06, 'num_tokens': 1523613.0, 'completions/mean_length': 35.0, 'completions/min_length': 26.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.49755197763442993, 'rewards/meter/std': 0.40188562870025635, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9942089319229126, 'rewards/repeat_soft/std': 0.007229667156934738, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.44823718070983887, 'rewards/total_composite/std': 0.14406037330627441, 'reward': 0.44823718070983887, 'reward_std': 0.14406037330627441, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18669536709785461, 'sampling/sampling_logp_difference/max': 2.0277328491210938, 'sampling/importance_sampling_ratio/min': 0.23620277643203735, 'sampling/importance_sampling_ratio/mean': 0.9968456029891968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.049236349761486, 'clip_ratio/low_mean': 0.0733456676825881, 'clip_ratio/low_min': 0.0733456676825881, 'clip_ratio/high_mean': 0.09544275049120188, 'clip_ratio/high_max': 0.09544275049120188, 'clip_ratio/region_mean': 0.16878841817378998, 'reward_total_mean': 0.44823718070983887, 'reward_meter_mean': 0.49755197763442993, 'reward_meter_std': 0.40188562870025635, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9942089319229126, 'reward_repeat_soft_std': 0.007229667156934738, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.44823718070983887, 'reward_total_composite_std': 0.14406037330627441, 'epoch': 0.09}

 26%|██▌       | 862/3300 [2:11:06<6:09:28,  9.09s/it]INFO 04-13 09:32:04 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 863/3300 [2:11:17<6:38:24,  9.81s/it]2026-04-13 09:32:15,262 | INFO | train_grpo_train | metrics_logged mode=train step=863

                                                      
{'loss': -0.1046, 'grad_norm': 2.331528902053833, 'learning_rate': 7.3878787878787885e-06, 'num_tokens': 1525192.0, 'completions/mean_length': 161.375, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.665990948677063, 'rewards/meter/std': 0.4026106297969818, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9687302112579346, 'rewards/repeat_soft/std': 0.023419933393597603, 'rewards/judge_quality/mean': 0.33125001192092896, 'rewards/judge_quality/std': 0.1738995909690857, 'rewards/total_composite/mean': 0.4068011939525604, 'rewards/total_composite/std': 0.2627447545528412, 'reward': 0.4068011939525604, 'reward_std': 0.2627447247505188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15849895775318146, 'sampling/sampling_logp_difference/max': 1.3615283966064453, 'sampling/importance_sampling_ratio/min': 0.25626879930496216, 'sampling/importance_sampling_ratio/mean': 1.0268831253051758, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9701281487941742, 'clip_ratio/low_mean': 0.012019230984151363, 'clip_ratio/low_min': 0.012019230984151363, 'clip_ratio/high_mean': 0.10455992724746466, 'clip_ratio/high_max': 0.10455992724746466, 'clip_ratio/region_mean': 0.11657915823161602, 'reward_total_mean': 0.4068011939525604, 'reward_meter_mean': 0.665990948677063, 'reward_meter_std': 0.4026106297969818, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9687302112579346, 'reward_repeat_soft_std': 0.023419933393597603, 'reward_judge_quality_mean': 0.33125001192092896, 'reward_judge_quality_std': 0.1738995909690857, 'reward_total_composite_mean': 0.4068011939525604, 'reward_total_composite_std': 0.2627447545528412, 'epoch': 0.09}

 26%|██▌       | 863/3300 [2:11:17<6:38:24,  9.81s/it]INFO 04-13 09:32:15 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 864/3300 [2:11:28<6:57:40, 10.29s/it]2026-04-13 09:32:26,667 | INFO | train_grpo_train | metrics_logged mode=train step=864

                                                      
{'loss': -0.0993, 'grad_norm': 4.113894939422607, 'learning_rate': 7.384848484848486e-06, 'num_tokens': 1526937.0, 'completions/mean_length': 94.125, 'completions/min_length': 27.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.42857360839844, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6087803840637207, 'rewards/meter/std': 0.2994439899921417, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9791477918624878, 'rewards/repeat_soft/std': 0.018508000299334526, 'rewards/judge_quality/mean': 0.8149999976158142, 'rewards/judge_quality/std': 0.30928489565849304, 'rewards/total_composite/mean': 0.6250215768814087, 'rewards/total_composite/std': 0.3098732531070709, 'reward': 0.6250215768814087, 'reward_std': 0.3098732531070709, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1697976291179657, 'sampling/sampling_logp_difference/max': 1.9107728004455566, 'sampling/importance_sampling_ratio/min': 0.14796599745750427, 'sampling/importance_sampling_ratio/mean': 0.9970444440841675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8175250664353371, 'clip_ratio/low_mean': 0.02083333395421505, 'clip_ratio/low_min': 0.02083333395421505, 'clip_ratio/high_mean': 0.08237126097083092, 'clip_ratio/high_max': 0.08237126097083092, 'clip_ratio/region_mean': 0.10320459492504597, 'reward_total_mean': 0.6250215768814087, 'reward_meter_mean': 0.6087803840637207, 'reward_meter_std': 0.2994439899921417, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9791477918624878, 'reward_repeat_soft_std': 0.018508000299334526, 'reward_judge_quality_mean': 0.8149999976158142, 'reward_judge_quality_std': 0.30928489565849304, 'reward_total_composite_mean': 0.6250215768814087, 'reward_total_composite_std': 0.3098732531070709, 'epoch': 0.09}

 26%|██▌       | 864/3300 [2:11:28<6:57:40, 10.29s/it]INFO 04-13 09:32:26 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 865/3300 [2:11:39<7:04:54, 10.47s/it]2026-04-13 09:32:37,562 | INFO | train_grpo_train | metrics_logged mode=train step=865

                                                      
{'loss': -0.0878, 'grad_norm': 2.10207462310791, 'learning_rate': 7.381818181818182e-06, 'num_tokens': 1528418.0, 'completions/mean_length': 159.125, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.613077700138092, 'rewards/meter/std': 0.496181458234787, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9650115966796875, 'rewards/repeat_soft/std': 0.026163186877965927, 'rewards/judge_quality/mean': 0.4012500047683716, 'rewards/judge_quality/std': 0.27351874113082886, 'rewards/total_composite/mean': 0.4715662896633148, 'rewards/total_composite/std': 0.33212050795555115, 'reward': 0.4715662896633148, 'reward_std': 0.33212053775787354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11449142545461655, 'sampling/sampling_logp_difference/max': 1.433730959892273, 'sampling/importance_sampling_ratio/min': 0.23841774463653564, 'sampling/importance_sampling_ratio/mean': 1.0337467193603516, 'sampling/importance_sampling_ratio/max': 1.96662175655365, 'entropy': 0.6000302210450172, 'clip_ratio/low_mean': 0.01944444514811039, 'clip_ratio/low_min': 0.01944444514811039, 'clip_ratio/high_mean': 0.07210694067180157, 'clip_ratio/high_max': 0.07210694067180157, 'clip_ratio/region_mean': 0.09155138581991196, 'reward_total_mean': 0.4715662896633148, 'reward_meter_mean': 0.613077700138092, 'reward_meter_std': 0.496181458234787, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9650115966796875, 'reward_repeat_soft_std': 0.026163186877965927, 'reward_judge_quality_mean': 0.4012500047683716, 'reward_judge_quality_std': 0.27351874113082886, 'reward_total_composite_mean': 0.4715662896633148, 'reward_total_composite_std': 0.33212050795555115, 'epoch': 0.09}

 26%|██▌       | 865/3300 [2:11:39<7:04:54, 10.47s/it]INFO 04-13 09:32:37 [block_pool.py:378] Successfully reset prefix cache

 26%|██▌       | 866/3300 [2:11:45<6:10:14,  9.13s/it]2026-04-13 09:32:43,556 | INFO | train_grpo_train | metrics_logged mode=train step=866

                                                      
{'loss': -0.0487, 'grad_norm': 16.4656925201416, 'learning_rate': 7.378787878787879e-06, 'num_tokens': 1530311.0, 'completions/mean_length': 40.625, 'completions/min_length': 31.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8736456632614136, 'rewards/meter/std': 0.2258608043193817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.922467827796936, 'rewards/repeat_soft/std': 0.048922378569841385, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6859979033470154, 'rewards/total_composite/std': 0.1750345379114151, 'reward': 0.6859979033470154, 'reward_std': 0.1750345379114151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18711188435554504, 'sampling/sampling_logp_difference/max': 1.5942096710205078, 'sampling/importance_sampling_ratio/min': 0.20306895673274994, 'sampling/importance_sampling_ratio/mean': 1.008025884628296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.359382875263691, 'clip_ratio/low_mean': 0.0766486581414938, 'clip_ratio/low_min': 0.0766486581414938, 'clip_ratio/high_mean': 0.06952030956745148, 'clip_ratio/high_max': 0.06952030956745148, 'clip_ratio/region_mean': 0.14616896770894527, 'reward_total_mean': 0.6859979033470154, 'reward_meter_mean': 0.8736456632614136, 'reward_meter_std': 0.2258608043193817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.922467827796936, 'reward_repeat_soft_std': 0.048922378569841385, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6859979033470154, 'reward_total_composite_std': 0.1750345379114151, 'epoch': 0.09}

 26%|██▌       | 866/3300 [2:11:45<6:10:14,  9.13s/it]INFO 04-13 09:32:43 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 867/3300 [2:11:52<5:35:37,  8.28s/it]2026-04-13 09:32:49,850 | INFO | train_grpo_train | metrics_logged mode=train step=867

                                                      
{'loss': -0.0763, 'grad_norm': 10.90440559387207, 'learning_rate': 7.375757575757576e-06, 'num_tokens': 1532103.0, 'completions/mean_length': 58.0, 'completions/min_length': 47.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8593844175338745, 'rewards/meter/std': 0.3025902509689331, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9123940467834473, 'rewards/repeat_soft/std': 0.07598268985748291, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5321553945541382, 'rewards/total_composite/std': 0.09145176410675049, 'reward': 0.5321553945541382, 'reward_std': 0.09145176410675049, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1319979429244995, 'sampling/sampling_logp_difference/max': 1.164151668548584, 'sampling/importance_sampling_ratio/min': 0.31218740344047546, 'sampling/importance_sampling_ratio/mean': 1.0285383462905884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.027083732187748, 'clip_ratio/low_mean': 0.029255319386720657, 'clip_ratio/low_min': 0.029255319386720657, 'clip_ratio/high_mean': 0.13843377400189638, 'clip_ratio/high_max': 0.13843377400189638, 'clip_ratio/region_mean': 0.16768909338861704, 'reward_total_mean': 0.5321553945541382, 'reward_meter_mean': 0.8593844175338745, 'reward_meter_std': 0.3025902509689331, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9123940467834473, 'reward_repeat_soft_std': 0.07598268985748291, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5321553945541382, 'reward_total_composite_std': 0.09145176410675049, 'epoch': 0.09}

 26%|██▋       | 867/3300 [2:11:52<5:35:37,  8.28s/it]INFO 04-13 09:32:50 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 868/3300 [2:12:03<6:09:23,  9.11s/it]2026-04-13 09:33:00,915 | INFO | train_grpo_train | metrics_logged mode=train step=868

                                                      
{'loss': -0.137, 'grad_norm': 3.0424578189849854, 'learning_rate': 7.372727272727274e-06, 'num_tokens': 1533748.0, 'completions/mean_length': 107.625, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.85714340209961, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8029947280883789, 'rewards/meter/std': 0.3323410749435425, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8788893818855286, 'rewards/repeat_soft/std': 0.2116212695837021, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.4934343099594116, 'rewards/total_composite/std': 0.21000738441944122, 'reward': 0.4934343099594116, 'reward_std': 0.21000738441944122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16232752799987793, 'sampling/sampling_logp_difference/max': 1.2833247184753418, 'sampling/importance_sampling_ratio/min': 0.29169297218322754, 'sampling/importance_sampling_ratio/mean': 1.0410609245300293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0614218637347221, 'clip_ratio/low_mean': 0.03357380721718073, 'clip_ratio/low_min': 0.03357380721718073, 'clip_ratio/high_mean': 0.10312587767839432, 'clip_ratio/high_max': 0.10312587767839432, 'clip_ratio/region_mean': 0.13669968489557505, 'reward_total_mean': 0.4934343099594116, 'reward_meter_mean': 0.8029947280883789, 'reward_meter_std': 0.3323410749435425, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8788893818855286, 'reward_repeat_soft_std': 0.2116212695837021, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.4934343099594116, 'reward_total_composite_std': 0.21000738441944122, 'epoch': 0.09}

 26%|██▋       | 868/3300 [2:12:03<6:09:23,  9.11s/it]INFO 04-13 09:33:01 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 869/3300 [2:12:08<5:27:37,  8.09s/it]2026-04-13 09:33:06,603 | INFO | train_grpo_train | metrics_logged mode=train step=869

                                                      
{'loss': -0.2605, 'grad_norm': 16.398597717285156, 'learning_rate': 7.36969696969697e-06, 'num_tokens': 1535196.0, 'completions/mean_length': 24.0, 'completions/min_length': 7.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 7.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.503116250038147, 'rewards/meter/std': 0.44779786467552185, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8853303790092468, 'rewards/repeat_soft/std': 0.1647690385580063, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.38540714979171753, 'rewards/total_composite/std': 0.26252666115760803, 'reward': 0.38540714979171753, 'reward_std': 0.26252666115760803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1784038543701172, 'sampling/sampling_logp_difference/max': 1.265294075012207, 'sampling/importance_sampling_ratio/min': 0.2821563184261322, 'sampling/importance_sampling_ratio/mean': 1.0248316526412964, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2060170471668243, 'clip_ratio/low_mean': 0.03300865925848484, 'clip_ratio/low_min': 0.03300865925848484, 'clip_ratio/high_mean': 0.0912698432803154, 'clip_ratio/high_max': 0.0912698432803154, 'clip_ratio/region_mean': 0.12427850253880024, 'reward_total_mean': 0.38540714979171753, 'reward_meter_mean': 0.503116250038147, 'reward_meter_std': 0.44779786467552185, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8853303790092468, 'reward_repeat_soft_std': 0.1647690385580063, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.38540714979171753, 'reward_total_composite_std': 0.26252666115760803, 'epoch': 0.09}

 26%|██▋       | 869/3300 [2:12:08<5:27:37,  8.09s/it]INFO 04-13 09:33:06 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 870/3300 [2:12:21<6:18:42,  9.35s/it]2026-04-13 09:33:18,904 | INFO | train_grpo_train | metrics_logged mode=train step=870

                                                      
{'loss': -0.0392, 'grad_norm': 1.135308861732483, 'learning_rate': 7.3666666666666676e-06, 'num_tokens': 1536690.0, 'completions/mean_length': 396.75, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.75, 'completions/mean_terminated_length': 51.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.2545756697654724, 'rewards/meter/std': 0.32993805408477783, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 0.5, 'rewards/hard_gate/std': 0.5345224738121033, 'rewards/repeat_soft/mean': 0.9848628044128418, 'rewards/repeat_soft/std': 0.016018152236938477, 'rewards/judge_quality/mean': 0.14249999821186066, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.1625368744134903, 'rewards/total_composite/std': 0.18057850003242493, 'reward': 0.1625368744134903, 'reward_std': 0.18057848513126373, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1754874289035797, 'sampling/sampling_logp_difference/max': 1.0462274551391602, 'sampling/importance_sampling_ratio/min': 0.35126039385795593, 'sampling/importance_sampling_ratio/mean': 1.0462087392807007, 'sampling/importance_sampling_ratio/max': 1.856386423110962, 'entropy': 0.4270182102918625, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.043608009815216064, 'clip_ratio/high_max': 0.043608009815216064, 'clip_ratio/region_mean': 0.043608009815216064, 'reward_total_mean': 0.1625368744134903, 'reward_meter_mean': 0.2545756697654724, 'reward_meter_std': 0.32993805408477783, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 0.5, 'reward_hard_gate_std': 0.5345224738121033, 'reward_repeat_soft_mean': 0.9848628044128418, 'reward_repeat_soft_std': 0.016018152236938477, 'reward_judge_quality_mean': 0.14249999821186066, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.1625368744134903, 'reward_total_composite_std': 0.18057850003242493, 'epoch': 0.09}

 26%|██▋       | 870/3300 [2:12:21<6:18:42,  9.35s/it]INFO 04-13 09:33:19 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 871/3300 [2:12:27<5:47:11,  8.58s/it]2026-04-13 09:33:25,673 | INFO | train_grpo_train | metrics_logged mode=train step=871

                                                      
{'loss': 0.0146, 'grad_norm': 9.108240127563477, 'learning_rate': 7.363636363636364e-06, 'num_tokens': 1539043.0, 'completions/mean_length': 95.125, 'completions/min_length': 77.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.125, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.6485616564750671, 'rewards/meter/std': 0.36094608902931213, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8932536840438843, 'rewards/repeat_soft/std': 0.05243608355522156, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43055039644241333, 'rewards/total_composite/std': 0.19202134013175964, 'reward': 0.43055039644241333, 'reward_std': 0.19202131032943726, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16381889581680298, 'sampling/sampling_logp_difference/max': 2.3126883506774902, 'sampling/importance_sampling_ratio/min': 0.09899476170539856, 'sampling/importance_sampling_ratio/mean': 1.0176433324813843, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0679646357893944, 'clip_ratio/low_mean': 0.04217723850160837, 'clip_ratio/low_min': 0.04217723850160837, 'clip_ratio/high_mean': 0.09624836035072803, 'clip_ratio/high_max': 0.09624836035072803, 'clip_ratio/region_mean': 0.1384255988523364, 'reward_total_mean': 0.43055039644241333, 'reward_meter_mean': 0.6485616564750671, 'reward_meter_std': 0.36094608902931213, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8932536840438843, 'reward_repeat_soft_std': 0.05243608355522156, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43055039644241333, 'reward_total_composite_std': 0.19202134013175964, 'epoch': 0.09}

 26%|██▋       | 871/3300 [2:12:27<5:47:11,  8.58s/it]INFO 04-13 09:33:25 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 872/3300 [2:12:34<5:17:51,  7.85s/it]2026-04-13 09:33:31,848 | INFO | train_grpo_train | metrics_logged mode=train step=872

                                                      
{'loss': -0.0083, 'grad_norm': 17.871868133544922, 'learning_rate': 7.360606060606061e-06, 'num_tokens': 1540625.0, 'completions/mean_length': 36.75, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.327788382768631, 'rewards/meter/std': 0.35086458921432495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9688414335250854, 'rewards/repeat_soft/std': 0.0360298790037632, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.22984080016613007, 'rewards/total_composite/mean': 0.459040105342865, 'rewards/total_composite/std': 0.12171979993581772, 'reward': 0.459040105342865, 'reward_std': 0.12171980738639832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18134641647338867, 'sampling/sampling_logp_difference/max': 1.570383071899414, 'sampling/importance_sampling_ratio/min': 0.20796550810337067, 'sampling/importance_sampling_ratio/mean': 1.0140125751495361, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.123292826116085, 'clip_ratio/low_mean': 0.07075076550245285, 'clip_ratio/low_min': 0.07075076550245285, 'clip_ratio/high_mean': 0.061073604971170425, 'clip_ratio/high_max': 0.061073604971170425, 'clip_ratio/region_mean': 0.13182437047362328, 'reward_total_mean': 0.459040105342865, 'reward_meter_mean': 0.327788382768631, 'reward_meter_std': 0.35086458921432495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9688414335250854, 'reward_repeat_soft_std': 0.0360298790037632, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.22984080016613007, 'reward_total_composite_mean': 0.459040105342865, 'reward_total_composite_std': 0.12171979993581772, 'epoch': 0.09}

 26%|██▋       | 872/3300 [2:12:34<5:17:51,  7.85s/it]INFO 04-13 09:33:32 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 873/3300 [2:12:45<6:04:00,  9.00s/it]2026-04-13 09:33:43,515 | INFO | train_grpo_train | metrics_logged mode=train step=873

                                                      
{'loss': -0.1484, 'grad_norm': 3.168926477432251, 'learning_rate': 7.357575757575758e-06, 'num_tokens': 1542634.0, 'completions/mean_length': 126.125, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9259101748466492, 'rewards/meter/std': 0.0754835233092308, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9551552534103394, 'rewards/repeat_soft/std': 0.029804719612002373, 'rewards/judge_quality/mean': 0.6937500238418579, 'rewards/judge_quality/std': 0.3114453852176666, 'rewards/total_composite/mean': 0.6748001575469971, 'rewards/total_composite/std': 0.29786404967308044, 'reward': 0.6748001575469971, 'reward_std': 0.29786401987075806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16060665249824524, 'sampling/sampling_logp_difference/max': 1.4686574935913086, 'sampling/importance_sampling_ratio/min': 0.23023438453674316, 'sampling/importance_sampling_ratio/mean': 1.017259120941162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0811131298542023, 'clip_ratio/low_mean': 0.035211266949772835, 'clip_ratio/low_min': 0.035211266949772835, 'clip_ratio/high_mean': 0.10306689236313105, 'clip_ratio/high_max': 0.10306689236313105, 'clip_ratio/region_mean': 0.13827815931290388, 'reward_total_mean': 0.6748001575469971, 'reward_meter_mean': 0.9259101748466492, 'reward_meter_std': 0.0754835233092308, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9551552534103394, 'reward_repeat_soft_std': 0.029804719612002373, 'reward_judge_quality_mean': 0.6937500238418579, 'reward_judge_quality_std': 0.3114453852176666, 'reward_total_composite_mean': 0.6748001575469971, 'reward_total_composite_std': 0.29786404967308044, 'epoch': 0.09}

 26%|██▋       | 873/3300 [2:12:45<6:04:00,  9.00s/it]INFO 04-13 09:33:43 [block_pool.py:378] Successfully reset prefix cache

 26%|██▋       | 874/3300 [2:12:52<5:31:42,  8.20s/it]2026-04-13 09:33:49,864 | INFO | train_grpo_train | metrics_logged mode=train step=874

                                                      
{'loss': 0.0187, 'grad_norm': 12.982820510864258, 'learning_rate': 7.354545454545456e-06, 'num_tokens': 1544620.0, 'completions/mean_length': 65.25, 'completions/min_length': 54.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.5792921781539917, 'rewards/meter/std': 0.3442385494709015, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9295393228530884, 'rewards/repeat_soft/std': 0.05003296583890915, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.20642796158790588, 'rewards/total_composite/mean': 0.530864953994751, 'rewards/total_composite/std': 0.17147348821163177, 'reward': 0.530864953994751, 'reward_std': 0.17147347331047058, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16094931960105896, 'sampling/sampling_logp_difference/max': 2.14949369430542, 'sampling/importance_sampling_ratio/min': 0.11654314398765564, 'sampling/importance_sampling_ratio/mean': 1.0058163404464722, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.842586986720562, 'clip_ratio/low_mean': 0.053718000650405884, 'clip_ratio/low_min': 0.053718000650405884, 'clip_ratio/high_mean': 0.08179479092359543, 'clip_ratio/high_max': 0.08179479092359543, 'clip_ratio/region_mean': 0.1355127915740013, 'reward_total_mean': 0.530864953994751, 'reward_meter_mean': 0.5792921781539917, 'reward_meter_std': 0.3442385494709015, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9295393228530884, 'reward_repeat_soft_std': 0.05003296583890915, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.20642796158790588, 'reward_total_composite_mean': 0.530864953994751, 'reward_total_composite_std': 0.17147348821163177, 'epoch': 0.09}

 26%|██▋       | 874/3300 [2:12:52<5:31:42,  8.20s/it]INFO 04-13 09:33:50 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 875/3300 [2:13:06<6:45:05, 10.02s/it]2026-04-13 09:34:04,129 | INFO | train_grpo_train | metrics_logged mode=train step=875

                                                      
{'loss': 0.0378, 'grad_norm': 12.344579696655273, 'learning_rate': 7.351515151515151e-06, 'num_tokens': 1546339.0, 'completions/mean_length': 46.875, 'completions/min_length': 45.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8060281872749329, 'rewards/meter/std': 0.2820912301540375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9610169529914856, 'rewards/repeat_soft/std': 0.038200557231903076, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6316392421722412, 'rewards/total_composite/std': 0.1540820449590683, 'reward': 0.6316392421722412, 'reward_std': 0.1540820151567459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15313468873500824, 'sampling/sampling_logp_difference/max': 2.062655448913574, 'sampling/importance_sampling_ratio/min': 0.12711596488952637, 'sampling/importance_sampling_ratio/mean': 1.0250355005264282, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0811816975474358, 'clip_ratio/low_mean': 0.09521636413410306, 'clip_ratio/low_min': 0.09521636413410306, 'clip_ratio/high_mean': 0.02960222028195858, 'clip_ratio/high_max': 0.02960222028195858, 'clip_ratio/region_mean': 0.12481858441606164, 'reward_total_mean': 0.6316392421722412, 'reward_meter_mean': 0.8060281872749329, 'reward_meter_std': 0.2820912301540375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9610169529914856, 'reward_repeat_soft_std': 0.038200557231903076, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6316392421722412, 'reward_total_composite_std': 0.1540820449590683, 'epoch': 0.09}

 27%|██▋       | 875/3300 [2:13:06<6:45:05, 10.02s/it]INFO 04-13 09:34:04 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 876/3300 [2:13:17<7:04:15, 10.50s/it]2026-04-13 09:34:15,746 | INFO | train_grpo_train | metrics_logged mode=train step=876

                                                      
{'loss': -0.1188, 'grad_norm': 2.8642337322235107, 'learning_rate': 7.348484848484849e-06, 'num_tokens': 1548107.0, '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': 48.0, 'rewards/meter/mean': 0.8618472814559937, 'rewards/meter/std': 0.13803237676620483, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.87885582447052, 'rewards/repeat_soft/std': 0.11330028623342514, 'rewards/judge_quality/mean': 0.4300000071525574, 'rewards/judge_quality/std': 0.24454039335250854, 'rewards/total_composite/mean': 0.5300813913345337, 'rewards/total_composite/std': 0.23540335893630981, 'reward': 0.5300813913345337, 'reward_std': 0.23540334403514862, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12949153780937195, 'sampling/sampling_logp_difference/max': 1.6033679246902466, 'sampling/importance_sampling_ratio/min': 0.2012176811695099, 'sampling/importance_sampling_ratio/mean': 1.0226432085037231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6162024512887001, 'clip_ratio/low_mean': 0.027820121496915817, 'clip_ratio/low_min': 0.027820121496915817, 'clip_ratio/high_mean': 0.0726192258298397, 'clip_ratio/high_max': 0.0726192258298397, 'clip_ratio/region_mean': 0.10043934732675552, 'reward_total_mean': 0.5300813913345337, 'reward_meter_mean': 0.8618472814559937, 'reward_meter_std': 0.13803237676620483, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.87885582447052, 'reward_repeat_soft_std': 0.11330028623342514, 'reward_judge_quality_mean': 0.4300000071525574, 'reward_judge_quality_std': 0.24454039335250854, 'reward_total_composite_mean': 0.5300813913345337, 'reward_total_composite_std': 0.23540335893630981, 'epoch': 0.09}

 27%|██▋       | 876/3300 [2:13:17<7:04:15, 10.50s/it]INFO 04-13 09:34:15 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 877/3300 [2:13:28<7:10:43, 10.67s/it]2026-04-13 09:34:26,800 | INFO | train_grpo_train | metrics_logged mode=train step=877

                                                      
{'loss': -0.0845, 'grad_norm': 3.6687779426574707, 'learning_rate': 7.345454545454546e-06, 'num_tokens': 1549526.0, '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': 50.0, 'rewards/meter/mean': 0.439591646194458, 'rewards/meter/std': 0.34563952684402466, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8043549060821533, 'rewards/repeat_soft/std': 0.12364349514245987, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.3206244111061096, 'rewards/total_composite/mean': 0.47407668828964233, 'rewards/total_composite/std': 0.268888384103775, 'reward': 0.47407668828964233, 'reward_std': 0.268888384103775, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09451233595609665, 'sampling/sampling_logp_difference/max': 1.1827497482299805, 'sampling/importance_sampling_ratio/min': 0.4081783890724182, 'sampling/importance_sampling_ratio/mean': 1.0186411142349243, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5319491885602474, 'clip_ratio/low_mean': 0.0352316303178668, 'clip_ratio/low_min': 0.0352316303178668, 'clip_ratio/high_mean': 0.04231360787525773, 'clip_ratio/high_max': 0.04231360787525773, 'clip_ratio/region_mean': 0.07754523819312453, 'reward_total_mean': 0.47407668828964233, 'reward_meter_mean': 0.439591646194458, 'reward_meter_std': 0.34563952684402466, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8043549060821533, 'reward_repeat_soft_std': 0.12364349514245987, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.3206244111061096, 'reward_total_composite_mean': 0.47407668828964233, 'reward_total_composite_std': 0.268888384103775, 'epoch': 0.09}

 27%|██▋       | 877/3300 [2:13:29<7:10:43, 10.67s/it]INFO 04-13 09:34:27 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 878/3300 [2:13:37<6:41:49,  9.95s/it]2026-04-13 09:34:35,093 | INFO | train_grpo_train | metrics_logged mode=train step=878

                                                      
{'loss': 0.888, 'grad_norm': 15.310219764709473, 'learning_rate': 7.342424242424243e-06, 'num_tokens': 1551335.0, 'completions/mean_length': 59.125, 'completions/min_length': 33.0, 'completions/max_length': 219.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 219.0, 'rewards/meter/mean': 0.3386475443840027, 'rewards/meter/std': 0.3047406077384949, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8161276578903198, 'rewards/repeat_soft/std': 0.26778846979141235, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.3847002387046814, 'rewards/total_composite/std': 0.1626531183719635, 'reward': 0.3847002387046814, 'reward_std': 0.1626531183719635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12099866569042206, 'sampling/sampling_logp_difference/max': 2.6160330772399902, 'sampling/importance_sampling_ratio/min': 0.07309223711490631, 'sampling/importance_sampling_ratio/mean': 1.0092995166778564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0028022192418575, 'clip_ratio/low_mean': 0.02445640368387103, 'clip_ratio/low_min': 0.02445640368387103, 'clip_ratio/high_mean': 0.11491053458303213, 'clip_ratio/high_max': 0.11491053458303213, 'clip_ratio/region_mean': 0.13936693826690316, 'reward_total_mean': 0.3847002387046814, 'reward_meter_mean': 0.3386475443840027, 'reward_meter_std': 0.3047406077384949, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8161276578903198, 'reward_repeat_soft_std': 0.26778846979141235, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.3847002387046814, 'reward_total_composite_std': 0.1626531183719635, 'epoch': 0.09}

 27%|██▋       | 878/3300 [2:13:37<6:41:49,  9.95s/it]INFO 04-13 09:34:35 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 879/3300 [2:13:43<5:54:22,  8.78s/it]2026-04-13 09:34:41,139 | INFO | train_grpo_train | metrics_logged mode=train step=879

                                                      
{'loss': 0.0035, 'grad_norm': 15.814071655273438, 'learning_rate': 7.3393939393939395e-06, 'num_tokens': 1552783.0, 'completions/mean_length': 39.0, 'completions/min_length': 31.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.34972092509269714, 'rewards/meter/std': 0.42591872811317444, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9957960844039917, 'rewards/repeat_soft/std': 0.006417128257453442, 'rewards/judge_quality/mean': 0.6812499761581421, 'rewards/judge_quality/std': 0.2231871634721756, 'rewards/total_composite/mean': 0.48811185359954834, 'rewards/total_composite/std': 0.2064460664987564, 'reward': 0.48811185359954834, 'reward_std': 0.2064460664987564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1526462584733963, 'sampling/sampling_logp_difference/max': 2.4089221954345703, 'sampling/importance_sampling_ratio/min': 0.2115991711616516, 'sampling/importance_sampling_ratio/mean': 1.0094772577285767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7610408887267113, 'clip_ratio/low_mean': 0.1086641252040863, 'clip_ratio/low_min': 0.1086641252040863, 'clip_ratio/high_mean': 0.054252199828624725, 'clip_ratio/high_max': 0.054252199828624725, 'clip_ratio/region_mean': 0.16291632503271103, 'reward_total_mean': 0.48811185359954834, 'reward_meter_mean': 0.34972092509269714, 'reward_meter_std': 0.42591872811317444, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9957960844039917, 'reward_repeat_soft_std': 0.006417128257453442, 'reward_judge_quality_mean': 0.6812499761581421, 'reward_judge_quality_std': 0.2231871634721756, 'reward_total_composite_mean': 0.48811185359954834, 'reward_total_composite_std': 0.2064460664987564, 'epoch': 0.09}

 27%|██▋       | 879/3300 [2:13:43<5:54:22,  8.78s/it]INFO 04-13 09:34:41 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 880/3300 [2:13:50<5:30:17,  8.19s/it]2026-04-13 09:34:47,978 | INFO | train_grpo_train | metrics_logged mode=train step=880

                                                      
{'loss': 0.0044, 'grad_norm': 10.868603706359863, 'learning_rate': 7.336363636363637e-06, 'num_tokens': 1554682.0, 'completions/mean_length': 72.375, 'completions/min_length': 64.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9660510420799255, 'rewards/meter/std': 0.03841925039887428, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8535215258598328, 'rewards/repeat_soft/std': 0.12191952764987946, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.631879448890686, 'rewards/total_composite/std': 0.12384099513292313, 'reward': 0.631879448890686, 'reward_std': 0.12384098768234253, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15149742364883423, 'sampling/sampling_logp_difference/max': 2.774995803833008, 'sampling/importance_sampling_ratio/min': 0.06234974041581154, 'sampling/importance_sampling_ratio/mean': 1.021327018737793, 'sampling/importance_sampling_ratio/max': 1.9698742628097534, 'entropy': 1.0442497804760933, 'clip_ratio/low_mean': 0.09328743256628513, 'clip_ratio/low_min': 0.09328743256628513, 'clip_ratio/high_mean': 0.01666666753590107, 'clip_ratio/high_max': 0.01666666753590107, 'clip_ratio/region_mean': 0.1099541001021862, 'reward_total_mean': 0.631879448890686, 'reward_meter_mean': 0.9660510420799255, 'reward_meter_std': 0.03841925039887428, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8535215258598328, 'reward_repeat_soft_std': 0.12191952764987946, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.631879448890686, 'reward_total_composite_std': 0.12384099513292313, 'epoch': 0.09}

 27%|██▋       | 880/3300 [2:13:50<5:30:17,  8.19s/it]INFO 04-13 09:34:48 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 881/3300 [2:13:57<5:23:53,  8.03s/it]2026-04-13 09:34:55,649 | INFO | train_grpo_train | metrics_logged mode=train step=881

                                                      
{'loss': -0.0457, 'grad_norm': 8.910069465637207, 'learning_rate': 7.333333333333333e-06, 'num_tokens': 1557133.0, 'completions/mean_length': 78.375, 'completions/min_length': 66.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.9773383140563965, 'rewards/meter/std': 0.019956661388278008, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8784147500991821, 'rewards/repeat_soft/std': 0.08408968150615692, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5673255920410156, 'rewards/total_composite/std': 0.019805608317255974, 'reward': 0.5673255920410156, 'reward_std': 0.019805599004030228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1358868032693863, 'sampling/sampling_logp_difference/max': 1.8858757019042969, 'sampling/importance_sampling_ratio/min': 0.253724604845047, 'sampling/importance_sampling_ratio/mean': 1.027084469795227, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1675440669059753, 'clip_ratio/low_mean': 0.06719556078314781, 'clip_ratio/low_min': 0.06719556078314781, 'clip_ratio/high_mean': 0.055186279118061066, 'clip_ratio/high_max': 0.055186279118061066, 'clip_ratio/region_mean': 0.12238183990120888, 'reward_total_mean': 0.5673255920410156, 'reward_meter_mean': 0.9773383140563965, 'reward_meter_std': 0.019956661388278008, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8784147500991821, 'reward_repeat_soft_std': 0.08408968150615692, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5673255920410156, 'reward_total_composite_std': 0.019805608317255974, 'epoch': 0.09}

 27%|██▋       | 881/3300 [2:13:57<5:23:53,  8.03s/it]INFO 04-13 09:34:56 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 882/3300 [2:14:04<5:12:51,  7.76s/it]2026-04-13 09:35:02,746 | INFO | train_grpo_train | metrics_logged mode=train step=882

                                                      
{'loss': 0.0344, 'grad_norm': 9.84181022644043, 'learning_rate': 7.330303030303031e-06, 'num_tokens': 1559327.0, 'completions/mean_length': 93.25, 'completions/min_length': 76.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.25, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9091979265213013, 'rewards/meter/std': 0.159868061542511, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8407905101776123, 'rewards/repeat_soft/std': 0.06750401109457016, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5818697214126587, 'rewards/total_composite/std': 0.10077501088380814, 'reward': 0.5818697214126587, 'reward_std': 0.10077502578496933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12708139419555664, 'sampling/sampling_logp_difference/max': 3.6555018424987793, 'sampling/importance_sampling_ratio/min': 0.025848522782325745, 'sampling/importance_sampling_ratio/mean': 1.0060118436813354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8384061381220818, 'clip_ratio/low_mean': 0.07911044172942638, 'clip_ratio/low_min': 0.07911044172942638, 'clip_ratio/high_mean': 0.029835755936801434, 'clip_ratio/high_max': 0.029835755936801434, 'clip_ratio/region_mean': 0.10894619766622782, 'reward_total_mean': 0.5818697214126587, 'reward_meter_mean': 0.9091979265213013, 'reward_meter_std': 0.159868061542511, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8407905101776123, 'reward_repeat_soft_std': 0.06750401109457016, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5818697214126587, 'reward_total_composite_std': 0.10077501088380814, 'epoch': 0.09}

 27%|██▋       | 882/3300 [2:14:04<5:12:51,  7.76s/it]INFO 04-13 09:35:02 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 883/3300 [2:14:11<4:55:56,  7.35s/it]2026-04-13 09:35:09,150 | INFO | train_grpo_train | metrics_logged mode=train step=883

                                                      
{'loss': 0.0462, 'grad_norm': 28.119281768798828, 'learning_rate': 7.3272727272727285e-06, 'num_tokens': 1561064.0, 'completions/mean_length': 45.125, 'completions/min_length': 41.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.4719432592391968, 'rewards/meter/std': 0.4077528417110443, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.984849750995636, 'rewards/repeat_soft/std': 0.013406476937234402, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.4829104244709015, 'rewards/total_composite/std': 0.11041922122240067, 'reward': 0.4829104244709015, 'reward_std': 0.11041922122240067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18426677584648132, 'sampling/sampling_logp_difference/max': 1.7394871711730957, 'sampling/importance_sampling_ratio/min': 0.17561043798923492, 'sampling/importance_sampling_ratio/mean': 1.00497305393219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1275340244174004, 'clip_ratio/low_mean': 0.07357970625162125, 'clip_ratio/low_min': 0.07357970625162125, 'clip_ratio/high_mean': 0.0731347594410181, 'clip_ratio/high_max': 0.0731347594410181, 'clip_ratio/region_mean': 0.14671446569263935, 'reward_total_mean': 0.4829104244709015, 'reward_meter_mean': 0.4719432592391968, 'reward_meter_std': 0.4077528417110443, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.984849750995636, 'reward_repeat_soft_std': 0.013406476937234402, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.4829104244709015, 'reward_total_composite_std': 0.11041922122240067, 'epoch': 0.09}

 27%|██▋       | 883/3300 [2:14:11<4:55:56,  7.35s/it]INFO 04-13 09:35:09 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 884/3300 [2:14:22<5:45:33,  8.58s/it]2026-04-13 09:35:20,586 | INFO | train_grpo_train | metrics_logged mode=train step=884

                                                      
{'loss': -0.1798, 'grad_norm': 3.204331874847412, 'learning_rate': 7.324242424242425e-06, 'num_tokens': 1563447.0, 'completions/mean_length': 160.875, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 110.71429443359375, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.6648643612861633, 'rewards/meter/std': 0.432773232460022, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.263523131608963, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7227896451950073, 'rewards/repeat_soft/std': 0.21885265409946442, 'rewards/judge_quality/mean': 0.3399999737739563, 'rewards/judge_quality/std': 0.15052290260791779, 'rewards/total_composite/mean': 0.3958885669708252, 'rewards/total_composite/std': 0.2000521868467331, 'reward': 0.3958885669708252, 'reward_std': 0.2000521868467331, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12298968434333801, 'sampling/sampling_logp_difference/max': 1.3381690979003906, 'sampling/importance_sampling_ratio/min': 0.2623255252838135, 'sampling/importance_sampling_ratio/mean': 1.0034453868865967, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6497861333191395, 'clip_ratio/low_mean': 0.0012886597542092204, 'clip_ratio/low_min': 0.0012886597542092204, 'clip_ratio/high_mean': 0.10129512473940849, 'clip_ratio/high_max': 0.10129512473940849, 'clip_ratio/region_mean': 0.10258378449361771, 'reward_total_mean': 0.3958885669708252, 'reward_meter_mean': 0.6648643612861633, 'reward_meter_std': 0.432773232460022, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.263523131608963, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7227896451950073, 'reward_repeat_soft_std': 0.21885265409946442, 'reward_judge_quality_mean': 0.3399999737739563, 'reward_judge_quality_std': 0.15052290260791779, 'reward_total_composite_mean': 0.3958885669708252, 'reward_total_composite_std': 0.2000521868467331, 'epoch': 0.09}

 27%|██▋       | 884/3300 [2:14:22<5:45:33,  8.58s/it]INFO 04-13 09:35:20 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 885/3300 [2:14:34<6:19:47,  9.44s/it]2026-04-13 09:35:32,044 | INFO | train_grpo_train | metrics_logged mode=train step=885

                                                      
{'loss': -0.093, 'grad_norm': 5.706384658813477, 'learning_rate': 7.321212121212122e-06, 'num_tokens': 1565490.0, 'completions/mean_length': 141.375, 'completions/min_length': 78.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 88.42857360839844, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.7723007202148438, 'rewards/meter/std': 0.30487778782844543, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9757391214370728, 'rewards/repeat_soft/std': 0.018861595541238785, 'rewards/judge_quality/mean': 0.5612499713897705, 'rewards/judge_quality/std': 0.3223324716091156, 'rewards/total_composite/mean': 0.532701849937439, 'rewards/total_composite/std': 0.344864159822464, 'reward': 0.532701849937439, 'reward_std': 0.344864159822464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1807408481836319, 'sampling/sampling_logp_difference/max': 2.2748515605926514, 'sampling/importance_sampling_ratio/min': 0.10281217098236084, 'sampling/importance_sampling_ratio/mean': 1.0091204643249512, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0115512311458588, 'clip_ratio/low_mean': 0.01692708395421505, 'clip_ratio/low_min': 0.01692708395421505, 'clip_ratio/high_mean': 0.15578365325927734, 'clip_ratio/high_max': 0.15578365325927734, 'clip_ratio/region_mean': 0.1727107372134924, 'reward_total_mean': 0.532701849937439, 'reward_meter_mean': 0.7723007202148438, 'reward_meter_std': 0.30487778782844543, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9757391214370728, 'reward_repeat_soft_std': 0.018861595541238785, 'reward_judge_quality_mean': 0.5612499713897705, 'reward_judge_quality_std': 0.3223324716091156, 'reward_total_composite_mean': 0.532701849937439, 'reward_total_composite_std': 0.344864159822464, 'epoch': 0.09}

 27%|██▋       | 885/3300 [2:14:34<6:19:47,  9.44s/it]INFO 04-13 09:35:32 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 886/3300 [2:14:40<5:44:49,  8.57s/it]2026-04-13 09:35:38,596 | INFO | train_grpo_train | metrics_logged mode=train step=886

                                                      
{'loss': 0.1349, 'grad_norm': 22.54247283935547, 'learning_rate': 7.3181818181818186e-06, 'num_tokens': 1567018.0, 'completions/mean_length': 24.0, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.7209341526031494, 'rewards/meter/std': 0.404031366109848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9467312693595886, 'rewards/repeat_soft/std': 0.03042098507285118, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.21224987506866455, 'rewards/total_composite/mean': 0.5589122772216797, 'rewards/total_composite/std': 0.18516050279140472, 'reward': 0.5589122772216797, 'reward_std': 0.18516047298908234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19285623729228973, 'sampling/sampling_logp_difference/max': 1.099189043045044, 'sampling/importance_sampling_ratio/min': 0.33314111828804016, 'sampling/importance_sampling_ratio/mean': 1.0386865139007568, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4280219674110413, 'clip_ratio/low_mean': 0.07509615365415812, 'clip_ratio/low_min': 0.07509615365415812, 'clip_ratio/high_mean': 0.09872482949867845, 'clip_ratio/high_max': 0.09872482949867845, 'clip_ratio/region_mean': 0.17382098315283656, 'reward_total_mean': 0.5589122772216797, 'reward_meter_mean': 0.7209341526031494, 'reward_meter_std': 0.404031366109848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9467312693595886, 'reward_repeat_soft_std': 0.03042098507285118, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.21224987506866455, 'reward_total_composite_mean': 0.5589122772216797, 'reward_total_composite_std': 0.18516050279140472, 'epoch': 0.09}

 27%|██▋       | 886/3300 [2:14:40<5:44:49,  8.57s/it]INFO 04-13 09:35:38 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 887/3300 [2:14:47<5:17:37,  7.90s/it]2026-04-13 09:35:44,894 | INFO | train_grpo_train | metrics_logged mode=train step=887

                                                      
{'loss': -0.0112, 'grad_norm': 22.737993240356445, 'learning_rate': 7.315151515151516e-06, 'num_tokens': 1568391.0, 'completions/mean_length': 27.625, 'completions/min_length': 22.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.5176418423652649, 'rewards/meter/std': 0.4330809414386749, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9610613584518433, 'rewards/repeat_soft/std': 0.004068940877914429, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.522869348526001, 'rewards/total_composite/std': 0.17281579971313477, 'reward': 0.522869348526001, 'reward_std': 0.17281579971313477, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.20167115330696106, 'sampling/sampling_logp_difference/max': 1.5116100311279297, 'sampling/importance_sampling_ratio/min': 0.22055459022521973, 'sampling/importance_sampling_ratio/mean': 1.0319517850875854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.278345376253128, 'clip_ratio/low_mean': 0.07225524540990591, 'clip_ratio/low_min': 0.07225524540990591, 'clip_ratio/high_mean': 0.11453500390052795, 'clip_ratio/high_max': 0.11453500390052795, 'clip_ratio/region_mean': 0.18679024931043386, 'reward_total_mean': 0.522869348526001, 'reward_meter_mean': 0.5176418423652649, 'reward_meter_std': 0.4330809414386749, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9610613584518433, 'reward_repeat_soft_std': 0.004068940877914429, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.522869348526001, 'reward_total_composite_std': 0.17281579971313477, 'epoch': 0.09}

 27%|██▋       | 887/3300 [2:14:47<5:17:37,  7.90s/it]INFO 04-13 09:35:45 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 888/3300 [2:14:54<5:07:31,  7.65s/it]2026-04-13 09:35:51,969 | INFO | train_grpo_train | metrics_logged mode=train step=888

                                                      
{'loss': 0.0598, 'grad_norm': 10.03502082824707, 'learning_rate': 7.312121212121212e-06, 'num_tokens': 1570540.0, 'completions/mean_length': 96.625, 'completions/min_length': 87.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.625, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9755557179450989, 'rewards/meter/std': 0.02128848060965538, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9463392496109009, 'rewards/repeat_soft/std': 0.04638994112610817, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.1505940854549408, 'rewards/total_composite/mean': 0.758562445640564, 'rewards/total_composite/std': 0.09291400015354156, 'reward': 0.758562445640564, 'reward_std': 0.09291400015354156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13130463659763336, 'sampling/sampling_logp_difference/max': 2.592618942260742, 'sampling/importance_sampling_ratio/min': 0.07482381910085678, 'sampling/importance_sampling_ratio/mean': 1.0128514766693115, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.900644063949585, 'clip_ratio/low_mean': 0.021150990389287472, 'clip_ratio/low_min': 0.021150990389287472, 'clip_ratio/high_mean': 0.09297308884561062, 'clip_ratio/high_max': 0.09297308884561062, 'clip_ratio/region_mean': 0.11412407923489809, 'reward_total_mean': 0.758562445640564, 'reward_meter_mean': 0.9755557179450989, 'reward_meter_std': 0.02128848060965538, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9463392496109009, 'reward_repeat_soft_std': 0.04638994112610817, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.1505940854549408, 'reward_total_composite_mean': 0.758562445640564, 'reward_total_composite_std': 0.09291400015354156, 'epoch': 0.09}

 27%|██▋       | 888/3300 [2:14:54<5:07:31,  7.65s/it]INFO 04-13 09:35:52 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 889/3300 [2:15:00<4:50:19,  7.23s/it]2026-04-13 09:35:58,206 | INFO | train_grpo_train | metrics_logged mode=train step=889

                                                      
{'loss': -0.0072, 'grad_norm': 14.553848266601562, 'learning_rate': 7.30909090909091e-06, 'num_tokens': 1572149.0, 'completions/mean_length': 48.125, 'completions/min_length': 39.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.31065088510513306, 'rewards/meter/std': 0.2979934513568878, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9635153412818909, 'rewards/repeat_soft/std': 0.04018327593803406, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.22890658676624298, 'rewards/total_composite/mean': 0.465585321187973, 'rewards/total_composite/std': 0.10353792458772659, 'reward': 0.465585321187973, 'reward_std': 0.1035379096865654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11885638535022736, 'sampling/sampling_logp_difference/max': 1.5418891906738281, 'sampling/importance_sampling_ratio/min': 0.21397648751735687, 'sampling/importance_sampling_ratio/mean': 1.0027971267700195, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5610169097781181, 'clip_ratio/low_mean': 0.0477488050237298, 'clip_ratio/low_min': 0.0477488050237298, 'clip_ratio/high_mean': 0.0557402353733778, 'clip_ratio/high_max': 0.0557402353733778, 'clip_ratio/region_mean': 0.1034890403971076, 'reward_total_mean': 0.465585321187973, 'reward_meter_mean': 0.31065088510513306, 'reward_meter_std': 0.2979934513568878, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9635153412818909, 'reward_repeat_soft_std': 0.04018327593803406, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.22890658676624298, 'reward_total_composite_mean': 0.465585321187973, 'reward_total_composite_std': 0.10353792458772659, 'epoch': 0.09}

 27%|██▋       | 889/3300 [2:15:00<4:50:19,  7.23s/it]INFO 04-13 09:35:58 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 890/3300 [2:15:06<4:42:25,  7.03s/it]2026-04-13 09:36:04,788 | INFO | train_grpo_train | metrics_logged mode=train step=890

                                                      
{'loss': -0.0118, 'grad_norm': 11.42885971069336, 'learning_rate': 7.306060606060607e-06, 'num_tokens': 1573790.0, 'completions/mean_length': 55.125, 'completions/min_length': 50.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.672086238861084, 'rewards/meter/std': 0.4170806109905243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6968518495559692, 'rewards/repeat_soft/std': 0.13996635377407074, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.49258965253829956, 'rewards/total_composite/std': 0.10857101529836655, 'reward': 0.49258965253829956, 'reward_std': 0.10857101529836655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13281306624412537, 'sampling/sampling_logp_difference/max': 1.5750236511230469, 'sampling/importance_sampling_ratio/min': 0.207002654671669, 'sampling/importance_sampling_ratio/mean': 1.0106837749481201, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9696046486496925, 'clip_ratio/low_mean': 0.051568303257226944, 'clip_ratio/low_min': 0.051568303257226944, 'clip_ratio/high_mean': 0.07031287346035242, 'clip_ratio/high_max': 0.07031287346035242, 'clip_ratio/region_mean': 0.12188117671757936, 'reward_total_mean': 0.49258965253829956, 'reward_meter_mean': 0.672086238861084, 'reward_meter_std': 0.4170806109905243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6968518495559692, 'reward_repeat_soft_std': 0.13996635377407074, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.49258965253829956, 'reward_total_composite_std': 0.10857101529836655, 'epoch': 0.09}

 27%|██▋       | 890/3300 [2:15:07<4:42:25,  7.03s/it]INFO 04-13 09:36:05 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 891/3300 [2:15:13<4:31:37,  6.77s/it]2026-04-13 09:36:10,929 | INFO | train_grpo_train | metrics_logged mode=train step=891

                                                      
{'loss': -0.0165, 'grad_norm': 19.228656768798828, 'learning_rate': 7.303030303030304e-06, 'num_tokens': 1575320.0, 'completions/mean_length': 39.25, 'completions/min_length': 32.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.8273656368255615, 'rewards/meter/std': 0.2690790593624115, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9214805364608765, 'rewards/repeat_soft/std': 0.09300963580608368, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.27994900941848755, 'rewards/total_composite/mean': 0.6273754239082336, 'rewards/total_composite/std': 0.28964903950691223, 'reward': 0.6273754239082336, 'reward_std': 0.28964903950691223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15418635308742523, 'sampling/sampling_logp_difference/max': 2.8024725914001465, 'sampling/importance_sampling_ratio/min': 0.06065988913178444, 'sampling/importance_sampling_ratio/mean': 1.008238673210144, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6868798024952412, 'clip_ratio/low_mean': 0.10877399798482656, 'clip_ratio/low_min': 0.10877399798482656, 'clip_ratio/high_mean': 0.05127991968765855, 'clip_ratio/high_max': 0.05127991968765855, 'clip_ratio/region_mean': 0.1600539176724851, 'reward_total_mean': 0.6273754239082336, 'reward_meter_mean': 0.8273656368255615, 'reward_meter_std': 0.2690790593624115, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9214805364608765, 'reward_repeat_soft_std': 0.09300963580608368, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.27994900941848755, 'reward_total_composite_mean': 0.6273754239082336, 'reward_total_composite_std': 0.28964903950691223, 'epoch': 0.09}

 27%|██▋       | 891/3300 [2:15:13<4:31:37,  6.77s/it]INFO 04-13 09:36:11 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 892/3300 [2:15:19<4:25:54,  6.63s/it]2026-04-13 09:36:17,225 | INFO | train_grpo_train | metrics_logged mode=train step=892

                                                      
{'loss': 0.0577, 'grad_norm': 13.683673858642578, 'learning_rate': 7.3e-06, 'num_tokens': 1577103.0, 'completions/mean_length': 47.875, 'completions/min_length': 40.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6726229786872864, 'rewards/meter/std': 0.36972475051879883, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9640060663223267, 'rewards/repeat_soft/std': 0.04401865229010582, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5704629421234131, 'rewards/total_composite/std': 0.1644839495420456, 'reward': 0.5704629421234131, 'reward_std': 0.1644839495420456, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17061322927474976, 'sampling/sampling_logp_difference/max': 1.981736183166504, 'sampling/importance_sampling_ratio/min': 0.1378297358751297, 'sampling/importance_sampling_ratio/mean': 0.9951289892196655, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1203052997589111, 'clip_ratio/low_mean': 0.09358753263950348, 'clip_ratio/low_min': 0.09358753263950348, 'clip_ratio/high_mean': 0.09015802852809429, 'clip_ratio/high_max': 0.09015802852809429, 'clip_ratio/region_mean': 0.18374556116759777, 'reward_total_mean': 0.5704629421234131, 'reward_meter_mean': 0.6726229786872864, 'reward_meter_std': 0.36972475051879883, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9640060663223267, 'reward_repeat_soft_std': 0.04401865229010582, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5704629421234131, 'reward_total_composite_std': 0.1644839495420456, 'epoch': 0.09}

 27%|██▋       | 892/3300 [2:15:19<4:25:54,  6.63s/it]INFO 04-13 09:36:17 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 893/3300 [2:15:25<4:20:30,  6.49s/it]2026-04-13 09:36:23,411 | INFO | train_grpo_train | metrics_logged mode=train step=893

                                                      
{'loss': 0.0362, 'grad_norm': 21.54730224609375, 'learning_rate': 7.296969696969698e-06, 'num_tokens': 1578534.0, 'completions/mean_length': 31.875, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.875, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6040737628936768, 'rewards/meter/std': 0.4211108982563019, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9561255574226379, 'rewards/repeat_soft/std': 0.028681941330432892, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.06902381032705307, 'rewards/total_composite/mean': 0.524419903755188, 'rewards/total_composite/std': 0.12749241292476654, 'reward': 0.524419903755188, 'reward_std': 0.12749241292476654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1891103833913803, 'sampling/sampling_logp_difference/max': 1.9380428791046143, 'sampling/importance_sampling_ratio/min': 0.14398548007011414, 'sampling/importance_sampling_ratio/mean': 1.0233358144760132, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1149725764989853, 'clip_ratio/low_mean': 0.04424442304298282, 'clip_ratio/low_min': 0.04424442304298282, 'clip_ratio/high_mean': 0.10814394149929285, 'clip_ratio/high_max': 0.10814394149929285, 'clip_ratio/region_mean': 0.15238836454227567, 'reward_total_mean': 0.524419903755188, 'reward_meter_mean': 0.6040737628936768, 'reward_meter_std': 0.4211108982563019, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9561255574226379, 'reward_repeat_soft_std': 0.028681941330432892, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.06902381032705307, 'reward_total_composite_mean': 0.524419903755188, 'reward_total_composite_std': 0.12749241292476654, 'epoch': 0.09}

 27%|██▋       | 893/3300 [2:15:25<4:20:30,  6.49s/it]INFO 04-13 09:36:23 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 894/3300 [2:15:31<4:11:45,  6.28s/it]2026-04-13 09:36:29,186 | INFO | train_grpo_train | metrics_logged mode=train step=894

                                                      
{'loss': 0.0664, 'grad_norm': 18.897003173828125, 'learning_rate': 7.293939393939394e-06, 'num_tokens': 1580012.0, 'completions/mean_length': 25.75, 'completions/min_length': 19.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.660660982131958, 'rewards/meter/std': 0.4115402400493622, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9652037620544434, 'rewards/repeat_soft/std': 0.007647482678294182, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.5076031684875488, 'rewards/total_composite/std': 0.3156152367591858, 'reward': 0.5076031684875488, 'reward_std': 0.3156152069568634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18812528252601624, 'sampling/sampling_logp_difference/max': 1.7578144073486328, 'sampling/importance_sampling_ratio/min': 0.17242130637168884, 'sampling/importance_sampling_ratio/mean': 0.9939836263656616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1995478235185146, 'clip_ratio/low_mean': 0.0933492835611105, 'clip_ratio/low_min': 0.0933492835611105, 'clip_ratio/high_mean': 0.11678571626543999, 'clip_ratio/high_max': 0.11678571626543999, 'clip_ratio/region_mean': 0.21013499982655048, 'reward_total_mean': 0.5076031684875488, 'reward_meter_mean': 0.660660982131958, 'reward_meter_std': 0.4115402400493622, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9652037620544434, 'reward_repeat_soft_std': 0.007647482678294182, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5076031684875488, 'reward_total_composite_std': 0.3156152367591858, 'epoch': 0.09}

 27%|██▋       | 894/3300 [2:15:31<4:11:45,  6.28s/it]INFO 04-13 09:36:29 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 895/3300 [2:15:37<4:12:40,  6.30s/it]2026-04-13 09:36:35,548 | INFO | train_grpo_train | metrics_logged mode=train step=895

                                                      
{'loss': 0.0457, 'grad_norm': 19.292024612426758, 'learning_rate': 7.290909090909092e-06, 'num_tokens': 1581940.0, 'completions/mean_length': 72.0, 'completions/min_length': 58.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.280703067779541, 'rewards/meter/std': 0.2876078188419342, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9423829913139343, 'rewards/repeat_soft/std': 0.025699563324451447, 'rewards/judge_quality/mean': 0.5612499713897705, 'rewards/judge_quality/std': 0.25614938139915466, 'rewards/total_composite/mean': 0.4261462688446045, 'rewards/total_composite/std': 0.24164758622646332, 'reward': 0.4261462688446045, 'reward_std': 0.24164758622646332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.158335879445076, 'sampling/sampling_logp_difference/max': 2.8191614151000977, 'sampling/importance_sampling_ratio/min': 0.05965594947338104, 'sampling/importance_sampling_ratio/mean': 1.0279756784439087, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9015419781208038, 'clip_ratio/low_mean': 0.07797630783170462, 'clip_ratio/low_min': 0.07797630783170462, 'clip_ratio/high_mean': 0.05729399900883436, 'clip_ratio/high_max': 0.05729399900883436, 'clip_ratio/region_mean': 0.13527030684053898, 'reward_total_mean': 0.4261462688446045, 'reward_meter_mean': 0.280703067779541, 'reward_meter_std': 0.2876078188419342, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9423829913139343, 'reward_repeat_soft_std': 0.025699563324451447, 'reward_judge_quality_mean': 0.5612499713897705, 'reward_judge_quality_std': 0.25614938139915466, 'reward_total_composite_mean': 0.4261462688446045, 'reward_total_composite_std': 0.24164758622646332, 'epoch': 0.09}

 27%|██▋       | 895/3300 [2:15:37<4:12:40,  6.30s/it]INFO 04-13 09:36:35 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 896/3300 [2:15:43<4:11:27,  6.28s/it]2026-04-13 09:36:41,760 | INFO | train_grpo_train | metrics_logged mode=train step=896

                                                      
{'loss': 0.1358, 'grad_norm': 20.5400447845459, 'learning_rate': 7.287878787878789e-06, 'num_tokens': 1583598.0, 'completions/mean_length': 36.25, 'completions/min_length': 29.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.25, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6870586276054382, 'rewards/meter/std': 0.300336092710495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921655058860779, 'rewards/repeat_soft/std': 0.010928143747150898, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.6126821041107178, 'rewards/total_composite/std': 0.18589404225349426, 'reward': 0.6126821041107178, 'reward_std': 0.18589404225349426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18025875091552734, 'sampling/sampling_logp_difference/max': 2.0302717685699463, 'sampling/importance_sampling_ratio/min': 0.13129983842372894, 'sampling/importance_sampling_ratio/mean': 1.0002015829086304, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.904352992773056, 'clip_ratio/low_mean': 0.10969074815511703, 'clip_ratio/low_min': 0.10969074815511703, 'clip_ratio/high_mean': 0.031870429404079914, 'clip_ratio/high_max': 0.031870429404079914, 'clip_ratio/region_mean': 0.14156117755919695, 'reward_total_mean': 0.6126821041107178, 'reward_meter_mean': 0.6870586276054382, 'reward_meter_std': 0.300336092710495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921655058860779, 'reward_repeat_soft_std': 0.010928143747150898, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.6126821041107178, 'reward_total_composite_std': 0.18589404225349426, 'epoch': 0.09}

 27%|██▋       | 896/3300 [2:15:43<4:11:27,  6.28s/it]INFO 04-13 09:36:42 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 897/3300 [2:15:50<4:15:37,  6.38s/it]2026-04-13 09:36:48,390 | INFO | train_grpo_train | metrics_logged mode=train step=897

                                                      
{'loss': 0.0058, 'grad_norm': 13.415826797485352, 'learning_rate': 7.284848484848486e-06, 'num_tokens': 1585191.0, 'completions/mean_length': 50.125, 'completions/min_length': 42.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8849648237228394, 'rewards/meter/std': 0.11630762368440628, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9979231953620911, 'rewards/repeat_soft/std': 0.004281661007553339, 'rewards/judge_quality/mean': 0.7775000333786011, 'rewards/judge_quality/std': 0.21359175443649292, 'rewards/total_composite/mean': 0.7969489693641663, 'rewards/total_composite/std': 0.1330748051404953, 'reward': 0.7969489693641663, 'reward_std': 0.1330747753381729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1257050335407257, 'sampling/sampling_logp_difference/max': 1.7089028358459473, 'sampling/importance_sampling_ratio/min': 0.18106433749198914, 'sampling/importance_sampling_ratio/mean': 1.0011931657791138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8672522008419037, 'clip_ratio/low_mean': 0.02392399311065674, 'clip_ratio/low_min': 0.02392399311065674, 'clip_ratio/high_mean': 0.09560730308294296, 'clip_ratio/high_max': 0.09560730308294296, 'clip_ratio/region_mean': 0.1195312961935997, 'reward_total_mean': 0.7969489693641663, 'reward_meter_mean': 0.8849648237228394, 'reward_meter_std': 0.11630762368440628, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9979231953620911, 'reward_repeat_soft_std': 0.004281661007553339, 'reward_judge_quality_mean': 0.7775000333786011, 'reward_judge_quality_std': 0.21359175443649292, 'reward_total_composite_mean': 0.7969489693641663, 'reward_total_composite_std': 0.1330748051404953, 'epoch': 0.09}

 27%|██▋       | 897/3300 [2:15:50<4:15:37,  6.38s/it]INFO 04-13 09:36:48 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 898/3300 [2:16:01<5:15:46,  7.89s/it]2026-04-13 09:36:59,789 | INFO | train_grpo_train | metrics_logged mode=train step=898

                                                      
{'loss': -0.06, 'grad_norm': 2.393796920776367, 'learning_rate': 7.281818181818182e-06, 'num_tokens': 1586526.0, 'completions/mean_length': 148.875, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.83333396911621, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.44433021545410156, 'rewards/meter/std': 0.4191552996635437, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8499475717544556, 'rewards/repeat_soft/std': 0.3267112970352173, 'rewards/judge_quality/mean': 0.30124998092651367, 'rewards/judge_quality/std': 0.18349288403987885, 'rewards/total_composite/mean': 0.3642628490924835, 'rewards/total_composite/std': 0.2603684365749359, 'reward': 0.3642628490924835, 'reward_std': 0.2603684067726135, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1771446168422699, 'sampling/sampling_logp_difference/max': 1.3258414268493652, 'sampling/importance_sampling_ratio/min': 0.3767942190170288, 'sampling/importance_sampling_ratio/mean': 1.0350534915924072, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8279794976115227, 'clip_ratio/low_mean': 0.024193547666072845, 'clip_ratio/low_min': 0.024193547666072845, 'clip_ratio/high_mean': 0.10722934734076262, 'clip_ratio/high_max': 0.10722934734076262, 'clip_ratio/region_mean': 0.13142289500683546, 'reward_total_mean': 0.3642628490924835, 'reward_meter_mean': 0.44433021545410156, 'reward_meter_std': 0.4191552996635437, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8499475717544556, 'reward_repeat_soft_std': 0.3267112970352173, 'reward_judge_quality_mean': 0.30124998092651367, 'reward_judge_quality_std': 0.18349288403987885, 'reward_total_composite_mean': 0.3642628490924835, 'reward_total_composite_std': 0.2603684365749359, 'epoch': 0.09}

 27%|██▋       | 898/3300 [2:16:02<5:15:46,  7.89s/it]INFO 04-13 09:37:00 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 899/3300 [2:16:08<4:56:22,  7.41s/it]2026-04-13 09:37:06,073 | INFO | train_grpo_train | metrics_logged mode=train step=899

                                                      
{'loss': 0.0371, 'grad_norm': 11.645864486694336, 'learning_rate': 7.2787878787878795e-06, 'num_tokens': 1588243.0, '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': 0.8100303411483765, 'rewards/meter/std': 0.32428187131881714, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9260385036468506, 'rewards/repeat_soft/std': 0.07360906898975372, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.7051819562911987, 'rewards/total_composite/std': 0.19023671746253967, 'reward': 0.7051819562911987, 'reward_std': 0.19023671746253967, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16436229646205902, 'sampling/sampling_logp_difference/max': 1.3809700012207031, 'sampling/importance_sampling_ratio/min': 0.25133463740348816, 'sampling/importance_sampling_ratio/mean': 1.0121458768844604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1463303491473198, 'clip_ratio/low_mean': 0.12049462832510471, 'clip_ratio/low_min': 0.12049462832510471, 'clip_ratio/high_mean': 0.052040016278624535, 'clip_ratio/high_max': 0.052040016278624535, 'clip_ratio/region_mean': 0.17253464460372925, 'reward_total_mean': 0.7051819562911987, 'reward_meter_mean': 0.8100303411483765, 'reward_meter_std': 0.32428187131881714, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9260385036468506, 'reward_repeat_soft_std': 0.07360906898975372, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.7051819562911987, 'reward_total_composite_std': 0.19023671746253967, 'epoch': 0.09}

 27%|██▋       | 899/3300 [2:16:08<4:56:22,  7.41s/it]INFO 04-13 09:37:06 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 900/3300 [2:16:15<4:51:01,  7.28s/it]2026-04-13 09:37:13,045 | INFO | train_grpo_train | metrics_logged mode=train step=900

                                                      
{'loss': 0.0353, 'grad_norm': 7.044389724731445, 'learning_rate': 7.275757575757576e-06, 'num_tokens': 1590605.0, 'completions/mean_length': 107.25, 'completions/min_length': 92.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.25, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.8630037307739258, 'rewards/meter/std': 0.20947736501693726, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9194906949996948, 'rewards/repeat_soft/std': 0.062074482440948486, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5716500878334045, 'rewards/total_composite/std': 0.12632188200950623, 'reward': 0.5716500878334045, 'reward_std': 0.12632188200950623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13405700027942657, 'sampling/sampling_logp_difference/max': 1.6423745155334473, 'sampling/importance_sampling_ratio/min': 0.1935199648141861, 'sampling/importance_sampling_ratio/mean': 1.0059252977371216, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0670483708381653, 'clip_ratio/low_mean': 0.08348380029201508, 'clip_ratio/low_min': 0.08348380029201508, 'clip_ratio/high_mean': 0.02966239769011736, 'clip_ratio/high_max': 0.02966239769011736, 'clip_ratio/region_mean': 0.11314619798213243, 'reward_total_mean': 0.5716500878334045, 'reward_meter_mean': 0.8630037307739258, 'reward_meter_std': 0.20947736501693726, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9194906949996948, 'reward_repeat_soft_std': 0.062074482440948486, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5716500878334045, 'reward_total_composite_std': 0.12632188200950623, 'epoch': 0.09}

 27%|██▋       | 900/3300 [2:16:15<4:51:01,  7.28s/it]INFO 04-13 09:37:13 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.85s/it]

 30%|███       | 3/10 [00:07<00:18,  2.69s/it]

 40%|████      | 4/10 [00:11<00:18,  3.08s/it]

 50%|█████     | 5/10 [00:15<00:18,  3.60s/it]

 60%|██████    | 6/10 [00:24<00:21,  5.25s/it]

 70%|███████   | 7/10 [00:33<00:19,  6.36s/it]

 80%|████████  | 8/10 [00:37<00:11,  5.59s/it]

 90%|█████████ | 9/10 [00:45<00:06,  6.59s/it]

100%|██████████| 10/10 [00:49<00:00,  5.55s/it]2026-04-13 09:38:06,570 | INFO | train_grpo_train | metrics_logged mode=eval step=900

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 53.4727, 'eval_samples_per_second': 1.496, 'eval_steps_per_second': 0.187, 'eval_num_tokens': 1590605.0, 'eval_completions/mean_length': 97.8125, 'eval_completions/min_length': 36.7, 'eval_completions/max_length': 239.6, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 76.43095321655274, 'eval_completions/min_terminated_length': 36.7, 'eval_completions/max_terminated_length': 122.0, 'eval_rewards/meter/mean': 0.7157208025455475, 'eval_rewards/meter/std': 0.3429990842938423, 'eval_rewards/count_adherence/mean': 0.9489583134651184, 'eval_rewards/count_adherence/std': 0.0947537835687399, 'eval_rewards/hard_gate/mean': 0.925, 'eval_rewards/hard_gate/std': 0.18771235942840575, 'eval_rewards/repeat_soft/mean': 0.9064767360687256, 'eval_rewards/repeat_soft/std': 0.10691137947142124, 'eval_rewards/judge_quality/mean': 0.48524999618530273, 'eval_rewards/judge_quality/std': 0.19364597499370576, 'eval_rewards/total_composite/mean': 0.5320614576339722, 'eval_rewards/total_composite/std': 0.19630679711699486, 'eval_reward': 0.5320614576339722, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08199086189270019, 'eval_sampling/sampling_logp_difference/max': 1.0795791625976563, 'eval_sampling/importance_sampling_ratio/min': 0.3488084301352501, 'eval_sampling/importance_sampling_ratio/mean': 1.0217031955718994, 'eval_sampling/importance_sampling_ratio/max': 1.4728089570999146, 'eval_entropy': 0.963804829120636, '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': 0.5320614576339722, 'eval_reward_meter_mean': 0.7157208025455475, 'eval_reward_meter_std': 0.3429990842938423, 'eval_reward_count_adherence_mean': 0.9489583134651184, 'eval_reward_count_adherence_std': 0.0947537835687399, 'eval_reward_hard_gate_mean': 0.925, 'eval_reward_hard_gate_std': 0.18771235942840575, 'eval_reward_repeat_soft_mean': 0.9064767360687256, 'eval_reward_repeat_soft_std': 0.10691137947142124, 'eval_reward_judge_quality_mean': 0.48524999618530273, 'eval_reward_judge_quality_std': 0.19364597499370576, 'eval_reward_total_composite_mean': 0.5320614576339722, 'eval_reward_total_composite_std': 0.19630679711699486, 'epoch': 0.09}

 27%|██▋       | 900/3300 [2:17:08<4:51:01,  7.28s/it]

100%|██████████| 10/10 [00:49<00:00,  5.55s/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 [2:17:17<15:56:03, 23.91s/it]2026-04-13 09:38:15,778 | INFO | train_grpo_train | metrics_logged mode=train step=901

                                                       
{'loss': 0.0298, 'grad_norm': 12.812335968017578, 'learning_rate': 7.272727272727273e-06, 'num_tokens': 1592260.0, 'completions/mean_length': 46.875, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.08536005020141602, 'rewards/meter/std': 0.08806665241718292, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7610586881637573, 'rewards/repeat_soft/std': 0.18405461311340332, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.3542100787162781, 'rewards/total_composite/std': 0.06443716585636139, 'reward': 0.3542100787162781, 'reward_std': 0.06443716585636139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13556799292564392, 'sampling/sampling_logp_difference/max': 1.5642199516296387, 'sampling/importance_sampling_ratio/min': 0.20925118029117584, 'sampling/importance_sampling_ratio/mean': 1.0086678266525269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7705042622983456, 'clip_ratio/low_mean': 0.051945476327091455, 'clip_ratio/low_min': 0.051945476327091455, 'clip_ratio/high_mean': 0.100767582654953, 'clip_ratio/high_max': 0.100767582654953, 'clip_ratio/region_mean': 0.15271305898204446, 'reward_total_mean': 0.3542100787162781, 'reward_meter_mean': 0.08536005020141602, 'reward_meter_std': 0.08806665241718292, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7610586881637573, 'reward_repeat_soft_std': 0.18405461311340332, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.3542100787162781, 'reward_total_composite_std': 0.06443716585636139, 'epoch': 0.09}

 27%|██▋       | 901/3300 [2:17:18<15:56:03, 23.91s/it]INFO 04-13 09:38:16 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 902/3300 [2:17:24<12:27:35, 18.71s/it]2026-04-13 09:38:22,332 | INFO | train_grpo_train | metrics_logged mode=train step=902

                                                       
{'loss': 0.0679, 'grad_norm': 14.534664154052734, 'learning_rate': 7.26969696969697e-06, 'num_tokens': 1593970.0, 'completions/mean_length': 56.75, 'completions/min_length': 39.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.4406425356864929, 'rewards/meter/std': 0.4224933087825775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9830940365791321, 'rewards/repeat_soft/std': 0.013649843633174896, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.13265828788280487, 'rewards/total_composite/mean': 0.49974721670150757, 'rewards/total_composite/std': 0.16015437245368958, 'reward': 0.49974721670150757, 'reward_std': 0.16015438735485077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1860671043395996, 'sampling/sampling_logp_difference/max': 3.8892030715942383, 'sampling/importance_sampling_ratio/min': 0.02046164683997631, 'sampling/importance_sampling_ratio/mean': 1.0121606588363647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9524694234132767, 'clip_ratio/low_mean': 0.11109669506549835, 'clip_ratio/low_min': 0.11109669506549835, 'clip_ratio/high_mean': 0.05636154394596815, 'clip_ratio/high_max': 0.05636154394596815, 'clip_ratio/region_mean': 0.1674582390114665, 'reward_total_mean': 0.49974721670150757, 'reward_meter_mean': 0.4406425356864929, 'reward_meter_std': 0.4224933087825775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9830940365791321, 'reward_repeat_soft_std': 0.013649843633174896, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.13265828788280487, 'reward_total_composite_mean': 0.49974721670150757, 'reward_total_composite_std': 0.16015437245368958, 'epoch': 0.09}

 27%|██▋       | 902/3300 [2:17:24<12:27:35, 18.71s/it]INFO 04-13 09:38:22 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 903/3300 [2:17:36<11:06:01, 16.67s/it]2026-04-13 09:38:34,263 | INFO | train_grpo_train | metrics_logged mode=train step=903

                                                       
{'loss': -0.1015, 'grad_norm': 3.356720447540283, 'learning_rate': 7.266666666666668e-06, 'num_tokens': 1595627.0, 'completions/mean_length': 95.125, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.57143020629883, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6811458468437195, 'rewards/meter/std': 0.3611838221549988, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9799220561981201, 'rewards/repeat_soft/std': 0.016584882512688637, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.23439893126487732, 'rewards/total_composite/mean': 0.4835904538631439, 'rewards/total_composite/std': 0.21161921322345734, 'reward': 0.4835904538631439, 'reward_std': 0.21161922812461853, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16999176144599915, 'sampling/sampling_logp_difference/max': 2.0852737426757812, 'sampling/importance_sampling_ratio/min': 0.12427309155464172, 'sampling/importance_sampling_ratio/mean': 1.0145553350448608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8613782152533531, 'clip_ratio/low_mean': 0.015243902802467346, 'clip_ratio/low_min': 0.015243902802467346, 'clip_ratio/high_mean': 0.10294332588091493, 'clip_ratio/high_max': 0.10294332588091493, 'clip_ratio/region_mean': 0.11818722868338227, 'reward_total_mean': 0.4835904538631439, 'reward_meter_mean': 0.6811458468437195, 'reward_meter_std': 0.3611838221549988, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9799220561981201, 'reward_repeat_soft_std': 0.016584882512688637, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.23439893126487732, 'reward_total_composite_mean': 0.4835904538631439, 'reward_total_composite_std': 0.21161921322345734, 'epoch': 0.09}

 27%|██▋       | 903/3300 [2:17:36<11:06:01, 16.67s/it]INFO 04-13 09:38:34 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 904/3300 [2:17:42<9:04:05, 13.62s/it] 2026-04-13 09:38:40,771 | INFO | train_grpo_train | metrics_logged mode=train step=904

                                                      
{'loss': 0.0736, 'grad_norm': 14.554370880126953, 'learning_rate': 7.263636363636364e-06, 'num_tokens': 1597503.0, 'completions/mean_length': 64.5, 'completions/min_length': 57.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.778179407119751, 'rewards/meter/std': 0.2762916684150696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.877153754234314, 'rewards/repeat_soft/std': 0.18072569370269775, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.194054514169693, 'rewards/total_composite/mean': 0.5995876789093018, 'rewards/total_composite/std': 0.10906385630369186, 'reward': 0.5995876789093018, 'reward_std': 0.10906385630369186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1650829166173935, 'sampling/sampling_logp_difference/max': 1.500819206237793, 'sampling/importance_sampling_ratio/min': 0.24453456699848175, 'sampling/importance_sampling_ratio/mean': 1.0208650827407837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.057773917913437, 'clip_ratio/low_mean': 0.08504586480557919, 'clip_ratio/low_min': 0.08504586480557919, 'clip_ratio/high_mean': 0.0739332064986229, 'clip_ratio/high_max': 0.0739332064986229, 'clip_ratio/region_mean': 0.15897907130420208, 'reward_total_mean': 0.5995876789093018, 'reward_meter_mean': 0.778179407119751, 'reward_meter_std': 0.2762916684150696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.877153754234314, 'reward_repeat_soft_std': 0.18072569370269775, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.194054514169693, 'reward_total_composite_mean': 0.5995876789093018, 'reward_total_composite_std': 0.10906385630369186, 'epoch': 0.09}

 27%|██▋       | 904/3300 [2:17:43<9:04:05, 13.62s/it]INFO 04-13 09:38:40 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 905/3300 [2:17:48<7:29:50, 11.27s/it]2026-04-13 09:38:46,549 | INFO | train_grpo_train | metrics_logged mode=train step=905

                                                      
{'loss': 0.0897, 'grad_norm': 18.605655670166016, 'learning_rate': 7.260606060606061e-06, 'num_tokens': 1598909.0, 'completions/mean_length': 24.75, 'completions/min_length': 21.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8512672185897827, 'rewards/meter/std': 0.25897178053855896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9602519273757935, 'rewards/repeat_soft/std': 0.006358357612043619, 'rewards/judge_quality/mean': 0.7437499761581421, 'rewards/judge_quality/std': 0.2432481348514557, 'rewards/total_composite/mean': 0.7538845539093018, 'rewards/total_composite/std': 0.19257384538650513, 'reward': 0.7538845539093018, 'reward_std': 0.19257384538650513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12659122049808502, 'sampling/sampling_logp_difference/max': 1.3429498672485352, 'sampling/importance_sampling_ratio/min': 0.26107439398765564, 'sampling/importance_sampling_ratio/mean': 1.024914026260376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6850458532571793, 'clip_ratio/low_mean': 0.06517857313156128, 'clip_ratio/low_min': 0.06517857313156128, 'clip_ratio/high_mean': 0.07068452564999461, 'clip_ratio/high_max': 0.07068452564999461, 'clip_ratio/region_mean': 0.1358630987815559, 'reward_total_mean': 0.7538845539093018, 'reward_meter_mean': 0.8512672185897827, 'reward_meter_std': 0.25897178053855896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9602519273757935, 'reward_repeat_soft_std': 0.006358357612043619, 'reward_judge_quality_mean': 0.7437499761581421, 'reward_judge_quality_std': 0.2432481348514557, 'reward_total_composite_mean': 0.7538845539093018, 'reward_total_composite_std': 0.19257384538650513, 'epoch': 0.09}

 27%|██▋       | 905/3300 [2:17:48<7:29:50, 11.27s/it]INFO 04-13 09:38:46 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 906/3300 [2:18:00<7:32:04, 11.33s/it]2026-04-13 09:38:58,019 | INFO | train_grpo_train | metrics_logged mode=train step=906

                                                      
{'loss': -0.1552, 'grad_norm': 3.5159833431243896, 'learning_rate': 7.257575757575758e-06, 'num_tokens': 1601376.0, 'completions/mean_length': 166.375, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 117.00000762939453, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.703851044178009, 'rewards/meter/std': 0.32434767484664917, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9248175024986267, 'rewards/repeat_soft/std': 0.07298389822244644, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.3435710668563843, 'rewards/total_composite/mean': 0.598581850528717, 'rewards/total_composite/std': 0.291879266500473, 'reward': 0.598581850528717, 'reward_std': 0.291879266500473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13965491950511932, 'sampling/sampling_logp_difference/max': 1.4687843322753906, 'sampling/importance_sampling_ratio/min': 0.23020517826080322, 'sampling/importance_sampling_ratio/mean': 1.0273159742355347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9906918928027153, 'clip_ratio/low_mean': 0.02160807466134429, 'clip_ratio/low_min': 0.02160807466134429, 'clip_ratio/high_mean': 0.08824530150741339, 'clip_ratio/high_max': 0.08824530150741339, 'clip_ratio/region_mean': 0.10985337616875768, 'reward_total_mean': 0.598581850528717, 'reward_meter_mean': 0.703851044178009, 'reward_meter_std': 0.32434767484664917, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9248175024986267, 'reward_repeat_soft_std': 0.07298389822244644, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.3435710668563843, 'reward_total_composite_mean': 0.598581850528717, 'reward_total_composite_std': 0.291879266500473, 'epoch': 0.09}

 27%|██▋       | 906/3300 [2:18:00<7:32:04, 11.33s/it]INFO 04-13 09:38:58 [block_pool.py:378] Successfully reset prefix cache

 27%|██▋       | 907/3300 [2:18:07<6:39:34, 10.02s/it]2026-04-13 09:39:04,975 | INFO | train_grpo_train | metrics_logged mode=train step=907

                                                      
{'loss': 0.0255, 'grad_norm': 7.501486778259277, 'learning_rate': 7.254545454545455e-06, 'num_tokens': 1603558.0, 'completions/mean_length': 104.75, 'completions/min_length': 96.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.8907060623168945, 'rewards/meter/std': 0.26040831208229065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7417531609535217, 'rewards/repeat_soft/std': 0.1804305613040924, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906257808208466, 'rewards/total_composite/mean': 0.5654819011688232, 'rewards/total_composite/std': 0.1330239325761795, 'reward': 0.5654819011688232, 'reward_std': 0.1330239474773407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10251452028751373, 'sampling/sampling_logp_difference/max': 1.8215961456298828, 'sampling/importance_sampling_ratio/min': 0.16176734864711761, 'sampling/importance_sampling_ratio/mean': 1.0123924016952515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7758409716188908, 'clip_ratio/low_mean': 0.022410391829907894, 'clip_ratio/low_min': 0.022410391829907894, 'clip_ratio/high_mean': 0.08412217628210783, 'clip_ratio/high_max': 0.08412217628210783, 'clip_ratio/region_mean': 0.10653256811201572, 'reward_total_mean': 0.5654819011688232, 'reward_meter_mean': 0.8907060623168945, 'reward_meter_std': 0.26040831208229065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7417531609535217, 'reward_repeat_soft_std': 0.1804305613040924, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906257808208466, 'reward_total_composite_mean': 0.5654819011688232, 'reward_total_composite_std': 0.1330239325761795, 'epoch': 0.09}

 27%|██▋       | 907/3300 [2:18:07<6:39:34, 10.02s/it]INFO 04-13 09:39:05 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 908/3300 [2:18:13<6:01:12,  9.06s/it]2026-04-13 09:39:11,799 | INFO | train_grpo_train | metrics_logged mode=train step=908

                                                      
{'loss': 0.0777, 'grad_norm': 9.802550315856934, 'learning_rate': 7.251515151515151e-06, 'num_tokens': 1605881.0, 'completions/mean_length': 106.375, 'completions/min_length': 87.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.375, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.9819270372390747, 'rewards/meter/std': 0.01999370940029621, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9211548566818237, 'rewards/repeat_soft/std': 0.050338804721832275, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7271592617034912, 'rewards/total_composite/std': 0.16884341835975647, 'reward': 0.7271592617034912, 'reward_std': 0.16884341835975647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16331268846988678, 'sampling/sampling_logp_difference/max': 1.952540397644043, 'sampling/importance_sampling_ratio/min': 0.14191310107707977, 'sampling/importance_sampling_ratio/mean': 1.0145978927612305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2415838986635208, 'clip_ratio/low_mean': 0.08216759376227856, 'clip_ratio/low_min': 0.08216759376227856, 'clip_ratio/high_mean': 0.05946306139230728, 'clip_ratio/high_max': 0.05946306139230728, 'clip_ratio/region_mean': 0.14163065515458584, 'reward_total_mean': 0.7271592617034912, 'reward_meter_mean': 0.9819270372390747, 'reward_meter_std': 0.01999370940029621, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9211548566818237, 'reward_repeat_soft_std': 0.050338804721832275, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7271592617034912, 'reward_total_composite_std': 0.16884341835975647, 'epoch': 0.09}

 28%|██▊       | 908/3300 [2:18:14<6:01:12,  9.06s/it]INFO 04-13 09:39:12 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 909/3300 [2:18:20<5:24:51,  8.15s/it]2026-04-13 09:39:17,833 | INFO | train_grpo_train | metrics_logged mode=train step=909

                                                      
{'loss': 0.0426, 'grad_norm': 18.87795066833496, 'learning_rate': 7.2484848484848495e-06, 'num_tokens': 1607426.0, 'completions/mean_length': 34.125, 'completions/min_length': 30.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.33157655596733093, 'rewards/meter/std': 0.31648188829421997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9801974296569824, 'rewards/repeat_soft/std': 0.012491202913224697, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5453124046325684, 'rewards/total_composite/std': 0.19037026166915894, 'reward': 0.5453124046325684, 'reward_std': 0.19037026166915894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16925941407680511, 'sampling/sampling_logp_difference/max': 1.7289962768554688, 'sampling/importance_sampling_ratio/min': 0.17746244370937347, 'sampling/importance_sampling_ratio/mean': 1.0000395774841309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7677922323346138, 'clip_ratio/low_mean': 0.06933283852413297, 'clip_ratio/low_min': 0.06933283852413297, 'clip_ratio/high_mean': 0.05304053891450167, 'clip_ratio/high_max': 0.05304053891450167, 'clip_ratio/region_mean': 0.12237337743863463, 'reward_total_mean': 0.5453124046325684, 'reward_meter_mean': 0.33157655596733093, 'reward_meter_std': 0.31648188829421997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9801974296569824, 'reward_repeat_soft_std': 0.012491202913224697, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5453124046325684, 'reward_total_composite_std': 0.19037026166915894, 'epoch': 0.09}

 28%|██▊       | 909/3300 [2:18:20<5:24:51,  8.15s/it]INFO 04-13 09:39:18 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 910/3300 [2:18:31<6:01:06,  9.07s/it]2026-04-13 09:39:29,036 | INFO | train_grpo_train | metrics_logged mode=train step=910

                                                      
{'loss': -0.1004, 'grad_norm': 3.7054879665374756, 'learning_rate': 7.245454545454546e-06, 'num_tokens': 1609176.0, 'completions/mean_length': 110.75, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.42857360839844, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6746935844421387, 'rewards/meter/std': 0.38369810581207275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9529876112937927, 'rewards/repeat_soft/std': 0.03730596601963043, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.5209311246871948, 'rewards/total_composite/std': 0.25986728072166443, 'reward': 0.5209311246871948, 'reward_std': 0.25986728072166443, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1810074895620346, 'sampling/sampling_logp_difference/max': 1.452235221862793, 'sampling/importance_sampling_ratio/min': 0.2340465486049652, 'sampling/importance_sampling_ratio/mean': 1.0295379161834717, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4939635396003723, 'clip_ratio/low_mean': 0.03707153536379337, 'clip_ratio/low_min': 0.03707153536379337, 'clip_ratio/high_mean': 0.0893766162917018, 'clip_ratio/high_max': 0.0893766162917018, 'clip_ratio/region_mean': 0.12644815165549517, 'reward_total_mean': 0.5209311246871948, 'reward_meter_mean': 0.6746935844421387, 'reward_meter_std': 0.38369810581207275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9529876112937927, 'reward_repeat_soft_std': 0.03730596601963043, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.5209311246871948, 'reward_total_composite_std': 0.25986728072166443, 'epoch': 0.09}

 28%|██▊       | 910/3300 [2:18:31<6:01:06,  9.07s/it]INFO 04-13 09:39:29 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 911/3300 [2:18:37<5:24:15,  8.14s/it]2026-04-13 09:39:35,024 | INFO | train_grpo_train | metrics_logged mode=train step=911

                                                      
{'loss': 0.0266, 'grad_norm': 14.23844051361084, 'learning_rate': 7.242424242424243e-06, 'num_tokens': 1610922.0, 'completions/mean_length': 39.25, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.1473059356212616, 'rewards/meter/std': 0.3147366940975189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7999280691146851, 'rewards/repeat_soft/std': 0.11304853856563568, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.3703829348087311, 'rewards/total_composite/std': 0.2418767809867859, 'reward': 0.3703829348087311, 'reward_std': 0.2418767660856247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1460590809583664, 'sampling/sampling_logp_difference/max': 2.1889023780822754, 'sampling/importance_sampling_ratio/min': 0.11203965544700623, 'sampling/importance_sampling_ratio/mean': 0.9989128708839417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4969974309206009, 'clip_ratio/low_mean': 0.10410348512232304, 'clip_ratio/low_min': 0.10410348512232304, 'clip_ratio/high_mean': 0.03947368450462818, 'clip_ratio/high_max': 0.03947368450462818, 'clip_ratio/region_mean': 0.14357716962695122, 'reward_total_mean': 0.3703829348087311, 'reward_meter_mean': 0.1473059356212616, 'reward_meter_std': 0.3147366940975189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7999280691146851, 'reward_repeat_soft_std': 0.11304853856563568, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.3703829348087311, 'reward_total_composite_std': 0.2418767809867859, 'epoch': 0.09}

 28%|██▊       | 911/3300 [2:18:37<5:24:15,  8.14s/it]INFO 04-13 09:39:35 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 912/3300 [2:18:43<4:56:09,  7.44s/it]2026-04-13 09:39:40,827 | INFO | train_grpo_train | metrics_logged mode=train step=912

                                                      
{'loss': 0.09, 'grad_norm': 20.343666076660156, 'learning_rate': 7.2393939393939404e-06, 'num_tokens': 1612603.0, 'completions/mean_length': 36.125, 'completions/min_length': 30.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5597535371780396, 'rewards/meter/std': 0.3825490474700928, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9814321398735046, 'rewards/repeat_soft/std': 0.018087225034832954, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5003030300140381, 'rewards/total_composite/std': 0.10381560027599335, 'reward': 0.5003030300140381, 'reward_std': 0.10381558537483215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17298763990402222, 'sampling/sampling_logp_difference/max': 1.3021645545959473, 'sampling/importance_sampling_ratio/min': 0.27194252610206604, 'sampling/importance_sampling_ratio/mean': 1.0064977407455444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.075918272137642, 'clip_ratio/low_mean': 0.08903274312615395, 'clip_ratio/low_min': 0.08903274312615395, 'clip_ratio/high_mean': 0.07369153341278434, 'clip_ratio/high_max': 0.07369153341278434, 'clip_ratio/region_mean': 0.16272427653893828, 'reward_total_mean': 0.5003030300140381, 'reward_meter_mean': 0.5597535371780396, 'reward_meter_std': 0.3825490474700928, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9814321398735046, 'reward_repeat_soft_std': 0.018087225034832954, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5003030300140381, 'reward_total_composite_std': 0.10381560027599335, 'epoch': 0.09}

 28%|██▊       | 912/3300 [2:18:43<4:56:09,  7.44s/it]INFO 04-13 09:39:41 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 913/3300 [2:18:49<4:41:21,  7.07s/it]2026-04-13 09:39:47,070 | INFO | train_grpo_train | metrics_logged mode=train step=913

                                                      
{'loss': 0.0452, 'grad_norm': 14.120256423950195, 'learning_rate': 7.236363636363637e-06, 'num_tokens': 1614162.0, 'completions/mean_length': 46.875, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8159139752388, 'rewards/meter/std': 0.24263866245746613, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9638544321060181, 'rewards/repeat_soft/std': 0.030508361756801605, 'rewards/judge_quality/mean': 0.643750011920929, 'rewards/judge_quality/std': 0.18133927881717682, 'rewards/total_composite/mean': 0.6741479635238647, 'rewards/total_composite/std': 0.12217270582914352, 'reward': 0.6741479635238647, 'reward_std': 0.12217270582914352, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1827840507030487, 'sampling/sampling_logp_difference/max': 1.6252827644348145, 'sampling/importance_sampling_ratio/min': 0.21006464958190918, 'sampling/importance_sampling_ratio/mean': 1.019508957862854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4227289482951164, 'clip_ratio/low_mean': 0.10298752877861261, 'clip_ratio/low_min': 0.10298752877861261, 'clip_ratio/high_mean': 0.08171163313090801, 'clip_ratio/high_max': 0.08171163313090801, 'clip_ratio/region_mean': 0.18469916190952063, 'reward_total_mean': 0.6741479635238647, 'reward_meter_mean': 0.8159139752388, 'reward_meter_std': 0.24263866245746613, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9638544321060181, 'reward_repeat_soft_std': 0.030508361756801605, 'reward_judge_quality_mean': 0.643750011920929, 'reward_judge_quality_std': 0.18133927881717682, 'reward_total_composite_mean': 0.6741479635238647, 'reward_total_composite_std': 0.12217270582914352, 'epoch': 0.09}

 28%|██▊       | 913/3300 [2:18:49<4:41:21,  7.07s/it]INFO 04-13 09:39:47 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 914/3300 [2:18:56<4:38:34,  7.01s/it]2026-04-13 09:39:53,918 | INFO | train_grpo_train | metrics_logged mode=train step=914

                                                      
{'loss': 0.0263, 'grad_norm': 12.239043235778809, 'learning_rate': 7.233333333333334e-06, 'num_tokens': 1615772.0, 'completions/mean_length': 47.25, 'completions/min_length': 46.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9544693827629089, 'rewards/meter/std': 0.05370640382170677, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.966535747051239, 'rewards/repeat_soft/std': 0.016960065811872482, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.6542842388153076, 'rewards/total_composite/std': 0.11274627596139908, 'reward': 0.6542842388153076, 'reward_std': 0.11274627596139908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11110243201255798, 'sampling/sampling_logp_difference/max': 1.6137062311172485, 'sampling/importance_sampling_ratio/min': 0.19914816319942474, 'sampling/importance_sampling_ratio/mean': 1.0147902965545654, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.70686886459589, 'clip_ratio/low_mean': 0.06612967816181481, 'clip_ratio/low_min': 0.06612967816181481, 'clip_ratio/high_mean': 0.02445652149617672, 'clip_ratio/high_max': 0.02445652149617672, 'clip_ratio/region_mean': 0.09058619965799153, 'reward_total_mean': 0.6542842388153076, 'reward_meter_mean': 0.9544693827629089, 'reward_meter_std': 0.05370640382170677, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.966535747051239, 'reward_repeat_soft_std': 0.016960065811872482, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.6542842388153076, 'reward_total_composite_std': 0.11274627596139908, 'epoch': 0.09}

 28%|██▊       | 914/3300 [2:18:56<4:38:34,  7.01s/it]INFO 04-13 09:39:54 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 915/3300 [2:19:02<4:36:06,  6.95s/it]2026-04-13 09:40:00,694 | INFO | train_grpo_train | metrics_logged mode=train step=915

                                                      
{'loss': 0.0788, 'grad_norm': 9.88178539276123, 'learning_rate': 7.2303030303030305e-06, 'num_tokens': 1617758.0, 'completions/mean_length': 85.25, 'completions/min_length': 76.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.25, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.672178328037262, 'rewards/meter/std': 0.30288252234458923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8553141355514526, 'rewards/repeat_soft/std': 0.09193862974643707, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5740668773651123, 'rewards/total_composite/std': 0.15459750592708588, 'reward': 0.5740668773651123, 'reward_std': 0.15459749102592468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16788269579410553, 'sampling/sampling_logp_difference/max': 3.2447891235351562, 'sampling/importance_sampling_ratio/min': 0.03897678107023239, 'sampling/importance_sampling_ratio/mean': 0.9828201532363892, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6544415801763535, 'clip_ratio/low_mean': 0.07083723042160273, 'clip_ratio/low_min': 0.07083723042160273, 'clip_ratio/high_mean': 0.07165043521672487, 'clip_ratio/high_max': 0.07165043521672487, 'clip_ratio/region_mean': 0.1424876656383276, 'reward_total_mean': 0.5740668773651123, 'reward_meter_mean': 0.672178328037262, 'reward_meter_std': 0.30288252234458923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8553141355514526, 'reward_repeat_soft_std': 0.09193862974643707, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5740668773651123, 'reward_total_composite_std': 0.15459750592708588, 'epoch': 0.09}

 28%|██▊       | 915/3300 [2:19:02<4:36:06,  6.95s/it]INFO 04-13 09:40:00 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 916/3300 [2:19:08<4:22:51,  6.62s/it]2026-04-13 09:40:06,537 | INFO | train_grpo_train | metrics_logged mode=train step=916

                                                      
{'loss': 0.0469, 'grad_norm': 20.423137664794922, 'learning_rate': 7.227272727272729e-06, 'num_tokens': 1619310.0, 'completions/mean_length': 33.0, 'completions/min_length': 31.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5603809356689453, 'rewards/meter/std': 0.36889195442199707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9368240237236023, 'rewards/repeat_soft/std': 0.0935758501291275, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720350325107574, 'rewards/total_composite/mean': 0.5057446956634521, 'rewards/total_composite/std': 0.08553271740674973, 'reward': 0.5057446956634521, 'reward_std': 0.08553270995616913, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17337189614772797, 'sampling/sampling_logp_difference/max': 4.319033145904541, 'sampling/importance_sampling_ratio/min': 0.0133127486333251, 'sampling/importance_sampling_ratio/mean': 0.9902936816215515, 'sampling/importance_sampling_ratio/max': 1.9824931621551514, 'entropy': 0.8912190422415733, 'clip_ratio/low_mean': 0.06076621077954769, 'clip_ratio/low_min': 0.06076621077954769, 'clip_ratio/high_mean': 0.09892127197235823, 'clip_ratio/high_max': 0.09892127197235823, 'clip_ratio/region_mean': 0.15968748275190592, 'reward_total_mean': 0.5057446956634521, 'reward_meter_mean': 0.5603809356689453, 'reward_meter_std': 0.36889195442199707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9368240237236023, 'reward_repeat_soft_std': 0.0935758501291275, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720350325107574, 'reward_total_composite_mean': 0.5057446956634521, 'reward_total_composite_std': 0.08553271740674973, 'epoch': 0.09}

 28%|██▊       | 916/3300 [2:19:08<4:22:51,  6.62s/it]INFO 04-13 09:40:06 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 917/3300 [2:19:19<5:14:40,  7.92s/it]2026-04-13 09:40:17,513 | INFO | train_grpo_train | metrics_logged mode=train step=917

                                                      
{'loss': -0.0847, 'grad_norm': 4.606402397155762, 'learning_rate': 7.224242424242425e-06, 'num_tokens': 1620756.0, 'completions/mean_length': 104.75, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 46.57143020629883, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9420249462127686, 'rewards/meter/std': 0.08945825695991516, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9198586940765381, 'rewards/repeat_soft/std': 0.018826227635145187, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.45015519857406616, 'rewards/total_composite/std': 0.279334157705307, 'reward': 0.45015519857406616, 'reward_std': 0.279334157705307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12947139143943787, 'sampling/sampling_logp_difference/max': 1.295595407485962, 'sampling/importance_sampling_ratio/min': 0.27373483777046204, 'sampling/importance_sampling_ratio/mean': 1.0317928791046143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9875844419002533, 'clip_ratio/low_mean': 0.007978723384439945, 'clip_ratio/low_min': 0.007978723384439945, 'clip_ratio/high_mean': 0.08037575893104076, 'clip_ratio/high_max': 0.08037575893104076, 'clip_ratio/region_mean': 0.08835448231548071, 'reward_total_mean': 0.45015519857406616, 'reward_meter_mean': 0.9420249462127686, 'reward_meter_std': 0.08945825695991516, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9198586940765381, 'reward_repeat_soft_std': 0.018826227635145187, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.45015519857406616, 'reward_total_composite_std': 0.279334157705307, 'epoch': 0.09}

 28%|██▊       | 917/3300 [2:19:19<5:14:40,  7.92s/it]INFO 04-13 09:40:17 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 918/3300 [2:19:30<5:54:45,  8.94s/it]2026-04-13 09:40:28,811 | INFO | train_grpo_train | metrics_logged mode=train step=918

                                                      
{'loss': -0.143, 'grad_norm': 5.479053974151611, 'learning_rate': 7.221212121212122e-06, 'num_tokens': 1622431.0, 'completions/mean_length': 121.375, 'completions/min_length': 60.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 65.5714340209961, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7508893609046936, 'rewards/meter/std': 0.2547018229961395, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9491066932678223, 'rewards/repeat_soft/std': 0.02991720475256443, 'rewards/judge_quality/mean': 0.3787499964237213, 'rewards/judge_quality/std': 0.2713688910007477, 'rewards/total_composite/mean': 0.4926729202270508, 'rewards/total_composite/std': 0.25319477915763855, 'reward': 0.4926729202270508, 'reward_std': 0.25319477915763855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16436295211315155, 'sampling/sampling_logp_difference/max': 1.2689733505249023, 'sampling/importance_sampling_ratio/min': 0.28112009167671204, 'sampling/importance_sampling_ratio/mean': 1.0305107831954956, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.070231020450592, 'clip_ratio/low_mean': 0.06362179573625326, 'clip_ratio/low_min': 0.06362179573625326, 'clip_ratio/high_mean': 0.07916897907853127, 'clip_ratio/high_max': 0.07916897907853127, 'clip_ratio/region_mean': 0.14279077481478453, 'reward_total_mean': 0.4926729202270508, 'reward_meter_mean': 0.7508893609046936, 'reward_meter_std': 0.2547018229961395, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9491066932678223, 'reward_repeat_soft_std': 0.02991720475256443, 'reward_judge_quality_mean': 0.3787499964237213, 'reward_judge_quality_std': 0.2713688910007477, 'reward_total_composite_mean': 0.4926729202270508, 'reward_total_composite_std': 0.25319477915763855, 'epoch': 0.09}

 28%|██▊       | 918/3300 [2:19:31<5:54:45,  8.94s/it]INFO 04-13 09:40:29 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 919/3300 [2:19:37<5:19:48,  8.06s/it]2026-04-13 09:40:34,822 | INFO | train_grpo_train | metrics_logged mode=train step=919

                                                      
{'loss': 0.0613, 'grad_norm': 14.65757942199707, 'learning_rate': 7.218181818181819e-06, 'num_tokens': 1623905.0, 'completions/mean_length': 43.25, 'completions/min_length': 40.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9844621419906616, 'rewards/meter/std': 0.010709667578339577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9381470680236816, 'rewards/repeat_soft/std': 0.046228960156440735, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6191074252128601, 'rewards/total_composite/std': 0.015671368688344955, 'reward': 0.6191074252128601, 'reward_std': 0.015671366825699806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16086256504058838, 'sampling/sampling_logp_difference/max': 1.6122045516967773, 'sampling/importance_sampling_ratio/min': 0.1994474232196808, 'sampling/importance_sampling_ratio/mean': 0.9934461116790771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8492658138275146, 'clip_ratio/low_mean': 0.07349553424865007, 'clip_ratio/low_min': 0.07349553424865007, 'clip_ratio/high_mean': 0.0476880082860589, 'clip_ratio/high_max': 0.0476880082860589, 'clip_ratio/region_mean': 0.12118354253470898, 'reward_total_mean': 0.6191074252128601, 'reward_meter_mean': 0.9844621419906616, 'reward_meter_std': 0.010709667578339577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9381470680236816, 'reward_repeat_soft_std': 0.046228960156440735, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6191074252128601, 'reward_total_composite_std': 0.015671368688344955, 'epoch': 0.09}

 28%|██▊       | 919/3300 [2:19:37<5:19:48,  8.06s/it]INFO 04-13 09:40:35 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 920/3300 [2:19:48<5:56:32,  8.99s/it]2026-04-13 09:40:45,985 | INFO | train_grpo_train | metrics_logged mode=train step=920

                                                      
{'loss': -0.1817, 'grad_norm': 2.979152202606201, 'learning_rate': 7.215151515151516e-06, 'num_tokens': 1625977.0, 'completions/mean_length': 155.0, 'completions/min_length': 96.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 104.00000762939453, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.767349362373352, 'rewards/meter/std': 0.2601783275604248, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9513424634933472, 'rewards/repeat_soft/std': 0.03866739943623543, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5252867341041565, 'rewards/total_composite/std': 0.23032616078853607, 'reward': 0.5252867341041565, 'reward_std': 0.23032614588737488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14671450853347778, 'sampling/sampling_logp_difference/max': 1.4239892959594727, 'sampling/importance_sampling_ratio/min': 0.24075166881084442, 'sampling/importance_sampling_ratio/mean': 1.009709119796753, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0711248964071274, 'clip_ratio/low_mean': 0.025735294446349144, 'clip_ratio/low_min': 0.025735294446349144, 'clip_ratio/high_mean': 0.11989929620176554, 'clip_ratio/high_max': 0.11989929620176554, 'clip_ratio/region_mean': 0.14563459064811468, 'reward_total_mean': 0.5252867341041565, 'reward_meter_mean': 0.767349362373352, 'reward_meter_std': 0.2601783275604248, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9513424634933472, 'reward_repeat_soft_std': 0.03866739943623543, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5252867341041565, 'reward_total_composite_std': 0.23032616078853607, 'epoch': 0.09}

 28%|██▊       | 920/3300 [2:19:48<5:56:32,  8.99s/it]INFO 04-13 09:40:46 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 921/3300 [2:19:54<5:22:36,  8.14s/it]2026-04-13 09:40:52,133 | INFO | train_grpo_train | metrics_logged mode=train step=921

                                                      
{'loss': -0.0443, 'grad_norm': 12.08022403717041, 'learning_rate': 7.212121212121212e-06, 'num_tokens': 1627581.0, 'completions/mean_length': 47.5, 'completions/min_length': 40.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6915476322174072, 'rewards/meter/std': 0.36313360929489136, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9853203296661377, 'rewards/repeat_soft/std': 0.022112060338258743, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6625669002532959, 'rewards/total_composite/std': 0.22150346636772156, 'reward': 0.6625669002532959, 'reward_std': 0.22150346636772156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1407664716243744, 'sampling/sampling_logp_difference/max': 2.5118842124938965, 'sampling/importance_sampling_ratio/min': 0.0811152532696724, 'sampling/importance_sampling_ratio/mean': 1.0195248126983643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0722940489649773, 'clip_ratio/low_mean': 0.10812500026077032, 'clip_ratio/low_min': 0.10812500026077032, 'clip_ratio/high_mean': 0.04378216154873371, 'clip_ratio/high_max': 0.04378216154873371, 'clip_ratio/region_mean': 0.15190716180950403, 'reward_total_mean': 0.6625669002532959, 'reward_meter_mean': 0.6915476322174072, 'reward_meter_std': 0.36313360929489136, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9853203296661377, 'reward_repeat_soft_std': 0.022112060338258743, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6625669002532959, 'reward_total_composite_std': 0.22150346636772156, 'epoch': 0.09}

 28%|██▊       | 921/3300 [2:19:54<5:22:36,  8.14s/it]INFO 04-13 09:40:52 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 922/3300 [2:20:01<5:05:21,  7.70s/it]2026-04-13 09:40:58,827 | INFO | train_grpo_train | metrics_logged mode=train step=922

                                                      
{'loss': 0.0334, 'grad_norm': 15.846829414367676, 'learning_rate': 7.2090909090909104e-06, 'num_tokens': 1629608.0, 'completions/mean_length': 87.375, 'completions/min_length': 76.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.375, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6068453788757324, 'rewards/meter/std': 0.21214169263839722, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9070358276367188, 'rewards/repeat_soft/std': 0.07087033987045288, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5358442068099976, 'rewards/total_composite/std': 0.13801956176757812, 'reward': 0.5358442068099976, 'reward_std': 0.13801954686641693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17047323286533356, 'sampling/sampling_logp_difference/max': 3.229947566986084, 'sampling/importance_sampling_ratio/min': 0.03955957293510437, 'sampling/importance_sampling_ratio/mean': 0.9821562767028809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6564752086997032, 'clip_ratio/low_mean': 0.08379318378865719, 'clip_ratio/low_min': 0.08379318378865719, 'clip_ratio/high_mean': 0.05765899270772934, 'clip_ratio/high_max': 0.05765899270772934, 'clip_ratio/region_mean': 0.14145217649638653, 'reward_total_mean': 0.5358442068099976, 'reward_meter_mean': 0.6068453788757324, 'reward_meter_std': 0.21214169263839722, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9070358276367188, 'reward_repeat_soft_std': 0.07087033987045288, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5358442068099976, 'reward_total_composite_std': 0.13801956176757812, 'epoch': 0.09}

 28%|██▊       | 922/3300 [2:20:01<5:05:21,  7.70s/it]INFO 04-13 09:40:59 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 923/3300 [2:20:07<4:46:52,  7.24s/it]2026-04-13 09:41:04,990 | INFO | train_grpo_train | metrics_logged mode=train step=923

                                                      
{'loss': 0.0264, 'grad_norm': 14.256244659423828, 'learning_rate': 7.206060606060606e-06, 'num_tokens': 1631080.0, 'completions/mean_length': 47.0, 'completions/min_length': 39.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.5419894456863403, 'rewards/meter/std': 0.36938440799713135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9782922267913818, 'rewards/repeat_soft/std': 0.02354535646736622, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.1302744448184967, 'rewards/total_composite/mean': 0.580936074256897, 'rewards/total_composite/std': 0.15119551122188568, 'reward': 0.580936074256897, 'reward_std': 0.15119551122188568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13457390666007996, 'sampling/sampling_logp_difference/max': 1.823359489440918, 'sampling/importance_sampling_ratio/min': 0.16148234903812408, 'sampling/importance_sampling_ratio/mean': 1.002490758895874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.971940703690052, 'clip_ratio/low_mean': 0.04533275682479143, 'clip_ratio/low_min': 0.04533275682479143, 'clip_ratio/high_mean': 0.11088750511407852, 'clip_ratio/high_max': 0.11088750511407852, 'clip_ratio/region_mean': 0.15622026193886995, 'reward_total_mean': 0.580936074256897, 'reward_meter_mean': 0.5419894456863403, 'reward_meter_std': 0.36938440799713135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9782922267913818, 'reward_repeat_soft_std': 0.02354535646736622, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.1302744448184967, 'reward_total_composite_mean': 0.580936074256897, 'reward_total_composite_std': 0.15119551122188568, 'epoch': 0.09}

 28%|██▊       | 923/3300 [2:20:07<4:46:52,  7.24s/it]INFO 04-13 09:41:05 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 924/3300 [2:20:13<4:31:02,  6.84s/it]2026-04-13 09:41:10,909 | INFO | train_grpo_train | metrics_logged mode=train step=924

                                                      
{'loss': 0.0329, 'grad_norm': 13.811881065368652, 'learning_rate': 7.203030303030304e-06, 'num_tokens': 1632629.0, 'completions/mean_length': 42.625, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.669313907623291, 'rewards/meter/std': 0.3524772822856903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9745993614196777, 'rewards/repeat_soft/std': 0.02978009358048439, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5712888240814209, 'rewards/total_composite/std': 0.1664804220199585, 'reward': 0.5712888240814209, 'reward_std': 0.1664804220199585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1601911038160324, 'sampling/sampling_logp_difference/max': 1.909719467163086, 'sampling/importance_sampling_ratio/min': 0.1481219381093979, 'sampling/importance_sampling_ratio/mean': 1.032793402671814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0413456037640572, 'clip_ratio/low_mean': 0.08925975020974874, 'clip_ratio/low_min': 0.08925975020974874, 'clip_ratio/high_mean': 0.07106302492320538, 'clip_ratio/high_max': 0.07106302492320538, 'clip_ratio/region_mean': 0.16032277513295412, 'reward_total_mean': 0.5712888240814209, 'reward_meter_mean': 0.669313907623291, 'reward_meter_std': 0.3524772822856903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9745993614196777, 'reward_repeat_soft_std': 0.02978009358048439, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5712888240814209, 'reward_total_composite_std': 0.1664804220199585, 'epoch': 0.09}

 28%|██▊       | 924/3300 [2:20:13<4:31:02,  6.84s/it]INFO 04-13 09:41:11 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 925/3300 [2:20:19<4:23:10,  6.65s/it]2026-04-13 09:41:17,098 | INFO | train_grpo_train | metrics_logged mode=train step=925

                                                      
{'loss': 0.0564, 'grad_norm': 13.910771369934082, 'learning_rate': 7.2000000000000005e-06, 'num_tokens': 1634440.0, 'completions/mean_length': 48.375, 'completions/min_length': 42.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7714744806289673, 'rewards/meter/std': 0.32325148582458496, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9265764951705933, 'rewards/repeat_soft/std': 0.12223569303750992, 'rewards/judge_quality/mean': 0.6599999666213989, 'rewards/judge_quality/std': 0.2855571210384369, 'rewards/total_composite/mean': 0.7035986185073853, 'rewards/total_composite/std': 0.24538688361644745, 'reward': 0.7035986185073853, 'reward_std': 0.24538686871528625, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16183127462863922, 'sampling/sampling_logp_difference/max': 1.7264628410339355, 'sampling/importance_sampling_ratio/min': 0.17791259288787842, 'sampling/importance_sampling_ratio/mean': 1.0172739028930664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8879731148481369, 'clip_ratio/low_mean': 0.08377291262149811, 'clip_ratio/low_min': 0.08377291262149811, 'clip_ratio/high_mean': 0.08953542355448008, 'clip_ratio/high_max': 0.08953542355448008, 'clip_ratio/region_mean': 0.17330833617597818, 'reward_total_mean': 0.7035986185073853, 'reward_meter_mean': 0.7714744806289673, 'reward_meter_std': 0.32325148582458496, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9265764951705933, 'reward_repeat_soft_std': 0.12223569303750992, 'reward_judge_quality_mean': 0.6599999666213989, 'reward_judge_quality_std': 0.2855571210384369, 'reward_total_composite_mean': 0.7035986185073853, 'reward_total_composite_std': 0.24538688361644745, 'epoch': 0.09}

 28%|██▊       | 925/3300 [2:20:19<4:23:10,  6.65s/it]INFO 04-13 09:41:17 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 926/3300 [2:20:26<4:28:10,  6.78s/it]2026-04-13 09:41:24,176 | INFO | train_grpo_train | metrics_logged mode=train step=926

                                                      
{'loss': 0.0661, 'grad_norm': 7.966991901397705, 'learning_rate': 7.196969696969698e-06, 'num_tokens': 1637037.0, 'completions/mean_length': 107.625, 'completions/min_length': 98.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.625, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.6245505809783936, 'rewards/meter/std': 0.3284919559955597, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9416639804840088, 'rewards/repeat_soft/std': 0.04125182330608368, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5026393532752991, 'rewards/total_composite/std': 0.1356973648071289, 'reward': 0.5026393532752991, 'reward_std': 0.1356973648071289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15753981471061707, 'sampling/sampling_logp_difference/max': 1.847491979598999, 'sampling/importance_sampling_ratio/min': 0.1576320081949234, 'sampling/importance_sampling_ratio/mean': 1.0120426416397095, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9671919271349907, 'clip_ratio/low_mean': 0.06904459372162819, 'clip_ratio/low_min': 0.06904459372162819, 'clip_ratio/high_mean': 0.05026556085795164, 'clip_ratio/high_max': 0.05026556085795164, 'clip_ratio/region_mean': 0.11931015457957983, 'reward_total_mean': 0.5026393532752991, 'reward_meter_mean': 0.6245505809783936, 'reward_meter_std': 0.3284919559955597, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9416639804840088, 'reward_repeat_soft_std': 0.04125182330608368, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5026393532752991, 'reward_total_composite_std': 0.1356973648071289, 'epoch': 0.09}

 28%|██▊       | 926/3300 [2:20:26<4:28:10,  6.78s/it]INFO 04-13 09:41:24 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 927/3300 [2:20:32<4:24:59,  6.70s/it]2026-04-13 09:41:30,694 | INFO | train_grpo_train | metrics_logged mode=train step=927

                                                      
{'loss': 0.0252, 'grad_norm': 9.569807052612305, 'learning_rate': 7.193939393939394e-06, 'num_tokens': 1639370.0, 'completions/mean_length': 99.625, 'completions/min_length': 91.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.625, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.5098446011543274, 'rewards/meter/std': 0.3082834482192993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8003026247024536, 'rewards/repeat_soft/std': 0.08848375082015991, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5001412630081177, 'rewards/total_composite/std': 0.13108134269714355, 'reward': 0.5001412630081177, 'reward_std': 0.13108134269714355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13828051090240479, 'sampling/sampling_logp_difference/max': 1.9262475967407227, 'sampling/importance_sampling_ratio/min': 0.14569388329982758, 'sampling/importance_sampling_ratio/mean': 0.9944371581077576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7524937465786934, 'clip_ratio/low_mean': 0.05981121212244034, 'clip_ratio/low_min': 0.05981121212244034, 'clip_ratio/high_mean': 0.07867600489407778, 'clip_ratio/high_max': 0.07867600489407778, 'clip_ratio/region_mean': 0.13848721701651812, 'reward_total_mean': 0.5001412630081177, 'reward_meter_mean': 0.5098446011543274, 'reward_meter_std': 0.3082834482192993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8003026247024536, 'reward_repeat_soft_std': 0.08848375082015991, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5001412630081177, 'reward_total_composite_std': 0.13108134269714355, 'epoch': 0.09}

 28%|██▊       | 927/3300 [2:20:32<4:24:59,  6.70s/it]INFO 04-13 09:41:30 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 928/3300 [2:20:38<4:17:00,  6.50s/it]2026-04-13 09:41:36,734 | INFO | train_grpo_train | metrics_logged mode=train step=928

                                                      
{'loss': 0.0807, 'grad_norm': 17.609262466430664, 'learning_rate': 7.1909090909090914e-06, 'num_tokens': 1640986.0, 'completions/mean_length': 38.0, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9794037342071533, 'rewards/meter/std': 0.027005111798644066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9568880796432495, 'rewards/repeat_soft/std': 0.047611843794584274, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7738229036331177, 'rewards/total_composite/std': 0.172917902469635, 'reward': 0.7738229036331177, 'reward_std': 0.17291788756847382, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15360701084136963, 'sampling/sampling_logp_difference/max': 1.5589860677719116, 'sampling/importance_sampling_ratio/min': 0.21034923195838928, 'sampling/importance_sampling_ratio/mean': 1.0201724767684937, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1102176755666733, 'clip_ratio/low_mean': 0.08575666323304176, 'clip_ratio/low_min': 0.08575666323304176, 'clip_ratio/high_mean': 0.07010342739522457, 'clip_ratio/high_max': 0.07010342739522457, 'clip_ratio/region_mean': 0.15586009062826633, 'reward_total_mean': 0.7738229036331177, 'reward_meter_mean': 0.9794037342071533, 'reward_meter_std': 0.027005111798644066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9568880796432495, 'reward_repeat_soft_std': 0.047611843794584274, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7738229036331177, 'reward_total_composite_std': 0.172917902469635, 'epoch': 0.09}

 28%|██▊       | 928/3300 [2:20:38<4:17:00,  6.50s/it]INFO 04-13 09:41:36 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 929/3300 [2:20:50<5:11:50,  7.89s/it]2026-04-13 09:41:47,867 | INFO | train_grpo_train | metrics_logged mode=train step=929

                                                      
{'loss': -0.083, 'grad_norm': 2.698056697845459, 'learning_rate': 7.187878787878788e-06, 'num_tokens': 1642428.0, 'completions/mean_length': 86.25, 'completions/min_length': 23.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.428571701049805, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8588230609893799, 'rewards/meter/std': 0.262309730052948, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9585226774215698, 'rewards/repeat_soft/std': 0.011249415576457977, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.13767844438552856, 'rewards/total_composite/mean': 0.5160880088806152, 'rewards/total_composite/std': 0.22122929990291595, 'reward': 0.5160880088806152, 'reward_std': 0.22122929990291595, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16538824141025543, 'sampling/sampling_logp_difference/max': 1.312190055847168, 'sampling/importance_sampling_ratio/min': 0.26922979950904846, 'sampling/importance_sampling_ratio/mean': 1.062637209892273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2604689002037048, 'clip_ratio/low_mean': 0.03125, 'clip_ratio/low_min': 0.03125, 'clip_ratio/high_mean': 0.09820446511730552, 'clip_ratio/high_max': 0.09820446511730552, 'clip_ratio/region_mean': 0.12945446511730552, 'reward_total_mean': 0.5160880088806152, 'reward_meter_mean': 0.8588230609893799, 'reward_meter_std': 0.262309730052948, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9585226774215698, 'reward_repeat_soft_std': 0.011249415576457977, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.13767844438552856, 'reward_total_composite_mean': 0.5160880088806152, 'reward_total_composite_std': 0.22122929990291595, 'epoch': 0.09}

 28%|██▊       | 929/3300 [2:20:50<5:11:50,  7.89s/it]INFO 04-13 09:41:48 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 930/3300 [2:20:56<4:51:25,  7.38s/it]2026-04-13 09:41:54,056 | INFO | train_grpo_train | metrics_logged mode=train step=930

                                                      
{'loss': -0.029, 'grad_norm': 13.70984172821045, 'learning_rate': 7.184848484848486e-06, 'num_tokens': 1644120.0, 'completions/mean_length': 46.5, 'completions/min_length': 39.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7361986041069031, 'rewards/meter/std': 0.3028372526168823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9762316942214966, 'rewards/repeat_soft/std': 0.024908561259508133, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5551185607910156, 'rewards/total_composite/std': 0.08748355507850647, 'reward': 0.5551185607910156, 'reward_std': 0.08748354762792587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1647631973028183, 'sampling/sampling_logp_difference/max': 1.473271131515503, 'sampling/importance_sampling_ratio/min': 0.28066200017929077, 'sampling/importance_sampling_ratio/mean': 1.0438170433044434, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.39552640914917, 'clip_ratio/low_mean': 0.06653808243572712, 'clip_ratio/low_min': 0.06653808243572712, 'clip_ratio/high_mean': 0.10128033626824617, 'clip_ratio/high_max': 0.10128033626824617, 'clip_ratio/region_mean': 0.1678184187039733, 'reward_total_mean': 0.5551185607910156, 'reward_meter_mean': 0.7361986041069031, 'reward_meter_std': 0.3028372526168823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9762316942214966, 'reward_repeat_soft_std': 0.024908561259508133, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5551185607910156, 'reward_total_composite_std': 0.08748355507850647, 'epoch': 0.09}

 28%|██▊       | 930/3300 [2:20:56<4:51:25,  7.38s/it]INFO 04-13 09:41:54 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 931/3300 [2:21:08<5:46:30,  8.78s/it]2026-04-13 09:42:06,090 | INFO | train_grpo_train | metrics_logged mode=train step=931

                                                      
{'loss': 0.0733, 'grad_norm': 20.699567794799805, 'learning_rate': 7.181818181818182e-06, 'num_tokens': 1645514.0, 'completions/mean_length': 25.25, 'completions/min_length': 23.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.3937675952911377, 'rewards/meter/std': 0.2996700704097748, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9385433197021484, 'rewards/repeat_soft/std': 0.04496556892991066, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.24859607219696045, 'rewards/total_composite/mean': 0.5503440499305725, 'rewards/total_composite/std': 0.18855828046798706, 'reward': 0.5503440499305725, 'reward_std': 0.18855829536914825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17943686246871948, 'sampling/sampling_logp_difference/max': 2.067169666290283, 'sampling/importance_sampling_ratio/min': 0.12654343247413635, 'sampling/importance_sampling_ratio/mean': 0.9980342984199524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7266267389059067, 'clip_ratio/low_mean': 0.09316770359873772, 'clip_ratio/low_min': 0.09316770359873772, 'clip_ratio/high_mean': 0.05708333197981119, 'clip_ratio/high_max': 0.05708333197981119, 'clip_ratio/region_mean': 0.1502510355785489, 'reward_total_mean': 0.5503440499305725, 'reward_meter_mean': 0.3937675952911377, 'reward_meter_std': 0.2996700704097748, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9385433197021484, 'reward_repeat_soft_std': 0.04496556892991066, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.24859607219696045, 'reward_total_composite_mean': 0.5503440499305725, 'reward_total_composite_std': 0.18855828046798706, 'epoch': 0.09}

 28%|██▊       | 931/3300 [2:21:08<5:46:30,  8.78s/it]INFO 04-13 09:42:06 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 932/3300 [2:21:15<5:23:20,  8.19s/it]2026-04-13 09:42:12,916 | INFO | train_grpo_train | metrics_logged mode=train step=932

                                                      
{'loss': 0.0617, 'grad_norm': 9.303230285644531, 'learning_rate': 7.17878787878788e-06, 'num_tokens': 1647840.0, 'completions/mean_length': 97.75, 'completions/min_length': 91.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9708800315856934, 'rewards/meter/std': 0.014429431408643723, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.855011522769928, 'rewards/repeat_soft/std': 0.09349798411130905, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.15638209879398346, 'rewards/total_composite/mean': 0.535617470741272, 'rewards/total_composite/std': 0.09823641926050186, 'reward': 0.535617470741272, 'reward_std': 0.09823642671108246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14550121128559113, 'sampling/sampling_logp_difference/max': 2.029776096343994, 'sampling/importance_sampling_ratio/min': 0.13136492669582367, 'sampling/importance_sampling_ratio/mean': 1.020607352256775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0373662635684013, 'clip_ratio/low_mean': 0.045313782058656216, 'clip_ratio/low_min': 0.045313782058656216, 'clip_ratio/high_mean': 0.08563645090907812, 'clip_ratio/high_max': 0.08563645090907812, 'clip_ratio/region_mean': 0.13095023296773434, 'reward_total_mean': 0.535617470741272, 'reward_meter_mean': 0.9708800315856934, 'reward_meter_std': 0.014429431408643723, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.855011522769928, 'reward_repeat_soft_std': 0.09349798411130905, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.15638209879398346, 'reward_total_composite_mean': 0.535617470741272, 'reward_total_composite_std': 0.09823641926050186, 'epoch': 0.09}

 28%|██▊       | 932/3300 [2:21:15<5:23:20,  8.19s/it]INFO 04-13 09:42:13 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 933/3300 [2:21:22<5:18:47,  8.08s/it]2026-04-13 09:42:20,741 | INFO | train_grpo_train | metrics_logged mode=train step=933

                                                      
{'loss': 0.0276, 'grad_norm': 25.0538330078125, 'learning_rate': 7.175757575757576e-06, 'num_tokens': 1649272.0, 'completions/mean_length': 23.0, 'completions/min_length': 18.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8249565958976746, 'rewards/meter/std': 0.32491472363471985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9525578022003174, 'rewards/repeat_soft/std': 0.014416005462408066, 'rewards/judge_quality/mean': 0.39750000834465027, 'rewards/judge_quality/std': 0.10110107809305191, 'rewards/total_composite/mean': 0.5508995056152344, 'rewards/total_composite/std': 0.09900034964084625, 'reward': 0.5508995056152344, 'reward_std': 0.09900033473968506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17802026867866516, 'sampling/sampling_logp_difference/max': 1.075533390045166, 'sampling/importance_sampling_ratio/min': 0.3726882338523865, 'sampling/importance_sampling_ratio/mean': 1.0378916263580322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3319678083062172, 'clip_ratio/low_mean': 0.06595849804580212, 'clip_ratio/low_min': 0.06595849804580212, 'clip_ratio/high_mean': 0.09105158830061555, 'clip_ratio/high_max': 0.09105158830061555, 'clip_ratio/region_mean': 0.15701008634641767, 'reward_total_mean': 0.5508995056152344, 'reward_meter_mean': 0.8249565958976746, 'reward_meter_std': 0.32491472363471985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9525578022003174, 'reward_repeat_soft_std': 0.014416005462408066, 'reward_judge_quality_mean': 0.39750000834465027, 'reward_judge_quality_std': 0.10110107809305191, 'reward_total_composite_mean': 0.5508995056152344, 'reward_total_composite_std': 0.09900034964084625, 'epoch': 0.09}

 28%|██▊       | 933/3300 [2:21:22<5:18:47,  8.08s/it]INFO 04-13 09:42:20 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 934/3300 [2:21:28<4:53:57,  7.45s/it]2026-04-13 09:42:26,731 | INFO | train_grpo_train | metrics_logged mode=train step=934

                                                      
{'loss': 0.0493, 'grad_norm': 15.443513870239258, 'learning_rate': 7.172727272727273e-06, 'num_tokens': 1651053.0, 'completions/mean_length': 48.625, 'completions/min_length': 45.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7085595726966858, 'rewards/meter/std': 0.3993377387523651, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9652116298675537, 'rewards/repeat_soft/std': 0.03574341535568237, 'rewards/judge_quality/mean': 0.6112499833106995, 'rewards/judge_quality/std': 0.15037456154823303, 'rewards/total_composite/mean': 0.6232345104217529, 'rewards/total_composite/std': 0.1727263331413269, 'reward': 0.6232345104217529, 'reward_std': 0.1727263331413269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1562386006116867, 'sampling/sampling_logp_difference/max': 1.3455955982208252, 'sampling/importance_sampling_ratio/min': 0.26038458943367004, 'sampling/importance_sampling_ratio/mean': 1.0110880136489868, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1191803216934204, 'clip_ratio/low_mean': 0.09817763511091471, 'clip_ratio/low_min': 0.09817763511091471, 'clip_ratio/high_mean': 0.061666665598750114, 'clip_ratio/high_max': 0.061666665598750114, 'clip_ratio/region_mean': 0.15984430070966482, 'reward_total_mean': 0.6232345104217529, 'reward_meter_mean': 0.7085595726966858, 'reward_meter_std': 0.3993377387523651, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9652116298675537, 'reward_repeat_soft_std': 0.03574341535568237, 'reward_judge_quality_mean': 0.6112499833106995, 'reward_judge_quality_std': 0.15037456154823303, 'reward_total_composite_mean': 0.6232345104217529, 'reward_total_composite_std': 0.1727263331413269, 'epoch': 0.09}

 28%|██▊       | 934/3300 [2:21:28<4:53:57,  7.45s/it]INFO 04-13 09:42:26 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 935/3300 [2:21:34<4:35:14,  6.98s/it]2026-04-13 09:42:32,613 | INFO | train_grpo_train | metrics_logged mode=train step=935

                                                      
{'loss': 0.0255, 'grad_norm': 24.60488510131836, 'learning_rate': 7.16969696969697e-06, 'num_tokens': 1652549.0, 'completions/mean_length': 21.0, 'completions/min_length': 19.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9167644381523132, 'rewards/meter/std': 0.07410956174135208, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9464983940124512, 'rewards/repeat_soft/std': 0.04525930806994438, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.586983323097229, 'rewards/total_composite/std': 0.04325531795620918, 'reward': 0.586983323097229, 'reward_std': 0.04325530305504799, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14468401670455933, 'sampling/sampling_logp_difference/max': 1.8507564067840576, 'sampling/importance_sampling_ratio/min': 0.1571182757616043, 'sampling/importance_sampling_ratio/mean': 1.015199899673462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9049945846199989, 'clip_ratio/low_mean': 0.04805492050945759, 'clip_ratio/low_min': 0.04805492050945759, 'clip_ratio/high_mean': 0.06471861619502306, 'clip_ratio/high_max': 0.06471861619502306, 'clip_ratio/region_mean': 0.11277353670448065, 'reward_total_mean': 0.586983323097229, 'reward_meter_mean': 0.9167644381523132, 'reward_meter_std': 0.07410956174135208, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9464983940124512, 'reward_repeat_soft_std': 0.04525930806994438, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.586983323097229, 'reward_total_composite_std': 0.04325531795620918, 'epoch': 0.09}

 28%|██▊       | 935/3300 [2:21:34<4:35:14,  6.98s/it]INFO 04-13 09:42:32 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 936/3300 [2:21:41<4:26:10,  6.76s/it]2026-04-13 09:42:38,839 | INFO | train_grpo_train | metrics_logged mode=train step=936

                                                      
{'loss': 0.0007, 'grad_norm': 11.984716415405273, 'learning_rate': 7.166666666666667e-06, 'num_tokens': 1654466.0, 'completions/mean_length': 62.625, 'completions/min_length': 56.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9602702856063843, 'rewards/meter/std': 0.019833838567137718, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9538745880126953, 'rewards/repeat_soft/std': 0.014454740099608898, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5918318629264832, 'rewards/total_composite/std': 0.03716662898659706, 'reward': 0.5918318629264832, 'reward_std': 0.03716662526130676, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15173695981502533, 'sampling/sampling_logp_difference/max': 2.329094409942627, 'sampling/importance_sampling_ratio/min': 0.09738390147686005, 'sampling/importance_sampling_ratio/mean': 1.0248721837997437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1424496248364449, 'clip_ratio/low_mean': 0.0366459209471941, 'clip_ratio/low_min': 0.0366459209471941, 'clip_ratio/high_mean': 0.12486749235540628, 'clip_ratio/high_max': 0.12486749235540628, 'clip_ratio/region_mean': 0.16151341330260038, 'reward_total_mean': 0.5918318629264832, 'reward_meter_mean': 0.9602702856063843, 'reward_meter_std': 0.019833838567137718, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9538745880126953, 'reward_repeat_soft_std': 0.014454740099608898, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5918318629264832, 'reward_total_composite_std': 0.03716662898659706, 'epoch': 0.09}

 28%|██▊       | 936/3300 [2:21:41<4:26:10,  6.76s/it]INFO 04-13 09:42:39 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 937/3300 [2:21:47<4:23:58,  6.70s/it]2026-04-13 09:42:45,420 | INFO | train_grpo_train | metrics_logged mode=train step=937

                                                      
{'loss': 0.012, 'grad_norm': 11.2230863571167, 'learning_rate': 7.163636363636363e-06, 'num_tokens': 1656597.0, 'completions/mean_length': 96.375, 'completions/min_length': 91.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.375, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.10800344496965408, 'rewards/meter/std': 0.11452758312225342, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8984007835388184, 'rewards/repeat_soft/std': 0.12824523448944092, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.3662221133708954, 'rewards/total_composite/std': 0.040435004979372025, 'reward': 0.3662221133708954, 'reward_std': 0.04043500870466232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1392218917608261, 'sampling/sampling_logp_difference/max': 1.3230323791503906, 'sampling/importance_sampling_ratio/min': 0.2663264870643616, 'sampling/importance_sampling_ratio/mean': 1.0213192701339722, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9289983585476875, 'clip_ratio/low_mean': 0.10917292069643736, 'clip_ratio/low_min': 0.10917292069643736, 'clip_ratio/high_mean': 0.031905731186270714, 'clip_ratio/high_max': 0.031905731186270714, 'clip_ratio/region_mean': 0.14107865188270807, 'reward_total_mean': 0.3662221133708954, 'reward_meter_mean': 0.10800344496965408, 'reward_meter_std': 0.11452758312225342, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8984007835388184, 'reward_repeat_soft_std': 0.12824523448944092, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.3662221133708954, 'reward_total_composite_std': 0.040435004979372025, 'epoch': 0.09}

 28%|██▊       | 937/3300 [2:21:47<4:23:58,  6.70s/it]INFO 04-13 09:42:45 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 938/3300 [2:21:54<4:22:22,  6.66s/it]2026-04-13 09:42:51,998 | INFO | train_grpo_train | metrics_logged mode=train step=938

                                                      
{'loss': 0.0011, 'grad_norm': 9.277302742004395, 'learning_rate': 7.1606060606060615e-06, 'num_tokens': 1658376.0, 'completions/mean_length': 63.375, 'completions/min_length': 55.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8871058225631714, 'rewards/meter/std': 0.12349243462085724, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9183275699615479, 'rewards/repeat_soft/std': 0.05043455958366394, 'rewards/judge_quality/mean': 0.47749996185302734, 'rewards/judge_quality/std': 0.16263456642627716, 'rewards/total_composite/mean': 0.6136891841888428, 'rewards/total_composite/std': 0.1065220907330513, 'reward': 0.6136891841888428, 'reward_std': 0.1065220981836319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1586700975894928, 'sampling/sampling_logp_difference/max': 1.6482105255126953, 'sampling/importance_sampling_ratio/min': 0.19239388406276703, 'sampling/importance_sampling_ratio/mean': 1.0129984617233276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0819894894957542, 'clip_ratio/low_mean': 0.12331399507820606, 'clip_ratio/low_min': 0.12331399507820606, 'clip_ratio/high_mean': 0.014492753893136978, 'clip_ratio/high_max': 0.014492753893136978, 'clip_ratio/region_mean': 0.13780674897134304, 'reward_total_mean': 0.6136891841888428, 'reward_meter_mean': 0.8871058225631714, 'reward_meter_std': 0.12349243462085724, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9183275699615479, 'reward_repeat_soft_std': 0.05043455958366394, 'reward_judge_quality_mean': 0.47749996185302734, 'reward_judge_quality_std': 0.16263456642627716, 'reward_total_composite_mean': 0.6136891841888428, 'reward_total_composite_std': 0.1065220907330513, 'epoch': 0.09}

 28%|██▊       | 938/3300 [2:21:54<4:22:22,  6.66s/it]INFO 04-13 09:42:52 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 939/3300 [2:22:00<4:13:41,  6.45s/it]2026-04-13 09:42:57,935 | INFO | train_grpo_train | metrics_logged mode=train step=939

                                                      
{'loss': 0.0903, 'grad_norm': 14.62889289855957, 'learning_rate': 7.157575757575758e-06, 'num_tokens': 1660098.0, 'completions/mean_length': 47.25, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7967574596405029, 'rewards/meter/std': 0.32078492641448975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9507006406784058, 'rewards/repeat_soft/std': 0.043021660298109055, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.566234827041626, 'rewards/total_composite/std': 0.08880805224180222, 'reward': 0.566234827041626, 'reward_std': 0.08880805224180222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14013466238975525, 'sampling/sampling_logp_difference/max': 1.2640771865844727, 'sampling/importance_sampling_ratio/min': 0.3426993191242218, 'sampling/importance_sampling_ratio/mean': 1.023146629333496, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.096999168395996, 'clip_ratio/low_mean': 0.034519231878221035, 'clip_ratio/low_min': 0.034519231878221035, 'clip_ratio/high_mean': 0.14281363226473331, 'clip_ratio/high_max': 0.14281363226473331, 'clip_ratio/region_mean': 0.17733286414295435, 'reward_total_mean': 0.566234827041626, 'reward_meter_mean': 0.7967574596405029, 'reward_meter_std': 0.32078492641448975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9507006406784058, 'reward_repeat_soft_std': 0.043021660298109055, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.566234827041626, 'reward_total_composite_std': 0.08880805224180222, 'epoch': 0.09}

 28%|██▊       | 939/3300 [2:22:00<4:13:41,  6.45s/it]INFO 04-13 09:42:58 [block_pool.py:378] Successfully reset prefix cache

 28%|██▊       | 940/3300 [2:22:06<4:07:09,  6.28s/it]2026-04-13 09:43:03,836 | INFO | train_grpo_train | metrics_logged mode=train step=940

                                                      
{'loss': -0.025, 'grad_norm': 11.421670913696289, 'learning_rate': 7.154545454545455e-06, 'num_tokens': 1661672.0, 'completions/mean_length': 45.75, 'completions/min_length': 40.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5459588766098022, 'rewards/meter/std': 0.4084034562110901, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9860960245132446, 'rewards/repeat_soft/std': 0.015939654782414436, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.5541008114814758, 'rewards/total_composite/std': 0.17735260725021362, 'reward': 0.5541008114814758, 'reward_std': 0.17735260725021362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13485819101333618, 'sampling/sampling_logp_difference/max': 1.647550106048584, 'sampling/importance_sampling_ratio/min': 0.19252097606658936, 'sampling/importance_sampling_ratio/mean': 0.9911300539970398, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7478036507964134, 'clip_ratio/low_mean': 0.08339780941605568, 'clip_ratio/low_min': 0.08339780941605568, 'clip_ratio/high_mean': 0.05372035503387451, 'clip_ratio/high_max': 0.05372035503387451, 'clip_ratio/region_mean': 0.1371181644499302, 'reward_total_mean': 0.5541008114814758, 'reward_meter_mean': 0.5459588766098022, 'reward_meter_std': 0.4084034562110901, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9860960245132446, 'reward_repeat_soft_std': 0.015939654782414436, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.5541008114814758, 'reward_total_composite_std': 0.17735260725021362, 'epoch': 0.09}

 28%|██▊       | 940/3300 [2:22:06<4:07:09,  6.28s/it]INFO 04-13 09:43:04 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 941/3300 [2:22:16<5:01:11,  7.66s/it]2026-04-13 09:43:14,714 | INFO | train_grpo_train | metrics_logged mode=train step=941

                                                      
{'loss': -0.069, 'grad_norm': 2.9967703819274902, 'learning_rate': 7.151515151515152e-06, 'num_tokens': 1662933.0, 'completions/mean_length': 85.625, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.71428680419922, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6023797988891602, 'rewards/meter/std': 0.36645641922950745, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9555096626281738, 'rewards/repeat_soft/std': 0.01863044500350952, 'rewards/judge_quality/mean': 0.36000001430511475, 'rewards/judge_quality/std': 0.140813946723938, 'rewards/total_composite/mean': 0.4410613775253296, 'rewards/total_composite/std': 0.20302338898181915, 'reward': 0.4410613775253296, 'reward_std': 0.20302337408065796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1412070393562317, 'sampling/sampling_logp_difference/max': 0.9118883609771729, 'sampling/importance_sampling_ratio/min': 0.401764839887619, 'sampling/importance_sampling_ratio/mean': 0.9974734783172607, 'sampling/importance_sampling_ratio/max': 1.8005191087722778, 'entropy': 0.8800483867526054, 'clip_ratio/low_mean': 0.04326923284679651, 'clip_ratio/low_min': 0.04326923284679651, 'clip_ratio/high_mean': 0.08372922521084547, 'clip_ratio/high_max': 0.08372922521084547, 'clip_ratio/region_mean': 0.12699845805764198, 'reward_total_mean': 0.4410613775253296, 'reward_meter_mean': 0.6023797988891602, 'reward_meter_std': 0.36645641922950745, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9555096626281738, 'reward_repeat_soft_std': 0.01863044500350952, 'reward_judge_quality_mean': 0.36000001430511475, 'reward_judge_quality_std': 0.140813946723938, 'reward_total_composite_mean': 0.4410613775253296, 'reward_total_composite_std': 0.20302338898181915, 'epoch': 0.09}

 29%|██▊       | 941/3300 [2:22:16<5:01:11,  7.66s/it]INFO 04-13 09:43:14 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 942/3300 [2:22:22<4:38:08,  7.08s/it]2026-04-13 09:43:20,426 | INFO | train_grpo_train | metrics_logged mode=train step=942

                                                      
{'loss': 0.0174, 'grad_norm': 18.894412994384766, 'learning_rate': 7.148484848484849e-06, 'num_tokens': 1664417.0, 'completions/mean_length': 23.5, 'completions/min_length': 19.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6397724747657776, 'rewards/meter/std': 0.4632761776447296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9264246225357056, 'rewards/repeat_soft/std': 0.03450237587094307, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2921227812767029, 'rewards/total_composite/mean': 0.5938670635223389, 'rewards/total_composite/std': 0.23742662370204926, 'reward': 0.5938670635223389, 'reward_std': 0.23742660880088806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16714201867580414, 'sampling/sampling_logp_difference/max': 1.317133903503418, 'sampling/importance_sampling_ratio/min': 0.2679020166397095, 'sampling/importance_sampling_ratio/mean': 1.0264168977737427, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0534771531820297, 'clip_ratio/low_mean': 0.09427609574049711, 'clip_ratio/low_min': 0.09427609574049711, 'clip_ratio/high_mean': 0.05899122916162014, 'clip_ratio/high_max': 0.05899122916162014, 'clip_ratio/region_mean': 0.15326732490211725, 'reward_total_mean': 0.5938670635223389, 'reward_meter_mean': 0.6397724747657776, 'reward_meter_std': 0.4632761776447296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9264246225357056, 'reward_repeat_soft_std': 0.03450237587094307, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2921227812767029, 'reward_total_composite_mean': 0.5938670635223389, 'reward_total_composite_std': 0.23742662370204926, 'epoch': 0.09}

 29%|██▊       | 942/3300 [2:22:22<4:38:08,  7.08s/it]INFO 04-13 09:43:20 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 943/3300 [2:22:29<4:36:44,  7.04s/it]2026-04-13 09:43:27,397 | INFO | train_grpo_train | metrics_logged mode=train step=943

                                                      
{'loss': 0.0319, 'grad_norm': 11.56407356262207, 'learning_rate': 7.145454545454547e-06, 'num_tokens': 1666637.0, 'completions/mean_length': 73.5, 'completions/min_length': 68.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.7723885774612427, 'rewards/meter/std': 0.23276889324188232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9819412231445312, 'rewards/repeat_soft/std': 0.020347705110907555, 'rewards/judge_quality/mean': 0.6950000524520874, 'rewards/judge_quality/std': 0.1908627301454544, 'rewards/total_composite/mean': 0.6922646760940552, 'rewards/total_composite/std': 0.12439776957035065, 'reward': 0.6922646760940552, 'reward_std': 0.12439776957035065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12646839022636414, 'sampling/sampling_logp_difference/max': 1.4681923389434814, 'sampling/importance_sampling_ratio/min': 0.23034149408340454, 'sampling/importance_sampling_ratio/mean': 1.0263735055923462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7399717159569263, 'clip_ratio/low_mean': 0.06865095626562834, 'clip_ratio/low_min': 0.06865095626562834, 'clip_ratio/high_mean': 0.05612659826874733, 'clip_ratio/high_max': 0.05612659826874733, 'clip_ratio/region_mean': 0.12477755453437567, 'reward_total_mean': 0.6922646760940552, 'reward_meter_mean': 0.7723885774612427, 'reward_meter_std': 0.23276889324188232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9819412231445312, 'reward_repeat_soft_std': 0.020347705110907555, 'reward_judge_quality_mean': 0.6950000524520874, 'reward_judge_quality_std': 0.1908627301454544, 'reward_total_composite_mean': 0.6922646760940552, 'reward_total_composite_std': 0.12439776957035065, 'epoch': 0.09}

 29%|██▊       | 943/3300 [2:22:29<4:36:44,  7.04s/it]INFO 04-13 09:43:27 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 944/3300 [2:22:40<5:22:15,  8.21s/it]2026-04-13 09:43:38,317 | INFO | train_grpo_train | metrics_logged mode=train step=944

                                                      
{'loss': -0.0716, 'grad_norm': 2.4013564586639404, 'learning_rate': 7.142424242424243e-06, 'num_tokens': 1668010.0, 'completions/mean_length': 83.625, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.428571701049805, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6109882593154907, 'rewards/meter/std': 0.44488978385925293, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9552083015441895, 'rewards/repeat_soft/std': 0.02062394842505455, 'rewards/judge_quality/mean': 0.2462500035762787, 'rewards/judge_quality/std': 0.09913014620542526, 'rewards/total_composite/mean': 0.41316545009613037, 'rewards/total_composite/std': 0.18364334106445312, 'reward': 0.41316545009613037, 'reward_std': 0.18364332616329193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1917419284582138, 'sampling/sampling_logp_difference/max': 1.7236121892929077, 'sampling/importance_sampling_ratio/min': 0.17842049896717072, 'sampling/importance_sampling_ratio/mean': 0.9943436980247498, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0650127157568932, 'clip_ratio/low_mean': 0.03967391233891249, 'clip_ratio/low_min': 0.03967391233891249, 'clip_ratio/high_mean': 0.11422209721058607, 'clip_ratio/high_max': 0.11422209721058607, 'clip_ratio/region_mean': 0.15389600954949856, 'reward_total_mean': 0.41316545009613037, 'reward_meter_mean': 0.6109882593154907, 'reward_meter_std': 0.44488978385925293, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9552083015441895, 'reward_repeat_soft_std': 0.02062394842505455, 'reward_judge_quality_mean': 0.2462500035762787, 'reward_judge_quality_std': 0.09913014620542526, 'reward_total_composite_mean': 0.41316545009613037, 'reward_total_composite_std': 0.18364334106445312, 'epoch': 0.09}

 29%|██▊       | 944/3300 [2:22:40<5:22:15,  8.21s/it]INFO 04-13 09:43:38 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 945/3300 [2:22:46<4:57:17,  7.57s/it]2026-04-13 09:43:44,412 | INFO | train_grpo_train | metrics_logged mode=train step=945

                                                      
{'loss': -0.0008, 'grad_norm': 11.407129287719727, 'learning_rate': 7.1393939393939405e-06, 'num_tokens': 1669694.0, 'completions/mean_length': 55.5, 'completions/min_length': 39.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4709435701370239, 'rewards/meter/std': 0.409063458442688, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9676225185394287, 'rewards/repeat_soft/std': 0.028310758993029594, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.10260014235973358, 'rewards/total_composite/mean': 0.47954416275024414, 'rewards/total_composite/std': 0.11127520352602005, 'reward': 0.47954416275024414, 'reward_std': 0.11127523332834244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16715635359287262, 'sampling/sampling_logp_difference/max': 1.3540549278259277, 'sampling/importance_sampling_ratio/min': 0.25819119811058044, 'sampling/importance_sampling_ratio/mean': 1.025413990020752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4451678767800331, 'clip_ratio/low_mean': 0.11211142037063837, 'clip_ratio/low_min': 0.11211142037063837, 'clip_ratio/high_mean': 0.059318218380212784, 'clip_ratio/high_max': 0.059318218380212784, 'clip_ratio/region_mean': 0.17142963875085115, 'reward_total_mean': 0.47954416275024414, 'reward_meter_mean': 0.4709435701370239, 'reward_meter_std': 0.409063458442688, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9676225185394287, 'reward_repeat_soft_std': 0.028310758993029594, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.10260014235973358, 'reward_total_composite_mean': 0.47954416275024414, 'reward_total_composite_std': 0.11127520352602005, 'epoch': 0.09}

 29%|██▊       | 945/3300 [2:22:46<4:57:17,  7.57s/it]INFO 04-13 09:43:44 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 946/3300 [2:22:53<4:52:30,  7.46s/it]2026-04-13 09:43:51,589 | INFO | train_grpo_train | metrics_logged mode=train step=946

                                                      
{'loss': 0.0241, 'grad_norm': 9.67296314239502, 'learning_rate': 7.136363636363637e-06, 'num_tokens': 1672116.0, 'completions/mean_length': 103.75, 'completions/min_length': 94.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.75, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.955755889415741, 'rewards/meter/std': 0.0636955127120018, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8535679578781128, 'rewards/repeat_soft/std': 0.04481218382716179, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.5544466972351074, 'rewards/total_composite/std': 0.07041594386100769, 'reward': 0.5544466972351074, 'reward_std': 0.07041595131158829, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13135242462158203, 'sampling/sampling_logp_difference/max': 2.8089582920074463, 'sampling/importance_sampling_ratio/min': 0.06026774272322655, 'sampling/importance_sampling_ratio/mean': 1.0094482898712158, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9232650846242905, 'clip_ratio/low_mean': 0.0405056644231081, 'clip_ratio/low_min': 0.0405056644231081, 'clip_ratio/high_mean': 0.09079292230308056, 'clip_ratio/high_max': 0.09079292230308056, 'clip_ratio/region_mean': 0.13129858672618866, 'reward_total_mean': 0.5544466972351074, 'reward_meter_mean': 0.955755889415741, 'reward_meter_std': 0.0636955127120018, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8535679578781128, 'reward_repeat_soft_std': 0.04481218382716179, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.5544466972351074, 'reward_total_composite_std': 0.07041594386100769, 'epoch': 0.1}

 29%|██▊       | 946/3300 [2:22:53<4:52:30,  7.46s/it]INFO 04-13 09:43:51 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 947/3300 [2:23:00<4:43:49,  7.24s/it]2026-04-13 09:43:58,319 | INFO | train_grpo_train | metrics_logged mode=train step=947

                                                      
{'loss': -0.0478, 'grad_norm': 10.617412567138672, 'learning_rate': 7.133333333333334e-06, 'num_tokens': 1674288.0, 'completions/mean_length': 87.5, 'completions/min_length': 74.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.5, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.5281727313995361, 'rewards/meter/std': 0.2483045905828476, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9489359855651855, 'rewards/repeat_soft/std': 0.024047967046499252, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.48799845576286316, 'rewards/total_composite/std': 0.06793003529310226, 'reward': 0.48799845576286316, 'reward_std': 0.06793004274368286, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15323258936405182, 'sampling/sampling_logp_difference/max': 2.0040841102600098, 'sampling/importance_sampling_ratio/min': 0.13478368520736694, 'sampling/importance_sampling_ratio/mean': 0.9981544017791748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9498008489608765, 'clip_ratio/low_mean': 0.049493699334561825, 'clip_ratio/low_min': 0.049493699334561825, 'clip_ratio/high_mean': 0.09747860673815012, 'clip_ratio/high_max': 0.09747860673815012, 'clip_ratio/region_mean': 0.14697230607271194, 'reward_total_mean': 0.48799845576286316, 'reward_meter_mean': 0.5281727313995361, 'reward_meter_std': 0.2483045905828476, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9489359855651855, 'reward_repeat_soft_std': 0.024047967046499252, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.48799845576286316, 'reward_total_composite_std': 0.06793003529310226, 'epoch': 0.1}

 29%|██▊       | 947/3300 [2:23:00<4:43:49,  7.24s/it]INFO 04-13 09:43:58 [block_pool.py:378] Successfully reset prefix cache

 29%|██▊       | 948/3300 [2:23:11<5:25:33,  8.31s/it]2026-04-13 09:44:09,116 | INFO | train_grpo_train | metrics_logged mode=train step=948

                                                      
{'loss': -0.0637, 'grad_norm': 1.6304961442947388, 'learning_rate': 7.130303030303031e-06, 'num_tokens': 1675619.0, 'completions/mean_length': 147.375, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 25.83333396911621, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6447250247001648, 'rewards/meter/std': 0.39510872960090637, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9657489061355591, 'rewards/repeat_soft/std': 0.014413577504456043, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.18314221501350403, 'rewards/total_composite/mean': 0.4207678437232971, 'rewards/total_composite/std': 0.27831757068634033, 'reward': 0.4207678437232971, 'reward_std': 0.27831754088401794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13598209619522095, 'sampling/sampling_logp_difference/max': 1.261570930480957, 'sampling/importance_sampling_ratio/min': 0.28320878744125366, 'sampling/importance_sampling_ratio/mean': 1.0316158533096313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8389666080474854, 'clip_ratio/low_mean': 0.004999999888241291, 'clip_ratio/low_min': 0.004999999888241291, 'clip_ratio/high_mean': 0.08720085490494967, 'clip_ratio/high_max': 0.08720085490494967, 'clip_ratio/region_mean': 0.09220085479319096, 'reward_total_mean': 0.4207678437232971, 'reward_meter_mean': 0.6447250247001648, 'reward_meter_std': 0.39510872960090637, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9657489061355591, 'reward_repeat_soft_std': 0.014413577504456043, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.18314221501350403, 'reward_total_composite_mean': 0.4207678437232971, 'reward_total_composite_std': 0.27831757068634033, 'epoch': 0.1}

 29%|██▊       | 948/3300 [2:23:11<5:25:33,  8.31s/it]INFO 04-13 09:44:09 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 949/3300 [2:23:16<4:52:52,  7.47s/it]2026-04-13 09:44:14,653 | INFO | train_grpo_train | metrics_logged mode=train step=949

                                                      
{'loss': -0.0106, 'grad_norm': 21.439594268798828, 'learning_rate': 7.127272727272728e-06, 'num_tokens': 1676994.0, 'completions/mean_length': 23.875, 'completions/min_length': 16.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9674949645996094, 'rewards/meter/std': 0.03753045201301575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9585226774215698, 'rewards/repeat_soft/std': 0.011249415576457977, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22403763234615326, 'rewards/total_composite/mean': 0.6150607466697693, 'rewards/total_composite/std': 0.14532624185085297, 'reward': 0.6150607466697693, 'reward_std': 0.14532622694969177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14376404881477356, 'sampling/sampling_logp_difference/max': 1.6640040874481201, 'sampling/importance_sampling_ratio/min': 0.39237943291664124, 'sampling/importance_sampling_ratio/mean': 1.0260133743286133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9445859491825104, 'clip_ratio/low_mean': 0.049259479623287916, 'clip_ratio/low_min': 0.049259479623287916, 'clip_ratio/high_mean': 0.01993534481152892, 'clip_ratio/high_max': 0.01993534481152892, 'clip_ratio/region_mean': 0.06919482443481684, 'reward_total_mean': 0.6150607466697693, 'reward_meter_mean': 0.9674949645996094, 'reward_meter_std': 0.03753045201301575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9585226774215698, 'reward_repeat_soft_std': 0.011249415576457977, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22403763234615326, 'reward_total_composite_mean': 0.6150607466697693, 'reward_total_composite_std': 0.14532624185085297, 'epoch': 0.1}

 29%|██▉       | 949/3300 [2:23:16<4:52:52,  7.47s/it]INFO 04-13 09:44:14 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 950/3300 [2:23:23<4:41:10,  7.18s/it]2026-04-13 09:44:21,142 | INFO | train_grpo_train | metrics_logged mode=train step=950

                                                      
{'loss': 0.0616, 'grad_norm': 10.279980659484863, 'learning_rate': 7.124242424242424e-06, 'num_tokens': 1678891.0, 'completions/mean_length': 73.125, 'completions/min_length': 65.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.864984393119812, 'rewards/meter/std': 0.13354484736919403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8937699794769287, 'rewards/repeat_soft/std': 0.06704690307378769, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.5357272028923035, 'rewards/total_composite/std': 0.06531858444213867, 'reward': 0.5357272028923035, 'reward_std': 0.06531858444213867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1296449452638626, 'sampling/sampling_logp_difference/max': 1.644510269165039, 'sampling/importance_sampling_ratio/min': 0.19310709834098816, 'sampling/importance_sampling_ratio/mean': 1.0150951147079468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7161218076944351, 'clip_ratio/low_mean': 0.03791075386106968, 'clip_ratio/low_min': 0.03791075386106968, 'clip_ratio/high_mean': 0.06450229743495584, 'clip_ratio/high_max': 0.06450229743495584, 'clip_ratio/region_mean': 0.10241305129602551, 'reward_total_mean': 0.5357272028923035, 'reward_meter_mean': 0.864984393119812, 'reward_meter_std': 0.13354484736919403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8937699794769287, 'reward_repeat_soft_std': 0.06704690307378769, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.5357272028923035, 'reward_total_composite_std': 0.06531858444213867, 'epoch': 0.1}

 29%|██▉       | 950/3300 [2:23:23<4:41:10,  7.18s/it]INFO 04-13 09:44:21 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:13,  1.63s/it]

 30%|███       | 3/10 [00:07<00:18,  2.58s/it]

 40%|████      | 4/10 [00:11<00:19,  3.33s/it]

 50%|█████     | 5/10 [00:15<00:17,  3.43s/it]

 60%|██████    | 6/10 [00:25<00:21,  5.49s/it]

 70%|███████   | 7/10 [00:29<00:15,  5.10s/it]

 80%|████████  | 8/10 [00:38<00:12,  6.39s/it]

 90%|█████████ | 9/10 [00:42<00:05,  5.71s/it]

100%|██████████| 10/10 [00:46<00:00,  5.27s/it]2026-04-13 09:45:12,003 | INFO | train_grpo_train | metrics_logged mode=eval step=950

                                                      

                                               
{'eval_loss': nan, 'eval_runtime': 50.8034, 'eval_samples_per_second': 1.575, 'eval_steps_per_second': 0.197, 'eval_num_tokens': 1678891.0, 'eval_completions/mean_length': 82.2, 'eval_completions/min_length': 32.6, 'eval_completions/max_length': 196.9, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 71.36071472167968, 'eval_completions/min_terminated_length': 32.6, 'eval_completions/max_terminated_length': 117.7, 'eval_rewards/meter/mean': 0.6677819490432739, 'eval_rewards/meter/std': 0.3473196476697922, 'eval_rewards/count_adherence/mean': 0.9433333158493042, 'eval_rewards/count_adherence/std': 0.0914482433348894, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.1767766922712326, 'eval_rewards/repeat_soft/mean': 0.950187748670578, 'eval_rewards/repeat_soft/std': 0.0460129925981164, 'eval_rewards/judge_quality/mean': 0.4873750001192093, 'eval_rewards/judge_quality/std': 0.20080990493297576, 'eval_rewards/total_composite/mean': 0.5184092968702316, 'eval_rewards/total_composite/std': 0.18421085253357888, 'eval_reward': 0.5184092968702316, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08459883593022824, 'eval_sampling/sampling_logp_difference/max': 1.1512426376342773, 'eval_sampling/importance_sampling_ratio/min': 0.32473981827497483, 'eval_sampling/importance_sampling_ratio/mean': 1.0215644121170044, 'eval_sampling/importance_sampling_ratio/max': 1.4665158629417419, 'eval_entropy': 0.9833505392074585, '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': 0.5184092968702316, 'eval_reward_meter_mean': 0.6677819490432739, 'eval_reward_meter_std': 0.3473196476697922, 'eval_reward_count_adherence_mean': 0.9433333158493042, 'eval_reward_count_adherence_std': 0.0914482433348894, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.1767766922712326, 'eval_reward_repeat_soft_mean': 0.950187748670578, 'eval_reward_repeat_soft_std': 0.0460129925981164, 'eval_reward_judge_quality_mean': 0.4873750001192093, 'eval_reward_judge_quality_std': 0.20080990493297576, 'eval_reward_total_composite_mean': 0.5184092968702316, 'eval_reward_total_composite_std': 0.18421085253357888, 'epoch': 0.1}

 29%|██▉       | 950/3300 [2:24:14<4:41:10,  7.18s/it]

100%|██████████| 10/10 [00:47<00:00,  5.27s/it]

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

 29%|██▉       | 951/3300 [2:24:24<15:21:06, 23.53s/it]2026-04-13 09:45:22,817 | INFO | train_grpo_train | metrics_logged mode=train step=951

                                                       
{'loss': 0.0392, 'grad_norm': 12.941825866699219, 'learning_rate': 7.121212121212122e-06, 'num_tokens': 1680537.0, 'completions/mean_length': 50.75, 'completions/min_length': 46.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8829889297485352, 'rewards/meter/std': 0.169756680727005, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9108825922012329, 'rewards/repeat_soft/std': 0.09247001260519028, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.7023867964744568, 'rewards/total_composite/std': 0.09584323316812515, 'reward': 0.7023867964744568, 'reward_std': 0.09584324061870575, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15451835095882416, 'sampling/sampling_logp_difference/max': 1.7626707553863525, 'sampling/importance_sampling_ratio/min': 0.17158597707748413, 'sampling/importance_sampling_ratio/mean': 0.9980716705322266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.186119705438614, 'clip_ratio/low_mean': 0.05622141156345606, 'clip_ratio/low_min': 0.05622141156345606, 'clip_ratio/high_mean': 0.08247832767665386, 'clip_ratio/high_max': 0.08247832767665386, 'clip_ratio/region_mean': 0.13869973924010992, 'reward_total_mean': 0.7023867964744568, 'reward_meter_mean': 0.8829889297485352, 'reward_meter_std': 0.169756680727005, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9108825922012329, 'reward_repeat_soft_std': 0.09247001260519028, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.7023867964744568, 'reward_total_composite_std': 0.09584323316812515, 'epoch': 0.1}

 29%|██▉       | 951/3300 [2:24:25<15:21:06, 23.53s/it]INFO 04-13 09:45:23 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 952/3300 [2:24:31<11:59:39, 18.39s/it]2026-04-13 09:45:29,230 | INFO | train_grpo_train | metrics_logged mode=train step=952

                                                       
{'loss': 0.0533, 'grad_norm': 21.523618698120117, 'learning_rate': 7.118181818181819e-06, 'num_tokens': 1681936.0, 'completions/mean_length': 36.875, 'completions/min_length': 32.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5032766461372375, 'rewards/meter/std': 0.41357243061065674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9757304191589355, 'rewards/repeat_soft/std': 0.02462381310760975, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.6085785627365112, 'rewards/total_composite/std': 0.22026771306991577, 'reward': 0.6085785627365112, 'reward_std': 0.22026769816875458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16467581689357758, 'sampling/sampling_logp_difference/max': 1.5548391342163086, 'sampling/importance_sampling_ratio/min': 0.21122334897518158, 'sampling/importance_sampling_ratio/mean': 1.0285831689834595, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8324597105383873, 'clip_ratio/low_mean': 0.08368434477597475, 'clip_ratio/low_min': 0.08368434477597475, 'clip_ratio/high_mean': 0.03679234907031059, 'clip_ratio/high_max': 0.03679234907031059, 'clip_ratio/region_mean': 0.12047669384628534, 'reward_total_mean': 0.6085785627365112, 'reward_meter_mean': 0.5032766461372375, 'reward_meter_std': 0.41357243061065674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9757304191589355, 'reward_repeat_soft_std': 0.02462381310760975, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.6085785627365112, 'reward_total_composite_std': 0.22026771306991577, 'epoch': 0.1}

 29%|██▉       | 952/3300 [2:24:31<11:59:39, 18.39s/it]INFO 04-13 09:45:29 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 953/3300 [2:24:37<9:34:16, 14.68s/it] 2026-04-13 09:45:35,243 | INFO | train_grpo_train | metrics_logged mode=train step=953

                                                      
{'loss': 0.0129, 'grad_norm': 23.9711971282959, 'learning_rate': 7.115151515151516e-06, 'num_tokens': 1683227.0, 'completions/mean_length': 22.375, 'completions/min_length': 18.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.6797617673873901, 'rewards/meter/std': 0.3961387574672699, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9248495101928711, 'rewards/repeat_soft/std': 0.04800232872366905, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.5333417654037476, 'rewards/total_composite/std': 0.11237462610006332, 'reward': 0.5333417654037476, 'reward_std': 0.11237462610006332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1608961820602417, 'sampling/sampling_logp_difference/max': 1.4233884811401367, 'sampling/importance_sampling_ratio/min': 0.24089635908603668, 'sampling/importance_sampling_ratio/mean': 1.0058501958847046, 'sampling/importance_sampling_ratio/max': 1.7986422777175903, 'entropy': 1.142195574939251, 'clip_ratio/low_mean': 0.037638889625668526, 'clip_ratio/low_min': 0.037638889625668526, 'clip_ratio/high_mean': 0.11143544130027294, 'clip_ratio/high_max': 0.11143544130027294, 'clip_ratio/region_mean': 0.14907433092594147, 'reward_total_mean': 0.5333417654037476, 'reward_meter_mean': 0.6797617673873901, 'reward_meter_std': 0.3961387574672699, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9248495101928711, 'reward_repeat_soft_std': 0.04800232872366905, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.5333417654037476, 'reward_total_composite_std': 0.11237462610006332, 'epoch': 0.1}

 29%|██▉       | 953/3300 [2:24:37<9:34:16, 14.68s/it]INFO 04-13 09:45:35 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 954/3300 [2:24:48<8:54:31, 13.67s/it]2026-04-13 09:45:46,560 | INFO | train_grpo_train | metrics_logged mode=train step=954

                                                      
{'loss': -0.082, 'grad_norm': 3.38273286819458, 'learning_rate': 7.1121212121212125e-06, 'num_tokens': 1684714.0, 'completions/mean_length': 95.875, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.42857360839844, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6658740043640137, 'rewards/meter/std': 0.37135615944862366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9352182745933533, 'rewards/repeat_soft/std': 0.09705035388469696, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.3381884694099426, 'rewards/total_composite/mean': 0.569175124168396, 'rewards/total_composite/std': 0.29871445894241333, 'reward': 0.569175124168396, 'reward_std': 0.29871445894241333, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13497833907604218, 'sampling/sampling_logp_difference/max': 2.0144739151000977, 'sampling/importance_sampling_ratio/min': 0.13339056074619293, 'sampling/importance_sampling_ratio/mean': 1.0093296766281128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7048365473747253, 'clip_ratio/low_mean': 0.04988268483430147, 'clip_ratio/low_min': 0.04988268483430147, 'clip_ratio/high_mean': 0.06866123713552952, 'clip_ratio/high_max': 0.06866123713552952, 'clip_ratio/region_mean': 0.11854392196983099, 'reward_total_mean': 0.569175124168396, 'reward_meter_mean': 0.6658740043640137, 'reward_meter_std': 0.37135615944862366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9352182745933533, 'reward_repeat_soft_std': 0.09705035388469696, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.3381884694099426, 'reward_total_composite_mean': 0.569175124168396, 'reward_total_composite_std': 0.29871445894241333, 'epoch': 0.1}

 29%|██▉       | 954/3300 [2:24:48<8:54:31, 13.67s/it]INFO 04-13 09:45:46 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 955/3300 [2:24:55<7:35:36, 11.66s/it]2026-04-13 09:45:53,519 | INFO | train_grpo_train | metrics_logged mode=train step=955

                                                      
{'loss': 0.0058, 'grad_norm': 11.505900382995605, 'learning_rate': 7.10909090909091e-06, 'num_tokens': 1686808.0, 'completions/mean_length': 79.75, 'completions/min_length': 72.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.1853046715259552, 'rewards/meter/std': 0.05850255861878395, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8830885887145996, 'rewards/repeat_soft/std': 0.044910505414009094, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.34127578139305115, 'rewards/total_composite/std': 0.021256819367408752, 'reward': 0.34127578139305115, 'reward_std': 0.0212568212300539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15823736786842346, 'sampling/sampling_logp_difference/max': 1.647247314453125, 'sampling/importance_sampling_ratio/min': 0.19257928431034088, 'sampling/importance_sampling_ratio/mean': 1.0165362358093262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.031076304614544, 'clip_ratio/low_mean': 0.05299842543900013, 'clip_ratio/low_min': 0.05299842543900013, 'clip_ratio/high_mean': 0.09335256181657314, 'clip_ratio/high_max': 0.09335256181657314, 'clip_ratio/region_mean': 0.14635098725557327, 'reward_total_mean': 0.34127578139305115, 'reward_meter_mean': 0.1853046715259552, 'reward_meter_std': 0.05850255861878395, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8830885887145996, 'reward_repeat_soft_std': 0.044910505414009094, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.34127578139305115, 'reward_total_composite_std': 0.021256819367408752, 'epoch': 0.1}

 29%|██▉       | 955/3300 [2:24:55<7:35:36, 11.66s/it]INFO 04-13 09:45:53 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 956/3300 [2:25:02<6:36:25, 10.15s/it]2026-04-13 09:46:00,142 | INFO | train_grpo_train | metrics_logged mode=train step=956

                                                      
{'loss': 0.0546, 'grad_norm': 8.847649574279785, 'learning_rate': 7.106060606060606e-06, 'num_tokens': 1688962.0, 'completions/mean_length': 93.25, 'completions/min_length': 85.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.25, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.9277536273002625, 'rewards/meter/std': 0.129310742020607, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9281121492385864, 'rewards/repeat_soft/std': 0.056305643171072006, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5524935722351074, 'rewards/total_composite/std': 0.03305196017026901, 'reward': 0.5524935722351074, 'reward_std': 0.03305196762084961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1509913206100464, 'sampling/sampling_logp_difference/max': 1.9407892227172852, 'sampling/importance_sampling_ratio/min': 0.14359058439731598, 'sampling/importance_sampling_ratio/mean': 1.0155034065246582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2515875771641731, 'clip_ratio/low_mean': 0.019607843831181526, 'clip_ratio/low_min': 0.019607843831181526, 'clip_ratio/high_mean': 0.1317813778296113, 'clip_ratio/high_max': 0.1317813778296113, 'clip_ratio/region_mean': 0.15138922166079283, 'reward_total_mean': 0.5524935722351074, 'reward_meter_mean': 0.9277536273002625, 'reward_meter_std': 0.129310742020607, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9281121492385864, 'reward_repeat_soft_std': 0.056305643171072006, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5524935722351074, 'reward_total_composite_std': 0.03305196017026901, 'epoch': 0.1}

 29%|██▉       | 956/3300 [2:25:02<6:36:25, 10.15s/it]INFO 04-13 09:46:00 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 957/3300 [2:25:07<5:43:52,  8.81s/it]2026-04-13 09:46:05,820 | INFO | train_grpo_train | metrics_logged mode=train step=957

                                                      
{'loss': -0.0271, 'grad_norm': 19.07744026184082, 'learning_rate': 7.103030303030304e-06, 'num_tokens': 1690193.0, 'completions/mean_length': 24.875, 'completions/min_length': 20.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.456256628036499, 'rewards/meter/std': 0.4533495604991913, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9200776815414429, 'rewards/repeat_soft/std': 0.051106806844472885, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4714059829711914, 'rewards/total_composite/std': 0.13499127328395844, 'reward': 0.4714059829711914, 'reward_std': 0.13499125838279724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16945433616638184, 'sampling/sampling_logp_difference/max': 0.8939728736877441, 'sampling/importance_sampling_ratio/min': 0.4090275168418884, 'sampling/importance_sampling_ratio/mean': 1.0692710876464844, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5706066191196442, 'clip_ratio/low_mean': 0.10254121199250221, 'clip_ratio/low_min': 0.10254121199250221, 'clip_ratio/high_mean': 0.06304713990539312, 'clip_ratio/high_max': 0.06304713990539312, 'clip_ratio/region_mean': 0.16558835189789534, 'reward_total_mean': 0.4714059829711914, 'reward_meter_mean': 0.456256628036499, 'reward_meter_std': 0.4533495604991913, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9200776815414429, 'reward_repeat_soft_std': 0.051106806844472885, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4714059829711914, 'reward_total_composite_std': 0.13499127328395844, 'epoch': 0.1}

 29%|██▉       | 957/3300 [2:25:08<5:43:52,  8.81s/it]INFO 04-13 09:46:06 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 958/3300 [2:25:13<5:09:57,  7.94s/it]2026-04-13 09:46:11,741 | INFO | train_grpo_train | metrics_logged mode=train step=958

                                                      
{'loss': 0.0107, 'grad_norm': 13.06394100189209, 'learning_rate': 7.100000000000001e-06, 'num_tokens': 1691815.0, 'completions/mean_length': 42.75, 'completions/min_length': 40.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6433264017105103, 'rewards/meter/std': 0.31973007321357727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9257329106330872, 'rewards/repeat_soft/std': 0.08968224376440048, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.11667262762784958, 'rewards/total_composite/mean': 0.53730309009552, 'rewards/total_composite/std': 0.1151580959558487, 'reward': 0.53730309009552, 'reward_std': 0.1151580885052681, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14228375256061554, 'sampling/sampling_logp_difference/max': 2.067209005355835, 'sampling/importance_sampling_ratio/min': 0.1265384554862976, 'sampling/importance_sampling_ratio/mean': 1.0139461755752563, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.035200595855713, 'clip_ratio/low_mean': 0.08142944891005754, 'clip_ratio/low_min': 0.08142944891005754, 'clip_ratio/high_mean': 0.07406162843108177, 'clip_ratio/high_max': 0.07406162843108177, 'clip_ratio/region_mean': 0.15549107734113932, 'reward_total_mean': 0.53730309009552, 'reward_meter_mean': 0.6433264017105103, 'reward_meter_std': 0.31973007321357727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9257329106330872, 'reward_repeat_soft_std': 0.08968224376440048, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.11667262762784958, 'reward_total_composite_mean': 0.53730309009552, 'reward_total_composite_std': 0.1151580959558487, 'epoch': 0.1}

 29%|██▉       | 958/3300 [2:25:13<5:09:57,  7.94s/it]INFO 04-13 09:46:11 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 959/3300 [2:25:25<5:49:41,  8.96s/it]2026-04-13 09:46:23,189 | INFO | train_grpo_train | metrics_logged mode=train step=959

                                                      
{'loss': -0.056, 'grad_norm': 1.32066810131073, 'learning_rate': 7.096969696969698e-06, 'num_tokens': 1693141.0, 'completions/mean_length': 144.75, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 22.33333396911621, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.5230863690376282, 'rewards/meter/std': 0.36974969506263733, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.4087499976158142, 'rewards/judge_quality/std': 0.27445724606513977, 'rewards/total_composite/mean': 0.43389588594436646, 'rewards/total_composite/std': 0.27731138467788696, 'reward': 0.43389588594436646, 'reward_std': 0.27731141448020935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1253676563501358, 'sampling/sampling_logp_difference/max': 1.0962228775024414, 'sampling/importance_sampling_ratio/min': 0.33413076400756836, 'sampling/importance_sampling_ratio/mean': 1.0182522535324097, 'sampling/importance_sampling_ratio/max': 1.468483567237854, 'entropy': 0.8347451686859131, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09764190902933478, 'clip_ratio/high_max': 0.09764190902933478, 'clip_ratio/region_mean': 0.09764190902933478, 'reward_total_mean': 0.43389588594436646, 'reward_meter_mean': 0.5230863690376282, 'reward_meter_std': 0.36974969506263733, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.4087499976158142, 'reward_judge_quality_std': 0.27445724606513977, 'reward_total_composite_mean': 0.43389588594436646, 'reward_total_composite_std': 0.27731138467788696, 'epoch': 0.1}

 29%|██▉       | 959/3300 [2:25:25<5:49:41,  8.96s/it]INFO 04-13 09:46:23 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 960/3300 [2:25:31<5:19:38,  8.20s/it]2026-04-13 09:46:29,705 | INFO | train_grpo_train | metrics_logged mode=train step=960

                                                      
{'loss': 0.0433, 'grad_norm': 9.702564239501953, 'learning_rate': 7.093939393939394e-06, 'num_tokens': 1694803.0, 'completions/mean_length': 42.75, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8904345035552979, 'rewards/meter/std': 0.18871240317821503, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9875833988189697, 'rewards/repeat_soft/std': 0.006782568525522947, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.18431341648101807, 'rewards/total_composite/mean': 0.6410684585571289, 'rewards/total_composite/std': 0.06271853297948837, 'reward': 0.6410684585571289, 'reward_std': 0.06271851807832718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16253142058849335, 'sampling/sampling_logp_difference/max': 1.612654209136963, 'sampling/importance_sampling_ratio/min': 0.1993577778339386, 'sampling/importance_sampling_ratio/mean': 1.015665888786316, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9276193827390671, 'clip_ratio/low_mean': 0.12437493074685335, 'clip_ratio/low_min': 0.12437493074685335, 'clip_ratio/high_mean': 0.027439024299383163, 'clip_ratio/high_max': 0.027439024299383163, 'clip_ratio/region_mean': 0.15181395504623652, 'reward_total_mean': 0.6410684585571289, 'reward_meter_mean': 0.8904345035552979, 'reward_meter_std': 0.18871240317821503, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9875833988189697, 'reward_repeat_soft_std': 0.006782568525522947, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.18431341648101807, 'reward_total_composite_mean': 0.6410684585571289, 'reward_total_composite_std': 0.06271853297948837, 'epoch': 0.1}

 29%|██▉       | 960/3300 [2:25:31<5:19:38,  8.20s/it]INFO 04-13 09:46:29 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 961/3300 [2:25:38<4:58:28,  7.66s/it]2026-04-13 09:46:35,901 | INFO | train_grpo_train | metrics_logged mode=train step=961

                                                      
{'loss': 0.0405, 'grad_norm': 13.252286911010742, 'learning_rate': 7.0909090909090916e-06, 'num_tokens': 1696555.0, 'completions/mean_length': 37.0, 'completions/min_length': 32.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.47514355182647705, 'rewards/meter/std': 0.3408444821834564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9729394316673279, 'rewards/repeat_soft/std': 0.029924456030130386, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.614441990852356, 'rewards/total_composite/std': 0.21273763477802277, 'reward': 0.614441990852356, 'reward_std': 0.21273763477802277, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1581917554140091, 'sampling/sampling_logp_difference/max': 1.3821625709533691, 'sampling/importance_sampling_ratio/min': 0.27115383744239807, 'sampling/importance_sampling_ratio/mean': 1.0278681516647339, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9341058731079102, 'clip_ratio/low_mean': 0.07212148141115904, 'clip_ratio/low_min': 0.07212148141115904, 'clip_ratio/high_mean': 0.07070096489042044, 'clip_ratio/high_max': 0.07070096489042044, 'clip_ratio/region_mean': 0.14282244630157948, 'reward_total_mean': 0.614441990852356, 'reward_meter_mean': 0.47514355182647705, 'reward_meter_std': 0.3408444821834564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9729394316673279, 'reward_repeat_soft_std': 0.029924456030130386, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.614441990852356, 'reward_total_composite_std': 0.21273763477802277, 'epoch': 0.1}

 29%|██▉       | 961/3300 [2:25:38<4:58:28,  7.66s/it]INFO 04-13 09:46:36 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 962/3300 [2:25:43<4:37:39,  7.13s/it]2026-04-13 09:46:41,888 | INFO | train_grpo_train | metrics_logged mode=train step=962

                                                      
{'loss': -0.0148, 'grad_norm': 14.771444320678711, 'learning_rate': 7.087878787878788e-06, 'num_tokens': 1698103.0, 'completions/mean_length': 40.5, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7526412010192871, 'rewards/meter/std': 0.4042608439922333, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9692273139953613, 'rewards/repeat_soft/std': 0.039702706038951874, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.6720747947692871, 'rewards/total_composite/std': 0.22597767412662506, 'reward': 0.6720747947692871, 'reward_std': 0.22597767412662506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14504121243953705, 'sampling/sampling_logp_difference/max': 1.164337158203125, 'sampling/importance_sampling_ratio/min': 0.31212949752807617, 'sampling/importance_sampling_ratio/mean': 1.0402779579162598, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1788065060973167, 'clip_ratio/low_mean': 0.08813421241939068, 'clip_ratio/low_min': 0.08813421241939068, 'clip_ratio/high_mean': 0.06852390244603157, 'clip_ratio/high_max': 0.06852390244603157, 'clip_ratio/region_mean': 0.15665811486542225, 'reward_total_mean': 0.6720747947692871, 'reward_meter_mean': 0.7526412010192871, 'reward_meter_std': 0.4042608439922333, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9692273139953613, 'reward_repeat_soft_std': 0.039702706038951874, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.6720747947692871, 'reward_total_composite_std': 0.22597767412662506, 'epoch': 0.1}

 29%|██▉       | 962/3300 [2:25:44<4:37:39,  7.13s/it]INFO 04-13 09:46:42 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 963/3300 [2:25:50<4:28:08,  6.88s/it]2026-04-13 09:46:48,294 | INFO | train_grpo_train | metrics_logged mode=train step=963

                                                      
{'loss': -0.0079, 'grad_norm': 12.6827974319458, 'learning_rate': 7.084848484848485e-06, 'num_tokens': 1699921.0, 'completions/mean_length': 50.25, 'completions/min_length': 44.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.25, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6226515173912048, 'rewards/meter/std': 0.3655487895011902, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9648508429527283, 'rewards/repeat_soft/std': 0.0903712809085846, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.5672609806060791, 'rewards/total_composite/std': 0.1732962727546692, 'reward': 0.5672609806060791, 'reward_std': 0.1732962727546692, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1483352780342102, 'sampling/sampling_logp_difference/max': 1.4136152267456055, 'sampling/importance_sampling_ratio/min': 0.24326223134994507, 'sampling/importance_sampling_ratio/mean': 1.0176331996917725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3203382641077042, 'clip_ratio/low_mean': 0.10926157608628273, 'clip_ratio/low_min': 0.10926157608628273, 'clip_ratio/high_mean': 0.04935689456760883, 'clip_ratio/high_max': 0.04935689456760883, 'clip_ratio/region_mean': 0.15861847065389156, 'reward_total_mean': 0.5672609806060791, 'reward_meter_mean': 0.6226515173912048, 'reward_meter_std': 0.3655487895011902, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9648508429527283, 'reward_repeat_soft_std': 0.0903712809085846, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.5672609806060791, 'reward_total_composite_std': 0.1732962727546692, 'epoch': 0.1}

 29%|██▉       | 963/3300 [2:25:50<4:28:08,  6.88s/it]INFO 04-13 09:46:48 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 964/3300 [2:25:56<4:22:44,  6.75s/it]2026-04-13 09:46:54,649 | INFO | train_grpo_train | metrics_logged mode=train step=964

                                                      
{'loss': 0.0154, 'grad_norm': 13.873187065124512, 'learning_rate': 7.081818181818182e-06, 'num_tokens': 1701561.0, 'completions/mean_length': 41.0, 'completions/min_length': 37.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9538505673408508, 'rewards/meter/std': 0.05888332799077034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9307888150215149, 'rewards/repeat_soft/std': 0.06291955709457397, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.10260014235973358, 'rewards/total_composite/mean': 0.6308960914611816, 'rewards/total_composite/std': 0.07448755949735641, 'reward': 0.6308960914611816, 'reward_std': 0.07448754459619522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12618441879749298, 'sampling/sampling_logp_difference/max': 1.232431411743164, 'sampling/importance_sampling_ratio/min': 0.29158276319503784, 'sampling/importance_sampling_ratio/mean': 1.0101318359375, 'sampling/importance_sampling_ratio/max': 1.8735744953155518, 'entropy': 1.036969743669033, 'clip_ratio/low_mean': 0.054261722369119525, 'clip_ratio/low_min': 0.054261722369119525, 'clip_ratio/high_mean': 0.025240384973585606, 'clip_ratio/high_max': 0.025240384973585606, 'clip_ratio/region_mean': 0.07950210734270513, 'reward_total_mean': 0.6308960914611816, 'reward_meter_mean': 0.9538505673408508, 'reward_meter_std': 0.05888332799077034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9307888150215149, 'reward_repeat_soft_std': 0.06291955709457397, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.10260014235973358, 'reward_total_composite_mean': 0.6308960914611816, 'reward_total_composite_std': 0.07448755949735641, 'epoch': 0.1}

 29%|██▉       | 964/3300 [2:25:56<4:22:44,  6.75s/it]INFO 04-13 09:46:54 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 965/3300 [2:26:08<5:16:05,  8.12s/it]2026-04-13 09:47:05,861 | INFO | train_grpo_train | metrics_logged mode=train step=965

                                                      
{'loss': -0.096, 'grad_norm': 2.3022987842559814, 'learning_rate': 7.07878787878788e-06, 'num_tokens': 1703041.0, 'completions/mean_length': 160.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 42.66666793823242, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.4400330185890198, 'rewards/meter/std': 0.31648334860801697, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9835957288742065, 'rewards/repeat_soft/std': 0.015569952316582203, 'rewards/judge_quality/mean': 0.3762499988079071, 'rewards/judge_quality/std': 0.2775370180606842, 'rewards/total_composite/mean': 0.36539924144744873, 'rewards/total_composite/std': 0.23746727406978607, 'reward': 0.36539924144744873, 'reward_std': 0.23746725916862488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14925897121429443, 'sampling/sampling_logp_difference/max': 1.7131309509277344, 'sampling/importance_sampling_ratio/min': 0.18030039966106415, 'sampling/importance_sampling_ratio/mean': 1.0325299501419067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5369725301861763, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09194457437843084, 'clip_ratio/high_max': 0.09194457437843084, 'clip_ratio/region_mean': 0.09194457437843084, 'reward_total_mean': 0.36539924144744873, 'reward_meter_mean': 0.4400330185890198, 'reward_meter_std': 0.31648334860801697, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9835957288742065, 'reward_repeat_soft_std': 0.015569952316582203, 'reward_judge_quality_mean': 0.3762499988079071, 'reward_judge_quality_std': 0.2775370180606842, 'reward_total_composite_mean': 0.36539924144744873, 'reward_total_composite_std': 0.23746727406978607, 'epoch': 0.1}

 29%|██▉       | 965/3300 [2:26:08<5:16:05,  8.12s/it]INFO 04-13 09:47:06 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 966/3300 [2:26:14<4:53:31,  7.55s/it]2026-04-13 09:47:12,063 | INFO | train_grpo_train | metrics_logged mode=train step=966

                                                      
{'loss': 0.0334, 'grad_norm': 13.992204666137695, 'learning_rate': 7.075757575757576e-06, 'num_tokens': 1704582.0, 'completions/mean_length': 35.625, 'completions/min_length': 31.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.745686411857605, 'rewards/meter/std': 0.34231388568878174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9419390559196472, 'rewards/repeat_soft/std': 0.057287752628326416, 'rewards/judge_quality/mean': 0.39750000834465027, 'rewards/judge_quality/std': 0.10110107809305191, 'rewards/total_composite/mean': 0.5078482031822205, 'rewards/total_composite/std': 0.21228386461734772, 'reward': 0.5078482031822205, 'reward_std': 0.21228386461734772, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14791765809059143, 'sampling/sampling_logp_difference/max': 1.932054042816162, 'sampling/importance_sampling_ratio/min': 0.14485037326812744, 'sampling/importance_sampling_ratio/mean': 0.9893924593925476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.854806199669838, 'clip_ratio/low_mean': 0.020003555342555046, 'clip_ratio/low_min': 0.020003555342555046, 'clip_ratio/high_mean': 0.11828875355422497, 'clip_ratio/high_max': 0.11828875355422497, 'clip_ratio/region_mean': 0.13829230889678001, 'reward_total_mean': 0.5078482031822205, 'reward_meter_mean': 0.745686411857605, 'reward_meter_std': 0.34231388568878174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9419390559196472, 'reward_repeat_soft_std': 0.057287752628326416, 'reward_judge_quality_mean': 0.39750000834465027, 'reward_judge_quality_std': 0.10110107809305191, 'reward_total_composite_mean': 0.5078482031822205, 'reward_total_composite_std': 0.21228386461734772, 'epoch': 0.1}

 29%|██▉       | 966/3300 [2:26:14<4:53:31,  7.55s/it]INFO 04-13 09:47:12 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 967/3300 [2:26:21<4:44:21,  7.31s/it]2026-04-13 09:47:18,833 | INFO | train_grpo_train | metrics_logged mode=train step=967

                                                      
{'loss': 0.0013, 'grad_norm': 9.609872817993164, 'learning_rate': 7.072727272727273e-06, 'num_tokens': 1706536.0, 'completions/mean_length': 70.25, 'completions/min_length': 64.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.25, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5268456935882568, 'rewards/meter/std': 0.2801974415779114, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9801520109176636, 'rewards/repeat_soft/std': 0.029769711196422577, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.4549916684627533, 'rewards/total_composite/std': 0.22096389532089233, 'reward': 0.4549916684627533, 'reward_std': 0.22096388041973114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13744767010211945, 'sampling/sampling_logp_difference/max': 1.8090505599975586, 'sampling/importance_sampling_ratio/min': 0.16380958259105682, 'sampling/importance_sampling_ratio/mean': 1.023453950881958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6548539251089096, 'clip_ratio/low_mean': 0.041847581043839455, 'clip_ratio/low_min': 0.041847581043839455, 'clip_ratio/high_mean': 0.06069137854501605, 'clip_ratio/high_max': 0.06069137854501605, 'clip_ratio/region_mean': 0.1025389595888555, 'reward_total_mean': 0.4549916684627533, 'reward_meter_mean': 0.5268456935882568, 'reward_meter_std': 0.2801974415779114, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9801520109176636, 'reward_repeat_soft_std': 0.029769711196422577, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.4549916684627533, 'reward_total_composite_std': 0.22096389532089233, 'epoch': 0.1}

 29%|██▉       | 967/3300 [2:26:21<4:44:21,  7.31s/it]INFO 04-13 09:47:19 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 968/3300 [2:26:27<4:33:06,  7.03s/it]2026-04-13 09:47:25,192 | INFO | train_grpo_train | metrics_logged mode=train step=968

                                                      
{'loss': -0.0144, 'grad_norm': 13.045398712158203, 'learning_rate': 7.06969696969697e-06, 'num_tokens': 1708262.0, 'completions/mean_length': 50.75, 'completions/min_length': 40.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.5692801475524902, 'rewards/meter/std': 0.3768622875213623, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9812266826629639, 'rewards/repeat_soft/std': 0.024538792669773102, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5042853951454163, 'rewards/total_composite/std': 0.10177413374185562, 'reward': 0.5042853951454163, 'reward_std': 0.10177412629127502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15098918974399567, 'sampling/sampling_logp_difference/max': 1.3089690208435059, 'sampling/importance_sampling_ratio/min': 0.27009838819503784, 'sampling/importance_sampling_ratio/mean': 1.0211642980575562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2397021055221558, 'clip_ratio/low_mean': 0.05614245776087046, 'clip_ratio/low_min': 0.05614245776087046, 'clip_ratio/high_mean': 0.06984651274979115, 'clip_ratio/high_max': 0.06984651274979115, 'clip_ratio/region_mean': 0.1259889705106616, 'reward_total_mean': 0.5042853951454163, 'reward_meter_mean': 0.5692801475524902, 'reward_meter_std': 0.3768622875213623, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9812266826629639, 'reward_repeat_soft_std': 0.024538792669773102, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5042853951454163, 'reward_total_composite_std': 0.10177413374185562, 'epoch': 0.1}

 29%|██▉       | 968/3300 [2:26:27<4:33:06,  7.03s/it]INFO 04-13 09:47:25 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 969/3300 [2:26:33<4:25:00,  6.82s/it]2026-04-13 09:47:31,533 | INFO | train_grpo_train | metrics_logged mode=train step=969

                                                      
{'loss': -0.0054, 'grad_norm': 13.439970970153809, 'learning_rate': 7.066666666666667e-06, 'num_tokens': 1709864.0, 'completions/mean_length': 50.25, 'completions/min_length': 47.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.25, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8206380009651184, 'rewards/meter/std': 0.33159759640693665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9912235736846924, 'rewards/repeat_soft/std': 0.008905306458473206, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.534211277961731, 'rewards/total_composite/std': 0.21942123770713806, 'reward': 0.534211277961731, 'reward_std': 0.21942123770713806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1369597315788269, 'sampling/sampling_logp_difference/max': 1.4509609937667847, 'sampling/importance_sampling_ratio/min': 0.2343449741601944, 'sampling/importance_sampling_ratio/mean': 1.0040875673294067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9222754910588264, 'clip_ratio/low_mean': 0.02849544119089842, 'clip_ratio/low_min': 0.02849544119089842, 'clip_ratio/high_mean': 0.10500968247652054, 'clip_ratio/high_max': 0.10500968247652054, 'clip_ratio/region_mean': 0.13350512366741896, 'reward_total_mean': 0.534211277961731, 'reward_meter_mean': 0.8206380009651184, 'reward_meter_std': 0.33159759640693665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9912235736846924, 'reward_repeat_soft_std': 0.008905306458473206, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.534211277961731, 'reward_total_composite_std': 0.21942123770713806, 'epoch': 0.1}

 29%|██▉       | 969/3300 [2:26:33<4:25:00,  6.82s/it]INFO 04-13 09:47:31 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 970/3300 [2:26:44<5:15:26,  8.12s/it]2026-04-13 09:47:42,695 | INFO | train_grpo_train | metrics_logged mode=train step=970

                                                      
{'loss': -0.1169, 'grad_norm': 2.641662836074829, 'learning_rate': 7.063636363636365e-06, 'num_tokens': 1711459.0, 'completions/mean_length': 102.375, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.85714340209961, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6911942958831787, 'rewards/meter/std': 0.3835291266441345, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9280762672424316, 'rewards/repeat_soft/std': 0.045472707599401474, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.4747639000415802, 'rewards/total_composite/std': 0.21479584276676178, 'reward': 0.4747639000415802, 'reward_std': 0.21479582786560059, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13141244649887085, 'sampling/sampling_logp_difference/max': 1.3056834936141968, 'sampling/importance_sampling_ratio/min': 0.27098727226257324, 'sampling/importance_sampling_ratio/mean': 1.0224891901016235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7379521951079369, 'clip_ratio/low_mean': 0.033444148022681475, 'clip_ratio/low_min': 0.033444148022681475, 'clip_ratio/high_mean': 0.09196178521960974, 'clip_ratio/high_max': 0.09196178521960974, 'clip_ratio/region_mean': 0.1254059332422912, 'reward_total_mean': 0.4747639000415802, 'reward_meter_mean': 0.6911942958831787, 'reward_meter_std': 0.3835291266441345, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9280762672424316, 'reward_repeat_soft_std': 0.045472707599401474, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.4747639000415802, 'reward_total_composite_std': 0.21479584276676178, 'epoch': 0.1}

 29%|██▉       | 970/3300 [2:26:44<5:15:26,  8.12s/it]INFO 04-13 09:47:42 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 971/3300 [2:26:50<4:49:18,  7.45s/it]2026-04-13 09:47:48,582 | INFO | train_grpo_train | metrics_logged mode=train step=971

                                                      
{'loss': 0.0848, 'grad_norm': 14.68692684173584, 'learning_rate': 7.060606060606061e-06, 'num_tokens': 1712834.0, 'completions/mean_length': 25.875, 'completions/min_length': 22.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.875, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.14169715344905853, 'rewards/meter/std': 0.2590996026992798, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9521924257278442, 'rewards/repeat_soft/std': 0.018484456464648247, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.3912311792373657, 'rewards/total_composite/std': 0.07280620187520981, 'reward': 0.3912311792373657, 'reward_std': 0.07280620187520981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1792679876089096, 'sampling/sampling_logp_difference/max': 2.050839424133301, 'sampling/importance_sampling_ratio/min': 0.12862688302993774, 'sampling/importance_sampling_ratio/mean': 1.0485748052597046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4391480088233948, 'clip_ratio/low_mean': 0.10998993832617998, 'clip_ratio/low_min': 0.10998993832617998, 'clip_ratio/high_mean': 0.04471343941986561, 'clip_ratio/high_max': 0.04471343941986561, 'clip_ratio/region_mean': 0.1547033777460456, 'reward_total_mean': 0.3912311792373657, 'reward_meter_mean': 0.14169715344905853, 'reward_meter_std': 0.2590996026992798, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9521924257278442, 'reward_repeat_soft_std': 0.018484456464648247, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.3912311792373657, 'reward_total_composite_std': 0.07280620187520981, 'epoch': 0.1}

 29%|██▉       | 971/3300 [2:26:50<4:49:18,  7.45s/it]INFO 04-13 09:47:48 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 972/3300 [2:26:57<4:38:54,  7.19s/it]2026-04-13 09:47:55,152 | INFO | train_grpo_train | metrics_logged mode=train step=972

                                                      
{'loss': 0.065, 'grad_norm': 13.530476570129395, 'learning_rate': 7.057575757575759e-06, 'num_tokens': 1714822.0, 'completions/mean_length': 75.5, 'completions/min_length': 67.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.5, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.43695682287216187, 'rewards/meter/std': 0.3026445209980011, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9667505025863647, 'rewards/repeat_soft/std': 0.025144891813397408, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4592045247554779, 'rewards/total_composite/std': 0.0866062343120575, 'reward': 0.4592045247554779, 'reward_std': 0.08660624176263809, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16183550655841827, 'sampling/sampling_logp_difference/max': 1.2516584396362305, 'sampling/importance_sampling_ratio/min': 0.2860300540924072, 'sampling/importance_sampling_ratio/mean': 1.0253628492355347, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4893262386322021, 'clip_ratio/low_mean': 0.10574120655655861, 'clip_ratio/low_min': 0.10574120655655861, 'clip_ratio/high_mean': 0.06007573939859867, 'clip_ratio/high_max': 0.06007573939859867, 'clip_ratio/region_mean': 0.16581694595515728, 'reward_total_mean': 0.4592045247554779, 'reward_meter_mean': 0.43695682287216187, 'reward_meter_std': 0.3026445209980011, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9667505025863647, 'reward_repeat_soft_std': 0.025144891813397408, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4592045247554779, 'reward_total_composite_std': 0.0866062343120575, 'epoch': 0.1}

 29%|██▉       | 972/3300 [2:26:57<4:38:54,  7.19s/it]INFO 04-13 09:47:55 [block_pool.py:378] Successfully reset prefix cache

 29%|██▉       | 973/3300 [2:27:03<4:24:01,  6.81s/it]2026-04-13 09:48:01,072 | INFO | train_grpo_train | metrics_logged mode=train step=973

                                                      
{'loss': 0.091, 'grad_norm': 11.856672286987305, 'learning_rate': 7.054545454545455e-06, 'num_tokens': 1716496.0, 'completions/mean_length': 44.25, 'completions/min_length': 31.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9789201021194458, 'rewards/meter/std': 0.013528779149055481, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.938413679599762, 'rewards/repeat_soft/std': 0.04366263374686241, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.194054514169693, 'rewards/total_composite/mean': 0.7009687423706055, 'rewards/total_composite/std': 0.13211536407470703, 'reward': 0.7009687423706055, 'reward_std': 0.13211536407470703, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14514295756816864, 'sampling/sampling_logp_difference/max': 1.3580703735351562, 'sampling/importance_sampling_ratio/min': 0.25715652108192444, 'sampling/importance_sampling_ratio/mean': 1.0207626819610596, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2455740720033646, 'clip_ratio/low_mean': 0.06271467031911016, 'clip_ratio/low_min': 0.06271467031911016, 'clip_ratio/high_mean': 0.06903463043272495, 'clip_ratio/high_max': 0.06903463043272495, 'clip_ratio/region_mean': 0.1317493007518351, 'reward_total_mean': 0.7009687423706055, 'reward_meter_mean': 0.9789201021194458, 'reward_meter_std': 0.013528779149055481, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.938413679599762, 'reward_repeat_soft_std': 0.04366263374686241, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.194054514169693, 'reward_total_composite_mean': 0.7009687423706055, 'reward_total_composite_std': 0.13211536407470703, 'epoch': 0.1}

 29%|██▉       | 973/3300 [2:27:03<4:24:01,  6.81s/it]INFO 04-13 09:48:01 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 974/3300 [2:27:09<4:15:50,  6.60s/it]2026-04-13 09:48:07,185 | INFO | train_grpo_train | metrics_logged mode=train step=974

                                                      
{'loss': 0.0094, 'grad_norm': 12.079497337341309, 'learning_rate': 7.0515151515151525e-06, 'num_tokens': 1718125.0, 'completions/mean_length': 51.625, 'completions/min_length': 40.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.920598030090332, 'rewards/meter/std': 0.11000407487154007, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9650930762290955, 'rewards/repeat_soft/std': 0.047561854124069214, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.664365291595459, 'rewards/total_composite/std': 0.0953650251030922, 'reward': 0.664365291595459, 'reward_std': 0.09536503255367279, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13764303922653198, 'sampling/sampling_logp_difference/max': 1.8628220558166504, 'sampling/importance_sampling_ratio/min': 0.1552339345216751, 'sampling/importance_sampling_ratio/mean': 1.0276992321014404, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1694976165890694, 'clip_ratio/low_mean': 0.0747088547796011, 'clip_ratio/low_min': 0.0747088547796011, 'clip_ratio/high_mean': 0.03707483038306236, 'clip_ratio/high_max': 0.03707483038306236, 'clip_ratio/region_mean': 0.11178368516266346, 'reward_total_mean': 0.664365291595459, 'reward_meter_mean': 0.920598030090332, 'reward_meter_std': 0.11000407487154007, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9650930762290955, 'reward_repeat_soft_std': 0.047561854124069214, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.664365291595459, 'reward_total_composite_std': 0.0953650251030922, 'epoch': 0.1}

 30%|██▉       | 974/3300 [2:27:09<4:15:50,  6.60s/it]INFO 04-13 09:48:07 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 975/3300 [2:27:20<5:08:16,  7.96s/it]2026-04-13 09:48:18,308 | INFO | train_grpo_train | metrics_logged mode=train step=975

                                                      
{'loss': -0.1509, 'grad_norm': 2.8824329376220703, 'learning_rate': 7.048484848484849e-06, 'num_tokens': 1719952.0, 'completions/mean_length': 124.375, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.709470272064209, 'rewards/meter/std': 0.3198787271976471, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9301737546920776, 'rewards/repeat_soft/std': 0.04480618238449097, 'rewards/judge_quality/mean': 0.44874998927116394, 'rewards/judge_quality/std': 0.2105392962694168, 'rewards/total_composite/mean': 0.5060956478118896, 'rewards/total_composite/std': 0.23397000133991241, 'reward': 0.5060956478118896, 'reward_std': 0.23396998643875122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14583653211593628, 'sampling/sampling_logp_difference/max': 1.4082255363464355, 'sampling/importance_sampling_ratio/min': 0.24457688629627228, 'sampling/importance_sampling_ratio/mean': 1.031238317489624, 'sampling/importance_sampling_ratio/max': 1.9890636205673218, 'entropy': 1.0118075907230377, 'clip_ratio/low_mean': 0.021739130839705467, 'clip_ratio/low_min': 0.021739130839705467, 'clip_ratio/high_mean': 0.10367511957883835, 'clip_ratio/high_max': 0.10367511957883835, 'clip_ratio/region_mean': 0.12541425041854382, 'reward_total_mean': 0.5060956478118896, 'reward_meter_mean': 0.709470272064209, 'reward_meter_std': 0.3198787271976471, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9301737546920776, 'reward_repeat_soft_std': 0.04480618238449097, 'reward_judge_quality_mean': 0.44874998927116394, 'reward_judge_quality_std': 0.2105392962694168, 'reward_total_composite_mean': 0.5060956478118896, 'reward_total_composite_std': 0.23397000133991241, 'epoch': 0.1}

 30%|██▉       | 975/3300 [2:27:20<5:08:16,  7.96s/it]INFO 04-13 09:48:18 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 976/3300 [2:27:26<4:46:55,  7.41s/it]2026-04-13 09:48:24,436 | INFO | train_grpo_train | metrics_logged mode=train step=976

                                                      
{'loss': 0.0768, 'grad_norm': 12.867921829223633, 'learning_rate': 7.045454545454546e-06, 'num_tokens': 1721600.0, 'completions/mean_length': 52.0, 'completions/min_length': 46.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7738261222839355, 'rewards/meter/std': 0.3358989655971527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9791769981384277, 'rewards/repeat_soft/std': 0.017892280593514442, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.5318248271942139, 'rewards/total_composite/std': 0.26320815086364746, 'reward': 0.5318248271942139, 'reward_std': 0.26320815086364746, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.160322904586792, 'sampling/sampling_logp_difference/max': 1.9544696807861328, 'sampling/importance_sampling_ratio/min': 0.1416395604610443, 'sampling/importance_sampling_ratio/mean': 1.0299495458602905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2601183354854584, 'clip_ratio/low_mean': 0.046510683838278055, 'clip_ratio/low_min': 0.046510683838278055, 'clip_ratio/high_mean': 0.078507199883461, 'clip_ratio/high_max': 0.078507199883461, 'clip_ratio/region_mean': 0.12501788372173905, 'reward_total_mean': 0.5318248271942139, 'reward_meter_mean': 0.7738261222839355, 'reward_meter_std': 0.3358989655971527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9791769981384277, 'reward_repeat_soft_std': 0.017892280593514442, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.5318248271942139, 'reward_total_composite_std': 0.26320815086364746, 'epoch': 0.1}

 30%|██▉       | 976/3300 [2:27:26<4:46:55,  7.41s/it]INFO 04-13 09:48:24 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 977/3300 [2:27:32<4:30:09,  6.98s/it]2026-04-13 09:48:30,407 | INFO | train_grpo_train | metrics_logged mode=train step=977

                                                      
{'loss': 0.0766, 'grad_norm': 16.257875442504883, 'learning_rate': 7.0424242424242426e-06, 'num_tokens': 1723047.0, 'completions/mean_length': 30.875, 'completions/min_length': 25.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.875, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9210174083709717, 'rewards/meter/std': 0.06121338531374931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9290193319320679, 'rewards/repeat_soft/std': 0.05646272376179695, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.7440645694732666, 'rewards/total_composite/std': 0.17584004998207092, 'reward': 0.7440645694732666, 'reward_std': 0.17584003508090973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12017352879047394, 'sampling/sampling_logp_difference/max': 1.0964205265045166, 'sampling/importance_sampling_ratio/min': 0.44648265838623047, 'sampling/importance_sampling_ratio/mean': 1.02653968334198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7153010219335556, 'clip_ratio/low_mean': 0.07673509418964386, 'clip_ratio/low_min': 0.07673509418964386, 'clip_ratio/high_mean': 0.05919354781508446, 'clip_ratio/high_max': 0.05919354781508446, 'clip_ratio/region_mean': 0.13592864200472832, 'reward_total_mean': 0.7440645694732666, 'reward_meter_mean': 0.9210174083709717, 'reward_meter_std': 0.06121338531374931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9290193319320679, 'reward_repeat_soft_std': 0.05646272376179695, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.7440645694732666, 'reward_total_composite_std': 0.17584004998207092, 'epoch': 0.1}

 30%|██▉       | 977/3300 [2:27:32<4:30:09,  6.98s/it]INFO 04-13 09:48:30 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 978/3300 [2:27:38<4:21:19,  6.75s/it]2026-04-13 09:48:36,634 | INFO | train_grpo_train | metrics_logged mode=train step=978

                                                      
{'loss': 0.1118, 'grad_norm': 12.845317840576172, 'learning_rate': 7.039393939393941e-06, 'num_tokens': 1724862.0, 'completions/mean_length': 51.875, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.5170915126800537, 'rewards/meter/std': 0.4167918562889099, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9292737245559692, 'rewards/repeat_soft/std': 0.06074531748890877, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5086134672164917, 'rewards/total_composite/std': 0.20576469600200653, 'reward': 0.5086134672164917, 'reward_std': 0.20576469600200653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16256125271320343, 'sampling/sampling_logp_difference/max': 2.1671018600463867, 'sampling/importance_sampling_ratio/min': 0.11450900137424469, 'sampling/importance_sampling_ratio/mean': 1.0142372846603394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0791840329766273, 'clip_ratio/low_mean': 0.06558190565556288, 'clip_ratio/low_min': 0.06558190565556288, 'clip_ratio/high_mean': 0.07605824619531631, 'clip_ratio/high_max': 0.07605824619531631, 'clip_ratio/region_mean': 0.1416401518508792, 'reward_total_mean': 0.5086134672164917, 'reward_meter_mean': 0.5170915126800537, 'reward_meter_std': 0.4167918562889099, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9292737245559692, 'reward_repeat_soft_std': 0.06074531748890877, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5086134672164917, 'reward_total_composite_std': 0.20576469600200653, 'epoch': 0.1}

 30%|██▉       | 978/3300 [2:27:38<4:21:19,  6.75s/it]INFO 04-13 09:48:36 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 979/3300 [2:27:44<4:09:34,  6.45s/it]2026-04-13 09:48:42,388 | INFO | train_grpo_train | metrics_logged mode=train step=979

                                                      
{'loss': 0.2517, 'grad_norm': 17.47881317138672, 'learning_rate': 7.036363636363637e-06, 'num_tokens': 1726339.0, 'completions/mean_length': 25.625, 'completions/min_length': 20.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9364637136459351, 'rewards/meter/std': 0.11381543427705765, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9130376577377319, 'rewards/repeat_soft/std': 0.05440940707921982, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.16974246501922607, 'rewards/total_composite/mean': 0.6187773942947388, 'rewards/total_composite/std': 0.15915215015411377, 'reward': 0.6187773942947388, 'reward_std': 0.15915215015411377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13362014293670654, 'sampling/sampling_logp_difference/max': 1.0516290664672852, 'sampling/importance_sampling_ratio/min': 0.3493681252002716, 'sampling/importance_sampling_ratio/mean': 1.01193106174469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9321458712220192, 'clip_ratio/low_mean': 0.06847826205193996, 'clip_ratio/low_min': 0.06847826205193996, 'clip_ratio/high_mean': 0.10737812984734774, 'clip_ratio/high_max': 0.10737812984734774, 'clip_ratio/region_mean': 0.1758563918992877, 'reward_total_mean': 0.6187773942947388, 'reward_meter_mean': 0.9364637136459351, 'reward_meter_std': 0.11381543427705765, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9130376577377319, 'reward_repeat_soft_std': 0.05440940707921982, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.16974246501922607, 'reward_total_composite_mean': 0.6187773942947388, 'reward_total_composite_std': 0.15915215015411377, 'epoch': 0.1}

 30%|██▉       | 979/3300 [2:27:44<4:09:34,  6.45s/it]INFO 04-13 09:48:42 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 980/3300 [2:27:51<4:13:51,  6.57s/it]2026-04-13 09:48:49,218 | INFO | train_grpo_train | metrics_logged mode=train step=980

                                                      
{'loss': 0.0489, 'grad_norm': 9.139081954956055, 'learning_rate': 7.033333333333334e-06, 'num_tokens': 1728496.0, 'completions/mean_length': 90.625, 'completions/min_length': 75.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.625, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.7046166062355042, 'rewards/meter/std': 0.33246853947639465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8811008334159851, 'rewards/repeat_soft/std': 0.04722922295331955, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5273675918579102, 'rewards/total_composite/std': 0.08341158181428909, 'reward': 0.5273675918579102, 'reward_std': 0.0834115743637085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13319680094718933, 'sampling/sampling_logp_difference/max': 1.6853256225585938, 'sampling/importance_sampling_ratio/min': 0.18538405001163483, 'sampling/importance_sampling_ratio/mean': 0.9984402060508728, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9484239295125008, 'clip_ratio/low_mean': 0.056068320758640766, 'clip_ratio/low_min': 0.056068320758640766, 'clip_ratio/high_mean': 0.09317009709775448, 'clip_ratio/high_max': 0.09317009709775448, 'clip_ratio/region_mean': 0.14923841785639524, 'reward_total_mean': 0.5273675918579102, 'reward_meter_mean': 0.7046166062355042, 'reward_meter_std': 0.33246853947639465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8811008334159851, 'reward_repeat_soft_std': 0.04722922295331955, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5273675918579102, 'reward_total_composite_std': 0.08341158181428909, 'epoch': 0.1}

 30%|██▉       | 980/3300 [2:27:51<4:13:51,  6.57s/it]INFO 04-13 09:48:49 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 981/3300 [2:27:57<4:08:02,  6.42s/it]2026-04-13 09:48:55,291 | INFO | train_grpo_train | metrics_logged mode=train step=981

                                                      
{'loss': 0.0621, 'grad_norm': 12.545612335205078, 'learning_rate': 7.030303030303031e-06, 'num_tokens': 1730106.0, 'completions/mean_length': 52.25, 'completions/min_length': 48.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4504271447658539, 'rewards/meter/std': 0.4088876247406006, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9795470237731934, 'rewards/repeat_soft/std': 0.03572678565979004, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.48322439193725586, 'rewards/total_composite/std': 0.10845566540956497, 'reward': 0.48322439193725586, 'reward_std': 0.10845565795898438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16551512479782104, 'sampling/sampling_logp_difference/max': 3.132415771484375, 'sampling/importance_sampling_ratio/min': 0.04361231252551079, 'sampling/importance_sampling_ratio/mean': 1.012925386428833, 'sampling/importance_sampling_ratio/max': 1.972832441329956, 'entropy': 1.3978762775659561, 'clip_ratio/low_mean': 0.05751907266676426, 'clip_ratio/low_min': 0.05751907266676426, 'clip_ratio/high_mean': 0.07019991148263216, 'clip_ratio/high_max': 0.07019991148263216, 'clip_ratio/region_mean': 0.12771898414939642, 'reward_total_mean': 0.48322439193725586, 'reward_meter_mean': 0.4504271447658539, 'reward_meter_std': 0.4088876247406006, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9795470237731934, 'reward_repeat_soft_std': 0.03572678565979004, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.48322439193725586, 'reward_total_composite_std': 0.10845566540956497, 'epoch': 0.1}

 30%|██▉       | 981/3300 [2:27:57<4:08:02,  6.42s/it]INFO 04-13 09:48:55 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 982/3300 [2:28:03<4:02:10,  6.27s/it]2026-04-13 09:49:01,212 | INFO | train_grpo_train | metrics_logged mode=train step=982

                                                      
{'loss': 0.007, 'grad_norm': 12.460480690002441, 'learning_rate': 7.027272727272728e-06, 'num_tokens': 1731808.0, 'completions/mean_length': 50.75, 'completions/min_length': 47.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4085719585418701, 'rewards/meter/std': 0.3669445812702179, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877614378929138, 'rewards/repeat_soft/std': 0.015099510550498962, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.46311163902282715, 'rewards/total_composite/std': 0.10389380156993866, 'reward': 0.46311163902282715, 'reward_std': 0.10389378666877747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14943093061447144, 'sampling/sampling_logp_difference/max': 1.428863286972046, 'sampling/importance_sampling_ratio/min': 0.23958110809326172, 'sampling/importance_sampling_ratio/mean': 0.9974090456962585, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0441494435071945, 'clip_ratio/low_mean': 0.09914177004247904, 'clip_ratio/low_min': 0.09914177004247904, 'clip_ratio/high_mean': 0.03319543041288853, 'clip_ratio/high_max': 0.03319543041288853, 'clip_ratio/region_mean': 0.13233720045536757, 'reward_total_mean': 0.46311163902282715, 'reward_meter_mean': 0.4085719585418701, 'reward_meter_std': 0.3669445812702179, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877614378929138, 'reward_repeat_soft_std': 0.015099510550498962, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.46311163902282715, 'reward_total_composite_std': 0.10389380156993866, 'epoch': 0.1}

 30%|██▉       | 982/3300 [2:28:03<4:02:10,  6.27s/it]INFO 04-13 09:49:01 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 983/3300 [2:28:10<4:10:53,  6.50s/it]2026-04-13 09:49:08,244 | INFO | train_grpo_train | metrics_logged mode=train step=983

                                                      
{'loss': 0.0131, 'grad_norm': 9.582853317260742, 'learning_rate': 7.024242424242424e-06, 'num_tokens': 1734194.0, 'completions/mean_length': 115.25, 'completions/min_length': 106.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.25, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.5660096406936646, 'rewards/meter/std': 0.3683083951473236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9500572681427002, 'rewards/repeat_soft/std': 0.04081742838025093, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6809823513031006, 'rewards/total_composite/std': 0.21884137392044067, 'reward': 0.6809823513031006, 'reward_std': 0.21884137392044067, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.152125746011734, 'sampling/sampling_logp_difference/max': 1.720388412475586, 'sampling/importance_sampling_ratio/min': 0.17899660766124725, 'sampling/importance_sampling_ratio/mean': 1.0238882303237915, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0740713849663734, 'clip_ratio/low_mean': 0.10374108701944351, 'clip_ratio/low_min': 0.10374108701944351, 'clip_ratio/high_mean': 0.05396524164825678, 'clip_ratio/high_max': 0.05396524164825678, 'clip_ratio/region_mean': 0.1577063286677003, 'reward_total_mean': 0.6809823513031006, 'reward_meter_mean': 0.5660096406936646, 'reward_meter_std': 0.3683083951473236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9500572681427002, 'reward_repeat_soft_std': 0.04081742838025093, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6809823513031006, 'reward_total_composite_std': 0.21884137392044067, 'epoch': 0.1}

 30%|██▉       | 983/3300 [2:28:10<4:10:53,  6.50s/it]INFO 04-13 09:49:08 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 984/3300 [2:28:16<4:03:03,  6.30s/it]2026-04-13 09:49:14,073 | INFO | train_grpo_train | metrics_logged mode=train step=984

                                                      
{'loss': 0.0445, 'grad_norm': 14.900094032287598, 'learning_rate': 7.021212121212122e-06, 'num_tokens': 1735716.0, 'completions/mean_length': 38.25, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.1784151941537857, 'rewards/meter/std': 0.21362148225307465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9586414694786072, 'rewards/repeat_soft/std': 0.042823221534490585, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.24485784769058228, 'rewards/total_composite/mean': 0.41910111904144287, 'rewards/total_composite/std': 0.13274075090885162, 'reward': 0.41910111904144287, 'reward_std': 0.13274075090885162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14827175438404083, 'sampling/sampling_logp_difference/max': 1.7476634979248047, 'sampling/importance_sampling_ratio/min': 0.17418043315410614, 'sampling/importance_sampling_ratio/mean': 1.0331103801727295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8540159165859222, 'clip_ratio/low_mean': 0.1263057179749012, 'clip_ratio/low_min': 0.1263057179749012, 'clip_ratio/high_mean': 0.03193196281790733, 'clip_ratio/high_max': 0.03193196281790733, 'clip_ratio/region_mean': 0.15823768079280853, 'reward_total_mean': 0.41910111904144287, 'reward_meter_mean': 0.1784151941537857, 'reward_meter_std': 0.21362148225307465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9586414694786072, 'reward_repeat_soft_std': 0.042823221534490585, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.24485784769058228, 'reward_total_composite_mean': 0.41910111904144287, 'reward_total_composite_std': 0.13274075090885162, 'epoch': 0.1}

 30%|██▉       | 984/3300 [2:28:16<4:03:03,  6.30s/it]INFO 04-13 09:49:14 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 985/3300 [2:28:22<4:01:39,  6.26s/it]2026-04-13 09:49:20,257 | INFO | train_grpo_train | metrics_logged mode=train step=985

                                                      
{'loss': 0.0334, 'grad_norm': 11.730185508728027, 'learning_rate': 7.018181818181818e-06, 'num_tokens': 1737657.0, 'completions/mean_length': 66.625, 'completions/min_length': 62.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.625, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.6742357015609741, 'rewards/meter/std': 0.3289588689804077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9603241682052612, 'rewards/repeat_soft/std': 0.03282156214118004, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5172199010848999, 'rewards/total_composite/std': 0.08268414437770844, 'reward': 0.5172199010848999, 'reward_std': 0.08268412947654724, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12095610797405243, 'sampling/sampling_logp_difference/max': 1.2483692169189453, 'sampling/importance_sampling_ratio/min': 0.28697243332862854, 'sampling/importance_sampling_ratio/mean': 1.0194545984268188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9505049139261246, 'clip_ratio/low_mean': 0.0459558037109673, 'clip_ratio/low_min': 0.0459558037109673, 'clip_ratio/high_mean': 0.0553629444912076, 'clip_ratio/high_max': 0.0553629444912076, 'clip_ratio/region_mean': 0.1013187482021749, 'reward_total_mean': 0.5172199010848999, 'reward_meter_mean': 0.6742357015609741, 'reward_meter_std': 0.3289588689804077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9603241682052612, 'reward_repeat_soft_std': 0.03282156214118004, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5172199010848999, 'reward_total_composite_std': 0.08268414437770844, 'epoch': 0.1}

 30%|██▉       | 985/3300 [2:28:22<4:01:39,  6.26s/it]INFO 04-13 09:49:20 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 986/3300 [2:28:33<5:00:10,  7.78s/it]2026-04-13 09:49:31,585 | INFO | train_grpo_train | metrics_logged mode=train step=986

                                                      
{'loss': -0.1604, 'grad_norm': 3.1272151470184326, 'learning_rate': 7.015151515151516e-06, 'num_tokens': 1739559.0, 'completions/mean_length': 146.75, 'completions/min_length': 89.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 94.5714340209961, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.4578801393508911, 'rewards/meter/std': 0.35649222135543823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8577362895011902, 'rewards/repeat_soft/std': 0.08329711109399796, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.1922610104084015, 'rewards/total_composite/mean': 0.40382319688796997, 'rewards/total_composite/std': 0.19341367483139038, 'reward': 0.40382319688796997, 'reward_std': 0.19341367483139038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13537777960300446, 'sampling/sampling_logp_difference/max': 2.4759693145751953, 'sampling/importance_sampling_ratio/min': 0.08408144861459732, 'sampling/importance_sampling_ratio/mean': 1.0210380554199219, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8224976658821106, 'clip_ratio/low_mean': 0.025727187283337116, 'clip_ratio/low_min': 0.025727187283337116, 'clip_ratio/high_mean': 0.07636873703449965, 'clip_ratio/high_max': 0.07636873703449965, 'clip_ratio/region_mean': 0.10209592431783676, 'reward_total_mean': 0.40382319688796997, 'reward_meter_mean': 0.4578801393508911, 'reward_meter_std': 0.35649222135543823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8577362895011902, 'reward_repeat_soft_std': 0.08329711109399796, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.1922610104084015, 'reward_total_composite_mean': 0.40382319688796997, 'reward_total_composite_std': 0.19341367483139038, 'epoch': 0.1}

 30%|██▉       | 986/3300 [2:28:33<5:00:10,  7.78s/it]INFO 04-13 09:49:31 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 987/3300 [2:28:40<4:50:39,  7.54s/it]2026-04-13 09:49:38,557 | INFO | train_grpo_train | metrics_logged mode=train step=987

                                                      
{'loss': 0.0662, 'grad_norm': 8.873651504516602, 'learning_rate': 7.0121212121212126e-06, 'num_tokens': 1741599.0, 'completions/mean_length': 92.0, 'completions/min_length': 84.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.35385987162590027, 'rewards/meter/std': 0.2851884067058563, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8786883354187012, 'rewards/repeat_soft/std': 0.10092589259147644, 'rewards/judge_quality/mean': 0.7274999618530273, 'rewards/judge_quality/std': 0.2549930214881897, 'rewards/total_composite/mean': 0.46242237091064453, 'rewards/total_composite/std': 0.2586142420768738, 'reward': 0.46242237091064453, 'reward_std': 0.2586142420768738, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13277406990528107, 'sampling/sampling_logp_difference/max': 3.5209875106811523, 'sampling/importance_sampling_ratio/min': 0.029570220038294792, 'sampling/importance_sampling_ratio/mean': 1.0147207975387573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.792277567088604, 'clip_ratio/low_mean': 0.06794197112321854, 'clip_ratio/low_min': 0.06794197112321854, 'clip_ratio/high_mean': 0.07332353945821524, 'clip_ratio/high_max': 0.07332353945821524, 'clip_ratio/region_mean': 0.14126551058143377, 'reward_total_mean': 0.46242237091064453, 'reward_meter_mean': 0.35385987162590027, 'reward_meter_std': 0.2851884067058563, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8786883354187012, 'reward_repeat_soft_std': 0.10092589259147644, 'reward_judge_quality_mean': 0.7274999618530273, 'reward_judge_quality_std': 0.2549930214881897, 'reward_total_composite_mean': 0.46242237091064453, 'reward_total_composite_std': 0.2586142420768738, 'epoch': 0.1}

 30%|██▉       | 987/3300 [2:28:40<4:50:39,  7.54s/it]INFO 04-13 09:49:38 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 988/3300 [2:28:51<5:33:32,  8.66s/it]2026-04-13 09:49:49,823 | INFO | train_grpo_train | metrics_logged mode=train step=988

                                                      
{'loss': -0.1923, 'grad_norm': 2.268547296524048, 'learning_rate': 7.00909090909091e-06, 'num_tokens': 1743645.0, 'completions/mean_length': 150.75, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 99.14286041259766, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.966325044631958, 'rewards/meter/std': 0.019533870741724968, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9042581915855408, 'rewards/repeat_soft/std': 0.053632546216249466, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5403504967689514, 'rewards/total_composite/std': 0.22497345507144928, 'reward': 0.5403504967689514, 'reward_std': 0.22497345507144928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12875798344612122, 'sampling/sampling_logp_difference/max': 1.6117757558822632, 'sampling/importance_sampling_ratio/min': 0.19953297078609467, 'sampling/importance_sampling_ratio/mean': 1.023693323135376, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0199998915195465, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10916881449520588, 'clip_ratio/high_max': 0.10916881449520588, 'clip_ratio/region_mean': 0.10916881449520588, 'reward_total_mean': 0.5403504967689514, 'reward_meter_mean': 0.966325044631958, 'reward_meter_std': 0.019533870741724968, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9042581915855408, 'reward_repeat_soft_std': 0.053632546216249466, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5403504967689514, 'reward_total_composite_std': 0.22497345507144928, 'epoch': 0.1}

 30%|██▉       | 988/3300 [2:28:52<5:33:32,  8.66s/it]INFO 04-13 09:49:50 [block_pool.py:378] Successfully reset prefix cache

 30%|██▉       | 989/3300 [2:28:57<5:02:27,  7.85s/it]2026-04-13 09:49:55,795 | INFO | train_grpo_train | metrics_logged mode=train step=989

                                                      
{'loss': 0.0492, 'grad_norm': 15.93040657043457, 'learning_rate': 7.006060606060606e-06, 'num_tokens': 1745120.0, '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': 0.4848489761352539, 'rewards/meter/std': 0.3328888714313507, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9595526456832886, 'rewards/repeat_soft/std': 0.03687829524278641, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.5670564770698547, 'rewards/total_composite/std': 0.18828682601451874, 'reward': 0.5670564770698547, 'reward_std': 0.18828679621219635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1228448748588562, 'sampling/sampling_logp_difference/max': 1.2438445091247559, 'sampling/importance_sampling_ratio/min': 0.2882738411426544, 'sampling/importance_sampling_ratio/mean': 1.0155507326126099, 'sampling/importance_sampling_ratio/max': 1.9814753532409668, 'entropy': 0.8090120628476143, 'clip_ratio/low_mean': 0.08419212605804205, 'clip_ratio/low_min': 0.08419212605804205, 'clip_ratio/high_mean': 0.04381816182285547, 'clip_ratio/high_max': 0.04381816182285547, 'clip_ratio/region_mean': 0.12801028788089752, 'reward_total_mean': 0.5670564770698547, 'reward_meter_mean': 0.4848489761352539, 'reward_meter_std': 0.3328888714313507, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9595526456832886, 'reward_repeat_soft_std': 0.03687829524278641, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.5670564770698547, 'reward_total_composite_std': 0.18828682601451874, 'epoch': 0.1}

 30%|██▉       | 989/3300 [2:28:58<5:02:27,  7.85s/it]INFO 04-13 09:49:56 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 990/3300 [2:29:05<5:04:09,  7.90s/it]2026-04-13 09:50:03,807 | INFO | train_grpo_train | metrics_logged mode=train step=990

                                                      
{'loss': 0.0151, 'grad_norm': 8.560992240905762, 'learning_rate': 7.0030303030303035e-06, 'num_tokens': 1747617.0, 'completions/mean_length': 103.125, 'completions/min_length': 93.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.7816253900527954, 'rewards/meter/std': 0.2825448513031006, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9555389285087585, 'rewards/repeat_soft/std': 0.019325638189911842, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.4619530439376831, 'rewards/total_composite/std': 0.24237580597400665, 'reward': 0.4619530439376831, 'reward_std': 0.24237582087516785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1434517651796341, 'sampling/sampling_logp_difference/max': 1.60699462890625, 'sampling/importance_sampling_ratio/min': 0.20048925280570984, 'sampling/importance_sampling_ratio/mean': 1.0370699167251587, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.046807236969471, 'clip_ratio/low_mean': 0.04783136956393719, 'clip_ratio/low_min': 0.04783136956393719, 'clip_ratio/high_mean': 0.06494290893897414, 'clip_ratio/high_max': 0.06494290893897414, 'clip_ratio/region_mean': 0.11277427850291133, 'reward_total_mean': 0.4619530439376831, 'reward_meter_mean': 0.7816253900527954, 'reward_meter_std': 0.2825448513031006, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9555389285087585, 'reward_repeat_soft_std': 0.019325638189911842, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.4619530439376831, 'reward_total_composite_std': 0.24237580597400665, 'epoch': 0.1}

 30%|███       | 990/3300 [2:29:06<5:04:09,  7.90s/it]INFO 04-13 09:50:04 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 991/3300 [2:29:11<4:42:04,  7.33s/it]2026-04-13 09:50:09,806 | INFO | train_grpo_train | metrics_logged mode=train step=991

                                                      
{'loss': 0.014, 'grad_norm': 19.645828247070312, 'learning_rate': 7e-06, 'num_tokens': 1748957.0, 'completions/mean_length': 25.5, 'completions/min_length': 18.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7838623523712158, 'rewards/meter/std': 0.29824531078338623, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9447510838508606, 'rewards/repeat_soft/std': 0.03007800504565239, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5572038292884827, 'rewards/total_composite/std': 0.08105086535215378, 'reward': 0.5572038292884827, 'reward_std': 0.08105086535215378, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1598208248615265, 'sampling/sampling_logp_difference/max': 1.5274133682250977, 'sampling/importance_sampling_ratio/min': 0.2170964926481247, 'sampling/importance_sampling_ratio/mean': 1.0279059410095215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.116222970187664, 'clip_ratio/low_mean': 0.04640468209981918, 'clip_ratio/low_min': 0.04640468209981918, 'clip_ratio/high_mean': 0.08125012461096048, 'clip_ratio/high_max': 0.08125012461096048, 'clip_ratio/region_mean': 0.12765480671077967, 'reward_total_mean': 0.5572038292884827, 'reward_meter_mean': 0.7838623523712158, 'reward_meter_std': 0.29824531078338623, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9447510838508606, 'reward_repeat_soft_std': 0.03007800504565239, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5572038292884827, 'reward_total_composite_std': 0.08105086535215378, 'epoch': 0.1}

 30%|███       | 991/3300 [2:29:12<4:42:04,  7.33s/it]INFO 04-13 09:50:10 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 992/3300 [2:29:18<4:34:17,  7.13s/it]2026-04-13 09:50:16,472 | INFO | train_grpo_train | metrics_logged mode=train step=992

                                                      
{'loss': 0.0397, 'grad_norm': 9.281546592712402, 'learning_rate': 6.996969696969698e-06, 'num_tokens': 1751097.0, 'completions/mean_length': 98.5, 'completions/min_length': 88.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.5, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.799793541431427, 'rewards/meter/std': 0.3340491056442261, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9555978178977966, 'rewards/repeat_soft/std': 0.029661815613508224, 'rewards/judge_quality/mean': 0.3687499761581421, 'rewards/judge_quality/std': 0.186581090092659, 'rewards/total_composite/mean': 0.5163706541061401, 'rewards/total_composite/std': 0.14937683939933777, 'reward': 0.5163706541061401, 'reward_std': 0.14937682449817657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13694427907466888, 'sampling/sampling_logp_difference/max': 3.1758790016174316, 'sampling/importance_sampling_ratio/min': 0.041757382452487946, 'sampling/importance_sampling_ratio/mean': 1.0145255327224731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9991964176297188, 'clip_ratio/low_mean': 0.05222881771624088, 'clip_ratio/low_min': 0.05222881771624088, 'clip_ratio/high_mean': 0.05285879969596863, 'clip_ratio/high_max': 0.05285879969596863, 'clip_ratio/region_mean': 0.10508761741220951, 'reward_total_mean': 0.5163706541061401, 'reward_meter_mean': 0.799793541431427, 'reward_meter_std': 0.3340491056442261, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9555978178977966, 'reward_repeat_soft_std': 0.029661815613508224, 'reward_judge_quality_mean': 0.3687499761581421, 'reward_judge_quality_std': 0.186581090092659, 'reward_total_composite_mean': 0.5163706541061401, 'reward_total_composite_std': 0.14937683939933777, 'epoch': 0.1}

 30%|███       | 992/3300 [2:29:18<4:34:17,  7.13s/it]INFO 04-13 09:50:16 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 993/3300 [2:29:24<4:17:09,  6.69s/it]2026-04-13 09:50:22,129 | INFO | train_grpo_train | metrics_logged mode=train step=993

                                                      
{'loss': 0.0773, 'grad_norm': 35.79539489746094, 'learning_rate': 6.993939393939394e-06, 'num_tokens': 1752483.0, 'completions/mean_length': 20.25, 'completions/min_length': 17.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.6874057054519653, 'rewards/meter/std': 0.4219528138637543, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9263834953308105, 'rewards/repeat_soft/std': 0.0961892157793045, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5270617008209229, 'rewards/total_composite/std': 0.11126212030649185, 'reward': 0.5270617008209229, 'reward_std': 0.11126213520765305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.151984304189682, 'sampling/sampling_logp_difference/max': 1.055746078491211, 'sampling/importance_sampling_ratio/min': 0.34793272614479065, 'sampling/importance_sampling_ratio/mean': 1.059171199798584, 'sampling/importance_sampling_ratio/max': 1.8717169761657715, 'entropy': 1.2759125232696533, 'clip_ratio/low_mean': 0.0470467833802104, 'clip_ratio/low_min': 0.0470467833802104, 'clip_ratio/high_mean': 0.11265883315354586, 'clip_ratio/high_max': 0.11265883315354586, 'clip_ratio/region_mean': 0.15970561653375626, 'reward_total_mean': 0.5270617008209229, 'reward_meter_mean': 0.6874057054519653, 'reward_meter_std': 0.4219528138637543, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9263834953308105, 'reward_repeat_soft_std': 0.0961892157793045, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5270617008209229, 'reward_total_composite_std': 0.11126212030649185, 'epoch': 0.1}

 30%|███       | 993/3300 [2:29:24<4:17:09,  6.69s/it]INFO 04-13 09:50:22 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 994/3300 [2:29:36<5:20:43,  8.34s/it]2026-04-13 09:50:34,338 | INFO | train_grpo_train | metrics_logged mode=train step=994

                                                      
{'loss': -0.1729, 'grad_norm': 2.7367634773254395, 'learning_rate': 6.990909090909092e-06, 'num_tokens': 1754302.0, 'completions/mean_length': 129.375, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.71428680419922, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.7008242607116699, 'rewards/meter/std': 0.3392601013183594, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9867971539497375, 'rewards/repeat_soft/std': 0.004942748695611954, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.4817453920841217, 'rewards/total_composite/std': 0.2017514705657959, 'reward': 0.4817453920841217, 'reward_std': 0.2017514556646347, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1783118098974228, 'sampling/sampling_logp_difference/max': 1.7253952026367188, 'sampling/importance_sampling_ratio/min': 0.17810265719890594, 'sampling/importance_sampling_ratio/mean': 1.037008285522461, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2566193640232086, 'clip_ratio/low_mean': 0.02651515230536461, 'clip_ratio/low_min': 0.02651515230536461, 'clip_ratio/high_mean': 0.09103912860155106, 'clip_ratio/high_max': 0.09103912860155106, 'clip_ratio/region_mean': 0.11755428090691566, 'reward_total_mean': 0.4817453920841217, 'reward_meter_mean': 0.7008242607116699, 'reward_meter_std': 0.3392601013183594, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9867971539497375, 'reward_repeat_soft_std': 0.004942748695611954, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.4817453920841217, 'reward_total_composite_std': 0.2017514705657959, 'epoch': 0.1}

 30%|███       | 994/3300 [2:29:36<5:20:43,  8.34s/it]INFO 04-13 09:50:34 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 995/3300 [2:29:47<5:52:48,  9.18s/it]2026-04-13 09:50:45,479 | INFO | train_grpo_train | metrics_logged mode=train step=995

                                                      
{'loss': -0.0748, 'grad_norm': 2.902416944503784, 'learning_rate': 6.987878787878788e-06, 'num_tokens': 1755856.0, 'completions/mean_length': 94.25, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 34.57143020629883, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.604584813117981, 'rewards/meter/std': 0.32237282395362854, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9927259683609009, 'rewards/repeat_soft/std': 0.01288491953164339, 'rewards/judge_quality/mean': 0.627500057220459, 'rewards/judge_quality/std': 0.3366537392139435, 'rewards/total_composite/mean': 0.5558649301528931, 'rewards/total_composite/std': 0.28423136472702026, 'reward': 0.5558649301528931, 'reward_std': 0.28423136472702026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11629508435726166, 'sampling/sampling_logp_difference/max': 1.8717248439788818, 'sampling/importance_sampling_ratio/min': 0.15385805070400238, 'sampling/importance_sampling_ratio/mean': 0.9965556263923645, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4548973888158798, 'clip_ratio/low_mean': 0.04940191563218832, 'clip_ratio/low_min': 0.04940191563218832, 'clip_ratio/high_mean': 0.022852607304230332, 'clip_ratio/high_max': 0.022852607304230332, 'clip_ratio/region_mean': 0.07225452293641865, 'reward_total_mean': 0.5558649301528931, 'reward_meter_mean': 0.604584813117981, 'reward_meter_std': 0.32237282395362854, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9927259683609009, 'reward_repeat_soft_std': 0.01288491953164339, 'reward_judge_quality_mean': 0.627500057220459, 'reward_judge_quality_std': 0.3366537392139435, 'reward_total_composite_mean': 0.5558649301528931, 'reward_total_composite_std': 0.28423136472702026, 'epoch': 0.1}

 30%|███       | 995/3300 [2:29:47<5:52:48,  9.18s/it]INFO 04-13 09:50:45 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 996/3300 [2:29:53<5:12:47,  8.15s/it]2026-04-13 09:50:51,204 | INFO | train_grpo_train | metrics_logged mode=train step=996

                                                      
{'loss': 0.0452, 'grad_norm': 15.451910018920898, 'learning_rate': 6.984848484848485e-06, 'num_tokens': 1757382.0, 'completions/mean_length': 27.75, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6767911314964294, 'rewards/meter/std': 0.40244999527931213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.933674693107605, 'rewards/repeat_soft/std': 0.036771856248378754, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.6221144199371338, 'rewards/total_composite/std': 0.19765277206897736, 'reward': 0.6221144199371338, 'reward_std': 0.19765277206897736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1169843077659607, 'sampling/sampling_logp_difference/max': 1.4371109008789062, 'sampling/importance_sampling_ratio/min': 0.2376132607460022, 'sampling/importance_sampling_ratio/mean': 1.033090591430664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9760359674692154, 'clip_ratio/low_mean': 0.03587464150041342, 'clip_ratio/low_min': 0.03587464150041342, 'clip_ratio/high_mean': 0.03456959826871753, 'clip_ratio/high_max': 0.03456959826871753, 'clip_ratio/region_mean': 0.07044423976913095, 'reward_total_mean': 0.6221144199371338, 'reward_meter_mean': 0.6767911314964294, 'reward_meter_std': 0.40244999527931213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.933674693107605, 'reward_repeat_soft_std': 0.036771856248378754, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.6221144199371338, 'reward_total_composite_std': 0.19765277206897736, 'epoch': 0.1}

 30%|███       | 996/3300 [2:29:53<5:12:47,  8.15s/it]INFO 04-13 09:50:51 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 997/3300 [2:29:59<4:49:48,  7.55s/it]2026-04-13 09:50:57,364 | INFO | train_grpo_train | metrics_logged mode=train step=997

                                                      
{'loss': -0.0252, 'grad_norm': 14.92132568359375, 'learning_rate': 6.981818181818183e-06, 'num_tokens': 1758955.0, 'completions/mean_length': 46.625, 'completions/min_length': 37.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7483348846435547, 'rewards/meter/std': 0.3284272849559784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9753111004829407, 'rewards/repeat_soft/std': 0.03309660404920578, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.11160357296466827, 'rewards/total_composite/mean': 0.5082617998123169, 'rewards/total_composite/std': 0.24537086486816406, 'reward': 0.5082617998123169, 'reward_std': 0.24537083506584167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16312509775161743, 'sampling/sampling_logp_difference/max': 1.5516786575317383, 'sampling/importance_sampling_ratio/min': 0.21189197897911072, 'sampling/importance_sampling_ratio/mean': 1.0018374919891357, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.112735353410244, 'clip_ratio/low_mean': 0.07301124837249517, 'clip_ratio/low_min': 0.07301124837249517, 'clip_ratio/high_mean': 0.0901602627709508, 'clip_ratio/high_max': 0.0901602627709508, 'clip_ratio/region_mean': 0.16317151114344597, 'reward_total_mean': 0.5082617998123169, 'reward_meter_mean': 0.7483348846435547, 'reward_meter_std': 0.3284272849559784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9753111004829407, 'reward_repeat_soft_std': 0.03309660404920578, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.11160357296466827, 'reward_total_composite_mean': 0.5082617998123169, 'reward_total_composite_std': 0.24537086486816406, 'epoch': 0.1}

 30%|███       | 997/3300 [2:29:59<4:49:48,  7.55s/it]INFO 04-13 09:50:57 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 998/3300 [2:30:06<4:38:38,  7.26s/it]2026-04-13 09:51:03,956 | INFO | train_grpo_train | metrics_logged mode=train step=998

                                                      
{'loss': 0.0716, 'grad_norm': 10.448753356933594, 'learning_rate': 6.978787878787879e-06, 'num_tokens': 1761004.0, 'completions/mean_length': 86.125, 'completions/min_length': 71.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.6371596455574036, 'rewards/meter/std': 0.33203747868537903, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.894789457321167, 'rewards/repeat_soft/std': 0.037303801625967026, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.46316301822662354, 'rewards/total_composite/std': 0.09780687838792801, 'reward': 0.46316301822662354, 'reward_std': 0.09780686348676682, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15020348131656647, 'sampling/sampling_logp_difference/max': 2.2706642150878906, 'sampling/importance_sampling_ratio/min': 0.10324358195066452, 'sampling/importance_sampling_ratio/mean': 1.0024322271347046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0162644013762474, 'clip_ratio/low_mean': 0.04118903633207083, 'clip_ratio/low_min': 0.04118903633207083, 'clip_ratio/high_mean': 0.09987357072532177, 'clip_ratio/high_max': 0.09987357072532177, 'clip_ratio/region_mean': 0.1410626070573926, 'reward_total_mean': 0.46316301822662354, 'reward_meter_mean': 0.6371596455574036, 'reward_meter_std': 0.33203747868537903, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.894789457321167, 'reward_repeat_soft_std': 0.037303801625967026, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.46316301822662354, 'reward_total_composite_std': 0.09780687838792801, 'epoch': 0.1}

 30%|███       | 998/3300 [2:30:06<4:38:38,  7.26s/it]INFO 04-13 09:51:04 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 999/3300 [2:30:12<4:25:07,  6.91s/it]2026-04-13 09:51:10,051 | INFO | train_grpo_train | metrics_logged mode=train step=999

                                                      
{'loss': 0.0327, 'grad_norm': 12.499972343444824, 'learning_rate': 6.975757575757577e-06, 'num_tokens': 1762845.0, 'completions/mean_length': 56.125, 'completions/min_length': 49.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8660759329795837, 'rewards/meter/std': 0.3027213215827942, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9728718996047974, 'rewards/repeat_soft/std': 0.0173123087733984, 'rewards/judge_quality/mean': 0.44875001907348633, 'rewards/judge_quality/std': 0.21256513893604279, 'rewards/total_composite/mean': 0.6023823618888855, 'rewards/total_composite/std': 0.16437385976314545, 'reward': 0.6023823618888855, 'reward_std': 0.16437385976314545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15853369235992432, 'sampling/sampling_logp_difference/max': 1.2867321968078613, 'sampling/importance_sampling_ratio/min': 0.2761717736721039, 'sampling/importance_sampling_ratio/mean': 1.0485488176345825, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.425296112895012, 'clip_ratio/low_mean': 0.03546767123043537, 'clip_ratio/low_min': 0.03546767123043537, 'clip_ratio/high_mean': 0.09489201102405787, 'clip_ratio/high_max': 0.09489201102405787, 'clip_ratio/region_mean': 0.13035968225449324, 'reward_total_mean': 0.6023823618888855, 'reward_meter_mean': 0.8660759329795837, 'reward_meter_std': 0.3027213215827942, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9728718996047974, 'reward_repeat_soft_std': 0.0173123087733984, 'reward_judge_quality_mean': 0.44875001907348633, 'reward_judge_quality_std': 0.21256513893604279, 'reward_total_composite_mean': 0.6023823618888855, 'reward_total_composite_std': 0.16437385976314545, 'epoch': 0.1}

 30%|███       | 999/3300 [2:30:12<4:25:07,  6.91s/it]INFO 04-13 09:51:10 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1000/3300 [2:30:18<4:22:02,  6.84s/it]2026-04-13 09:51:16,716 | INFO | train_grpo_train | metrics_logged mode=train step=1000

                                                       
{'loss': 0.0485, 'grad_norm': 9.855438232421875, 'learning_rate': 6.9727272727272735e-06, 'num_tokens': 1764638.0, 'completions/mean_length': 76.125, 'completions/min_length': 67.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.2827584743499756, 'rewards/meter/std': 0.3035575747489929, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9801421165466309, 'rewards/repeat_soft/std': 0.015814922749996185, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.4346918761730194, 'rewards/total_composite/std': 0.08062057942152023, 'reward': 0.4346918761730194, 'reward_std': 0.08062057942152023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15195401012897491, 'sampling/sampling_logp_difference/max': 2.9526617527008057, 'sampling/importance_sampling_ratio/min': 0.052200574427843094, 'sampling/importance_sampling_ratio/mean': 1.0094891786575317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0911433175206184, 'clip_ratio/low_mean': 0.07451982609927654, 'clip_ratio/low_min': 0.07451982609927654, 'clip_ratio/high_mean': 0.049187893979251385, 'clip_ratio/high_max': 0.049187893979251385, 'clip_ratio/region_mean': 0.12370772007852793, 'reward_total_mean': 0.4346918761730194, 'reward_meter_mean': 0.2827584743499756, 'reward_meter_std': 0.3035575747489929, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9801421165466309, 'reward_repeat_soft_std': 0.015814922749996185, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.4346918761730194, 'reward_total_composite_std': 0.08062057942152023, 'epoch': 0.1}

 30%|███       | 1000/3300 [2:30:18<4:22:02,  6.84s/it]INFO 04-13 09:51:16 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:11<00:44,  5.60s/it]

 30%|███       | 3/10 [00:15<00:34,  4.87s/it]

 40%|████      | 4/10 [00:18<00:25,  4.26s/it]

 50%|█████     | 5/10 [00:22<00:21,  4.28s/it]

 60%|██████    | 6/10 [00:26<00:17,  4.25s/it]

 70%|███████   | 7/10 [00:30<00:12,  4.18s/it]

 80%|████████  | 8/10 [00:34<00:08,  4.08s/it]

 90%|█████████ | 9/10 [00:38<00:03,  3.86s/it]

100%|██████████| 10/10 [00:41<00:00,  3.62s/it]2026-04-13 09:52:01,608 | INFO | train_grpo_train | metrics_logged mode=eval step=1000

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 44.8301, 'eval_samples_per_second': 1.785, 'eval_steps_per_second': 0.223, 'eval_num_tokens': 1764638.0, 'eval_completions/mean_length': 70.9125, 'eval_completions/min_length': 35.5, 'eval_completions/max_length': 141.0, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 65.25357170104981, 'eval_completions/min_terminated_length': 35.5, 'eval_completions/max_terminated_length': 98.6, 'eval_rewards/meter/mean': 0.6456027030944824, 'eval_rewards/meter/std': 0.3476584076881409, 'eval_rewards/count_adherence/mean': 0.8720833361148834, 'eval_rewards/count_adherence/std': 0.158842521160841, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.9484191417694092, 'eval_rewards/repeat_soft/std': 0.04719752427190542, 'eval_rewards/judge_quality/mean': 0.46599999666213987, 'eval_rewards/judge_quality/std': 0.17726094285026192, 'eval_rewards/total_composite/mean': 0.5045325458049774, 'eval_rewards/total_composite/std': 0.15694166645407676, 'eval_reward': 0.5045325458049774, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08031205013394356, 'eval_sampling/sampling_logp_difference/max': 1.203929328918457, 'eval_sampling/importance_sampling_ratio/min': 0.3103459820151329, 'eval_sampling/importance_sampling_ratio/mean': 1.0168931245803834, 'eval_sampling/importance_sampling_ratio/max': 1.449403500556946, 'eval_entropy': 0.8986621379852295, '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': 0.5045325458049774, 'eval_reward_meter_mean': 0.6456027030944824, 'eval_reward_meter_std': 0.3476584076881409, 'eval_reward_count_adherence_mean': 0.8720833361148834, 'eval_reward_count_adherence_std': 0.158842521160841, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.9484191417694092, 'eval_reward_repeat_soft_std': 0.04719752427190542, 'eval_reward_judge_quality_mean': 0.46599999666213987, 'eval_reward_judge_quality_std': 0.17726094285026192, 'eval_reward_total_composite_mean': 0.5045325458049774, 'eval_reward_total_composite_std': 0.15694166645407676, 'epoch': 0.1}

 30%|███       | 1000/3300 [2:31:03<4:22:02,  6.84s/it]

100%|██████████| 10/10 [00:41<00:00,  3.62s/it]

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

 30%|███       | 1001/3300 [2:31:13<13:34:24, 21.25s/it]2026-04-13 09:52:11,607 | INFO | train_grpo_train | metrics_logged mode=train step=1001

                                                        
{'loss': 0.0217, 'grad_norm': 9.642687797546387, 'learning_rate': 6.969696969696971e-06, 'num_tokens': 1766852.0, 'completions/mean_length': 82.75, 'completions/min_length': 72.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.6661903262138367, 'rewards/meter/std': 0.30564743280410767, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8967072367668152, 'rewards/repeat_soft/std': 0.04659102112054825, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.4313625395298004, 'rewards/total_composite/std': 0.09541169553995132, 'reward': 0.4313625395298004, 'reward_std': 0.09541170299053192, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1348487138748169, 'sampling/sampling_logp_difference/max': 2.945591926574707, 'sampling/importance_sampling_ratio/min': 0.05257093533873558, 'sampling/importance_sampling_ratio/mean': 1.0046889781951904, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.808423787355423, 'clip_ratio/low_mean': 0.04772999323904514, 'clip_ratio/low_min': 0.04772999323904514, 'clip_ratio/high_mean': 0.09033445827662945, 'clip_ratio/high_max': 0.09033445827662945, 'clip_ratio/region_mean': 0.1380644515156746, 'reward_total_mean': 0.4313625395298004, 'reward_meter_mean': 0.6661903262138367, 'reward_meter_std': 0.30564743280410767, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8967072367668152, 'reward_repeat_soft_std': 0.04659102112054825, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.4313625395298004, 'reward_total_composite_std': 0.09541169553995132, 'epoch': 0.1}

 30%|███       | 1001/3300 [2:31:13<13:34:24, 21.25s/it]INFO 04-13 09:52:11 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1002/3300 [2:31:20<10:43:07, 16.79s/it]2026-04-13 09:52:17,992 | INFO | train_grpo_train | metrics_logged mode=train step=1002

                                                        
{'loss': -0.295, 'grad_norm': 5.131814002990723, 'learning_rate': 6.966666666666667e-06, 'num_tokens': 1768564.0, 'completions/mean_length': 62.0, 'completions/min_length': 10.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.0, 'completions/min_terminated_length': 10.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.849102258682251, 'rewards/meter/std': 0.34443584084510803, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9541321396827698, 'rewards/repeat_soft/std': 0.038377951830625534, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.514508068561554, 'rewards/total_composite/std': 0.21034754812717438, 'reward': 0.514508068561554, 'reward_std': 0.21034754812717438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13236641883850098, 'sampling/sampling_logp_difference/max': 2.604266405105591, 'sampling/importance_sampling_ratio/min': 0.07395736873149872, 'sampling/importance_sampling_ratio/mean': 0.992257833480835, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1153065040707588, 'clip_ratio/low_mean': 0.03750000149011612, 'clip_ratio/low_min': 0.03750000149011612, 'clip_ratio/high_mean': 0.11825434910133481, 'clip_ratio/high_max': 0.11825434910133481, 'clip_ratio/region_mean': 0.15575435059145093, 'reward_total_mean': 0.514508068561554, 'reward_meter_mean': 0.849102258682251, 'reward_meter_std': 0.34443584084510803, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9541321396827698, 'reward_repeat_soft_std': 0.038377951830625534, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.514508068561554, 'reward_total_composite_std': 0.21034754812717438, 'epoch': 0.1}

 30%|███       | 1002/3300 [2:31:20<10:43:07, 16.79s/it]INFO 04-13 09:52:18 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1003/3300 [2:31:26<8:42:29, 13.65s/it] 2026-04-13 09:52:24,298 | INFO | train_grpo_train | metrics_logged mode=train step=1003

                                                       
{'loss': 0.0668, 'grad_norm': 16.053508758544922, 'learning_rate': 6.963636363636364e-06, 'num_tokens': 1770278.0, 'completions/mean_length': 35.25, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.4937446713447571, 'rewards/meter/std': 0.2696358263492584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8978832960128784, 'rewards/repeat_soft/std': 0.08877812325954437, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.48789727687835693, 'rewards/total_composite/std': 0.0975428968667984, 'reward': 0.48789727687835693, 'reward_std': 0.0975428894162178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1792350709438324, 'sampling/sampling_logp_difference/max': 1.2945919036865234, 'sampling/importance_sampling_ratio/min': 0.274009644985199, 'sampling/importance_sampling_ratio/mean': 1.033050775527954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6600244417786598, 'clip_ratio/low_mean': 0.08472664840519428, 'clip_ratio/low_min': 0.08472664840519428, 'clip_ratio/high_mean': 0.07087546866387129, 'clip_ratio/high_max': 0.07087546866387129, 'clip_ratio/region_mean': 0.15560211706906557, 'reward_total_mean': 0.48789727687835693, 'reward_meter_mean': 0.4937446713447571, 'reward_meter_std': 0.2696358263492584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8978832960128784, 'reward_repeat_soft_std': 0.08877812325954437, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.48789727687835693, 'reward_total_composite_std': 0.0975428968667984, 'epoch': 0.1}

 30%|███       | 1003/3300 [2:31:26<8:42:29, 13.65s/it]INFO 04-13 09:52:24 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1004/3300 [2:31:37<8:15:37, 12.95s/it]2026-04-13 09:52:35,628 | INFO | train_grpo_train | metrics_logged mode=train step=1004

                                                       
{'loss': -0.095, 'grad_norm': 3.897873640060425, 'learning_rate': 6.960606060606061e-06, 'num_tokens': 1772071.0, 'completions/mean_length': 129.125, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.42857360839844, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.47224485874176025, 'rewards/meter/std': 0.3324452340602875, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9462336301803589, 'rewards/repeat_soft/std': 0.05075499042868614, 'rewards/judge_quality/mean': 0.45250001549720764, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.4135204553604126, 'rewards/total_composite/std': 0.21175077557563782, 'reward': 0.4135204553604126, 'reward_std': 0.21175077557563782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14971968531608582, 'sampling/sampling_logp_difference/max': 2.313939094543457, 'sampling/importance_sampling_ratio/min': 0.09887102246284485, 'sampling/importance_sampling_ratio/mean': 1.0206713676452637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9123754799365997, 'clip_ratio/low_mean': 0.026376837864518166, 'clip_ratio/low_min': 0.026376837864518166, 'clip_ratio/high_mean': 0.07666672486811876, 'clip_ratio/high_max': 0.07666672486811876, 'clip_ratio/region_mean': 0.10304356273263693, 'reward_total_mean': 0.4135204553604126, 'reward_meter_mean': 0.47224485874176025, 'reward_meter_std': 0.3324452340602875, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9462336301803589, 'reward_repeat_soft_std': 0.05075499042868614, 'reward_judge_quality_mean': 0.45250001549720764, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.4135204553604126, 'reward_total_composite_std': 0.21175077557563782, 'epoch': 0.1}

 30%|███       | 1004/3300 [2:31:37<8:15:37, 12.95s/it]INFO 04-13 09:52:35 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1005/3300 [2:31:43<6:52:34, 10.79s/it]2026-04-13 09:52:41,359 | INFO | train_grpo_train | metrics_logged mode=train step=1005

                                                       
{'loss': 0.0251, 'grad_norm': 23.581077575683594, 'learning_rate': 6.957575757575759e-06, 'num_tokens': 1773484.0, 'completions/mean_length': 21.625, 'completions/min_length': 20.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8936288356781006, 'rewards/meter/std': 0.2433493435382843, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9524593949317932, 'rewards/repeat_soft/std': 0.02408791147172451, 'rewards/judge_quality/mean': 0.4012500047683716, 'rewards/judge_quality/std': 0.10260012745857239, 'rewards/total_composite/mean': 0.5899405479431152, 'rewards/total_composite/std': 0.09260963648557663, 'reward': 0.5899405479431152, 'reward_std': 0.09260963648557663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1577124148607254, 'sampling/sampling_logp_difference/max': 2.8035335540771484, 'sampling/importance_sampling_ratio/min': 0.0605955645442009, 'sampling/importance_sampling_ratio/mean': 1.0209708213806152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7599133998155594, 'clip_ratio/low_mean': 0.021739130839705467, 'clip_ratio/low_min': 0.021739130839705467, 'clip_ratio/high_mean': 0.07247023982927203, 'clip_ratio/high_max': 0.07247023982927203, 'clip_ratio/region_mean': 0.0942093706689775, 'reward_total_mean': 0.5899405479431152, 'reward_meter_mean': 0.8936288356781006, 'reward_meter_std': 0.2433493435382843, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9524593949317932, 'reward_repeat_soft_std': 0.02408791147172451, 'reward_judge_quality_mean': 0.4012500047683716, 'reward_judge_quality_std': 0.10260012745857239, 'reward_total_composite_mean': 0.5899405479431152, 'reward_total_composite_std': 0.09260963648557663, 'epoch': 0.1}

 30%|███       | 1005/3300 [2:31:43<6:52:34, 10.79s/it]INFO 04-13 09:52:41 [block_pool.py:378] Successfully reset prefix cache

 30%|███       | 1006/3300 [2:31:49<6:01:40,  9.46s/it]2026-04-13 09:52:47,723 | INFO | train_grpo_train | metrics_logged mode=train step=1006

                                                       
{'loss': -0.0363, 'grad_norm': 12.378266334533691, 'learning_rate': 6.954545454545455e-06, 'num_tokens': 1775445.0, 'completions/mean_length': 73.125, 'completions/min_length': 67.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.24075129628181458, 'rewards/meter/std': 0.26709258556365967, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9653311371803284, 'rewards/repeat_soft/std': 0.016731778159737587, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.41278842091560364, 'rewards/total_composite/std': 0.07317467033863068, 'reward': 0.41278842091560364, 'reward_std': 0.07317467033863068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14169339835643768, 'sampling/sampling_logp_difference/max': 3.1516613960266113, 'sampling/importance_sampling_ratio/min': 0.04278099164366722, 'sampling/importance_sampling_ratio/mean': 1.0222159624099731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7837047949433327, 'clip_ratio/low_mean': 0.07796356175094843, 'clip_ratio/low_min': 0.07796356175094843, 'clip_ratio/high_mean': 0.03137254994362593, 'clip_ratio/high_max': 0.03137254994362593, 'clip_ratio/region_mean': 0.10933611169457436, 'reward_total_mean': 0.41278842091560364, 'reward_meter_mean': 0.24075129628181458, 'reward_meter_std': 0.26709258556365967, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9653311371803284, 'reward_repeat_soft_std': 0.016731778159737587, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.41278842091560364, 'reward_total_composite_std': 0.07317467033863068, 'epoch': 0.1}

 30%|███       | 1006/3300 [2:31:49<6:01:40,  9.46s/it]INFO 04-13 09:52:47 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1007/3300 [2:31:55<5:21:38,  8.42s/it]2026-04-13 09:52:53,711 | INFO | train_grpo_train | metrics_logged mode=train step=1007

                                                       
{'loss': 0.0718, 'grad_norm': 14.407511711120605, 'learning_rate': 6.951515151515153e-06, 'num_tokens': 1777019.0, 'completions/mean_length': 42.75, 'completions/min_length': 37.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9675606489181519, 'rewards/meter/std': 0.031776558607816696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9352800250053406, 'rewards/repeat_soft/std': 0.09628202021121979, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.8021653890609741, 'rewards/total_composite/std': 0.1718769669532776, 'reward': 0.8021653890609741, 'reward_std': 0.1718769520521164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16105450689792633, 'sampling/sampling_logp_difference/max': 1.6512861251831055, 'sampling/importance_sampling_ratio/min': 0.19180306792259216, 'sampling/importance_sampling_ratio/mean': 1.0006808042526245, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1489304453134537, 'clip_ratio/low_mean': 0.051672996021807194, 'clip_ratio/low_min': 0.051672996021807194, 'clip_ratio/high_mean': 0.10751363914459944, 'clip_ratio/high_max': 0.10751363914459944, 'clip_ratio/region_mean': 0.15918663516640663, 'reward_total_mean': 0.8021653890609741, 'reward_meter_mean': 0.9675606489181519, 'reward_meter_std': 0.031776558607816696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9352800250053406, 'reward_repeat_soft_std': 0.09628202021121979, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.8021653890609741, 'reward_total_composite_std': 0.1718769669532776, 'epoch': 0.1}

 31%|███       | 1007/3300 [2:31:55<5:21:38,  8.42s/it]INFO 04-13 09:52:53 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1008/3300 [2:32:02<4:57:35,  7.79s/it]2026-04-13 09:53:00,036 | INFO | train_grpo_train | metrics_logged mode=train step=1008

                                                       
{'loss': 0.0522, 'grad_norm': 9.007136344909668, 'learning_rate': 6.948484848484849e-06, 'num_tokens': 1778936.0, 'completions/mean_length': 63.625, 'completions/min_length': 53.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.625, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.79938805103302, 'rewards/meter/std': 0.3317202031612396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9361878037452698, 'rewards/repeat_soft/std': 0.08320317417383194, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6969414949417114, 'rewards/total_composite/std': 0.1936228722333908, 'reward': 0.6969414949417114, 'reward_std': 0.19362285733222961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15141348540782928, 'sampling/sampling_logp_difference/max': 1.6684417724609375, 'sampling/importance_sampling_ratio/min': 0.18854062259197235, 'sampling/importance_sampling_ratio/mean': 1.0258336067199707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2222661450505257, 'clip_ratio/low_mean': 0.08359825983643532, 'clip_ratio/low_min': 0.08359825983643532, 'clip_ratio/high_mean': 0.054921312257647514, 'clip_ratio/high_max': 0.054921312257647514, 'clip_ratio/region_mean': 0.13851957209408283, 'reward_total_mean': 0.6969414949417114, 'reward_meter_mean': 0.79938805103302, 'reward_meter_std': 0.3317202031612396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9361878037452698, 'reward_repeat_soft_std': 0.08320317417383194, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6969414949417114, 'reward_total_composite_std': 0.1936228722333908, 'epoch': 0.1}

 31%|███       | 1008/3300 [2:32:02<4:57:35,  7.79s/it]INFO 04-13 09:53:00 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1009/3300 [2:32:08<4:43:43,  7.43s/it]2026-04-13 09:53:06,624 | INFO | train_grpo_train | metrics_logged mode=train step=1009

                                                       
{'loss': -0.0071, 'grad_norm': 12.920982360839844, 'learning_rate': 6.945454545454546e-06, 'num_tokens': 1780722.0, 'completions/mean_length': 57.25, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6258684396743774, 'rewards/meter/std': 0.3698013424873352, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9881253242492676, 'rewards/repeat_soft/std': 0.008947896771132946, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5284538865089417, 'rewards/total_composite/std': 0.09727796167135239, 'reward': 0.5284538865089417, 'reward_std': 0.09727796912193298, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16133199632167816, 'sampling/sampling_logp_difference/max': 1.516340732574463, 'sampling/importance_sampling_ratio/min': 0.2195136696100235, 'sampling/importance_sampling_ratio/mean': 1.0304008722305298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3624932616949081, 'clip_ratio/low_mean': 0.037349299527704716, 'clip_ratio/low_min': 0.037349299527704716, 'clip_ratio/high_mean': 0.11407075077295303, 'clip_ratio/high_max': 0.11407075077295303, 'clip_ratio/region_mean': 0.15142005030065775, 'reward_total_mean': 0.5284538865089417, 'reward_meter_mean': 0.6258684396743774, 'reward_meter_std': 0.3698013424873352, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9881253242492676, 'reward_repeat_soft_std': 0.008947896771132946, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5284538865089417, 'reward_total_composite_std': 0.09727796167135239, 'epoch': 0.1}

 31%|███       | 1009/3300 [2:32:08<4:43:43,  7.43s/it]INFO 04-13 09:53:06 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1010/3300 [2:32:15<4:39:47,  7.33s/it]2026-04-13 09:53:13,723 | INFO | train_grpo_train | metrics_logged mode=train step=1010

                                                       
{'loss': 0.0691, 'grad_norm': 8.525898933410645, 'learning_rate': 6.942424242424243e-06, 'num_tokens': 1783280.0, 'completions/mean_length': 107.75, 'completions/min_length': 96.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.7836625576019287, 'rewards/meter/std': 0.23392094671726227, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9279478788375854, 'rewards/repeat_soft/std': 0.024648524820804596, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.46836158633232117, 'rewards/total_composite/std': 0.09341593831777573, 'reward': 0.46836158633232117, 'reward_std': 0.09341592341661453, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16590675711631775, 'sampling/sampling_logp_difference/max': 1.8753886222839355, 'sampling/importance_sampling_ratio/min': 0.1532953828573227, 'sampling/importance_sampling_ratio/mean': 1.0183576345443726, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.105267234146595, 'clip_ratio/low_mean': 0.04774219263345003, 'clip_ratio/low_min': 0.04774219263345003, 'clip_ratio/high_mean': 0.09435775689780712, 'clip_ratio/high_max': 0.09435775689780712, 'clip_ratio/region_mean': 0.14209994953125715, 'reward_total_mean': 0.46836158633232117, 'reward_meter_mean': 0.7836625576019287, 'reward_meter_std': 0.23392094671726227, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9279478788375854, 'reward_repeat_soft_std': 0.024648524820804596, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.46836158633232117, 'reward_total_composite_std': 0.09341593831777573, 'epoch': 0.1}

 31%|███       | 1010/3300 [2:32:15<4:39:47,  7.33s/it]INFO 04-13 09:53:13 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1011/3300 [2:32:22<4:28:40,  7.04s/it]2026-04-13 09:53:20,099 | INFO | train_grpo_train | metrics_logged mode=train step=1011

                                                       
{'loss': 0.0427, 'grad_norm': 12.076990127563477, 'learning_rate': 6.93939393939394e-06, 'num_tokens': 1785403.0, 'completions/mean_length': 69.375, 'completions/min_length': 65.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.375, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.9308876395225525, 'rewards/meter/std': 0.176197811961174, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9217841625213623, 'rewards/repeat_soft/std': 0.0469188392162323, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5286530256271362, 'rewards/total_composite/std': 0.05541308969259262, 'reward': 0.5286530256271362, 'reward_std': 0.05541309341788292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12863069772720337, 'sampling/sampling_logp_difference/max': 1.9285722970962524, 'sampling/importance_sampling_ratio/min': 0.14535556733608246, 'sampling/importance_sampling_ratio/mean': 1.018244743347168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9357447624206543, 'clip_ratio/low_mean': 0.0390593558549881, 'clip_ratio/low_min': 0.0390593558549881, 'clip_ratio/high_mean': 0.07747931219637394, 'clip_ratio/high_max': 0.07747931219637394, 'clip_ratio/region_mean': 0.11653866805136204, 'reward_total_mean': 0.5286530256271362, 'reward_meter_mean': 0.9308876395225525, 'reward_meter_std': 0.176197811961174, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9217841625213623, 'reward_repeat_soft_std': 0.0469188392162323, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5286530256271362, 'reward_total_composite_std': 0.05541308969259262, 'epoch': 0.1}

 31%|███       | 1011/3300 [2:32:22<4:28:40,  7.04s/it]INFO 04-13 09:53:20 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1012/3300 [2:32:28<4:19:39,  6.81s/it]2026-04-13 09:53:26,368 | INFO | train_grpo_train | metrics_logged mode=train step=1012

                                                       
{'loss': -0.0774, 'grad_norm': 15.02917194366455, 'learning_rate': 6.936363636363636e-06, 'num_tokens': 1786934.0, 'completions/mean_length': 40.375, 'completions/min_length': 32.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.7910438179969788, 'rewards/meter/std': 0.3508176803588867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.942015528678894, 'rewards/repeat_soft/std': 0.054328788071870804, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5620580911636353, 'rewards/total_composite/std': 0.09672543406486511, 'reward': 0.5620580911636353, 'reward_std': 0.09672542661428452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1535942405462265, 'sampling/sampling_logp_difference/max': 1.341993808746338, 'sampling/importance_sampling_ratio/min': 0.2613241374492645, 'sampling/importance_sampling_ratio/mean': 1.0322272777557373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2767742425203323, 'clip_ratio/low_mean': 0.022569444496184587, 'clip_ratio/low_min': 0.022569444496184587, 'clip_ratio/high_mean': 0.10155987646430731, 'clip_ratio/high_max': 0.10155987646430731, 'clip_ratio/region_mean': 0.1241293209604919, 'reward_total_mean': 0.5620580911636353, 'reward_meter_mean': 0.7910438179969788, 'reward_meter_std': 0.3508176803588867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.942015528678894, 'reward_repeat_soft_std': 0.054328788071870804, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5620580911636353, 'reward_total_composite_std': 0.09672543406486511, 'epoch': 0.1}

 31%|███       | 1012/3300 [2:32:28<4:19:39,  6.81s/it]INFO 04-13 09:53:26 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1013/3300 [2:32:35<4:21:58,  6.87s/it]2026-04-13 09:53:33,389 | INFO | train_grpo_train | metrics_logged mode=train step=1013

                                                       
{'loss': -0.0606, 'grad_norm': 10.494180679321289, 'learning_rate': 6.9333333333333344e-06, 'num_tokens': 1789010.0, 'completions/mean_length': 86.5, 'completions/min_length': 68.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9619123935699463, 'rewards/meter/std': 0.02976352907717228, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9517722725868225, 'rewards/repeat_soft/std': 0.04128618910908699, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5409913063049316, 'rewards/total_composite/std': 0.231267049908638, 'reward': 0.5409913063049316, 'reward_std': 0.2312670350074768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14850583672523499, 'sampling/sampling_logp_difference/max': 5.278156280517578, 'sampling/importance_sampling_ratio/min': 0.005101828370243311, 'sampling/importance_sampling_ratio/mean': 1.021235466003418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9896829351782799, 'clip_ratio/low_mean': 0.02757352963089943, 'clip_ratio/low_min': 0.02757352963089943, 'clip_ratio/high_mean': 0.1027968917042017, 'clip_ratio/high_max': 0.1027968917042017, 'clip_ratio/region_mean': 0.13037042133510113, 'reward_total_mean': 0.5409913063049316, 'reward_meter_mean': 0.9619123935699463, 'reward_meter_std': 0.02976352907717228, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9517722725868225, 'reward_repeat_soft_std': 0.04128618910908699, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5409913063049316, 'reward_total_composite_std': 0.231267049908638, 'epoch': 0.1}

 31%|███       | 1013/3300 [2:32:35<4:21:58,  6.87s/it]INFO 04-13 09:53:33 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1014/3300 [2:32:42<4:24:16,  6.94s/it]2026-04-13 09:53:40,469 | INFO | train_grpo_train | metrics_logged mode=train step=1014

                                                       
{'loss': -0.001, 'grad_norm': 7.80047607421875, 'learning_rate': 6.930303030303031e-06, 'num_tokens': 1791389.0, 'completions/mean_length': 104.375, 'completions/min_length': 93.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.375, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.900068998336792, 'rewards/meter/std': 0.18561285734176636, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9651224613189697, 'rewards/repeat_soft/std': 0.020892761647701263, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.613532543182373, 'rewards/total_composite/std': 0.2619045674800873, 'reward': 0.613532543182373, 'reward_std': 0.2619045674800873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13769012689590454, 'sampling/sampling_logp_difference/max': 1.6550350189208984, 'sampling/importance_sampling_ratio/min': 0.1910853534936905, 'sampling/importance_sampling_ratio/mean': 1.0003539323806763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9395716786384583, 'clip_ratio/low_mean': 0.04137605335563421, 'clip_ratio/low_min': 0.04137605335563421, 'clip_ratio/high_mean': 0.09778095968067646, 'clip_ratio/high_max': 0.09778095968067646, 'clip_ratio/region_mean': 0.13915701303631067, 'reward_total_mean': 0.613532543182373, 'reward_meter_mean': 0.900068998336792, 'reward_meter_std': 0.18561285734176636, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9651224613189697, 'reward_repeat_soft_std': 0.020892761647701263, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.613532543182373, 'reward_total_composite_std': 0.2619045674800873, 'epoch': 0.1}

 31%|███       | 1014/3300 [2:32:42<4:24:16,  6.94s/it]INFO 04-13 09:53:40 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1015/3300 [2:32:49<4:27:10,  7.02s/it]2026-04-13 09:53:47,674 | INFO | train_grpo_train | metrics_logged mode=train step=1015

                                                       
{'loss': -0.0082, 'grad_norm': 8.329638481140137, 'learning_rate': 6.927272727272728e-06, 'num_tokens': 1793857.0, 'completions/mean_length': 113.5, 'completions/min_length': 102.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.5, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9305871725082397, 'rewards/meter/std': 0.13200393319129944, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9458436369895935, 'rewards/repeat_soft/std': 0.05212901160120964, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5292602181434631, 'rewards/total_composite/std': 0.03420192375779152, 'reward': 0.5292602181434631, 'reward_std': 0.03420192748308182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14664679765701294, 'sampling/sampling_logp_difference/max': 1.6847548484802246, 'sampling/importance_sampling_ratio/min': 0.18548989295959473, 'sampling/importance_sampling_ratio/mean': 1.0259116888046265, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9773837104439735, 'clip_ratio/low_mean': 0.03706501983106136, 'clip_ratio/low_min': 0.03706501983106136, 'clip_ratio/high_mean': 0.08734916616231203, 'clip_ratio/high_max': 0.08734916616231203, 'clip_ratio/region_mean': 0.1244141859933734, 'reward_total_mean': 0.5292602181434631, 'reward_meter_mean': 0.9305871725082397, 'reward_meter_std': 0.13200393319129944, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9458436369895935, 'reward_repeat_soft_std': 0.05212901160120964, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5292602181434631, 'reward_total_composite_std': 0.03420192375779152, 'epoch': 0.1}

 31%|███       | 1015/3300 [2:32:49<4:27:10,  7.02s/it]INFO 04-13 09:53:47 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1016/3300 [2:32:56<4:22:05,  6.89s/it]2026-04-13 09:53:54,255 | INFO | train_grpo_train | metrics_logged mode=train step=1016

                                                       
{'loss': 0.0519, 'grad_norm': 17.249526977539062, 'learning_rate': 6.9242424242424245e-06, 'num_tokens': 1795389.0, 'completions/mean_length': 40.5, 'completions/min_length': 36.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.46202582120895386, 'rewards/meter/std': 0.35253575444221497, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9827362895011902, 'rewards/repeat_soft/std': 0.013210390694439411, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.48027363419532776, 'rewards/total_composite/std': 0.09445100277662277, 'reward': 0.48027363419532776, 'reward_std': 0.09445099532604218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1720263659954071, 'sampling/sampling_logp_difference/max': 1.777175784111023, 'sampling/importance_sampling_ratio/min': 0.1691150963306427, 'sampling/importance_sampling_ratio/mean': 1.011866807937622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9946768060326576, 'clip_ratio/low_mean': 0.12039735168218613, 'clip_ratio/low_min': 0.12039735168218613, 'clip_ratio/high_mean': 0.08580685593187809, 'clip_ratio/high_max': 0.08580685593187809, 'clip_ratio/region_mean': 0.20620420761406422, 'reward_total_mean': 0.48027363419532776, 'reward_meter_mean': 0.46202582120895386, 'reward_meter_std': 0.35253575444221497, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9827362895011902, 'reward_repeat_soft_std': 0.013210390694439411, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.48027363419532776, 'reward_total_composite_std': 0.09445100277662277, 'epoch': 0.1}

 31%|███       | 1016/3300 [2:32:56<4:22:05,  6.89s/it]INFO 04-13 09:53:54 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1017/3300 [2:33:03<4:18:38,  6.80s/it]2026-04-13 09:54:00,840 | INFO | train_grpo_train | metrics_logged mode=train step=1017

                                                       
{'loss': 0.0164, 'grad_norm': 17.912460327148438, 'learning_rate': 6.921212121212122e-06, 'num_tokens': 1796921.0, 'completions/mean_length': 36.5, 'completions/min_length': 30.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8674060702323914, 'rewards/meter/std': 0.16231243312358856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9547463059425354, 'rewards/repeat_soft/std': 0.047527726739645004, 'rewards/judge_quality/mean': 0.706250011920929, 'rewards/judge_quality/std': 0.19316445291042328, 'rewards/total_composite/mean': 0.7387485504150391, 'rewards/total_composite/std': 0.12863102555274963, 'reward': 0.7387485504150391, 'reward_std': 0.12863105535507202, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18040382862091064, 'sampling/sampling_logp_difference/max': 1.5863866806030273, 'sampling/importance_sampling_ratio/min': 0.20466379821300507, 'sampling/importance_sampling_ratio/mean': 1.0324939489364624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2933059111237526, 'clip_ratio/low_mean': 0.07632113993167877, 'clip_ratio/low_min': 0.07632113993167877, 'clip_ratio/high_mean': 0.12189750000834465, 'clip_ratio/high_max': 0.12189750000834465, 'clip_ratio/region_mean': 0.19821863994002342, 'reward_total_mean': 0.7387485504150391, 'reward_meter_mean': 0.8674060702323914, 'reward_meter_std': 0.16231243312358856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9547463059425354, 'reward_repeat_soft_std': 0.047527726739645004, 'reward_judge_quality_mean': 0.706250011920929, 'reward_judge_quality_std': 0.19316445291042328, 'reward_total_composite_mean': 0.7387485504150391, 'reward_total_composite_std': 0.12863102555274963, 'epoch': 0.1}

 31%|███       | 1017/3300 [2:33:03<4:18:38,  6.80s/it]INFO 04-13 09:54:01 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1018/3300 [2:33:09<4:14:20,  6.69s/it]2026-04-13 09:54:07,270 | INFO | train_grpo_train | metrics_logged mode=train step=1018

                                                       
{'loss': 0.0239, 'grad_norm': 9.51146125793457, 'learning_rate': 6.918181818181818e-06, 'num_tokens': 1798948.0, 'completions/mean_length': 82.375, 'completions/min_length': 70.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.8627983331680298, 'rewards/meter/std': 0.27659159898757935, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9733031988143921, 'rewards/repeat_soft/std': 0.023586789146065712, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5315394401550293, 'rewards/total_composite/std': 0.07479587197303772, 'reward': 0.5315394401550293, 'reward_std': 0.07479587942361832, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1614338755607605, 'sampling/sampling_logp_difference/max': 2.192643165588379, 'sampling/importance_sampling_ratio/min': 0.11162132769823074, 'sampling/importance_sampling_ratio/mean': 1.003227710723877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2144625186920166, 'clip_ratio/low_mean': 0.028445051982998848, 'clip_ratio/low_min': 0.028445051982998848, 'clip_ratio/high_mean': 0.15170305408537388, 'clip_ratio/high_max': 0.15170305408537388, 'clip_ratio/region_mean': 0.18014810606837273, 'reward_total_mean': 0.5315394401550293, 'reward_meter_mean': 0.8627983331680298, 'reward_meter_std': 0.27659159898757935, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9733031988143921, 'reward_repeat_soft_std': 0.023586789146065712, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5315394401550293, 'reward_total_composite_std': 0.07479587197303772, 'epoch': 0.1}

 31%|███       | 1018/3300 [2:33:09<4:14:20,  6.69s/it]INFO 04-13 09:54:07 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1019/3300 [2:33:16<4:16:14,  6.74s/it]2026-04-13 09:54:14,132 | INFO | train_grpo_train | metrics_logged mode=train step=1019

                                                       
{'loss': 0.013, 'grad_norm': 10.348309516906738, 'learning_rate': 6.915151515151515e-06, 'num_tokens': 1801016.0, 'completions/mean_length': 91.5, 'completions/min_length': 69.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.6451882123947144, 'rewards/meter/std': 0.2994335889816284, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8715066909790039, 'rewards/repeat_soft/std': 0.10998848080635071, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.47752851247787476, 'rewards/total_composite/std': 0.08254332095384598, 'reward': 0.47752851247787476, 'reward_std': 0.08254332095384598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1572115123271942, 'sampling/sampling_logp_difference/max': 2.785656213760376, 'sampling/importance_sampling_ratio/min': 0.06168859452009201, 'sampling/importance_sampling_ratio/mean': 1.021525502204895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1143873408436775, 'clip_ratio/low_mean': 0.05920429341495037, 'clip_ratio/low_min': 0.05920429341495037, 'clip_ratio/high_mean': 0.057869311422109604, 'clip_ratio/high_max': 0.057869311422109604, 'clip_ratio/region_mean': 0.11707360483705997, 'reward_total_mean': 0.47752851247787476, 'reward_meter_mean': 0.6451882123947144, 'reward_meter_std': 0.2994335889816284, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8715066909790039, 'reward_repeat_soft_std': 0.10998848080635071, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.47752851247787476, 'reward_total_composite_std': 0.08254332095384598, 'epoch': 0.1}

 31%|███       | 1019/3300 [2:33:16<4:16:14,  6.74s/it]INFO 04-13 09:54:14 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1020/3300 [2:33:22<4:13:36,  6.67s/it]2026-04-13 09:54:20,654 | INFO | train_grpo_train | metrics_logged mode=train step=1020

                                                       
{'loss': 0.0311, 'grad_norm': 8.36316204071045, 'learning_rate': 6.912121212121212e-06, 'num_tokens': 1803200.0, 'completions/mean_length': 94.0, 'completions/min_length': 84.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.769997239112854, 'rewards/meter/std': 0.2617892026901245, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9365980625152588, 'rewards/repeat_soft/std': 0.026574136689305305, 'rewards/judge_quality/mean': 0.5699999928474426, 'rewards/judge_quality/std': 0.16035676002502441, 'rewards/total_composite/mean': 0.546271026134491, 'rewards/total_composite/std': 0.1045205146074295, 'reward': 0.546271026134491, 'reward_std': 0.1045205146074295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14200660586357117, 'sampling/sampling_logp_difference/max': 1.6306532621383667, 'sampling/importance_sampling_ratio/min': 0.19580160081386566, 'sampling/importance_sampling_ratio/mean': 1.0257067680358887, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.14267548173666, 'clip_ratio/low_mean': 0.07997111836448312, 'clip_ratio/low_min': 0.07997111836448312, 'clip_ratio/high_mean': 0.029277191497385502, 'clip_ratio/high_max': 0.029277191497385502, 'clip_ratio/region_mean': 0.10924830986186862, 'reward_total_mean': 0.546271026134491, 'reward_meter_mean': 0.769997239112854, 'reward_meter_std': 0.2617892026901245, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9365980625152588, 'reward_repeat_soft_std': 0.026574136689305305, 'reward_judge_quality_mean': 0.5699999928474426, 'reward_judge_quality_std': 0.16035676002502441, 'reward_total_composite_mean': 0.546271026134491, 'reward_total_composite_std': 0.1045205146074295, 'epoch': 0.1}

 31%|███       | 1020/3300 [2:33:22<4:13:36,  6.67s/it]INFO 04-13 09:54:20 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1021/3300 [2:33:29<4:11:46,  6.63s/it]2026-04-13 09:54:27,180 | INFO | train_grpo_train | metrics_logged mode=train step=1021

                                                       
{'loss': 0.0667, 'grad_norm': 11.773134231567383, 'learning_rate': 6.90909090909091e-06, 'num_tokens': 1805389.0, 'completions/mean_length': 81.625, 'completions/min_length': 48.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 81.625, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7907544374465942, 'rewards/meter/std': 0.28326550126075745, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9537135362625122, 'rewards/repeat_soft/std': 0.015823470428586006, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.5364459156990051, 'rewards/total_composite/std': 0.06541776657104492, 'reward': 0.5364459156990051, 'reward_std': 0.06541776657104492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15752579271793365, 'sampling/sampling_logp_difference/max': 1.6947920322418213, 'sampling/importance_sampling_ratio/min': 0.18363742530345917, 'sampling/importance_sampling_ratio/mean': 1.0146045684814453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0685177594423294, 'clip_ratio/low_mean': 0.04912734962999821, 'clip_ratio/low_min': 0.04912734962999821, 'clip_ratio/high_mean': 0.10065225791186094, 'clip_ratio/high_max': 0.10065225791186094, 'clip_ratio/region_mean': 0.14977960754185915, 'reward_total_mean': 0.5364459156990051, 'reward_meter_mean': 0.7907544374465942, 'reward_meter_std': 0.28326550126075745, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9537135362625122, 'reward_repeat_soft_std': 0.015823470428586006, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.5364459156990051, 'reward_total_composite_std': 0.06541776657104492, 'epoch': 0.1}

 31%|███       | 1021/3300 [2:33:29<4:11:46,  6.63s/it]INFO 04-13 09:54:27 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1022/3300 [2:33:35<4:04:16,  6.43s/it]2026-04-13 09:54:33,158 | INFO | train_grpo_train | metrics_logged mode=train step=1022

                                                       
{'loss': 0.0067, 'grad_norm': 15.89610481262207, 'learning_rate': 6.906060606060606e-06, 'num_tokens': 1806966.0, 'completions/mean_length': 47.125, 'completions/min_length': 39.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.5483500957489014, 'rewards/meter/std': 0.42771151661872864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9534322619438171, 'rewards/repeat_soft/std': 0.053940366953611374, 'rewards/judge_quality/mean': 0.4762499928474426, 'rewards/judge_quality/std': 0.09941796213388443, 'rewards/total_composite/mean': 0.5007228851318359, 'rewards/total_composite/std': 0.11496507376432419, 'reward': 0.5007228851318359, 'reward_std': 0.11496507376432419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16265255212783813, 'sampling/sampling_logp_difference/max': 1.4988012313842773, 'sampling/importance_sampling_ratio/min': 0.22339780628681183, 'sampling/importance_sampling_ratio/mean': 1.0458287000656128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0818421095609665, 'clip_ratio/low_mean': 0.08918740693479776, 'clip_ratio/low_min': 0.08918740693479776, 'clip_ratio/high_mean': 0.06208793632686138, 'clip_ratio/high_max': 0.06208793632686138, 'clip_ratio/region_mean': 0.15127534326165915, 'reward_total_mean': 0.5007228851318359, 'reward_meter_mean': 0.5483500957489014, 'reward_meter_std': 0.42771151661872864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9534322619438171, 'reward_repeat_soft_std': 0.053940366953611374, 'reward_judge_quality_mean': 0.4762499928474426, 'reward_judge_quality_std': 0.09941796213388443, 'reward_total_composite_mean': 0.5007228851318359, 'reward_total_composite_std': 0.11496507376432419, 'epoch': 0.1}

 31%|███       | 1022/3300 [2:33:35<4:04:16,  6.43s/it]INFO 04-13 09:54:33 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1023/3300 [2:33:41<4:01:25,  6.36s/it]2026-04-13 09:54:39,347 | INFO | train_grpo_train | metrics_logged mode=train step=1023

                                                       
{'loss': 0.0205, 'grad_norm': 14.367979049682617, 'learning_rate': 6.903030303030304e-06, 'num_tokens': 1808554.0, 'completions/mean_length': 39.5, 'completions/min_length': 32.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8621435165405273, 'rewards/meter/std': 0.30848589539527893, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9868086576461792, 'rewards/repeat_soft/std': 0.010005839169025421, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.6074839234352112, 'rewards/total_composite/std': 0.10497891902923584, 'reward': 0.6074839234352112, 'reward_std': 0.10497891902923584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16451585292816162, 'sampling/sampling_logp_difference/max': 2.0143141746520996, 'sampling/importance_sampling_ratio/min': 0.13341186940670013, 'sampling/importance_sampling_ratio/mean': 1.035903811454773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4066985547542572, 'clip_ratio/low_mean': 0.008522727526724339, 'clip_ratio/low_min': 0.008522727526724339, 'clip_ratio/high_mean': 0.12929297517985106, 'clip_ratio/high_max': 0.12929297517985106, 'clip_ratio/region_mean': 0.1378157027065754, 'reward_total_mean': 0.6074839234352112, 'reward_meter_mean': 0.8621435165405273, 'reward_meter_std': 0.30848589539527893, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9868086576461792, 'reward_repeat_soft_std': 0.010005839169025421, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.6074839234352112, 'reward_total_composite_std': 0.10497891902923584, 'epoch': 0.1}

 31%|███       | 1023/3300 [2:33:41<4:01:25,  6.36s/it]INFO 04-13 09:54:39 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1024/3300 [2:33:47<3:52:46,  6.14s/it]2026-04-13 09:54:44,960 | INFO | train_grpo_train | metrics_logged mode=train step=1024

                                                       
{'loss': -0.0067, 'grad_norm': 17.783018112182617, 'learning_rate': 6.9e-06, 'num_tokens': 1809928.0, 'completions/mean_length': 25.75, 'completions/min_length': 22.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9660782814025879, 'rewards/meter/std': 0.06317251920700073, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9620115756988525, 'rewards/repeat_soft/std': 0.0013815105194225907, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.7001445293426514, 'rewards/total_composite/std': 0.1472204178571701, 'reward': 0.7001445293426514, 'reward_std': 0.1472204029560089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14203289151191711, 'sampling/sampling_logp_difference/max': 1.0836036205291748, 'sampling/importance_sampling_ratio/min': 0.35814791917800903, 'sampling/importance_sampling_ratio/mean': 1.026048183441162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9504631012678146, 'clip_ratio/low_mean': 0.08692905679345131, 'clip_ratio/low_min': 0.08692905679345131, 'clip_ratio/high_mean': 0.017241379246115685, 'clip_ratio/high_max': 0.017241379246115685, 'clip_ratio/region_mean': 0.104170436039567, 'reward_total_mean': 0.7001445293426514, 'reward_meter_mean': 0.9660782814025879, 'reward_meter_std': 0.06317251920700073, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9620115756988525, 'reward_repeat_soft_std': 0.0013815105194225907, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.7001445293426514, 'reward_total_composite_std': 0.1472204178571701, 'epoch': 0.1}

 31%|███       | 1024/3300 [2:33:47<3:52:46,  6.14s/it]INFO 04-13 09:54:45 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1025/3300 [2:33:53<3:50:51,  6.09s/it]2026-04-13 09:54:50,939 | INFO | train_grpo_train | metrics_logged mode=train step=1025

                                                       
{'loss': 0.0915, 'grad_norm': 16.288047790527344, 'learning_rate': 6.896969696969697e-06, 'num_tokens': 1811509.0, 'completions/mean_length': 45.625, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.736138105392456, 'rewards/meter/std': 0.43030664324760437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9859573841094971, 'rewards/repeat_soft/std': 0.009918374940752983, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5562318563461304, 'rewards/total_composite/std': 0.11651643365621567, 'reward': 0.5562318563461304, 'reward_std': 0.11651644110679626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15458832681179047, 'sampling/sampling_logp_difference/max': 1.1145048141479492, 'sampling/importance_sampling_ratio/min': 0.3280777037143707, 'sampling/importance_sampling_ratio/mean': 1.0327446460723877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4500950127840042, 'clip_ratio/low_mean': 0.05376512371003628, 'clip_ratio/low_min': 0.05376512371003628, 'clip_ratio/high_mean': 0.12403824273496866, 'clip_ratio/high_max': 0.12403824273496866, 'clip_ratio/region_mean': 0.17780336644500494, 'reward_total_mean': 0.5562318563461304, 'reward_meter_mean': 0.736138105392456, 'reward_meter_std': 0.43030664324760437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9859573841094971, 'reward_repeat_soft_std': 0.009918374940752983, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5562318563461304, 'reward_total_composite_std': 0.11651643365621567, 'epoch': 0.1}

 31%|███       | 1025/3300 [2:33:53<3:50:51,  6.09s/it]INFO 04-13 09:54:51 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1026/3300 [2:33:59<3:48:47,  6.04s/it]2026-04-13 09:54:56,858 | INFO | train_grpo_train | metrics_logged mode=train step=1026

                                                       
{'loss': 0.2184, 'grad_norm': 21.804195404052734, 'learning_rate': 6.893939393939395e-06, 'num_tokens': 1812732.0, 'completions/mean_length': 26.875, 'completions/min_length': 18.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9173821210861206, 'rewards/meter/std': 0.07125087082386017, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9528553485870361, 'rewards/repeat_soft/std': 0.030632054433226585, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.5611845254898071, 'rewards/total_composite/std': 0.08697989583015442, 'reward': 0.5611845254898071, 'reward_std': 0.08697989583015442, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18380378186702728, 'sampling/sampling_logp_difference/max': 0.9918375015258789, 'sampling/importance_sampling_ratio/min': 0.37089455127716064, 'sampling/importance_sampling_ratio/mean': 1.047836422920227, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.3717611581087112, 'clip_ratio/low_mean': 0.05922025255858898, 'clip_ratio/low_min': 0.05922025255858898, 'clip_ratio/high_mean': 0.12956710113212466, 'clip_ratio/high_max': 0.12956710113212466, 'clip_ratio/region_mean': 0.18878735369071364, 'reward_total_mean': 0.5611845254898071, 'reward_meter_mean': 0.9173821210861206, 'reward_meter_std': 0.07125087082386017, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9528553485870361, 'reward_repeat_soft_std': 0.030632054433226585, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.5611845254898071, 'reward_total_composite_std': 0.08697989583015442, 'epoch': 0.1}

 31%|███       | 1026/3300 [2:33:59<3:48:47,  6.04s/it]INFO 04-13 09:54:57 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1027/3300 [2:34:05<3:54:11,  6.18s/it]2026-04-13 09:55:03,373 | INFO | train_grpo_train | metrics_logged mode=train step=1027

                                                       
{'loss': 0.0335, 'grad_norm': 14.44566535949707, 'learning_rate': 6.890909090909092e-06, 'num_tokens': 1814942.0, 'completions/mean_length': 87.25, 'completions/min_length': 79.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.25, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.6665823459625244, 'rewards/meter/std': 0.319399356842041, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9894012808799744, 'rewards/repeat_soft/std': 0.005543151404708624, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.20672619342803955, 'rewards/total_composite/mean': 0.5615822672843933, 'rewards/total_composite/std': 0.1492321789264679, 'reward': 0.5615822672843933, 'reward_std': 0.1492321789264679, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18131457269191742, 'sampling/sampling_logp_difference/max': 2.362147331237793, 'sampling/importance_sampling_ratio/min': 0.0942176878452301, 'sampling/importance_sampling_ratio/mean': 1.0077650547027588, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8842276111245155, 'clip_ratio/low_mean': 0.10200242698192596, 'clip_ratio/low_min': 0.10200242698192596, 'clip_ratio/high_mean': 0.050042061135172844, 'clip_ratio/high_max': 0.050042061135172844, 'clip_ratio/region_mean': 0.1520444881170988, 'reward_total_mean': 0.5615822672843933, 'reward_meter_mean': 0.6665823459625244, 'reward_meter_std': 0.319399356842041, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9894012808799744, 'reward_repeat_soft_std': 0.005543151404708624, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.20672619342803955, 'reward_total_composite_mean': 0.5615822672843933, 'reward_total_composite_std': 0.1492321789264679, 'epoch': 0.1}

 31%|███       | 1027/3300 [2:34:05<3:54:11,  6.18s/it]INFO 04-13 09:55:03 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1028/3300 [2:34:11<3:56:36,  6.25s/it]2026-04-13 09:55:09,780 | INFO | train_grpo_train | metrics_logged mode=train step=1028

                                                       
{'loss': 0.0588, 'grad_norm': 8.865718841552734, 'learning_rate': 6.887878787878789e-06, 'num_tokens': 1817006.0, 'completions/mean_length': 75.0, 'completions/min_length': 68.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9105406999588013, 'rewards/meter/std': 0.11760742217302322, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9534875154495239, 'rewards/repeat_soft/std': 0.04239608719944954, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6329624652862549, 'rewards/total_composite/std': 0.07994944602251053, 'reward': 0.6329624652862549, 'reward_std': 0.07994943857192993, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15391583740711212, 'sampling/sampling_logp_difference/max': 1.6568903923034668, 'sampling/importance_sampling_ratio/min': 0.19073115289211273, 'sampling/importance_sampling_ratio/mean': 1.0367811918258667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.289549097418785, 'clip_ratio/low_mean': 0.09470611996948719, 'clip_ratio/low_min': 0.09470611996948719, 'clip_ratio/high_mean': 0.043478261679410934, 'clip_ratio/high_max': 0.043478261679410934, 'clip_ratio/region_mean': 0.13818438164889812, 'reward_total_mean': 0.6329624652862549, 'reward_meter_mean': 0.9105406999588013, 'reward_meter_std': 0.11760742217302322, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9534875154495239, 'reward_repeat_soft_std': 0.04239608719944954, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6329624652862549, 'reward_total_composite_std': 0.07994944602251053, 'epoch': 0.1}

 31%|███       | 1028/3300 [2:34:12<3:56:36,  6.25s/it]INFO 04-13 09:55:10 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1029/3300 [2:34:18<3:59:36,  6.33s/it]2026-04-13 09:55:16,300 | INFO | train_grpo_train | metrics_logged mode=train step=1029

                                                       
{'loss': 0.0274, 'grad_norm': 9.817951202392578, 'learning_rate': 6.8848484848484854e-06, 'num_tokens': 1818995.0, 'completions/mean_length': 73.625, 'completions/min_length': 52.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9692832231521606, 'rewards/meter/std': 0.03116937167942524, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9359431266784668, 'rewards/repeat_soft/std': 0.04442360997200012, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.5677253007888794, 'rewards/total_composite/std': 0.08369077742099762, 'reward': 0.5677253007888794, 'reward_std': 0.08369076997041702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15762563049793243, 'sampling/sampling_logp_difference/max': 1.8013087511062622, 'sampling/importance_sampling_ratio/min': 0.16508269309997559, 'sampling/importance_sampling_ratio/mean': 1.0174403190612793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.230010949075222, 'clip_ratio/low_mean': 0.11425292678177357, 'clip_ratio/low_min': 0.11425292678177357, 'clip_ratio/high_mean': 0.03355263266712427, 'clip_ratio/high_max': 0.03355263266712427, 'clip_ratio/region_mean': 0.14780555944889784, 'reward_total_mean': 0.5677253007888794, 'reward_meter_mean': 0.9692832231521606, 'reward_meter_std': 0.03116937167942524, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9359431266784668, 'reward_repeat_soft_std': 0.04442360997200012, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.5677253007888794, 'reward_total_composite_std': 0.08369077742099762, 'epoch': 0.1}

 31%|███       | 1029/3300 [2:34:18<3:59:36,  6.33s/it]INFO 04-13 09:55:16 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1030/3300 [2:34:24<3:56:14,  6.24s/it]2026-04-13 09:55:22,341 | INFO | train_grpo_train | metrics_logged mode=train step=1030

                                                       
{'loss': 0.109, 'grad_norm': 14.510078430175781, 'learning_rate': 6.881818181818183e-06, 'num_tokens': 1820599.0, 'completions/mean_length': 42.5, 'completions/min_length': 30.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7224732637405396, 'rewards/meter/std': 0.3413645029067993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9543820023536682, 'rewards/repeat_soft/std': 0.04431818425655365, 'rewards/judge_quality/mean': 0.6024999618530273, 'rewards/judge_quality/std': 0.1976107507944107, 'rewards/total_composite/mean': 0.6407123804092407, 'rewards/total_composite/std': 0.18194904923439026, 'reward': 0.6407123804092407, 'reward_std': 0.18194904923439026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1599501073360443, 'sampling/sampling_logp_difference/max': 1.005096435546875, 'sampling/importance_sampling_ratio/min': 0.36600935459136963, 'sampling/importance_sampling_ratio/mean': 1.0132290124893188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3126289695501328, 'clip_ratio/low_mean': 0.08708263840526342, 'clip_ratio/low_min': 0.08708263840526342, 'clip_ratio/high_mean': 0.05918329767882824, 'clip_ratio/high_max': 0.05918329767882824, 'clip_ratio/region_mean': 0.14626593608409166, 'reward_total_mean': 0.6407123804092407, 'reward_meter_mean': 0.7224732637405396, 'reward_meter_std': 0.3413645029067993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9543820023536682, 'reward_repeat_soft_std': 0.04431818425655365, 'reward_judge_quality_mean': 0.6024999618530273, 'reward_judge_quality_std': 0.1976107507944107, 'reward_total_composite_mean': 0.6407123804092407, 'reward_total_composite_std': 0.18194904923439026, 'epoch': 0.1}

 31%|███       | 1030/3300 [2:34:24<3:56:14,  6.24s/it]INFO 04-13 09:55:22 [block_pool.py:378] Successfully reset prefix cache

 31%|███       | 1031/3300 [2:34:30<3:58:04,  6.30s/it]2026-04-13 09:55:28,758 | INFO | train_grpo_train | metrics_logged mode=train step=1031

                                                       
{'loss': -0.0032, 'grad_norm': 13.197099685668945, 'learning_rate': 6.878787878787879e-06, 'num_tokens': 1822347.0, 'completions/mean_length': 62.5, 'completions/min_length': 55.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.9130160808563232, 'rewards/meter/std': 0.21048715710639954, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.930400013923645, 'rewards/repeat_soft/std': 0.02374812588095665, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5388133525848389, 'rewards/total_composite/std': 0.05753803625702858, 'reward': 0.5388133525848389, 'reward_std': 0.05753804370760918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15961933135986328, 'sampling/sampling_logp_difference/max': 1.3118410110473633, 'sampling/importance_sampling_ratio/min': 0.26932376623153687, 'sampling/importance_sampling_ratio/mean': 1.0412040948867798, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4632093608379364, 'clip_ratio/low_mean': 0.015909090638160706, 'clip_ratio/low_min': 0.015909090638160706, 'clip_ratio/high_mean': 0.11567418463528156, 'clip_ratio/high_max': 0.11567418463528156, 'clip_ratio/region_mean': 0.13158327527344227, 'reward_total_mean': 0.5388133525848389, 'reward_meter_mean': 0.9130160808563232, 'reward_meter_std': 0.21048715710639954, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.930400013923645, 'reward_repeat_soft_std': 0.02374812588095665, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5388133525848389, 'reward_total_composite_std': 0.05753803625702858, 'epoch': 0.1}

 31%|███       | 1031/3300 [2:34:30<3:58:04,  6.30s/it]INFO 04-13 09:55:28 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1032/3300 [2:34:42<4:54:18,  7.79s/it]2026-04-13 09:55:40,024 | INFO | train_grpo_train | metrics_logged mode=train step=1032

                                                       
{'loss': -0.1328, 'grad_norm': 2.621129035949707, 'learning_rate': 6.875757575757576e-06, 'num_tokens': 1824186.0, 'completions/mean_length': 117.875, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 61.57143020629883, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8732489347457886, 'rewards/meter/std': 0.3334799110889435, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9191628694534302, 'rewards/repeat_soft/std': 0.04863610491156578, 'rewards/judge_quality/mean': 0.3687499761581421, 'rewards/judge_quality/std': 0.1940867006778717, 'rewards/total_composite/mean': 0.4840266704559326, 'rewards/total_composite/std': 0.21617349982261658, 'reward': 0.4840266704559326, 'reward_std': 0.21617348492145538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15878494083881378, 'sampling/sampling_logp_difference/max': 2.8949594497680664, 'sampling/importance_sampling_ratio/min': 0.05530126392841339, 'sampling/importance_sampling_ratio/mean': 1.0299222469329834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0180036127567291, 'clip_ratio/low_mean': 0.024951783940196037, 'clip_ratio/low_min': 0.024951783940196037, 'clip_ratio/high_mean': 0.08247385453432798, 'clip_ratio/high_max': 0.08247385453432798, 'clip_ratio/region_mean': 0.10742563847452402, 'reward_total_mean': 0.4840266704559326, 'reward_meter_mean': 0.8732489347457886, 'reward_meter_std': 0.3334799110889435, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9191628694534302, 'reward_repeat_soft_std': 0.04863610491156578, 'reward_judge_quality_mean': 0.3687499761581421, 'reward_judge_quality_std': 0.1940867006778717, 'reward_total_composite_mean': 0.4840266704559326, 'reward_total_composite_std': 0.21617349982261658, 'epoch': 0.1}

 31%|███▏      | 1032/3300 [2:34:42<4:54:18,  7.79s/it]INFO 04-13 09:55:40 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1033/3300 [2:34:53<5:29:20,  8.72s/it]2026-04-13 09:55:50,911 | INFO | train_grpo_train | metrics_logged mode=train step=1033

                                                       
{'loss': -0.0549, 'grad_norm': 4.162497043609619, 'learning_rate': 6.872727272727273e-06, 'num_tokens': 1825457.0, 'completions/mean_length': 84.875, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.85714340209961, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.37773576378822327, 'rewards/meter/std': 0.4795103371143341, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9588949680328369, 'rewards/repeat_soft/std': 0.024640807881951332, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.262593537569046, 'rewards/total_composite/mean': 0.4405129849910736, 'rewards/total_composite/std': 0.26010143756866455, 'reward': 0.4405129849910736, 'reward_std': 0.26010146737098694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1446915566921234, 'sampling/sampling_logp_difference/max': 0.833672285079956, 'sampling/importance_sampling_ratio/min': 0.47228142619132996, 'sampling/importance_sampling_ratio/mean': 1.040169358253479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.445899561047554, 'clip_ratio/low_mean': 0.07106978353112936, 'clip_ratio/low_min': 0.07106978353112936, 'clip_ratio/high_mean': 0.06505952402949333, 'clip_ratio/high_max': 0.06505952402949333, 'clip_ratio/region_mean': 0.1361293075606227, 'reward_total_mean': 0.4405129849910736, 'reward_meter_mean': 0.37773576378822327, 'reward_meter_std': 0.4795103371143341, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9588949680328369, 'reward_repeat_soft_std': 0.024640807881951332, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.262593537569046, 'reward_total_composite_mean': 0.4405129849910736, 'reward_total_composite_std': 0.26010143756866455, 'epoch': 0.1}

 31%|███▏      | 1033/3300 [2:34:53<5:29:20,  8.72s/it]INFO 04-13 09:55:51 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1034/3300 [2:34:59<5:00:13,  7.95s/it]2026-04-13 09:55:57,068 | INFO | train_grpo_train | metrics_logged mode=train step=1034

                                                       
{'loss': 0.0458, 'grad_norm': 12.486868858337402, 'learning_rate': 6.869696969696971e-06, 'num_tokens': 1827302.0, 'completions/mean_length': 49.625, 'completions/min_length': 38.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9833406209945679, 'rewards/meter/std': 0.02095111645758152, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9626805186271667, 'rewards/repeat_soft/std': 0.032711006700992584, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6223706007003784, 'rewards/total_composite/std': 0.013778239488601685, 'reward': 0.6223706007003784, 'reward_std': 0.01377825252711773, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1560501903295517, 'sampling/sampling_logp_difference/max': 1.6480932235717773, 'sampling/importance_sampling_ratio/min': 0.19241644442081451, 'sampling/importance_sampling_ratio/mean': 1.0278218984603882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6125024110078812, 'clip_ratio/low_mean': 0.09322284162044525, 'clip_ratio/low_min': 0.09322284162044525, 'clip_ratio/high_mean': 0.05953734088689089, 'clip_ratio/high_max': 0.05953734088689089, 'clip_ratio/region_mean': 0.15276018250733614, 'reward_total_mean': 0.6223706007003784, 'reward_meter_mean': 0.9833406209945679, 'reward_meter_std': 0.02095111645758152, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9626805186271667, 'reward_repeat_soft_std': 0.032711006700992584, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6223706007003784, 'reward_total_composite_std': 0.013778239488601685, 'epoch': 0.1}

 31%|███▏      | 1034/3300 [2:34:59<5:00:13,  7.95s/it]INFO 04-13 09:55:57 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1035/3300 [2:35:06<4:48:58,  7.66s/it]2026-04-13 09:56:04,038 | INFO | train_grpo_train | metrics_logged mode=train step=1035

                                                       
{'loss': 0.0472, 'grad_norm': 8.736976623535156, 'learning_rate': 6.866666666666667e-06, 'num_tokens': 1829630.0, 'completions/mean_length': 96.0, 'completions/min_length': 82.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8666692972183228, 'rewards/meter/std': 0.3343394994735718, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8650614023208618, 'rewards/repeat_soft/std': 0.12002670764923096, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5006651878356934, 'rewards/total_composite/std': 0.08232749998569489, 'reward': 0.5006651878356934, 'reward_std': 0.08232749998569489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12479706853628159, 'sampling/sampling_logp_difference/max': 1.4670276641845703, 'sampling/importance_sampling_ratio/min': 0.2306099236011505, 'sampling/importance_sampling_ratio/mean': 1.0250110626220703, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.084361806511879, 'clip_ratio/low_mean': 0.019859813153743744, 'clip_ratio/low_min': 0.019859813153743744, 'clip_ratio/high_mean': 0.08743510372005403, 'clip_ratio/high_max': 0.08743510372005403, 'clip_ratio/region_mean': 0.10729491687379777, 'reward_total_mean': 0.5006651878356934, 'reward_meter_mean': 0.8666692972183228, 'reward_meter_std': 0.3343394994735718, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8650614023208618, 'reward_repeat_soft_std': 0.12002670764923096, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5006651878356934, 'reward_total_composite_std': 0.08232749998569489, 'epoch': 0.1}

 31%|███▏      | 1035/3300 [2:35:06<4:48:58,  7.66s/it]INFO 04-13 09:56:04 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1036/3300 [2:35:12<4:34:28,  7.27s/it]2026-04-13 09:56:10,422 | INFO | train_grpo_train | metrics_logged mode=train step=1036

                                                       
{'loss': 0.0382, 'grad_norm': 12.796359062194824, 'learning_rate': 6.8636363636363645e-06, 'num_tokens': 1831621.0, 'completions/mean_length': 70.875, 'completions/min_length': 57.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9110418558120728, 'rewards/meter/std': 0.198968306183815, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9122515916824341, 'rewards/repeat_soft/std': 0.07234229147434235, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5355521440505981, 'rewards/total_composite/std': 0.05388759821653366, 'reward': 0.5355521440505981, 'reward_std': 0.05388757586479187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15520618855953217, 'sampling/sampling_logp_difference/max': 1.6682729721069336, 'sampling/importance_sampling_ratio/min': 0.1885724514722824, 'sampling/importance_sampling_ratio/mean': 1.0142109394073486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2992232292890549, 'clip_ratio/low_mean': 0.021831881254911423, 'clip_ratio/low_min': 0.021831881254911423, 'clip_ratio/high_mean': 0.11172603955492377, 'clip_ratio/high_max': 0.11172603955492377, 'clip_ratio/region_mean': 0.1335579208098352, 'reward_total_mean': 0.5355521440505981, 'reward_meter_mean': 0.9110418558120728, 'reward_meter_std': 0.198968306183815, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9122515916824341, 'reward_repeat_soft_std': 0.07234229147434235, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5355521440505981, 'reward_total_composite_std': 0.05388759821653366, 'epoch': 0.1}

 31%|███▏      | 1036/3300 [2:35:12<4:34:28,  7.27s/it]INFO 04-13 09:56:10 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1037/3300 [2:35:18<4:21:13,  6.93s/it]2026-04-13 09:56:16,535 | INFO | train_grpo_train | metrics_logged mode=train step=1037

                                                       
{'loss': 0.0578, 'grad_norm': 14.418715476989746, 'learning_rate': 6.860606060606061e-06, 'num_tokens': 1833571.0, 'completions/mean_length': 62.75, 'completions/min_length': 54.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.4608146548271179, 'rewards/meter/std': 0.37401625514030457, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9413318037986755, 'rewards/repeat_soft/std': 0.058841411024332047, 'rewards/judge_quality/mean': 0.6237500309944153, 'rewards/judge_quality/std': 0.2232191562652588, 'rewards/total_composite/mean': 0.4488751292228699, 'rewards/total_composite/std': 0.11704400181770325, 'reward': 0.4488751292228699, 'reward_std': 0.11704400926828384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1682649403810501, 'sampling/sampling_logp_difference/max': 2.2751240730285645, 'sampling/importance_sampling_ratio/min': 0.1027841567993164, 'sampling/importance_sampling_ratio/mean': 1.0111799240112305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8135004937648773, 'clip_ratio/low_mean': 0.0947592044249177, 'clip_ratio/low_min': 0.0947592044249177, 'clip_ratio/high_mean': 0.06098484992980957, 'clip_ratio/high_max': 0.06098484992980957, 'clip_ratio/region_mean': 0.15574405435472727, 'reward_total_mean': 0.4488751292228699, 'reward_meter_mean': 0.4608146548271179, 'reward_meter_std': 0.37401625514030457, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9413318037986755, 'reward_repeat_soft_std': 0.058841411024332047, 'reward_judge_quality_mean': 0.6237500309944153, 'reward_judge_quality_std': 0.2232191562652588, 'reward_total_composite_mean': 0.4488751292228699, 'reward_total_composite_std': 0.11704400181770325, 'epoch': 0.1}

 31%|███▏      | 1037/3300 [2:35:18<4:21:13,  6.93s/it]INFO 04-13 09:56:16 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1038/3300 [2:35:32<5:34:00,  8.86s/it]2026-04-13 09:56:29,913 | INFO | train_grpo_train | metrics_logged mode=train step=1038

                                                       
{'loss': -0.1081, 'grad_norm': 3.557382822036743, 'learning_rate': 6.857575757575758e-06, 'num_tokens': 1835343.0, 'completions/mean_length': 101.5, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.85714340209961, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6340430974960327, 'rewards/meter/std': 0.4002546966075897, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9766391515731812, 'rewards/repeat_soft/std': 0.01620764099061489, 'rewards/judge_quality/mean': 0.581250011920929, 'rewards/judge_quality/std': 0.2953660190105438, 'rewards/total_composite/mean': 0.5815249681472778, 'rewards/total_composite/std': 0.2931891679763794, 'reward': 0.5815249681472778, 'reward_std': 0.293189138174057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19768273830413818, 'sampling/sampling_logp_difference/max': 1.267542839050293, 'sampling/importance_sampling_ratio/min': 0.2815225422382355, 'sampling/importance_sampling_ratio/mean': 1.0388178825378418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5156055986881256, 'clip_ratio/low_mean': 0.05267857201397419, 'clip_ratio/low_min': 0.05267857201397419, 'clip_ratio/high_mean': 0.08354822359979153, 'clip_ratio/high_max': 0.08354822359979153, 'clip_ratio/region_mean': 0.13622679561376572, 'reward_total_mean': 0.5815249681472778, 'reward_meter_mean': 0.6340430974960327, 'reward_meter_std': 0.4002546966075897, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9766391515731812, 'reward_repeat_soft_std': 0.01620764099061489, 'reward_judge_quality_mean': 0.581250011920929, 'reward_judge_quality_std': 0.2953660190105438, 'reward_total_composite_mean': 0.5815249681472778, 'reward_total_composite_std': 0.2931891679763794, 'epoch': 0.1}

 31%|███▏      | 1038/3300 [2:35:32<5:34:00,  8.86s/it]INFO 04-13 09:56:30 [block_pool.py:378] Successfully reset prefix cache

 31%|███▏      | 1039/3300 [2:35:38<5:03:07,  8.04s/it]2026-04-13 09:56:36,050 | INFO | train_grpo_train | metrics_logged mode=train step=1039

                                                       
{'loss': 0.0198, 'grad_norm': 17.626461029052734, 'learning_rate': 6.854545454545455e-06, 'num_tokens': 1836971.0, 'completions/mean_length': 34.5, 'completions/min_length': 25.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.42783617973327637, 'rewards/meter/std': 0.36010390520095825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9915820360183716, 'rewards/repeat_soft/std': 0.00942145474255085, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.4947046637535095, 'rewards/total_composite/std': 0.13748851418495178, 'reward': 0.4947046637535095, 'reward_std': 0.13748851418495178, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1619514524936676, 'sampling/sampling_logp_difference/max': 1.6502609252929688, 'sampling/importance_sampling_ratio/min': 0.19199980795383453, 'sampling/importance_sampling_ratio/mean': 1.031578540802002, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2084343880414963, 'clip_ratio/low_mean': 0.09601151198148727, 'clip_ratio/low_min': 0.09601151198148727, 'clip_ratio/high_mean': 0.06850469391793013, 'clip_ratio/high_max': 0.06850469391793013, 'clip_ratio/region_mean': 0.1645162058994174, 'reward_total_mean': 0.4947046637535095, 'reward_meter_mean': 0.42783617973327637, 'reward_meter_std': 0.36010390520095825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9915820360183716, 'reward_repeat_soft_std': 0.00942145474255085, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.4947046637535095, 'reward_total_composite_std': 0.13748851418495178, 'epoch': 0.1}

 31%|███▏      | 1039/3300 [2:35:38<5:03:07,  8.04s/it]INFO 04-13 09:56:36 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1040/3300 [2:35:44<4:42:55,  7.51s/it]2026-04-13 09:56:42,322 | INFO | train_grpo_train | metrics_logged mode=train step=1040

                                                       
{'loss': 0.0609, 'grad_norm': 10.139225006103516, 'learning_rate': 6.851515151515153e-06, 'num_tokens': 1839066.0, 'completions/mean_length': 80.875, 'completions/min_length': 74.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.7077836990356445, 'rewards/meter/std': 0.39633890986442566, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9848485589027405, 'rewards/repeat_soft/std': 0.015427964739501476, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5426092743873596, 'rewards/total_composite/std': 0.16634903848171234, 'reward': 0.5426092743873596, 'reward_std': 0.16634905338287354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17039991915225983, 'sampling/sampling_logp_difference/max': 1.3873834609985352, 'sampling/importance_sampling_ratio/min': 0.24972787499427795, 'sampling/importance_sampling_ratio/mean': 1.0405789613723755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.9115580022335052, 'clip_ratio/low_mean': 0.06445186212658882, 'clip_ratio/low_min': 0.06445186212658882, 'clip_ratio/high_mean': 0.09954247251152992, 'clip_ratio/high_max': 0.09954247251152992, 'clip_ratio/region_mean': 0.16399433463811874, 'reward_total_mean': 0.5426092743873596, 'reward_meter_mean': 0.7077836990356445, 'reward_meter_std': 0.39633890986442566, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9848485589027405, 'reward_repeat_soft_std': 0.015427964739501476, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5426092743873596, 'reward_total_composite_std': 0.16634903848171234, 'epoch': 0.1}

 32%|███▏      | 1040/3300 [2:35:44<4:42:55,  7.51s/it]INFO 04-13 09:56:42 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1041/3300 [2:35:51<4:35:21,  7.31s/it]2026-04-13 09:56:49,173 | INFO | train_grpo_train | metrics_logged mode=train step=1041

                                                       
{'loss': -0.0033, 'grad_norm': 8.641227722167969, 'learning_rate': 6.848484848484849e-06, 'num_tokens': 1841294.0, 'completions/mean_length': 91.5, 'completions/min_length': 78.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.5, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.972457766532898, 'rewards/meter/std': 0.04122280329465866, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9048149585723877, 'rewards/repeat_soft/std': 0.02271733246743679, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.5127357840538025, 'rewards/total_composite/std': 0.06192810833454132, 'reward': 0.5127357840538025, 'reward_std': 0.06192811205983162, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17517006397247314, 'sampling/sampling_logp_difference/max': 1.6087284088134766, 'sampling/importance_sampling_ratio/min': 0.20014193654060364, 'sampling/importance_sampling_ratio/mean': 1.0346416234970093, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7875735759735107, 'clip_ratio/low_mean': 0.027623284608125687, 'clip_ratio/low_min': 0.027623284608125687, 'clip_ratio/high_mean': 0.1168333850800991, 'clip_ratio/high_max': 0.1168333850800991, 'clip_ratio/region_mean': 0.1444566696882248, 'reward_total_mean': 0.5127357840538025, 'reward_meter_mean': 0.972457766532898, 'reward_meter_std': 0.04122280329465866, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9048149585723877, 'reward_repeat_soft_std': 0.02271733246743679, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.5127357840538025, 'reward_total_composite_std': 0.06192810833454132, 'epoch': 0.1}

 32%|███▏      | 1041/3300 [2:35:51<4:35:21,  7.31s/it]INFO 04-13 09:56:49 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1042/3300 [2:35:57<4:24:42,  7.03s/it]2026-04-13 09:56:55,551 | INFO | train_grpo_train | metrics_logged mode=train step=1042

                                                       
{'loss': 0.0234, 'grad_norm': 12.043434143066406, 'learning_rate': 6.845454545454546e-06, 'num_tokens': 1842849.0, 'completions/mean_length': 43.375, 'completions/min_length': 40.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6926597356796265, 'rewards/meter/std': 0.3405209481716156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9692997932434082, 'rewards/repeat_soft/std': 0.04518914222717285, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6476796865463257, 'rewards/total_composite/std': 0.226663738489151, 'reward': 0.6476796865463257, 'reward_std': 0.226663738489151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11147625744342804, 'sampling/sampling_logp_difference/max': 1.3472824096679688, 'sampling/importance_sampling_ratio/min': 0.25994572043418884, 'sampling/importance_sampling_ratio/mean': 1.0130397081375122, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7532584443688393, 'clip_ratio/low_mean': 0.061964747961610556, 'clip_ratio/low_min': 0.061964747961610556, 'clip_ratio/high_mean': 0.025568853598088026, 'clip_ratio/high_max': 0.025568853598088026, 'clip_ratio/region_mean': 0.08753360155969858, 'reward_total_mean': 0.6476796865463257, 'reward_meter_mean': 0.6926597356796265, 'reward_meter_std': 0.3405209481716156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9692997932434082, 'reward_repeat_soft_std': 0.04518914222717285, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6476796865463257, 'reward_total_composite_std': 0.226663738489151, 'epoch': 0.1}

 32%|███▏      | 1042/3300 [2:35:57<4:24:42,  7.03s/it]INFO 04-13 09:56:55 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1043/3300 [2:36:04<4:20:31,  6.93s/it]2026-04-13 09:57:02,225 | INFO | train_grpo_train | metrics_logged mode=train step=1043

                                                       
{'loss': 0.0315, 'grad_norm': 9.035298347473145, 'learning_rate': 6.842424242424243e-06, 'num_tokens': 1845063.0, 'completions/mean_length': 87.75, 'completions/min_length': 76.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.75, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.621731162071228, 'rewards/meter/std': 0.1971936672925949, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.965823769569397, 'rewards/repeat_soft/std': 0.02204710990190506, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4653659462928772, 'rewards/total_composite/std': 0.0575556755065918, 'reward': 0.4653659462928772, 'reward_std': 0.0575556717813015, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.172149658203125, 'sampling/sampling_logp_difference/max': 1.8061906099319458, 'sampling/importance_sampling_ratio/min': 0.16427874565124512, 'sampling/importance_sampling_ratio/mean': 1.033522605895996, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5045871436595917, 'clip_ratio/low_mean': 0.06947345472872257, 'clip_ratio/low_min': 0.06947345472872257, 'clip_ratio/high_mean': 0.0719126695767045, 'clip_ratio/high_max': 0.0719126695767045, 'clip_ratio/region_mean': 0.14138612430542707, 'reward_total_mean': 0.4653659462928772, 'reward_meter_mean': 0.621731162071228, 'reward_meter_std': 0.1971936672925949, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.965823769569397, 'reward_repeat_soft_std': 0.02204710990190506, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4653659462928772, 'reward_total_composite_std': 0.0575556755065918, 'epoch': 0.1}

 32%|███▏      | 1043/3300 [2:36:04<4:20:31,  6.93s/it]INFO 04-13 09:57:02 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1044/3300 [2:36:10<4:08:14,  6.60s/it]2026-04-13 09:57:08,076 | INFO | train_grpo_train | metrics_logged mode=train step=1044

                                                       
{'loss': 0.0675, 'grad_norm': 15.171388626098633, 'learning_rate': 6.83939393939394e-06, 'num_tokens': 1846734.0, 'completions/mean_length': 38.875, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.6489021182060242, 'rewards/meter/std': 0.3967968821525574, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9885673522949219, 'rewards/repeat_soft/std': 0.014239184558391571, 'rewards/judge_quality/mean': 0.5175000429153442, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.5683600902557373, 'rewards/total_composite/std': 0.1472955346107483, 'reward': 0.5683600902557373, 'reward_std': 0.1472955346107483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18410830199718475, 'sampling/sampling_logp_difference/max': 1.4863176345825195, 'sampling/importance_sampling_ratio/min': 0.22620409727096558, 'sampling/importance_sampling_ratio/mean': 1.034193754196167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.764886274933815, 'clip_ratio/low_mean': 0.06886911019682884, 'clip_ratio/low_min': 0.06886911019682884, 'clip_ratio/high_mean': 0.08515406399965286, 'clip_ratio/high_max': 0.08515406399965286, 'clip_ratio/region_mean': 0.1540231741964817, 'reward_total_mean': 0.5683600902557373, 'reward_meter_mean': 0.6489021182060242, 'reward_meter_std': 0.3967968821525574, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9885673522949219, 'reward_repeat_soft_std': 0.014239184558391571, 'reward_judge_quality_mean': 0.5175000429153442, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.5683600902557373, 'reward_total_composite_std': 0.1472955346107483, 'epoch': 0.1}

 32%|███▏      | 1044/3300 [2:36:10<4:08:14,  6.60s/it]INFO 04-13 09:57:08 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1045/3300 [2:36:21<5:00:50,  8.00s/it]2026-04-13 09:57:19,350 | INFO | train_grpo_train | metrics_logged mode=train step=1045

                                                       
{'loss': -0.0614, 'grad_norm': 1.2960211038589478, 'learning_rate': 6.8363636363636364e-06, 'num_tokens': 1848221.0, 'completions/mean_length': 209.875, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 28.600000381469727, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7898814678192139, 'rewards/meter/std': 0.33841297030448914, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9671218991279602, 'rewards/repeat_soft/std': 0.013286017812788486, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/total_composite/mean': 0.3426004648208618, 'rewards/total_composite/std': 0.29025980830192566, 'reward': 0.3426004648208618, 'reward_std': 0.29025980830192566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.22148950397968292, 'sampling/sampling_logp_difference/max': 1.2577190399169922, 'sampling/importance_sampling_ratio/min': 0.2843017578125, 'sampling/importance_sampling_ratio/mean': 1.036569595336914, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6734884232282639, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11819592490792274, 'clip_ratio/high_max': 0.11819592490792274, 'clip_ratio/region_mean': 0.11819592490792274, 'reward_total_mean': 0.3426004648208618, 'reward_meter_mean': 0.7898814678192139, 'reward_meter_std': 0.33841297030448914, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9671218991279602, 'reward_repeat_soft_std': 0.013286017812788486, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.3426004648208618, 'reward_total_composite_std': 0.29025980830192566, 'epoch': 0.1}

 32%|███▏      | 1045/3300 [2:36:21<5:00:50,  8.00s/it]INFO 04-13 09:57:19 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1046/3300 [2:36:32<5:39:14,  9.03s/it]2026-04-13 09:57:30,775 | INFO | train_grpo_train | metrics_logged mode=train step=1046

                                                       
{'loss': -0.1636, 'grad_norm': 2.6663920879364014, 'learning_rate': 6.833333333333334e-06, 'num_tokens': 1849932.0, 'completions/mean_length': 116.875, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 60.42857360839844, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.8841900825500488, 'rewards/meter/std': 0.2517623007297516, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.945469081401825, 'rewards/repeat_soft/std': 0.06652350723743439, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.22947145998477936, 'rewards/total_composite/mean': 0.5942699909210205, 'rewards/total_composite/std': 0.2586207389831543, 'reward': 0.5942699909210205, 'reward_std': 0.2586207091808319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1772335171699524, 'sampling/sampling_logp_difference/max': 1.5601367950439453, 'sampling/importance_sampling_ratio/min': 0.21010732650756836, 'sampling/importance_sampling_ratio/mean': 1.0624712705612183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7860264033079147, 'clip_ratio/low_mean': 0.013636363670229912, 'clip_ratio/low_min': 0.013636363670229912, 'clip_ratio/high_mean': 0.1267325459048152, 'clip_ratio/high_max': 0.1267325459048152, 'clip_ratio/region_mean': 0.1403689095750451, 'reward_total_mean': 0.5942699909210205, 'reward_meter_mean': 0.8841900825500488, 'reward_meter_std': 0.2517623007297516, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.945469081401825, 'reward_repeat_soft_std': 0.06652350723743439, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.22947145998477936, 'reward_total_composite_mean': 0.5942699909210205, 'reward_total_composite_std': 0.2586207389831543, 'epoch': 0.11}

 32%|███▏      | 1046/3300 [2:36:33<5:39:14,  9.03s/it]INFO 04-13 09:57:31 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1047/3300 [2:36:38<4:59:51,  7.99s/it]2026-04-13 09:57:36,324 | INFO | train_grpo_train | metrics_logged mode=train step=1047

                                                       
{'loss': 0.1043, 'grad_norm': 21.364267349243164, 'learning_rate': 6.83030303030303e-06, 'num_tokens': 1851278.0, 'completions/mean_length': 22.25, 'completions/min_length': 18.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.8054901957511902, 'rewards/meter/std': 0.3369612991809845, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9038182497024536, 'rewards/repeat_soft/std': 0.08919284492731094, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.6021766662597656, 'rewards/total_composite/std': 0.15866397321224213, 'reward': 0.6021766662597656, 'reward_std': 0.15866400301456451, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1650094836950302, 'sampling/sampling_logp_difference/max': 1.4395227432250977, 'sampling/importance_sampling_ratio/min': 0.23704086244106293, 'sampling/importance_sampling_ratio/mean': 0.9973371028900146, 'sampling/importance_sampling_ratio/max': 1.6196039915084839, 'entropy': 1.2577237635850906, 'clip_ratio/low_mean': 0.05333129828795791, 'clip_ratio/low_min': 0.05333129828795791, 'clip_ratio/high_mean': 0.07839912362396717, 'clip_ratio/high_max': 0.07839912362396717, 'clip_ratio/region_mean': 0.13173042191192508, 'reward_total_mean': 0.6021766662597656, 'reward_meter_mean': 0.8054901957511902, 'reward_meter_std': 0.3369612991809845, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9038182497024536, 'reward_repeat_soft_std': 0.08919284492731094, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.6021766662597656, 'reward_total_composite_std': 0.15866397321224213, 'epoch': 0.11}

 32%|███▏      | 1047/3300 [2:36:38<4:59:51,  7.99s/it]INFO 04-13 09:57:36 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1048/3300 [2:36:49<5:35:36,  8.94s/it]2026-04-13 09:57:47,495 | INFO | train_grpo_train | metrics_logged mode=train step=1048

                                                       
{'loss': -0.1363, 'grad_norm': 3.268489360809326, 'learning_rate': 6.827272727272728e-06, 'num_tokens': 1852978.0, 'completions/mean_length': 119.5, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.42857360839844, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7645368576049805, 'rewards/meter/std': 0.29501092433929443, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9716801643371582, 'rewards/repeat_soft/std': 0.014356398023664951, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.22947144508361816, 'rewards/total_composite/mean': 0.5657984614372253, 'rewards/total_composite/std': 0.2622303068637848, 'reward': 0.5657984614372253, 'reward_std': 0.2622303366661072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1616942435503006, 'sampling/sampling_logp_difference/max': 1.4891023635864258, 'sampling/importance_sampling_ratio/min': 0.2255750596523285, 'sampling/importance_sampling_ratio/mean': 1.0321862697601318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3364130854606628, 'clip_ratio/low_mean': 0.028233150951564312, 'clip_ratio/low_min': 0.028233150951564312, 'clip_ratio/high_mean': 0.08829259779304266, 'clip_ratio/high_max': 0.08829259779304266, 'clip_ratio/region_mean': 0.11652574874460697, 'reward_total_mean': 0.5657984614372253, 'reward_meter_mean': 0.7645368576049805, 'reward_meter_std': 0.29501092433929443, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9716801643371582, 'reward_repeat_soft_std': 0.014356398023664951, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.22947144508361816, 'reward_total_composite_mean': 0.5657984614372253, 'reward_total_composite_std': 0.2622303068637848, 'epoch': 0.11}

 32%|███▏      | 1048/3300 [2:36:49<5:35:36,  8.94s/it]INFO 04-13 09:57:47 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1049/3300 [2:37:00<5:56:52,  9.51s/it]2026-04-13 09:57:58,335 | INFO | train_grpo_train | metrics_logged mode=train step=1049

                                                       
{'loss': -0.0369, 'grad_norm': 5.236736297607422, 'learning_rate': 6.824242424242425e-06, 'num_tokens': 1854346.0, 'completions/mean_length': 85.0, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.000001907348633, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7998566627502441, 'rewards/meter/std': 0.3420889675617218, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9669623374938965, 'rewards/repeat_soft/std': 0.0133640943095088, 'rewards/judge_quality/mean': 0.41875001788139343, 'rewards/judge_quality/std': 0.24398112297058105, 'rewards/total_composite/mean': 0.4902343153953552, 'rewards/total_composite/std': 0.324945330619812, 'reward': 0.4902343153953552, 'reward_std': 0.324945330619812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17035700380802155, 'sampling/sampling_logp_difference/max': 0.9106950759887695, 'sampling/importance_sampling_ratio/min': 0.40224453806877136, 'sampling/importance_sampling_ratio/mean': 1.0246261358261108, 'sampling/importance_sampling_ratio/max': 1.9856619834899902, 'entropy': 1.6026954054832458, 'clip_ratio/low_mean': 0.01923076994717121, 'clip_ratio/low_min': 0.01923076994717121, 'clip_ratio/high_mean': 0.10076902061700821, 'clip_ratio/high_max': 0.10076902061700821, 'clip_ratio/region_mean': 0.11999979056417942, 'reward_total_mean': 0.4902343153953552, 'reward_meter_mean': 0.7998566627502441, 'reward_meter_std': 0.3420889675617218, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9669623374938965, 'reward_repeat_soft_std': 0.0133640943095088, 'reward_judge_quality_mean': 0.41875001788139343, 'reward_judge_quality_std': 0.24398112297058105, 'reward_total_composite_mean': 0.4902343153953552, 'reward_total_composite_std': 0.324945330619812, 'epoch': 0.11}

 32%|███▏      | 1049/3300 [2:37:00<5:56:52,  9.51s/it]INFO 04-13 09:57:58 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1050/3300 [2:37:11<6:14:18,  9.98s/it]2026-04-13 09:58:09,413 | INFO | train_grpo_train | metrics_logged mode=train step=1050

                                                       
{'loss': -0.1394, 'grad_norm': 1.9295040369033813, 'learning_rate': 6.821212121212122e-06, 'num_tokens': 1856068.0, 'completions/mean_length': 172.25, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7838911414146423, 'rewards/meter/std': 0.37247487902641296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.946242094039917, 'rewards/repeat_soft/std': 0.04032229632139206, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.45562565326690674, 'rewards/total_composite/std': 0.2813253104686737, 'reward': 0.45562565326690674, 'reward_std': 0.2813253104686737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14240796864032745, 'sampling/sampling_logp_difference/max': 1.0745148658752441, 'sampling/importance_sampling_ratio/min': 0.3414633572101593, 'sampling/importance_sampling_ratio/mean': 1.0431674718856812, 'sampling/importance_sampling_ratio/max': 1.7962925434112549, 'entropy': 1.164476990699768, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09937345143407583, 'clip_ratio/high_max': 0.09937345143407583, 'clip_ratio/region_mean': 0.09937345143407583, 'reward_total_mean': 0.45562565326690674, 'reward_meter_mean': 0.7838911414146423, 'reward_meter_std': 0.37247487902641296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.946242094039917, 'reward_repeat_soft_std': 0.04032229632139206, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.45562565326690674, 'reward_total_composite_std': 0.2813253104686737, 'epoch': 0.11}

 32%|███▏      | 1050/3300 [2:37:11<6:14:18,  9.98s/it]INFO 04-13 09:58:09 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.76s/it]

 30%|███       | 3/10 [00:07<00:18,  2.57s/it]

 40%|████      | 4/10 [00:16<00:29,  4.92s/it]

 50%|█████     | 5/10 [00:19<00:22,  4.51s/it]

 60%|██████    | 6/10 [00:28<00:23,  5.96s/it]

 70%|███████   | 7/10 [00:32<00:15,  5.11s/it]

 80%|████████  | 8/10 [00:35<00:09,  4.52s/it]

 90%|█████████ | 9/10 [00:38<00:04,  4.21s/it]

100%|██████████| 10/10 [00:41<00:00,  3.82s/it]2026-04-13 09:59:00,133 | INFO | train_grpo_train | metrics_logged mode=eval step=1050

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 50.6596, 'eval_samples_per_second': 1.579, 'eval_steps_per_second': 0.197, 'eval_num_tokens': 1856068.0, 'eval_completions/mean_length': 95.9, 'eval_completions/min_length': 33.5, 'eval_completions/max_length': 216.0, 'eval_completions/clipped_ratio': 0.0875, 'eval_completions/mean_terminated_length': 57.025000381469724, 'eval_completions/min_terminated_length': 33.5, 'eval_completions/max_terminated_length': 83.9, 'eval_rewards/meter/mean': 0.6779823750257492, 'eval_rewards/meter/std': 0.3323016606271267, 'eval_rewards/count_adherence/mean': 0.8643750131130219, 'eval_rewards/count_adherence/std': 0.14014482572674752, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.13509859144687653, 'eval_rewards/repeat_soft/mean': 0.9586742341518402, 'eval_rewards/repeat_soft/std': 0.043604453839361665, 'eval_rewards/judge_quality/mean': 0.4231249958276749, 'eval_rewards/judge_quality/std': 0.139389075525105, 'eval_rewards/total_composite/mean': 0.48171039670705795, 'eval_rewards/total_composite/std': 0.1573707213625312, 'eval_reward': 0.48171039670705795, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.11468239948153496, 'eval_sampling/sampling_logp_difference/max': 1.0133841991424561, 'eval_sampling/importance_sampling_ratio/min': 0.38035872131586074, 'eval_sampling/importance_sampling_ratio/mean': 1.0362835764884948, 'eval_sampling/importance_sampling_ratio/max': 1.5530759930610656, 'eval_entropy': 1.5260615825653077, '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': 0.48171039670705795, 'eval_reward_meter_mean': 0.6779823750257492, 'eval_reward_meter_std': 0.3323016606271267, 'eval_reward_count_adherence_mean': 0.8643750131130219, 'eval_reward_count_adherence_std': 0.14014482572674752, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.13509859144687653, 'eval_reward_repeat_soft_mean': 0.9586742341518402, 'eval_reward_repeat_soft_std': 0.043604453839361665, 'eval_reward_judge_quality_mean': 0.4231249958276749, 'eval_reward_judge_quality_std': 0.139389075525105, 'eval_reward_total_composite_mean': 0.48171039670705795, 'eval_reward_total_composite_std': 0.1573707213625312, 'epoch': 0.11}

 32%|███▏      | 1050/3300 [2:38:02<6:14:18,  9.98s/it]

100%|██████████| 10/10 [00:41<00:00,  3.82s/it]

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

 32%|███▏      | 1051/3300 [2:38:16<16:31:10, 26.44s/it]2026-04-13 09:59:14,266 | INFO | train_grpo_train | metrics_logged mode=train step=1051

                                                        
{'loss': -0.0744, 'grad_norm': 2.019414186477661, 'learning_rate': 6.818181818181818e-06, 'num_tokens': 1857417.0, 'completions/mean_length': 83.625, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.428571701049805, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8272644281387329, 'rewards/meter/std': 0.19974714517593384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9372775554656982, 'rewards/repeat_soft/std': 0.04599282518029213, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.14327171444892883, 'rewards/total_composite/mean': 0.47729331254959106, 'rewards/total_composite/std': 0.20679502189159393, 'reward': 0.47729331254959106, 'reward_std': 0.20679503679275513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1740417629480362, 'sampling/sampling_logp_difference/max': 1.1143274307250977, 'sampling/importance_sampling_ratio/min': 0.3281359076499939, 'sampling/importance_sampling_ratio/mean': 1.0323822498321533, 'sampling/importance_sampling_ratio/max': 1.8156473636627197, 'entropy': 1.4743000343441963, 'clip_ratio/low_mean': 0.0243055559694767, 'clip_ratio/low_min': 0.0243055559694767, 'clip_ratio/high_mean': 0.10145257040858269, 'clip_ratio/high_max': 0.10145257040858269, 'clip_ratio/region_mean': 0.1257581263780594, 'reward_total_mean': 0.47729331254959106, 'reward_meter_mean': 0.8272644281387329, 'reward_meter_std': 0.19974714517593384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9372775554656982, 'reward_repeat_soft_std': 0.04599282518029213, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.14327171444892883, 'reward_total_composite_mean': 0.47729331254959106, 'reward_total_composite_std': 0.20679502189159393, 'epoch': 0.11}

 32%|███▏      | 1051/3300 [2:38:16<16:31:10, 26.44s/it]INFO 04-13 09:59:14 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1052/3300 [2:38:23<12:52:18, 20.61s/it]2026-04-13 09:59:21,287 | INFO | train_grpo_train | metrics_logged mode=train step=1052

                                                        
{'loss': 0.0115, 'grad_norm': 10.280098915100098, 'learning_rate': 6.8151515151515155e-06, 'num_tokens': 1859368.0, 'completions/mean_length': 66.875, 'completions/min_length': 59.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.9164806604385376, 'rewards/meter/std': 0.20353005826473236, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9280893206596375, 'rewards/repeat_soft/std': 0.04672547057271004, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4956486225128174, 'rewards/total_composite/std': 0.062345072627067566, 'reward': 0.4956486225128174, 'reward_std': 0.062345072627067566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1618267148733139, 'sampling/sampling_logp_difference/max': 1.3241872787475586, 'sampling/importance_sampling_ratio/min': 0.2660190761089325, 'sampling/importance_sampling_ratio/mean': 1.0189158916473389, 'sampling/importance_sampling_ratio/max': 1.9141427278518677, 'entropy': 1.711425855755806, 'clip_ratio/low_mean': 0.037193973548710346, 'clip_ratio/low_min': 0.037193973548710346, 'clip_ratio/high_mean': 0.12622408755123615, 'clip_ratio/high_max': 0.12622408755123615, 'clip_ratio/region_mean': 0.1634180610999465, 'reward_total_mean': 0.4956486225128174, 'reward_meter_mean': 0.9164806604385376, 'reward_meter_std': 0.20353005826473236, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9280893206596375, 'reward_repeat_soft_std': 0.04672547057271004, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4956486225128174, 'reward_total_composite_std': 0.062345072627067566, 'epoch': 0.11}

 32%|███▏      | 1052/3300 [2:38:23<12:52:18, 20.61s/it]INFO 04-13 09:59:21 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1053/3300 [2:38:29<10:10:58, 16.31s/it]2026-04-13 09:59:27,561 | INFO | train_grpo_train | metrics_logged mode=train step=1053

                                                        
{'loss': -0.0086, 'grad_norm': 10.646592140197754, 'learning_rate': 6.812121212121212e-06, 'num_tokens': 1861314.0, 'completions/mean_length': 64.25, 'completions/min_length': 58.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.25, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.7658095359802246, 'rewards/meter/std': 0.3318875730037689, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.96340012550354, 'rewards/repeat_soft/std': 0.017175309360027313, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.452426552772522, 'rewards/total_composite/std': 0.08626030385494232, 'reward': 0.452426552772522, 'reward_std': 0.08626029640436172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1762935072183609, 'sampling/sampling_logp_difference/max': 1.2531490325927734, 'sampling/importance_sampling_ratio/min': 0.28560400009155273, 'sampling/importance_sampling_ratio/mean': 1.0375665426254272, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7278798073530197, 'clip_ratio/low_mean': 0.06385537888854742, 'clip_ratio/low_min': 0.06385537888854742, 'clip_ratio/high_mean': 0.10123779810965061, 'clip_ratio/high_max': 0.10123779810965061, 'clip_ratio/region_mean': 0.16509317699819803, 'reward_total_mean': 0.452426552772522, 'reward_meter_mean': 0.7658095359802246, 'reward_meter_std': 0.3318875730037689, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.96340012550354, 'reward_repeat_soft_std': 0.017175309360027313, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.452426552772522, 'reward_total_composite_std': 0.08626030385494232, 'epoch': 0.11}

 32%|███▏      | 1053/3300 [2:38:29<10:10:58, 16.31s/it]INFO 04-13 09:59:27 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1054/3300 [2:38:40<9:12:00, 14.75s/it] 2026-04-13 09:59:38,649 | INFO | train_grpo_train | metrics_logged mode=train step=1054

                                                       
{'loss': -0.1456, 'grad_norm': 3.234853506088257, 'learning_rate': 6.80909090909091e-06, 'num_tokens': 1863005.0, 'completions/mean_length': 108.375, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.71428680419922, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9629600048065186, 'rewards/meter/std': 0.04580124840140343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9219887852668762, 'rewards/repeat_soft/std': 0.07961937040090561, 'rewards/judge_quality/mean': 0.4725000262260437, 'rewards/judge_quality/std': 0.24364788830280304, 'rewards/total_composite/mean': 0.5899137258529663, 'rewards/total_composite/std': 0.2688349187374115, 'reward': 0.5899137258529663, 'reward_std': 0.2688349187374115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1691655069589615, 'sampling/sampling_logp_difference/max': 1.1366252899169922, 'sampling/importance_sampling_ratio/min': 0.3507244884967804, 'sampling/importance_sampling_ratio/mean': 1.0380213260650635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2046373188495636, 'clip_ratio/low_mean': 0.029059450142085552, 'clip_ratio/low_min': 0.029059450142085552, 'clip_ratio/high_mean': 0.08967863582074642, 'clip_ratio/high_max': 0.08967863582074642, 'clip_ratio/region_mean': 0.11873808596283197, 'reward_total_mean': 0.5899137258529663, 'reward_meter_mean': 0.9629600048065186, 'reward_meter_std': 0.04580124840140343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9219887852668762, 'reward_repeat_soft_std': 0.07961937040090561, 'reward_judge_quality_mean': 0.4725000262260437, 'reward_judge_quality_std': 0.24364788830280304, 'reward_total_composite_mean': 0.5899137258529663, 'reward_total_composite_std': 0.2688349187374115, 'epoch': 0.11}

 32%|███▏      | 1054/3300 [2:38:40<9:12:00, 14.75s/it]INFO 04-13 09:59:38 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1055/3300 [2:38:46<7:34:24, 12.14s/it]2026-04-13 09:59:44,721 | INFO | train_grpo_train | metrics_logged mode=train step=1055

                                                       
{'loss': 0.1372, 'grad_norm': 20.88136100769043, 'learning_rate': 6.806060606060607e-06, 'num_tokens': 1864564.0, 'completions/mean_length': 34.875, 'completions/min_length': 30.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.875, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.5309699773788452, 'rewards/meter/std': 0.39381441473960876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.992468535900116, 'rewards/repeat_soft/std': 0.01208141166716814, 'rewards/judge_quality/mean': 0.6850000023841858, 'rewards/judge_quality/std': 0.21837061643600464, 'rewards/total_composite/mean': 0.5953447818756104, 'rewards/total_composite/std': 0.21855324506759644, 'reward': 0.5953447818756104, 'reward_std': 0.21855324506759644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1475285142660141, 'sampling/sampling_logp_difference/max': 1.246315360069275, 'sampling/importance_sampling_ratio/min': 0.28756242990493774, 'sampling/importance_sampling_ratio/mean': 1.0113372802734375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9731053039431572, 'clip_ratio/low_mean': 0.05729138432070613, 'clip_ratio/low_min': 0.05729138432070613, 'clip_ratio/high_mean': 0.06666379282251, 'clip_ratio/high_max': 0.06666379282251, 'clip_ratio/region_mean': 0.12395517714321613, 'reward_total_mean': 0.5953447818756104, 'reward_meter_mean': 0.5309699773788452, 'reward_meter_std': 0.39381441473960876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.992468535900116, 'reward_repeat_soft_std': 0.01208141166716814, 'reward_judge_quality_mean': 0.6850000023841858, 'reward_judge_quality_std': 0.21837061643600464, 'reward_total_composite_mean': 0.5953447818756104, 'reward_total_composite_std': 0.21855324506759644, 'epoch': 0.11}

 32%|███▏      | 1055/3300 [2:38:46<7:34:24, 12.14s/it]INFO 04-13 09:59:44 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1056/3300 [2:38:58<7:24:36, 11.89s/it]2026-04-13 09:59:56,046 | INFO | train_grpo_train | metrics_logged mode=train step=1056

                                                       
{'loss': -0.1162, 'grad_norm': 3.0492923259735107, 'learning_rate': 6.803030303030304e-06, 'num_tokens': 1866155.0, 'completions/mean_length': 103.875, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 45.57143020629883, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7495633363723755, 'rewards/meter/std': 0.3189528286457062, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9526692628860474, 'rewards/repeat_soft/std': 0.07355757802724838, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.3223324716091156, 'rewards/total_composite/mean': 0.5910525918006897, 'rewards/total_composite/std': 0.2746005058288574, 'reward': 0.5910525918006897, 'reward_std': 0.2746005058288574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14629122614860535, 'sampling/sampling_logp_difference/max': 1.2436683177947998, 'sampling/importance_sampling_ratio/min': 0.28832462430000305, 'sampling/importance_sampling_ratio/mean': 1.023461937904358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8375407457351685, 'clip_ratio/low_mean': 0.040999338030815125, 'clip_ratio/low_min': 0.040999338030815125, 'clip_ratio/high_mean': 0.0836129542440176, 'clip_ratio/high_max': 0.0836129542440176, 'clip_ratio/region_mean': 0.12461229227483273, 'reward_total_mean': 0.5910525918006897, 'reward_meter_mean': 0.7495633363723755, 'reward_meter_std': 0.3189528286457062, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9526692628860474, 'reward_repeat_soft_std': 0.07355757802724838, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.3223324716091156, 'reward_total_composite_mean': 0.5910525918006897, 'reward_total_composite_std': 0.2746005058288574, 'epoch': 0.11}

 32%|███▏      | 1056/3300 [2:38:58<7:24:36, 11.89s/it]INFO 04-13 09:59:56 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1057/3300 [2:39:04<6:24:55, 10.30s/it]2026-04-13 10:00:02,630 | INFO | train_grpo_train | metrics_logged mode=train step=1057

                                                       
{'loss': -0.0128, 'grad_norm': 11.504645347595215, 'learning_rate': 6.800000000000001e-06, 'num_tokens': 1867795.0, 'completions/mean_length': 42.0, 'completions/min_length': 38.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5639481544494629, 'rewards/meter/std': 0.3438780605792999, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9573047161102295, 'rewards/repeat_soft/std': 0.0400361493229866, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5008670091629028, 'rewards/total_composite/std': 0.09752810746431351, 'reward': 0.5008670091629028, 'reward_std': 0.0975281149148941, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13496150076389313, 'sampling/sampling_logp_difference/max': 2.762829303741455, 'sampling/importance_sampling_ratio/min': 0.06311295181512833, 'sampling/importance_sampling_ratio/mean': 1.0118708610534668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0667807087302208, 'clip_ratio/low_mean': 0.07114577945321798, 'clip_ratio/low_min': 0.07114577945321798, 'clip_ratio/high_mean': 0.0737397177144885, 'clip_ratio/high_max': 0.0737397177144885, 'clip_ratio/region_mean': 0.1448854971677065, 'reward_total_mean': 0.5008670091629028, 'reward_meter_mean': 0.5639481544494629, 'reward_meter_std': 0.3438780605792999, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9573047161102295, 'reward_repeat_soft_std': 0.0400361493229866, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5008670091629028, 'reward_total_composite_std': 0.09752810746431351, 'epoch': 0.11}

 32%|███▏      | 1057/3300 [2:39:04<6:24:55, 10.30s/it]INFO 04-13 10:00:02 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1058/3300 [2:39:12<5:52:17,  9.43s/it]2026-04-13 10:00:09,998 | INFO | train_grpo_train | metrics_logged mode=train step=1058

                                                       
{'loss': -0.0093, 'grad_norm': 10.222298622131348, 'learning_rate': 6.796969696969697e-06, 'num_tokens': 1870232.0, 'completions/mean_length': 77.625, 'completions/min_length': 59.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.7992756366729736, 'rewards/meter/std': 0.2670969069004059, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8280136585235596, 'rewards/repeat_soft/std': 0.061355601996183395, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.4604276418685913, 'rewards/total_composite/std': 0.09280714392662048, 'reward': 0.4604276418685913, 'reward_std': 0.0928071141242981, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13437646627426147, 'sampling/sampling_logp_difference/max': 1.5356802940368652, 'sampling/importance_sampling_ratio/min': 0.21530917286872864, 'sampling/importance_sampling_ratio/mean': 0.9962149262428284, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0224271193146706, 'clip_ratio/low_mean': 0.036021286621689796, 'clip_ratio/low_min': 0.036021286621689796, 'clip_ratio/high_mean': 0.08083718456327915, 'clip_ratio/high_max': 0.08083718456327915, 'clip_ratio/region_mean': 0.11685847118496895, 'reward_total_mean': 0.4604276418685913, 'reward_meter_mean': 0.7992756366729736, 'reward_meter_std': 0.2670969069004059, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8280136585235596, 'reward_repeat_soft_std': 0.061355601996183395, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.4604276418685913, 'reward_total_composite_std': 0.09280714392662048, 'epoch': 0.11}

 32%|███▏      | 1058/3300 [2:39:12<5:52:17,  9.43s/it]INFO 04-13 10:00:10 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1059/3300 [2:39:18<5:17:31,  8.50s/it]2026-04-13 10:00:16,337 | INFO | train_grpo_train | metrics_logged mode=train step=1059

                                                       
{'loss': -0.0395, 'grad_norm': 18.02076530456543, 'learning_rate': 6.793939393939395e-06, 'num_tokens': 1871661.0, 'completions/mean_length': 21.625, 'completions/min_length': 18.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7551445960998535, 'rewards/meter/std': 0.3369503915309906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9593343734741211, 'rewards/repeat_soft/std': 0.008953613229095936, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5632452964782715, 'rewards/total_composite/std': 0.08104057610034943, 'reward': 0.5632452964782715, 'reward_std': 0.08104058355093002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15181419253349304, 'sampling/sampling_logp_difference/max': 0.9656963348388672, 'sampling/importance_sampling_ratio/min': 0.38071802258491516, 'sampling/importance_sampling_ratio/mean': 1.0368188619613647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3528812900185585, 'clip_ratio/low_mean': 0.040570175275206566, 'clip_ratio/low_min': 0.040570175275206566, 'clip_ratio/high_mean': 0.10240816045552492, 'clip_ratio/high_max': 0.10240816045552492, 'clip_ratio/region_mean': 0.1429783357307315, 'reward_total_mean': 0.5632452964782715, 'reward_meter_mean': 0.7551445960998535, 'reward_meter_std': 0.3369503915309906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9593343734741211, 'reward_repeat_soft_std': 0.008953613229095936, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5632452964782715, 'reward_total_composite_std': 0.08104057610034943, 'epoch': 0.11}

 32%|███▏      | 1059/3300 [2:39:18<5:17:31,  8.50s/it]INFO 04-13 10:00:16 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1060/3300 [2:39:24<4:52:13,  7.83s/it]2026-04-13 10:00:22,592 | INFO | train_grpo_train | metrics_logged mode=train step=1060

                                                       
{'loss': 0.0055, 'grad_norm': 10.958192825317383, 'learning_rate': 6.790909090909091e-06, 'num_tokens': 1873690.0, 'completions/mean_length': 73.625, 'completions/min_length': 67.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.8756072521209717, 'rewards/meter/std': 0.17839808762073517, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9417939186096191, 'rewards/repeat_soft/std': 0.03722754493355751, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.5927358865737915, 'rewards/total_composite/std': 0.12903288006782532, 'reward': 0.5927358865737915, 'reward_std': 0.12903286516666412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1629772037267685, 'sampling/sampling_logp_difference/max': 3.0517749786376953, 'sampling/importance_sampling_ratio/min': 0.047274935990571976, 'sampling/importance_sampling_ratio/mean': 1.0284122228622437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.303021028637886, 'clip_ratio/low_mean': 0.09073344804346561, 'clip_ratio/low_min': 0.09073344804346561, 'clip_ratio/high_mean': 0.03935917653143406, 'clip_ratio/high_max': 0.03935917653143406, 'clip_ratio/region_mean': 0.13009262457489967, 'reward_total_mean': 0.5927358865737915, 'reward_meter_mean': 0.8756072521209717, 'reward_meter_std': 0.17839808762073517, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9417939186096191, 'reward_repeat_soft_std': 0.03722754493355751, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.5927358865737915, 'reward_total_composite_std': 0.12903288006782532, 'epoch': 0.11}

 32%|███▏      | 1060/3300 [2:39:24<4:52:13,  7.83s/it]INFO 04-13 10:00:22 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1061/3300 [2:39:30<4:31:04,  7.26s/it]2026-04-13 10:00:28,543 | INFO | train_grpo_train | metrics_logged mode=train step=1061

                                                       
{'loss': 0.0683, 'grad_norm': 12.164383888244629, 'learning_rate': 6.787878787878789e-06, 'num_tokens': 1875233.0, 'completions/mean_length': 43.875, 'completions/min_length': 34.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9468818306922913, 'rewards/meter/std': 0.04013519361615181, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9603518843650818, 'rewards/repeat_soft/std': 0.029111694544553757, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.7183948159217834, 'rewards/total_composite/std': 0.15400682389736176, 'reward': 0.7183948159217834, 'reward_std': 0.15400682389736176, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13824084401130676, 'sampling/sampling_logp_difference/max': 1.3404159545898438, 'sampling/importance_sampling_ratio/min': 0.26173678040504456, 'sampling/importance_sampling_ratio/mean': 1.019260287284851, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2784786894917488, 'clip_ratio/low_mean': 0.06544010294601321, 'clip_ratio/low_min': 0.06544010294601321, 'clip_ratio/high_mean': 0.024099512491375208, 'clip_ratio/high_max': 0.024099512491375208, 'clip_ratio/region_mean': 0.08953961543738842, 'reward_total_mean': 0.7183948159217834, 'reward_meter_mean': 0.9468818306922913, 'reward_meter_std': 0.04013519361615181, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9603518843650818, 'reward_repeat_soft_std': 0.029111694544553757, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.7183948159217834, 'reward_total_composite_std': 0.15400682389736176, 'epoch': 0.11}

 32%|███▏      | 1061/3300 [2:39:30<4:31:04,  7.26s/it]INFO 04-13 10:00:28 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1062/3300 [2:39:36<4:19:23,  6.95s/it]2026-04-13 10:00:34,774 | INFO | train_grpo_train | metrics_logged mode=train step=1062

                                                       
{'loss': 0.0408, 'grad_norm': 10.422037124633789, 'learning_rate': 6.7848484848484855e-06, 'num_tokens': 1877171.0, 'completions/mean_length': 67.25, 'completions/min_length': 61.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.971000075340271, 'rewards/meter/std': 0.0235135480761528, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9678769111633301, 'rewards/repeat_soft/std': 0.014561736956238747, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5703381896018982, 'rewards/total_composite/std': 0.08263786137104034, 'reward': 0.5703381896018982, 'reward_std': 0.08263785392045975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1390233039855957, 'sampling/sampling_logp_difference/max': 1.2323393821716309, 'sampling/importance_sampling_ratio/min': 0.29160958528518677, 'sampling/importance_sampling_ratio/mean': 1.012794852256775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1702293306589127, 'clip_ratio/low_mean': 0.10602645669132471, 'clip_ratio/low_min': 0.10602645669132471, 'clip_ratio/high_mean': 0.02459016442298889, 'clip_ratio/high_max': 0.02459016442298889, 'clip_ratio/region_mean': 0.1306166211143136, 'reward_total_mean': 0.5703381896018982, 'reward_meter_mean': 0.971000075340271, 'reward_meter_std': 0.0235135480761528, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9678769111633301, 'reward_repeat_soft_std': 0.014561736956238747, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5703381896018982, 'reward_total_composite_std': 0.08263786137104034, 'epoch': 0.11}

 32%|███▏      | 1062/3300 [2:39:37<4:19:23,  6.95s/it]INFO 04-13 10:00:35 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1063/3300 [2:39:43<4:13:02,  6.79s/it]2026-04-13 10:00:41,165 | INFO | train_grpo_train | metrics_logged mode=train step=1063

                                                       
{'loss': 0.0279, 'grad_norm': 13.100746154785156, 'learning_rate': 6.781818181818183e-06, 'num_tokens': 1879210.0, 'completions/mean_length': 73.875, 'completions/min_length': 65.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8981308937072754, 'rewards/meter/std': 0.23610873520374298, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9531311988830566, 'rewards/repeat_soft/std': 0.042662639170885086, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5381593704223633, 'rewards/total_composite/std': 0.06434082239866257, 'reward': 0.5381593704223633, 'reward_std': 0.06434082239866257, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15877306461334229, 'sampling/sampling_logp_difference/max': 2.6850342750549316, 'sampling/importance_sampling_ratio/min': 0.23416869342327118, 'sampling/importance_sampling_ratio/mean': 1.0295064449310303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2262151837348938, 'clip_ratio/low_mean': 0.036093054339289665, 'clip_ratio/low_min': 0.036093054339289665, 'clip_ratio/high_mean': 0.10448102559894323, 'clip_ratio/high_max': 0.10448102559894323, 'clip_ratio/region_mean': 0.1405740799382329, 'reward_total_mean': 0.5381593704223633, 'reward_meter_mean': 0.8981308937072754, 'reward_meter_std': 0.23610873520374298, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9531311988830566, 'reward_repeat_soft_std': 0.042662639170885086, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5381593704223633, 'reward_total_composite_std': 0.06434082239866257, 'epoch': 0.11}

 32%|███▏      | 1063/3300 [2:39:43<4:13:02,  6.79s/it]INFO 04-13 10:00:41 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1064/3300 [2:39:51<4:33:07,  7.33s/it]2026-04-13 10:00:49,759 | INFO | train_grpo_train | metrics_logged mode=train step=1064

                                                       
{'loss': -0.0478, 'grad_norm': 18.498001098632812, 'learning_rate': 6.778787878787879e-06, 'num_tokens': 1880747.0, 'completions/mean_length': 23.125, 'completions/min_length': 18.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9713336229324341, 'rewards/meter/std': 0.020421607419848442, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9412878751754761, 'rewards/repeat_soft/std': 0.04843747243285179, 'rewards/judge_quality/mean': 0.581250011920929, 'rewards/judge_quality/std': 0.2968134582042694, 'rewards/total_composite/mean': 0.7095893621444702, 'rewards/total_composite/std': 0.18525193631649017, 'reward': 0.7095893621444702, 'reward_std': 0.18525192141532898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19308802485466003, 'sampling/sampling_logp_difference/max': 1.3790791034698486, 'sampling/importance_sampling_ratio/min': 0.25181034207344055, 'sampling/importance_sampling_ratio/mean': 1.010938048362732, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4194799214601517, 'clip_ratio/low_mean': 0.0904420162551105, 'clip_ratio/low_min': 0.0904420162551105, 'clip_ratio/high_mean': 0.03762464504688978, 'clip_ratio/high_max': 0.03762464504688978, 'clip_ratio/region_mean': 0.12806666130200028, 'reward_total_mean': 0.7095893621444702, 'reward_meter_mean': 0.9713336229324341, 'reward_meter_std': 0.020421607419848442, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9412878751754761, 'reward_repeat_soft_std': 0.04843747243285179, 'reward_judge_quality_mean': 0.581250011920929, 'reward_judge_quality_std': 0.2968134582042694, 'reward_total_composite_mean': 0.7095893621444702, 'reward_total_composite_std': 0.18525193631649017, 'epoch': 0.11}

 32%|███▏      | 1064/3300 [2:39:51<4:33:07,  7.33s/it]INFO 04-13 10:00:49 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1065/3300 [2:39:57<4:18:16,  6.93s/it]2026-04-13 10:00:55,769 | INFO | train_grpo_train | metrics_logged mode=train step=1065

                                                       
{'loss': 0.107, 'grad_norm': 10.864692687988281, 'learning_rate': 6.7757575757575765e-06, 'num_tokens': 1882443.0, 'completions/mean_length': 45.0, 'completions/min_length': 36.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9339120388031006, 'rewards/meter/std': 0.1331321746110916, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9318022131919861, 'rewards/repeat_soft/std': 0.04915972426533699, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5836708545684814, 'rewards/total_composite/std': 0.0491509884595871, 'reward': 0.5836708545684814, 'reward_std': 0.049150995910167694, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16631630063056946, 'sampling/sampling_logp_difference/max': 1.5867557525634766, 'sampling/importance_sampling_ratio/min': 0.20458826422691345, 'sampling/importance_sampling_ratio/mean': 1.01224684715271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.440430074930191, 'clip_ratio/low_mean': 0.04046474304050207, 'clip_ratio/low_min': 0.04046474304050207, 'clip_ratio/high_mean': 0.09686570474877954, 'clip_ratio/high_max': 0.09686570474877954, 'clip_ratio/region_mean': 0.1373304477892816, 'reward_total_mean': 0.5836708545684814, 'reward_meter_mean': 0.9339120388031006, 'reward_meter_std': 0.1331321746110916, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9318022131919861, 'reward_repeat_soft_std': 0.04915972426533699, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5836708545684814, 'reward_total_composite_std': 0.0491509884595871, 'epoch': 0.11}

 32%|███▏      | 1065/3300 [2:39:57<4:18:16,  6.93s/it]INFO 04-13 10:00:55 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1066/3300 [2:40:12<5:46:40,  9.31s/it]2026-04-13 10:01:10,630 | INFO | train_grpo_train | metrics_logged mode=train step=1066

                                                       
{'loss': -0.1539, 'grad_norm': 2.0058774948120117, 'learning_rate': 6.772727272727273e-06, 'num_tokens': 1884430.0, 'completions/mean_length': 119.375, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.28571701049805, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9014915227890015, 'rewards/meter/std': 0.13286340236663818, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9050803184509277, 'rewards/repeat_soft/std': 0.05679691955447197, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.4573853015899658, 'rewards/total_composite/std': 0.19074977934360504, 'reward': 0.4573853015899658, 'reward_std': 0.19074977934360504, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11312473565340042, 'sampling/sampling_logp_difference/max': 1.3569793701171875, 'sampling/importance_sampling_ratio/min': 0.25743722915649414, 'sampling/importance_sampling_ratio/mean': 1.0099263191223145, 'sampling/importance_sampling_ratio/max': 1.7764612436294556, 'entropy': 0.8119281008839607, 'clip_ratio/low_mean': 0.009803921915590763, 'clip_ratio/low_min': 0.009803921915590763, 'clip_ratio/high_mean': 0.07999108778312802, 'clip_ratio/high_max': 0.07999108778312802, 'clip_ratio/region_mean': 0.08979500969871879, 'reward_total_mean': 0.4573853015899658, 'reward_meter_mean': 0.9014915227890015, 'reward_meter_std': 0.13286340236663818, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9050803184509277, 'reward_repeat_soft_std': 0.05679691955447197, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.4573853015899658, 'reward_total_composite_std': 0.19074977934360504, 'epoch': 0.11}

 32%|███▏      | 1066/3300 [2:40:12<5:46:40,  9.31s/it]INFO 04-13 10:01:10 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1067/3300 [2:40:18<5:08:03,  8.28s/it]2026-04-13 10:01:16,496 | INFO | train_grpo_train | metrics_logged mode=train step=1067

                                                       
{'loss': 0.0287, 'grad_norm': 16.06044578552246, 'learning_rate': 6.76969696969697e-06, 'num_tokens': 1885995.0, 'completions/mean_length': 36.625, 'completions/min_length': 32.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8607951402664185, 'rewards/meter/std': 0.34211716055870056, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9544247984886169, 'rewards/repeat_soft/std': 0.0466926284134388, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.22984081506729126, 'rewards/total_composite/mean': 0.6810033321380615, 'rewards/total_composite/std': 0.19374404847621918, 'reward': 0.6810033321380615, 'reward_std': 0.19374403357505798, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15902628004550934, 'sampling/sampling_logp_difference/max': 2.501688003540039, 'sampling/importance_sampling_ratio/min': 0.08194655179977417, 'sampling/importance_sampling_ratio/mean': 0.9926603436470032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8786809593439102, 'clip_ratio/low_mean': 0.1110491082072258, 'clip_ratio/low_min': 0.1110491082072258, 'clip_ratio/high_mean': 0.06022810656577349, 'clip_ratio/high_max': 0.06022810656577349, 'clip_ratio/region_mean': 0.1712772147729993, 'reward_total_mean': 0.6810033321380615, 'reward_meter_mean': 0.8607951402664185, 'reward_meter_std': 0.34211716055870056, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9544247984886169, 'reward_repeat_soft_std': 0.0466926284134388, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.22984081506729126, 'reward_total_composite_mean': 0.6810033321380615, 'reward_total_composite_std': 0.19374404847621918, 'epoch': 0.11}

 32%|███▏      | 1067/3300 [2:40:18<5:08:03,  8.28s/it]INFO 04-13 10:01:16 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1068/3300 [2:40:25<4:47:23,  7.73s/it]2026-04-13 10:01:22,937 | INFO | train_grpo_train | metrics_logged mode=train step=1068

                                                       
{'loss': -0.023, 'grad_norm': 12.230772972106934, 'learning_rate': 6.7666666666666665e-06, 'num_tokens': 1887856.0, 'completions/mean_length': 72.625, 'completions/min_length': 63.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9264994859695435, 'rewards/meter/std': 0.0505947582423687, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9661044478416443, 'rewards/repeat_soft/std': 0.01599237136542797, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5478500127792358, 'rewards/total_composite/std': 0.014911556616425514, 'reward': 0.5478500127792358, 'reward_std': 0.014911546371877193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12288069725036621, 'sampling/sampling_logp_difference/max': 0.9862861633300781, 'sampling/importance_sampling_ratio/min': 0.3729592263698578, 'sampling/importance_sampling_ratio/mean': 1.034889817237854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0377146005630493, 'clip_ratio/low_mean': 0.0284205237403512, 'clip_ratio/low_min': 0.0284205237403512, 'clip_ratio/high_mean': 0.08064270671457052, 'clip_ratio/high_max': 0.08064270671457052, 'clip_ratio/region_mean': 0.10906323045492172, 'reward_total_mean': 0.5478500127792358, 'reward_meter_mean': 0.9264994859695435, 'reward_meter_std': 0.0505947582423687, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9661044478416443, 'reward_repeat_soft_std': 0.01599237136542797, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5478500127792358, 'reward_total_composite_std': 0.014911556616425514, 'epoch': 0.11}

 32%|███▏      | 1068/3300 [2:40:25<4:47:23,  7.73s/it]INFO 04-13 10:01:23 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1069/3300 [2:40:31<4:27:16,  7.19s/it]2026-04-13 10:01:28,872 | INFO | train_grpo_train | metrics_logged mode=train step=1069

                                                       
{'loss': 0.001, 'grad_norm': 14.955061912536621, 'learning_rate': 6.763636363636365e-06, 'num_tokens': 1889477.0, 'completions/mean_length': 47.625, 'completions/min_length': 43.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7776228189468384, 'rewards/meter/std': 0.27847403287887573, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9883979558944702, 'rewards/repeat_soft/std': 0.013899828307330608, 'rewards/judge_quality/mean': 0.8050000071525574, 'rewards/judge_quality/std': 0.16860775649547577, 'rewards/total_composite/mean': 0.7516942620277405, 'rewards/total_composite/std': 0.16910621523857117, 'reward': 0.7516942620277405, 'reward_std': 0.16910621523857117, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11631647497415543, 'sampling/sampling_logp_difference/max': 2.1365227699279785, 'sampling/importance_sampling_ratio/min': 0.11806467175483704, 'sampling/importance_sampling_ratio/mean': 1.0089021921157837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5627031028270721, 'clip_ratio/low_mean': 0.032379837008193135, 'clip_ratio/low_min': 0.032379837008193135, 'clip_ratio/high_mean': 0.08231146773323417, 'clip_ratio/high_max': 0.08231146773323417, 'clip_ratio/region_mean': 0.1146913047414273, 'reward_total_mean': 0.7516942620277405, 'reward_meter_mean': 0.7776228189468384, 'reward_meter_std': 0.27847403287887573, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9883979558944702, 'reward_repeat_soft_std': 0.013899828307330608, 'reward_judge_quality_mean': 0.8050000071525574, 'reward_judge_quality_std': 0.16860775649547577, 'reward_total_composite_mean': 0.7516942620277405, 'reward_total_composite_std': 0.16910621523857117, 'epoch': 0.11}

 32%|███▏      | 1069/3300 [2:40:31<4:27:16,  7.19s/it]INFO 04-13 10:01:29 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1070/3300 [2:40:37<4:13:40,  6.83s/it]2026-04-13 10:01:34,845 | INFO | train_grpo_train | metrics_logged mode=train step=1070

                                                       
{'loss': 0.0614, 'grad_norm': 11.298134803771973, 'learning_rate': 6.760606060606061e-06, 'num_tokens': 1891061.0, 'completions/mean_length': 41.0, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9851936101913452, 'rewards/meter/std': 0.007341459859162569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9215592741966248, 'rewards/repeat_soft/std': 0.05847397819161415, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6871336698532104, 'rewards/total_composite/std': 0.14895614981651306, 'reward': 0.6871336698532104, 'reward_std': 0.14895614981651306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1290155053138733, 'sampling/sampling_logp_difference/max': 1.2429022789001465, 'sampling/importance_sampling_ratio/min': 0.2885455787181854, 'sampling/importance_sampling_ratio/mean': 1.0117584466934204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8554912656545639, 'clip_ratio/low_mean': 0.0737000578083098, 'clip_ratio/low_min': 0.0737000578083098, 'clip_ratio/high_mean': 0.03482142882421613, 'clip_ratio/high_max': 0.03482142882421613, 'clip_ratio/region_mean': 0.10852148663252592, 'reward_total_mean': 0.6871336698532104, 'reward_meter_mean': 0.9851936101913452, 'reward_meter_std': 0.007341459859162569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9215592741966248, 'reward_repeat_soft_std': 0.05847397819161415, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6871336698532104, 'reward_total_composite_std': 0.14895614981651306, 'epoch': 0.11}

 32%|███▏      | 1070/3300 [2:40:37<4:13:40,  6.83s/it]INFO 04-13 10:01:35 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1071/3300 [2:40:43<4:10:02,  6.73s/it]2026-04-13 10:01:41,357 | INFO | train_grpo_train | metrics_logged mode=train step=1071

                                                       
{'loss': 0.0196, 'grad_norm': 10.647320747375488, 'learning_rate': 6.757575757575758e-06, 'num_tokens': 1893169.0, 'completions/mean_length': 67.5, 'completions/min_length': 59.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.5, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9314833283424377, 'rewards/meter/std': 0.09214688092470169, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9345080852508545, 'rewards/repeat_soft/std': 0.05088561773300171, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.4987291395664215, 'rewards/total_composite/std': 0.06691356748342514, 'reward': 0.4987291395664215, 'reward_std': 0.06691356748342514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14687930047512054, 'sampling/sampling_logp_difference/max': 2.208160877227783, 'sampling/importance_sampling_ratio/min': 0.10990258306264877, 'sampling/importance_sampling_ratio/mean': 1.0051355361938477, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2108322158455849, 'clip_ratio/low_mean': 0.06116915214806795, 'clip_ratio/low_min': 0.06116915214806795, 'clip_ratio/high_mean': 0.07562159188091755, 'clip_ratio/high_max': 0.07562159188091755, 'clip_ratio/region_mean': 0.1367907440289855, 'reward_total_mean': 0.4987291395664215, 'reward_meter_mean': 0.9314833283424377, 'reward_meter_std': 0.09214688092470169, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9345080852508545, 'reward_repeat_soft_std': 0.05088561773300171, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.4987291395664215, 'reward_total_composite_std': 0.06691356748342514, 'epoch': 0.11}

 32%|███▏      | 1071/3300 [2:40:43<4:10:02,  6.73s/it]INFO 04-13 10:01:41 [block_pool.py:378] Successfully reset prefix cache

 32%|███▏      | 1072/3300 [2:40:49<4:04:01,  6.57s/it]2026-04-13 10:01:47,556 | INFO | train_grpo_train | metrics_logged mode=train step=1072

                                                       
{'loss': 0.0086, 'grad_norm': 16.221986770629883, 'learning_rate': 6.754545454545455e-06, 'num_tokens': 1894673.0, 'completions/mean_length': 44.0, 'completions/min_length': 42.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8085959553718567, 'rewards/meter/std': 0.262198805809021, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9473362565040588, 'rewards/repeat_soft/std': 0.04573459178209305, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5633194446563721, 'rewards/total_composite/std': 0.06888439506292343, 'reward': 0.5633194446563721, 'reward_std': 0.06888439506292343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11900659650564194, 'sampling/sampling_logp_difference/max': 0.989473819732666, 'sampling/importance_sampling_ratio/min': 0.4080321788787842, 'sampling/importance_sampling_ratio/mean': 1.03138267993927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0347795113921165, 'clip_ratio/low_mean': 0.05752122774720192, 'clip_ratio/low_min': 0.05752122774720192, 'clip_ratio/high_mean': 0.06697582267224789, 'clip_ratio/high_max': 0.06697582267224789, 'clip_ratio/region_mean': 0.1244970504194498, 'reward_total_mean': 0.5633194446563721, 'reward_meter_mean': 0.8085959553718567, 'reward_meter_std': 0.262198805809021, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9473362565040588, 'reward_repeat_soft_std': 0.04573459178209305, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5633194446563721, 'reward_total_composite_std': 0.06888439506292343, 'epoch': 0.11}

 32%|███▏      | 1072/3300 [2:40:49<4:04:01,  6.57s/it]INFO 04-13 10:01:47 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1073/3300 [2:40:55<3:59:02,  6.44s/it]2026-04-13 10:01:53,691 | INFO | train_grpo_train | metrics_logged mode=train step=1073

                                                       
{'loss': 0.0238, 'grad_norm': 14.687736511230469, 'learning_rate': 6.751515151515152e-06, 'num_tokens': 1896182.0, 'completions/mean_length': 42.625, 'completions/min_length': 32.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.870096743106842, 'rewards/meter/std': 0.1931677907705307, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.984889805316925, 'rewards/repeat_soft/std': 0.010807708837091923, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5965648293495178, 'rewards/total_composite/std': 0.05712573602795601, 'reward': 0.5965648293495178, 'reward_std': 0.05712572857737541, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16201752424240112, 'sampling/sampling_logp_difference/max': 1.2331657409667969, 'sampling/importance_sampling_ratio/min': 0.2913687229156494, 'sampling/importance_sampling_ratio/mean': 1.0427515506744385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.352506972849369, 'clip_ratio/low_mean': 0.06444899551570415, 'clip_ratio/low_min': 0.06444899551570415, 'clip_ratio/high_mean': 0.08135287696495652, 'clip_ratio/high_max': 0.08135287696495652, 'clip_ratio/region_mean': 0.14580187248066068, 'reward_total_mean': 0.5965648293495178, 'reward_meter_mean': 0.870096743106842, 'reward_meter_std': 0.1931677907705307, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.984889805316925, 'reward_repeat_soft_std': 0.010807708837091923, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5965648293495178, 'reward_total_composite_std': 0.05712573602795601, 'epoch': 0.11}

 33%|███▎      | 1073/3300 [2:40:55<3:59:02,  6.44s/it]INFO 04-13 10:01:53 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1074/3300 [2:41:02<3:56:18,  6.37s/it]2026-04-13 10:01:59,899 | INFO | train_grpo_train | metrics_logged mode=train step=1074

                                                       
{'loss': 0.0522, 'grad_norm': 12.577512741088867, 'learning_rate': 6.748484848484848e-06, 'num_tokens': 1897835.0, 'completions/mean_length': 46.625, 'completions/min_length': 35.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6758618354797363, 'rewards/meter/std': 0.2913423776626587, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9415541291236877, 'rewards/repeat_soft/std': 0.04388896003365517, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.22890658676624298, 'rewards/total_composite/mean': 0.6123546361923218, 'rewards/total_composite/std': 0.1519799530506134, 'reward': 0.6123546361923218, 'reward_std': 0.1519799530506134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12512366473674774, 'sampling/sampling_logp_difference/max': 1.4607925415039062, 'sampling/importance_sampling_ratio/min': 0.2320522964000702, 'sampling/importance_sampling_ratio/mean': 1.0119976997375488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6220041923224926, 'clip_ratio/low_mean': 0.07702832575887442, 'clip_ratio/low_min': 0.07702832575887442, 'clip_ratio/high_mean': 0.059330222196877, 'clip_ratio/high_max': 0.059330222196877, 'clip_ratio/region_mean': 0.13635854795575142, 'reward_total_mean': 0.6123546361923218, 'reward_meter_mean': 0.6758618354797363, 'reward_meter_std': 0.2913423776626587, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9415541291236877, 'reward_repeat_soft_std': 0.04388896003365517, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.22890658676624298, 'reward_total_composite_mean': 0.6123546361923218, 'reward_total_composite_std': 0.1519799530506134, 'epoch': 0.11}

 33%|███▎      | 1074/3300 [2:41:02<3:56:18,  6.37s/it]INFO 04-13 10:02:00 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1075/3300 [2:41:08<3:53:23,  6.29s/it]2026-04-13 10:02:06,015 | INFO | train_grpo_train | metrics_logged mode=train step=1075

                                                       
{'loss': 0.0121, 'grad_norm': 27.427839279174805, 'learning_rate': 6.7454545454545465e-06, 'num_tokens': 1899412.0, 'completions/mean_length': 20.125, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8380264043807983, 'rewards/meter/std': 0.2947549521923065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8501101732254028, 'rewards/repeat_soft/std': 0.07873211055994034, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.5350409746170044, 'rewards/total_composite/std': 0.08704709261655807, 'reward': 0.5350409746170044, 'reward_std': 0.08704709261655807, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14486834406852722, 'sampling/sampling_logp_difference/max': 1.1462392807006836, 'sampling/importance_sampling_ratio/min': 0.31782978773117065, 'sampling/importance_sampling_ratio/mean': 1.034041166305542, 'sampling/importance_sampling_ratio/max': 1.9090057611465454, 'entropy': 1.1850376725196838, 'clip_ratio/low_mean': 0.03128482960164547, 'clip_ratio/low_min': 0.03128482960164547, 'clip_ratio/high_mean': 0.08520944183692336, 'clip_ratio/high_max': 0.08520944183692336, 'clip_ratio/region_mean': 0.11649427143856883, 'reward_total_mean': 0.5350409746170044, 'reward_meter_mean': 0.8380264043807983, 'reward_meter_std': 0.2947549521923065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8501101732254028, 'reward_repeat_soft_std': 0.07873211055994034, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.5350409746170044, 'reward_total_composite_std': 0.08704709261655807, 'epoch': 0.11}

 33%|███▎      | 1075/3300 [2:41:08<3:53:23,  6.29s/it]INFO 04-13 10:02:06 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1076/3300 [2:41:14<3:53:06,  6.29s/it]2026-04-13 10:02:12,289 | INFO | train_grpo_train | metrics_logged mode=train step=1076

                                                       
{'loss': 0.0427, 'grad_norm': 12.953285217285156, 'learning_rate': 6.742424242424243e-06, 'num_tokens': 1900961.0, 'completions/mean_length': 41.625, 'completions/min_length': 32.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.9870993494987488, 'rewards/meter/std': 0.0081932432949543, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.932524561882019, 'rewards/repeat_soft/std': 0.06071160361170769, 'rewards/judge_quality/mean': 0.38875001668930054, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.5891315340995789, 'rewards/total_composite/std': 0.05717764049768448, 'reward': 0.5891315340995789, 'reward_std': 0.05717764422297478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1406351923942566, 'sampling/sampling_logp_difference/max': 1.5530462265014648, 'sampling/importance_sampling_ratio/min': 0.21160240471363068, 'sampling/importance_sampling_ratio/mean': 1.0404138565063477, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1810264438390732, 'clip_ratio/low_mean': 0.03371212258934975, 'clip_ratio/low_min': 0.03371212258934975, 'clip_ratio/high_mean': 0.10204389691352844, 'clip_ratio/high_max': 0.10204389691352844, 'clip_ratio/region_mean': 0.1357560195028782, 'reward_total_mean': 0.5891315340995789, 'reward_meter_mean': 0.9870993494987488, 'reward_meter_std': 0.0081932432949543, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.932524561882019, 'reward_repeat_soft_std': 0.06071160361170769, 'reward_judge_quality_mean': 0.38875001668930054, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.5891315340995789, 'reward_total_composite_std': 0.05717764049768448, 'epoch': 0.11}

 33%|███▎      | 1076/3300 [2:41:14<3:53:06,  6.29s/it]INFO 04-13 10:02:12 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1077/3300 [2:41:26<4:59:40,  8.09s/it]2026-04-13 10:02:24,583 | INFO | train_grpo_train | metrics_logged mode=train step=1077

                                                       
{'loss': -0.1388, 'grad_norm': 2.1381711959838867, 'learning_rate': 6.73939393939394e-06, 'num_tokens': 1902642.0, 'completions/mean_length': 114.125, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.28571701049805, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9222038984298706, 'rewards/meter/std': 0.1504916250705719, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.976231575012207, 'rewards/repeat_soft/std': 0.02974354662001133, 'rewards/judge_quality/mean': 0.4150000214576721, 'rewards/judge_quality/std': 0.18031719326972961, 'rewards/total_composite/mean': 0.5495796799659729, 'rewards/total_composite/std': 0.23759342730045319, 'reward': 0.5495796799659729, 'reward_std': 0.23759342730045319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1703491061925888, 'sampling/sampling_logp_difference/max': 1.1962089538574219, 'sampling/importance_sampling_ratio/min': 0.30233821272850037, 'sampling/importance_sampling_ratio/mean': 1.0201829671859741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4231328517198563, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.14230037666857243, 'clip_ratio/high_max': 0.14230037666857243, 'clip_ratio/region_mean': 0.16015752032399178, 'reward_total_mean': 0.5495796799659729, 'reward_meter_mean': 0.9222038984298706, 'reward_meter_std': 0.1504916250705719, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.976231575012207, 'reward_repeat_soft_std': 0.02974354662001133, 'reward_judge_quality_mean': 0.4150000214576721, 'reward_judge_quality_std': 0.18031719326972961, 'reward_total_composite_mean': 0.5495796799659729, 'reward_total_composite_std': 0.23759342730045319, 'epoch': 0.11}

 33%|███▎      | 1077/3300 [2:41:26<4:59:40,  8.09s/it]INFO 04-13 10:02:24 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1078/3300 [2:41:37<5:32:43,  8.98s/it]2026-04-13 10:02:35,658 | INFO | train_grpo_train | metrics_logged mode=train step=1078

                                                       
{'loss': -0.1107, 'grad_norm': 3.0935230255126953, 'learning_rate': 6.7363636363636365e-06, 'num_tokens': 1904408.0, 'completions/mean_length': 102.75, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.28571701049805, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8180112838745117, 'rewards/meter/std': 0.3261679708957672, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9903794527053833, 'rewards/repeat_soft/std': 0.008472509682178497, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.21625052392482758, 'rewards/total_composite/mean': 0.5496870875358582, 'rewards/total_composite/std': 0.2644262909889221, 'reward': 0.5496870875358582, 'reward_std': 0.2644262909889221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1494397521018982, 'sampling/sampling_logp_difference/max': 1.6385045051574707, 'sampling/importance_sampling_ratio/min': 0.19427035748958588, 'sampling/importance_sampling_ratio/mean': 1.0164092779159546, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.141548976302147, 'clip_ratio/low_mean': 0.016025641933083534, 'clip_ratio/low_min': 0.016025641933083534, 'clip_ratio/high_mean': 0.10979404021054506, 'clip_ratio/high_max': 0.10979404021054506, 'clip_ratio/region_mean': 0.1258196821436286, 'reward_total_mean': 0.5496870875358582, 'reward_meter_mean': 0.8180112838745117, 'reward_meter_std': 0.3261679708957672, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9903794527053833, 'reward_repeat_soft_std': 0.008472509682178497, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.21625052392482758, 'reward_total_composite_mean': 0.5496870875358582, 'reward_total_composite_std': 0.2644262909889221, 'epoch': 0.11}

 33%|███▎      | 1078/3300 [2:41:37<5:32:43,  8.98s/it]INFO 04-13 10:02:35 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1079/3300 [2:41:44<5:06:36,  8.28s/it]2026-04-13 10:02:42,300 | INFO | train_grpo_train | metrics_logged mode=train step=1079

                                                       
{'loss': 0.0086, 'grad_norm': 10.065486907958984, 'learning_rate': 6.733333333333334e-06, 'num_tokens': 1906543.0, 'completions/mean_length': 85.875, 'completions/min_length': 60.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.7328099608421326, 'rewards/meter/std': 0.28351694345474243, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9198435544967651, 'rewards/repeat_soft/std': 0.05026472732424736, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.504819929599762, 'rewards/total_composite/std': 0.12204664200544357, 'reward': 0.504819929599762, 'reward_std': 0.12204664945602417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14232237637043, 'sampling/sampling_logp_difference/max': 1.5423803329467773, 'sampling/importance_sampling_ratio/min': 0.21387141942977905, 'sampling/importance_sampling_ratio/mean': 1.0154612064361572, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1735976114869118, 'clip_ratio/low_mean': 0.06142867170274258, 'clip_ratio/low_min': 0.06142867170274258, 'clip_ratio/high_mean': 0.05976541340351105, 'clip_ratio/high_max': 0.05976541340351105, 'clip_ratio/region_mean': 0.12119408510625362, 'reward_total_mean': 0.504819929599762, 'reward_meter_mean': 0.7328099608421326, 'reward_meter_std': 0.28351694345474243, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9198435544967651, 'reward_repeat_soft_std': 0.05026472732424736, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.504819929599762, 'reward_total_composite_std': 0.12204664200544357, 'epoch': 0.11}

 33%|███▎      | 1079/3300 [2:41:44<5:06:36,  8.28s/it]INFO 04-13 10:02:42 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1080/3300 [2:41:50<4:40:14,  7.57s/it]2026-04-13 10:02:48,222 | INFO | train_grpo_train | metrics_logged mode=train step=1080

                                                       
{'loss': 0.0736, 'grad_norm': 16.76507568359375, 'learning_rate': 6.73030303030303e-06, 'num_tokens': 1907997.0, 'completions/mean_length': 37.75, 'completions/min_length': 35.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7624353170394897, 'rewards/meter/std': 0.41185691952705383, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9511844515800476, 'rewards/repeat_soft/std': 0.039821553975343704, 'rewards/judge_quality/mean': 0.8737500309944153, 'rewards/judge_quality/std': 0.08601287752389908, 'rewards/total_composite/mean': 0.7842920422554016, 'rewards/total_composite/std': 0.24220286309719086, 'reward': 0.7842920422554016, 'reward_std': 0.24220287799835205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13939060270786285, 'sampling/sampling_logp_difference/max': 1.7247705459594727, 'sampling/importance_sampling_ratio/min': 0.1782139390707016, 'sampling/importance_sampling_ratio/mean': 1.0068621635437012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9707998856902122, 'clip_ratio/low_mean': 0.0301974443718791, 'clip_ratio/low_min': 0.0301974443718791, 'clip_ratio/high_mean': 0.09268568316474557, 'clip_ratio/high_max': 0.09268568316474557, 'clip_ratio/region_mean': 0.12288312753662467, 'reward_total_mean': 0.7842920422554016, 'reward_meter_mean': 0.7624353170394897, 'reward_meter_std': 0.41185691952705383, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9511844515800476, 'reward_repeat_soft_std': 0.039821553975343704, 'reward_judge_quality_mean': 0.8737500309944153, 'reward_judge_quality_std': 0.08601287752389908, 'reward_total_composite_mean': 0.7842920422554016, 'reward_total_composite_std': 0.24220286309719086, 'epoch': 0.11}

 33%|███▎      | 1080/3300 [2:41:50<4:40:14,  7.57s/it]INFO 04-13 10:02:48 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1081/3300 [2:41:56<4:29:03,  7.27s/it]2026-04-13 10:02:54,797 | INFO | train_grpo_train | metrics_logged mode=train step=1081

                                                       
{'loss': 0.1389, 'grad_norm': 10.372480392456055, 'learning_rate': 6.7272727272727275e-06, 'num_tokens': 1910092.0, 'completions/mean_length': 80.875, 'completions/min_length': 62.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.767886757850647, 'rewards/meter/std': 0.24709437787532806, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9749447107315063, 'rewards/repeat_soft/std': 0.014261425472795963, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.47303062677383423, 'rewards/total_composite/std': 0.06528296321630478, 'reward': 0.47303062677383423, 'reward_std': 0.06528296321630478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1676093339920044, 'sampling/sampling_logp_difference/max': 2.04561710357666, 'sampling/importance_sampling_ratio/min': 0.12930037081241608, 'sampling/importance_sampling_ratio/mean': 1.0062720775604248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4282553941011429, 'clip_ratio/low_mean': 0.06684344448149204, 'clip_ratio/low_min': 0.06684344448149204, 'clip_ratio/high_mean': 0.07917457446455956, 'clip_ratio/high_max': 0.07917457446455956, 'clip_ratio/region_mean': 0.1460180189460516, 'reward_total_mean': 0.47303062677383423, 'reward_meter_mean': 0.767886757850647, 'reward_meter_std': 0.24709437787532806, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9749447107315063, 'reward_repeat_soft_std': 0.014261425472795963, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.47303062677383423, 'reward_total_composite_std': 0.06528296321630478, 'epoch': 0.11}

 33%|███▎      | 1081/3300 [2:41:57<4:29:03,  7.27s/it]INFO 04-13 10:02:55 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1082/3300 [2:42:03<4:15:55,  6.92s/it]2026-04-13 10:03:00,897 | INFO | train_grpo_train | metrics_logged mode=train step=1082

                                                       
{'loss': 0.0677, 'grad_norm': 16.086347579956055, 'learning_rate': 6.724242424242424e-06, 'num_tokens': 1911612.0, 'completions/mean_length': 42.0, 'completions/min_length': 35.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5908030271530151, 'rewards/meter/std': 0.391478568315506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9693589806556702, 'rewards/repeat_soft/std': 0.029551459476351738, 'rewards/judge_quality/mean': 0.6137499809265137, 'rewards/judge_quality/std': 0.1927943229675293, 'rewards/total_composite/mean': 0.5864404439926147, 'rewards/total_composite/std': 0.1658206284046173, 'reward': 0.5864404439926147, 'reward_std': 0.1658206284046173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16322758793830872, 'sampling/sampling_logp_difference/max': 1.5092816352844238, 'sampling/importance_sampling_ratio/min': 0.22106872498989105, 'sampling/importance_sampling_ratio/mean': 1.0038870573043823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0525820776820183, 'clip_ratio/low_mean': 0.059773560613393784, 'clip_ratio/low_min': 0.059773560613393784, 'clip_ratio/high_mean': 0.08948134444653988, 'clip_ratio/high_max': 0.08948134444653988, 'clip_ratio/region_mean': 0.14925490505993366, 'reward_total_mean': 0.5864404439926147, 'reward_meter_mean': 0.5908030271530151, 'reward_meter_std': 0.391478568315506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9693589806556702, 'reward_repeat_soft_std': 0.029551459476351738, 'reward_judge_quality_mean': 0.6137499809265137, 'reward_judge_quality_std': 0.1927943229675293, 'reward_total_composite_mean': 0.5864404439926147, 'reward_total_composite_std': 0.1658206284046173, 'epoch': 0.11}

 33%|███▎      | 1082/3300 [2:42:03<4:15:55,  6.92s/it]INFO 04-13 10:03:01 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1083/3300 [2:42:10<4:16:25,  6.94s/it]2026-04-13 10:03:07,880 | INFO | train_grpo_train | metrics_logged mode=train step=1083

                                                       
{'loss': 0.0562, 'grad_norm': 9.393478393554688, 'learning_rate': 6.721212121212122e-06, 'num_tokens': 1913940.0, 'completions/mean_length': 89.0, 'completions/min_length': 79.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.0, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9875081777572632, 'rewards/meter/std': 0.012425047345459461, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9484192132949829, 'rewards/repeat_soft/std': 0.041187919676303864, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6330901980400085, 'rewards/total_composite/std': 0.09796515107154846, 'reward': 0.6330901980400085, 'reward_std': 0.09796513617038727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14874045550823212, 'sampling/sampling_logp_difference/max': 2.1184170246124268, 'sampling/importance_sampling_ratio/min': 0.12022178620100021, 'sampling/importance_sampling_ratio/mean': 1.0069530010223389, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.046143539249897, 'clip_ratio/low_mean': 0.07416177168488503, 'clip_ratio/low_min': 0.07416177168488503, 'clip_ratio/high_mean': 0.06384505704045296, 'clip_ratio/high_max': 0.06384505704045296, 'clip_ratio/region_mean': 0.13800682872533798, 'reward_total_mean': 0.6330901980400085, 'reward_meter_mean': 0.9875081777572632, 'reward_meter_std': 0.012425047345459461, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9484192132949829, 'reward_repeat_soft_std': 0.041187919676303864, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6330901980400085, 'reward_total_composite_std': 0.09796515107154846, 'epoch': 0.11}

 33%|███▎      | 1083/3300 [2:42:10<4:16:25,  6.94s/it]INFO 04-13 10:03:08 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1084/3300 [2:42:17<4:17:47,  6.98s/it]2026-04-13 10:03:14,955 | INFO | train_grpo_train | metrics_logged mode=train step=1084

                                                       
{'loss': 0.024, 'grad_norm': 9.318251609802246, 'learning_rate': 6.718181818181819e-06, 'num_tokens': 1916260.0, 'completions/mean_length': 117.0, 'completions/min_length': 99.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.0, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.5858095288276672, 'rewards/meter/std': 0.2507801949977875, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9901808500289917, 'rewards/repeat_soft/std': 0.003913509659469128, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.43121761083602905, 'rewards/total_composite/std': 0.1951574683189392, 'reward': 0.43121761083602905, 'reward_std': 0.1951574683189392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16685812175273895, 'sampling/sampling_logp_difference/max': 1.9139630794525146, 'sampling/importance_sampling_ratio/min': 0.14749470353126526, 'sampling/importance_sampling_ratio/mean': 1.0210882425308228, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4054247438907623, 'clip_ratio/low_mean': 0.04495614022016525, 'clip_ratio/low_min': 0.04495614022016525, 'clip_ratio/high_mean': 0.1365183051675558, 'clip_ratio/high_max': 0.1365183051675558, 'clip_ratio/region_mean': 0.18147444538772106, 'reward_total_mean': 0.43121761083602905, 'reward_meter_mean': 0.5858095288276672, 'reward_meter_std': 0.2507801949977875, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9901808500289917, 'reward_repeat_soft_std': 0.003913509659469128, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.43121761083602905, 'reward_total_composite_std': 0.1951574683189392, 'epoch': 0.11}

 33%|███▎      | 1084/3300 [2:42:17<4:17:47,  6.98s/it]INFO 04-13 10:03:15 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1085/3300 [2:42:23<4:15:14,  6.91s/it]2026-04-13 10:03:21,715 | INFO | train_grpo_train | metrics_logged mode=train step=1085

                                                       
{'loss': -0.0325, 'grad_norm': 13.496150016784668, 'learning_rate': 6.715151515151516e-06, 'num_tokens': 1918167.0, 'completions/mean_length': 74.375, 'completions/min_length': 63.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.375, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.6591826677322388, 'rewards/meter/std': 0.3186195194721222, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9551384449005127, 'rewards/repeat_soft/std': 0.022562414407730103, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.546758770942688, 'rewards/total_composite/std': 0.09543414413928986, 'reward': 0.546758770942688, 'reward_std': 0.09543413668870926, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14800848066806793, 'sampling/sampling_logp_difference/max': 1.3822741508483887, 'sampling/importance_sampling_ratio/min': 0.251007080078125, 'sampling/importance_sampling_ratio/mean': 1.0303698778152466, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.157143548130989, 'clip_ratio/low_mean': 0.04560697544366121, 'clip_ratio/low_min': 0.04560697544366121, 'clip_ratio/high_mean': 0.09135039150714874, 'clip_ratio/high_max': 0.09135039150714874, 'clip_ratio/region_mean': 0.13695736695080996, 'reward_total_mean': 0.546758770942688, 'reward_meter_mean': 0.6591826677322388, 'reward_meter_std': 0.3186195194721222, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9551384449005127, 'reward_repeat_soft_std': 0.022562414407730103, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.546758770942688, 'reward_total_composite_std': 0.09543414413928986, 'epoch': 0.11}

 33%|███▎      | 1085/3300 [2:42:23<4:15:14,  6.91s/it]INFO 04-13 10:03:21 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1086/3300 [2:42:30<4:13:25,  6.87s/it]2026-04-13 10:03:28,475 | INFO | train_grpo_train | metrics_logged mode=train step=1086

                                                       
{'loss': 0.14, 'grad_norm': 14.193161964416504, 'learning_rate': 6.712121212121213e-06, 'num_tokens': 1920085.0, 'completions/mean_length': 67.75, 'completions/min_length': 51.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.75, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.8095401525497437, 'rewards/meter/std': 0.2770191431045532, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9794746041297913, 'rewards/repeat_soft/std': 0.012927144765853882, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5207315683364868, 'rewards/total_composite/std': 0.07252418994903564, 'reward': 0.5207315683364868, 'reward_std': 0.07252418249845505, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18057994544506073, 'sampling/sampling_logp_difference/max': 1.7014508247375488, 'sampling/importance_sampling_ratio/min': 0.18241867423057556, 'sampling/importance_sampling_ratio/mean': 1.0027400255203247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.337526798248291, 'clip_ratio/low_mean': 0.05472476966679096, 'clip_ratio/low_min': 0.05472476966679096, 'clip_ratio/high_mean': 0.11002439074218273, 'clip_ratio/high_max': 0.11002439074218273, 'clip_ratio/region_mean': 0.1647491604089737, 'reward_total_mean': 0.5207315683364868, 'reward_meter_mean': 0.8095401525497437, 'reward_meter_std': 0.2770191431045532, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9794746041297913, 'reward_repeat_soft_std': 0.012927144765853882, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5207315683364868, 'reward_total_composite_std': 0.07252418994903564, 'epoch': 0.11}

 33%|███▎      | 1086/3300 [2:42:30<4:13:25,  6.87s/it]INFO 04-13 10:03:28 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1087/3300 [2:42:36<4:02:24,  6.57s/it]2026-04-13 10:03:34,356 | INFO | train_grpo_train | metrics_logged mode=train step=1087

                                                       
{'loss': 0.02, 'grad_norm': 24.208295822143555, 'learning_rate': 6.709090909090909e-06, 'num_tokens': 1921455.0, 'completions/mean_length': 22.25, 'completions/min_length': 19.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9190458059310913, 'rewards/meter/std': 0.11271469295024872, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8771973252296448, 'rewards/repeat_soft/std': 0.10158171504735947, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.12947696447372437, 'rewards/total_composite/mean': 0.5464772582054138, 'rewards/total_composite/std': 0.10019097477197647, 'reward': 0.5464772582054138, 'reward_std': 0.10019097477197647, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15317396819591522, 'sampling/sampling_logp_difference/max': 0.9672427177429199, 'sampling/importance_sampling_ratio/min': 0.38012972474098206, 'sampling/importance_sampling_ratio/mean': 0.9914935827255249, 'sampling/importance_sampling_ratio/max': 1.8522214889526367, 'entropy': 1.1571223884820938, 'clip_ratio/low_mean': 0.03112400509417057, 'clip_ratio/low_min': 0.03112400509417057, 'clip_ratio/high_mean': 0.09717909200116992, 'clip_ratio/high_max': 0.09717909200116992, 'clip_ratio/region_mean': 0.1283030970953405, 'reward_total_mean': 0.5464772582054138, 'reward_meter_mean': 0.9190458059310913, 'reward_meter_std': 0.11271469295024872, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8771973252296448, 'reward_repeat_soft_std': 0.10158171504735947, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.12947696447372437, 'reward_total_composite_mean': 0.5464772582054138, 'reward_total_composite_std': 0.10019097477197647, 'epoch': 0.11}

 33%|███▎      | 1087/3300 [2:42:36<4:02:24,  6.57s/it]INFO 04-13 10:03:34 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1088/3300 [2:42:43<4:04:55,  6.64s/it]2026-04-13 10:03:41,167 | INFO | train_grpo_train | metrics_logged mode=train step=1088

                                                       
{'loss': 0.0118, 'grad_norm': 9.507678985595703, 'learning_rate': 6.706060606060607e-06, 'num_tokens': 1923674.0, 'completions/mean_length': 99.375, 'completions/min_length': 91.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.8243648409843445, 'rewards/meter/std': 0.29020729660987854, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9430813789367676, 'rewards/repeat_soft/std': 0.056274957954883575, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.1642080694437027, 'rewards/total_composite/mean': 0.5826771855354309, 'rewards/total_composite/std': 0.15603899955749512, 'reward': 0.5826771855354309, 'reward_std': 0.15603899955749512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14087340235710144, 'sampling/sampling_logp_difference/max': 1.4260168075561523, 'sampling/importance_sampling_ratio/min': 0.24026404321193695, 'sampling/importance_sampling_ratio/mean': 1.0190045833587646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9441261067986488, 'clip_ratio/low_mean': 0.09247690811753273, 'clip_ratio/low_min': 0.09247690811753273, 'clip_ratio/high_mean': 0.03256148658692837, 'clip_ratio/high_max': 0.03256148658692837, 'clip_ratio/region_mean': 0.1250383947044611, 'reward_total_mean': 0.5826771855354309, 'reward_meter_mean': 0.8243648409843445, 'reward_meter_std': 0.29020729660987854, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9430813789367676, 'reward_repeat_soft_std': 0.056274957954883575, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.1642080694437027, 'reward_total_composite_mean': 0.5826771855354309, 'reward_total_composite_std': 0.15603899955749512, 'epoch': 0.11}

 33%|███▎      | 1088/3300 [2:42:43<4:04:55,  6.64s/it]INFO 04-13 10:03:41 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1089/3300 [2:42:49<4:01:04,  6.54s/it]2026-04-13 10:03:47,471 | INFO | train_grpo_train | metrics_logged mode=train step=1089

                                                       
{'loss': 0.0741, 'grad_norm': 11.498926162719727, 'learning_rate': 6.703030303030304e-06, 'num_tokens': 1925438.0, 'completions/mean_length': 48.5, 'completions/min_length': 39.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.551418125629425, 'rewards/meter/std': 0.3613598048686981, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9469898343086243, 'rewards/repeat_soft/std': 0.033088844269514084, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5358335375785828, 'rewards/total_composite/std': 0.17223727703094482, 'reward': 0.5358335375785828, 'reward_std': 0.17223727703094482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1180436760187149, 'sampling/sampling_logp_difference/max': 1.4316868782043457, 'sampling/importance_sampling_ratio/min': 0.23890559375286102, 'sampling/importance_sampling_ratio/mean': 1.0067400932312012, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6848305761814117, 'clip_ratio/low_mean': 0.07360310107469559, 'clip_ratio/low_min': 0.07360310107469559, 'clip_ratio/high_mean': 0.04411872383207083, 'clip_ratio/high_max': 0.04411872383207083, 'clip_ratio/region_mean': 0.11772182490676641, 'reward_total_mean': 0.5358335375785828, 'reward_meter_mean': 0.551418125629425, 'reward_meter_std': 0.3613598048686981, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9469898343086243, 'reward_repeat_soft_std': 0.033088844269514084, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5358335375785828, 'reward_total_composite_std': 0.17223727703094482, 'epoch': 0.11}

 33%|███▎      | 1089/3300 [2:42:49<4:01:04,  6.54s/it]INFO 04-13 10:03:47 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1090/3300 [2:42:55<3:53:47,  6.35s/it]2026-04-13 10:03:53,362 | INFO | train_grpo_train | metrics_logged mode=train step=1090

                                                       
{'loss': 0.0789, 'grad_norm': 23.13025665283203, 'learning_rate': 6.700000000000001e-06, 'num_tokens': 1926922.0, 'completions/mean_length': 26.5, 'completions/min_length': 24.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9310585260391235, 'rewards/meter/std': 0.08636492490768433, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9370556473731995, 'rewards/repeat_soft/std': 0.057211827486753464, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2921227812767029, 'rewards/total_composite/mean': 0.7428244352340698, 'rewards/total_composite/std': 0.20079685747623444, 'reward': 0.7428244352340698, 'reward_std': 0.20079685747623444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17366978526115417, 'sampling/sampling_logp_difference/max': 1.4532572031021118, 'sampling/importance_sampling_ratio/min': 0.2338074892759323, 'sampling/importance_sampling_ratio/mean': 1.015945315361023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9537479653954506, 'clip_ratio/low_mean': 0.06787037011235952, 'clip_ratio/low_min': 0.06787037011235952, 'clip_ratio/high_mean': 0.054067461751401424, 'clip_ratio/high_max': 0.054067461751401424, 'clip_ratio/region_mean': 0.12193783186376095, 'reward_total_mean': 0.7428244352340698, 'reward_meter_mean': 0.9310585260391235, 'reward_meter_std': 0.08636492490768433, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9370556473731995, 'reward_repeat_soft_std': 0.057211827486753464, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2921227812767029, 'reward_total_composite_mean': 0.7428244352340698, 'reward_total_composite_std': 0.20079685747623444, 'epoch': 0.11}

 33%|███▎      | 1090/3300 [2:42:55<3:53:47,  6.35s/it]INFO 04-13 10:03:53 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1091/3300 [2:43:01<3:48:37,  6.21s/it]2026-04-13 10:03:59,252 | INFO | train_grpo_train | metrics_logged mode=train step=1091

                                                       
{'loss': -0.0127, 'grad_norm': 20.685056686401367, 'learning_rate': 6.6969696969696975e-06, 'num_tokens': 1928540.0, 'completions/mean_length': 29.25, 'completions/min_length': 26.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.5196160078048706, 'rewards/meter/std': 0.4102964997291565, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.978695809841156, 'rewards/repeat_soft/std': 0.02221783623099327, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.11563489586114883, 'rewards/total_composite/mean': 0.5142635107040405, 'rewards/total_composite/std': 0.14864948391914368, 'reward': 0.5142635107040405, 'reward_std': 0.14864948391914368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16559728980064392, 'sampling/sampling_logp_difference/max': 1.7340381145477295, 'sampling/importance_sampling_ratio/min': 0.17656995356082916, 'sampling/importance_sampling_ratio/mean': 0.9995377063751221, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.006347917020321, 'clip_ratio/low_mean': 0.08393357694149017, 'clip_ratio/low_min': 0.08393357694149017, 'clip_ratio/high_mean': 0.062229438684880733, 'clip_ratio/high_max': 0.062229438684880733, 'clip_ratio/region_mean': 0.1461630156263709, 'reward_total_mean': 0.5142635107040405, 'reward_meter_mean': 0.5196160078048706, 'reward_meter_std': 0.4102964997291565, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.978695809841156, 'reward_repeat_soft_std': 0.02221783623099327, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.11563489586114883, 'reward_total_composite_mean': 0.5142635107040405, 'reward_total_composite_std': 0.14864948391914368, 'epoch': 0.11}

 33%|███▎      | 1091/3300 [2:43:01<3:48:37,  6.21s/it]INFO 04-13 10:03:59 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1092/3300 [2:43:08<3:54:33,  6.37s/it]2026-04-13 10:04:06,010 | INFO | train_grpo_train | metrics_logged mode=train step=1092

                                                       
{'loss': 0.0369, 'grad_norm': 11.980788230895996, 'learning_rate': 6.693939393939395e-06, 'num_tokens': 1930382.0, 'completions/mean_length': 58.25, 'completions/min_length': 51.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.6034842729568481, 'rewards/meter/std': 0.3847072124481201, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9206419587135315, 'rewards/repeat_soft/std': 0.0525461807847023, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.4950702488422394, 'rewards/total_composite/std': 0.1105421930551529, 'reward': 0.4950702488422394, 'reward_std': 0.1105421930551529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14994347095489502, 'sampling/sampling_logp_difference/max': 1.428484559059143, 'sampling/importance_sampling_ratio/min': 0.23967187106609344, 'sampling/importance_sampling_ratio/mean': 1.0109158754348755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9124277085065842, 'clip_ratio/low_mean': 0.06228564865887165, 'clip_ratio/low_min': 0.06228564865887165, 'clip_ratio/high_mean': 0.06121163163334131, 'clip_ratio/high_max': 0.06121163163334131, 'clip_ratio/region_mean': 0.12349728029221296, 'reward_total_mean': 0.4950702488422394, 'reward_meter_mean': 0.6034842729568481, 'reward_meter_std': 0.3847072124481201, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9206419587135315, 'reward_repeat_soft_std': 0.0525461807847023, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.4950702488422394, 'reward_total_composite_std': 0.1105421930551529, 'epoch': 0.11}

 33%|███▎      | 1092/3300 [2:43:08<3:54:33,  6.37s/it]INFO 04-13 10:04:06 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1093/3300 [2:43:14<3:58:39,  6.49s/it]2026-04-13 10:04:12,766 | INFO | train_grpo_train | metrics_logged mode=train step=1093

                                                       
{'loss': 0.0238, 'grad_norm': 11.41279125213623, 'learning_rate': 6.690909090909091e-06, 'num_tokens': 1932357.0, 'completions/mean_length': 70.875, 'completions/min_length': 54.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9844741821289062, 'rewards/meter/std': 0.018437247723340988, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9510754346847534, 'rewards/repeat_soft/std': 0.021513454616069794, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5364227294921875, 'rewards/total_composite/std': 0.010164810344576836, 'reward': 0.5364227294921875, 'reward_std': 0.010164814069867134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14469408988952637, 'sampling/sampling_logp_difference/max': 3.0628788471221924, 'sampling/importance_sampling_ratio/min': 0.04675290733575821, 'sampling/importance_sampling_ratio/mean': 1.026376724243164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9509404972195625, 'clip_ratio/low_mean': 0.09619823005050421, 'clip_ratio/low_min': 0.09619823005050421, 'clip_ratio/high_mean': 0.044693732634186745, 'clip_ratio/high_max': 0.044693732634186745, 'clip_ratio/region_mean': 0.14089196268469095, 'reward_total_mean': 0.5364227294921875, 'reward_meter_mean': 0.9844741821289062, 'reward_meter_std': 0.018437247723340988, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9510754346847534, 'reward_repeat_soft_std': 0.021513454616069794, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5364227294921875, 'reward_total_composite_std': 0.010164810344576836, 'epoch': 0.11}

 33%|███▎      | 1093/3300 [2:43:14<3:58:39,  6.49s/it]INFO 04-13 10:04:13 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1094/3300 [2:43:25<4:47:55,  7.83s/it]2026-04-13 10:04:23,729 | INFO | train_grpo_train | metrics_logged mode=train step=1094

                                                       
{'loss': -0.0871, 'grad_norm': 2.331590175628662, 'learning_rate': 6.687878787878788e-06, 'num_tokens': 1933729.0, 'completions/mean_length': 92.5, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 32.57143020629883, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.6815160512924194, 'rewards/meter/std': 0.2812528610229492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9843109250068665, 'rewards/repeat_soft/std': 0.011462407186627388, 'rewards/judge_quality/mean': 0.8112499713897705, 'rewards/judge_quality/std': 0.3075914680957794, 'rewards/total_composite/mean': 0.6813645362854004, 'rewards/total_composite/std': 0.31527772545814514, 'reward': 0.6813645362854004, 'reward_std': 0.31527772545814514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12275287508964539, 'sampling/sampling_logp_difference/max': 1.4223895072937012, 'sampling/importance_sampling_ratio/min': 0.24113713204860687, 'sampling/importance_sampling_ratio/mean': 1.011479139328003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5784760490059853, 'clip_ratio/low_mean': 0.011029412038624287, 'clip_ratio/low_min': 0.011029412038624287, 'clip_ratio/high_mean': 0.09784006420522928, 'clip_ratio/high_max': 0.09784006420522928, 'clip_ratio/region_mean': 0.10886947624385357, 'reward_total_mean': 0.6813645362854004, 'reward_meter_mean': 0.6815160512924194, 'reward_meter_std': 0.2812528610229492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9843109250068665, 'reward_repeat_soft_std': 0.011462407186627388, 'reward_judge_quality_mean': 0.8112499713897705, 'reward_judge_quality_std': 0.3075914680957794, 'reward_total_composite_mean': 0.6813645362854004, 'reward_total_composite_std': 0.31527772545814514, 'epoch': 0.11}

 33%|███▎      | 1094/3300 [2:43:25<4:47:55,  7.83s/it]INFO 04-13 10:04:23 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1095/3300 [2:43:32<4:32:22,  7.41s/it]2026-04-13 10:04:30,159 | INFO | train_grpo_train | metrics_logged mode=train step=1095

                                                       
{'loss': 0.0162, 'grad_norm': 14.497970581054688, 'learning_rate': 6.684848484848485e-06, 'num_tokens': 1935420.0, 'completions/mean_length': 49.375, 'completions/min_length': 47.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6142832040786743, 'rewards/meter/std': 0.3878927230834961, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9937841892242432, 'rewards/repeat_soft/std': 0.009870151989161968, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.19255799055099487, 'rewards/total_composite/mean': 0.4929462671279907, 'rewards/total_composite/std': 0.25837865471839905, 'reward': 0.4929462671279907, 'reward_std': 0.25837868452072144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18936829268932343, 'sampling/sampling_logp_difference/max': 2.8573012351989746, 'sampling/importance_sampling_ratio/min': 0.2412366271018982, 'sampling/importance_sampling_ratio/mean': 1.0276849269866943, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5982573479413986, 'clip_ratio/low_mean': 0.07016889192163944, 'clip_ratio/low_min': 0.07016889192163944, 'clip_ratio/high_mean': 0.1261708401143551, 'clip_ratio/high_max': 0.1261708401143551, 'clip_ratio/region_mean': 0.19633973203599453, 'reward_total_mean': 0.4929462671279907, 'reward_meter_mean': 0.6142832040786743, 'reward_meter_std': 0.3878927230834961, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9937841892242432, 'reward_repeat_soft_std': 0.009870151989161968, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.19255799055099487, 'reward_total_composite_mean': 0.4929462671279907, 'reward_total_composite_std': 0.25837865471839905, 'epoch': 0.11}

 33%|███▎      | 1095/3300 [2:43:32<4:32:22,  7.41s/it]INFO 04-13 10:04:30 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1096/3300 [2:43:39<4:25:20,  7.22s/it]2026-04-13 10:04:36,945 | INFO | train_grpo_train | metrics_logged mode=train step=1096

                                                       
{'loss': -0.0192, 'grad_norm': 11.968867301940918, 'learning_rate': 6.681818181818183e-06, 'num_tokens': 1937240.0, 'completions/mean_length': 57.5, 'completions/min_length': 53.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.5, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7465670704841614, 'rewards/meter/std': 0.24056415259838104, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9406276941299438, 'rewards/repeat_soft/std': 0.035218942910432816, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.516263484954834, 'rewards/total_composite/std': 0.05505356192588806, 'reward': 0.516263484954834, 'reward_std': 0.05505356192588806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1491585373878479, 'sampling/sampling_logp_difference/max': 1.6534686088562012, 'sampling/importance_sampling_ratio/min': 0.1913849115371704, 'sampling/importance_sampling_ratio/mean': 1.0211995840072632, 'sampling/importance_sampling_ratio/max': 1.9647976160049438, 'entropy': 1.1102144196629524, 'clip_ratio/low_mean': 0.040419649332761765, 'clip_ratio/low_min': 0.040419649332761765, 'clip_ratio/high_mean': 0.08970674313604832, 'clip_ratio/high_max': 0.08970674313604832, 'clip_ratio/region_mean': 0.13012639246881008, 'reward_total_mean': 0.516263484954834, 'reward_meter_mean': 0.7465670704841614, 'reward_meter_std': 0.24056415259838104, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9406276941299438, 'reward_repeat_soft_std': 0.035218942910432816, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.516263484954834, 'reward_total_composite_std': 0.05505356192588806, 'epoch': 0.11}

 33%|███▎      | 1096/3300 [2:43:39<4:25:20,  7.22s/it]INFO 04-13 10:04:37 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1097/3300 [2:43:50<5:07:32,  8.38s/it]2026-04-13 10:04:48,009 | INFO | train_grpo_train | metrics_logged mode=train step=1097

                                                       
{'loss': -0.0776, 'grad_norm': 1.4841713905334473, 'learning_rate': 6.678787878787879e-06, 'num_tokens': 1938710.0, 'completions/mean_length': 86.75, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 26.000001907348633, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8709642887115479, 'rewards/meter/std': 0.35193586349487305, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9601908326148987, 'rewards/repeat_soft/std': 0.04054465889930725, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.5243704319000244, 'rewards/total_composite/std': 0.21498365700244904, 'reward': 0.5243704319000244, 'reward_std': 0.21498365700244904, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14930590987205505, 'sampling/sampling_logp_difference/max': 1.9566144943237305, 'sampling/importance_sampling_ratio/min': 0.1413360983133316, 'sampling/importance_sampling_ratio/mean': 1.0057408809661865, 'sampling/importance_sampling_ratio/max': 1.6734319925308228, 'entropy': 0.8684336245059967, 'clip_ratio/low_mean': 0.014880952425301075, 'clip_ratio/low_min': 0.014880952425301075, 'clip_ratio/high_mean': 0.06891872081905603, 'clip_ratio/high_max': 0.06891872081905603, 'clip_ratio/region_mean': 0.08379967324435711, 'reward_total_mean': 0.5243704319000244, 'reward_meter_mean': 0.8709642887115479, 'reward_meter_std': 0.35193586349487305, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9601908326148987, 'reward_repeat_soft_std': 0.04054465889930725, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.5243704319000244, 'reward_total_composite_std': 0.21498365700244904, 'epoch': 0.11}

 33%|███▎      | 1097/3300 [2:43:50<5:07:32,  8.38s/it]INFO 04-13 10:04:48 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1098/3300 [2:43:56<4:40:20,  7.64s/it]2026-04-13 10:04:53,930 | INFO | train_grpo_train | metrics_logged mode=train step=1098

                                                       
{'loss': -0.0076, 'grad_norm': 13.859899520874023, 'learning_rate': 6.6757575757575766e-06, 'num_tokens': 1940332.0, 'completions/mean_length': 43.75, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8528269529342651, 'rewards/meter/std': 0.23338685929775238, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9464430809020996, 'rewards/repeat_soft/std': 0.03395868465304375, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.16974246501922607, 'rewards/total_composite/mean': 0.6254202127456665, 'rewards/total_composite/std': 0.13749991357326508, 'reward': 0.6254202127456665, 'reward_std': 0.13749991357326508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15951161086559296, 'sampling/sampling_logp_difference/max': 2.5978305339813232, 'sampling/importance_sampling_ratio/min': 0.07443489134311676, 'sampling/importance_sampling_ratio/mean': 1.0151758193969727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.996156208217144, 'clip_ratio/low_mean': 0.09627203363925219, 'clip_ratio/low_min': 0.09627203363925219, 'clip_ratio/high_mean': 0.043026004917919636, 'clip_ratio/high_max': 0.043026004917919636, 'clip_ratio/region_mean': 0.13929803855717182, 'reward_total_mean': 0.6254202127456665, 'reward_meter_mean': 0.8528269529342651, 'reward_meter_std': 0.23338685929775238, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9464430809020996, 'reward_repeat_soft_std': 0.03395868465304375, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.16974246501922607, 'reward_total_composite_mean': 0.6254202127456665, 'reward_total_composite_std': 0.13749991357326508, 'epoch': 0.11}

 33%|███▎      | 1098/3300 [2:43:56<4:40:20,  7.64s/it]INFO 04-13 10:04:54 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1099/3300 [2:44:02<4:31:16,  7.40s/it]2026-04-13 10:05:00,757 | INFO | train_grpo_train | metrics_logged mode=train step=1099

                                                       
{'loss': 0.0229, 'grad_norm': 11.417180061340332, 'learning_rate': 6.672727272727273e-06, 'num_tokens': 1942676.0, 'completions/mean_length': 96.0, 'completions/min_length': 76.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.0, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.4464265704154968, 'rewards/meter/std': 0.36981698870658875, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9601141214370728, 'rewards/repeat_soft/std': 0.04743463173508644, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.38410118222236633, 'rewards/total_composite/std': 0.19046717882156372, 'reward': 0.38410118222236633, 'reward_std': 0.19046717882156372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14962473511695862, 'sampling/sampling_logp_difference/max': 2.2815001010894775, 'sampling/importance_sampling_ratio/min': 0.10213088244199753, 'sampling/importance_sampling_ratio/mean': 0.9960708022117615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9751346260309219, 'clip_ratio/low_mean': 0.0620602834969759, 'clip_ratio/low_min': 0.0620602834969759, 'clip_ratio/high_mean': 0.07958950288593769, 'clip_ratio/high_max': 0.07958950288593769, 'clip_ratio/region_mean': 0.1416497863829136, 'reward_total_mean': 0.38410118222236633, 'reward_meter_mean': 0.4464265704154968, 'reward_meter_std': 0.36981698870658875, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9601141214370728, 'reward_repeat_soft_std': 0.04743463173508644, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.38410118222236633, 'reward_total_composite_std': 0.19046717882156372, 'epoch': 0.11}

 33%|███▎      | 1099/3300 [2:44:02<4:31:16,  7.40s/it]INFO 04-13 10:05:00 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1100/3300 [2:44:09<4:18:05,  7.04s/it]2026-04-13 10:05:06,975 | INFO | train_grpo_train | metrics_logged mode=train step=1100

                                                       
{'loss': 0.0269, 'grad_norm': 14.642799377441406, 'learning_rate': 6.66969696969697e-06, 'num_tokens': 1944623.0, 'completions/mean_length': 62.375, 'completions/min_length': 58.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.375, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7071465253829956, 'rewards/meter/std': 0.30786022543907166, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9604854583740234, 'rewards/repeat_soft/std': 0.026317918673157692, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.4587589502334595, 'rewards/total_composite/std': 0.09426096081733704, 'reward': 0.4587589502334595, 'reward_std': 0.09426096826791763, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16210375726222992, 'sampling/sampling_logp_difference/max': 1.7364449501037598, 'sampling/importance_sampling_ratio/min': 0.1761454939842224, 'sampling/importance_sampling_ratio/mean': 1.0446785688400269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2134244069457054, 'clip_ratio/low_mean': 0.09298294503241777, 'clip_ratio/low_min': 0.09298294503241777, 'clip_ratio/high_mean': 0.030166194774210453, 'clip_ratio/high_max': 0.030166194774210453, 'clip_ratio/region_mean': 0.12314913980662823, 'reward_total_mean': 0.4587589502334595, 'reward_meter_mean': 0.7071465253829956, 'reward_meter_std': 0.30786022543907166, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9604854583740234, 'reward_repeat_soft_std': 0.026317918673157692, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.4587589502334595, 'reward_total_composite_std': 0.09426096081733704, 'epoch': 0.11}

 33%|███▎      | 1100/3300 [2:44:09<4:18:05,  7.04s/it]INFO 04-13 10:05:07 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:02<00:11,  1.48s/it]

 30%|███       | 3/10 [00:07<00:19,  2.72s/it]

 40%|████      | 4/10 [00:10<00:16,  2.78s/it]

 50%|█████     | 5/10 [00:15<00:18,  3.77s/it]

 60%|██████    | 6/10 [00:19<00:14,  3.72s/it]

 70%|███████   | 7/10 [00:23<00:11,  3.94s/it]

 80%|████████  | 8/10 [00:27<00:07,  3.81s/it]

 90%|█████████ | 9/10 [00:31<00:03,  3.83s/it]

100%|██████████| 10/10 [00:34<00:00,  3.58s/it]2026-04-13 10:05:45,106 | INFO | train_grpo_train | metrics_logged mode=eval step=1100

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 38.0621, 'eval_samples_per_second': 2.102, 'eval_steps_per_second': 0.263, 'eval_num_tokens': 1944623.0, 'eval_completions/mean_length': 59.0875, 'eval_completions/min_length': 30.1, 'eval_completions/max_length': 87.9, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 59.0875, 'eval_completions/min_terminated_length': 30.1, 'eval_completions/max_terminated_length': 87.9, 'eval_rewards/meter/mean': 0.7739002108573914, 'eval_rewards/meter/std': 0.3174647897481918, 'eval_rewards/count_adherence/mean': 0.8747916758060456, 'eval_rewards/count_adherence/std': 0.13870810568332673, 'eval_rewards/hard_gate/mean': 0.9625, 'eval_rewards/hard_gate/std': 0.10606601536273956, 'eval_rewards/repeat_soft/mean': 0.9302830159664154, 'eval_rewards/repeat_soft/std': 0.05693140346556902, 'eval_rewards/judge_quality/mean': 0.47762500047683715, 'eval_rewards/judge_quality/std': 0.16730604767799379, 'eval_rewards/total_composite/mean': 0.5368325889110566, 'eval_rewards/total_composite/std': 0.1783295027911663, 'eval_reward': 0.5368325889110566, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08279976770281791, 'eval_sampling/sampling_logp_difference/max': 1.016366195678711, 'eval_sampling/importance_sampling_ratio/min': 0.3745600953698158, 'eval_sampling/importance_sampling_ratio/mean': 1.0196762204170227, 'eval_sampling/importance_sampling_ratio/max': 1.4289889454841613, 'eval_entropy': 0.9940823078155517, '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': 0.5368325889110566, 'eval_reward_meter_mean': 0.7739002108573914, 'eval_reward_meter_std': 0.3174647897481918, 'eval_reward_count_adherence_mean': 0.8747916758060456, 'eval_reward_count_adherence_std': 0.13870810568332673, 'eval_reward_hard_gate_mean': 0.9625, 'eval_reward_hard_gate_std': 0.10606601536273956, 'eval_reward_repeat_soft_mean': 0.9302830159664154, 'eval_reward_repeat_soft_std': 0.05693140346556902, 'eval_reward_judge_quality_mean': 0.47762500047683715, 'eval_reward_judge_quality_std': 0.16730604767799379, 'eval_reward_total_composite_mean': 0.5368325889110566, 'eval_reward_total_composite_std': 0.1783295027911663, 'epoch': 0.11}

 33%|███▎      | 1100/3300 [2:44:47<4:18:05,  7.04s/it]

100%|██████████| 10/10 [00:34<00:00,  3.58s/it]

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

 33%|███▎      | 1101/3300 [2:44:57<11:48:09, 19.32s/it]2026-04-13 10:05:54,951 | INFO | train_grpo_train | metrics_logged mode=train step=1101

                                                        
{'loss': 0.0135, 'grad_norm': 9.57143497467041, 'learning_rate': 6.666666666666667e-06, 'num_tokens': 1946331.0, 'completions/mean_length': 51.5, 'completions/min_length': 42.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9508030414581299, 'rewards/meter/std': 0.09549621492624283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9613487124443054, 'rewards/repeat_soft/std': 0.039299752563238144, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.6705681681632996, 'rewards/total_composite/std': 0.13356658816337585, 'reward': 0.6705681681632996, 'reward_std': 0.13356660306453705, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12682457268238068, 'sampling/sampling_logp_difference/max': 1.6577732563018799, 'sampling/importance_sampling_ratio/min': 0.19056284427642822, 'sampling/importance_sampling_ratio/mean': 1.0136646032333374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9868056029081345, 'clip_ratio/low_mean': 0.08438512543216348, 'clip_ratio/low_min': 0.08438512543216348, 'clip_ratio/high_mean': 0.05235273018479347, 'clip_ratio/high_max': 0.05235273018479347, 'clip_ratio/region_mean': 0.13673785561695695, 'reward_total_mean': 0.6705681681632996, 'reward_meter_mean': 0.9508030414581299, 'reward_meter_std': 0.09549621492624283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9613487124443054, 'reward_repeat_soft_std': 0.039299752563238144, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.6705681681632996, 'reward_total_composite_std': 0.13356658816337585, 'epoch': 0.11}

 33%|███▎      | 1101/3300 [2:44:57<11:48:09, 19.32s/it]INFO 04-13 10:05:55 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1102/3300 [2:45:03<9:29:31, 15.55s/it] 2026-04-13 10:06:01,689 | INFO | train_grpo_train | metrics_logged mode=train step=1102

                                                       
{'loss': 0.0406, 'grad_norm': 9.237506866455078, 'learning_rate': 6.663636363636365e-06, 'num_tokens': 1948547.0, 'completions/mean_length': 65.0, 'completions/min_length': 61.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.0, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9411215782165527, 'rewards/meter/std': 0.06898335367441177, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.910030722618103, 'rewards/repeat_soft/std': 0.0485076941549778, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5892764925956726, 'rewards/total_composite/std': 0.09233304858207703, 'reward': 0.5892764925956726, 'reward_std': 0.09233304113149643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12735266983509064, 'sampling/sampling_logp_difference/max': 1.9520063400268555, 'sampling/importance_sampling_ratio/min': 0.14198890328407288, 'sampling/importance_sampling_ratio/mean': 0.9912409782409668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6516686156392097, 'clip_ratio/low_mean': 0.08202955033630133, 'clip_ratio/low_min': 0.08202955033630133, 'clip_ratio/high_mean': 0.0356006845831871, 'clip_ratio/high_max': 0.0356006845831871, 'clip_ratio/region_mean': 0.11763023491948843, 'reward_total_mean': 0.5892764925956726, 'reward_meter_mean': 0.9411215782165527, 'reward_meter_std': 0.06898335367441177, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.910030722618103, 'reward_repeat_soft_std': 0.0485076941549778, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5892764925956726, 'reward_total_composite_std': 0.09233304858207703, 'epoch': 0.11}

 33%|███▎      | 1102/3300 [2:45:03<9:29:31, 15.55s/it]INFO 04-13 10:06:01 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1103/3300 [2:45:10<7:51:54, 12.89s/it]2026-04-13 10:06:08,377 | INFO | train_grpo_train | metrics_logged mode=train step=1103

                                                       
{'loss': 0.0492, 'grad_norm': 8.690917015075684, 'learning_rate': 6.660606060606061e-06, 'num_tokens': 1950482.0, 'completions/mean_length': 69.875, 'completions/min_length': 51.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9910345077514648, 'rewards/meter/std': 0.006791371386498213, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9173648357391357, 'rewards/repeat_soft/std': 0.03406515344977379, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5605899691581726, 'rewards/total_composite/std': 0.010260870680212975, 'reward': 0.5605899691581726, 'reward_std': 0.010260866023600101, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14378008246421814, 'sampling/sampling_logp_difference/max': 1.3936083316802979, 'sampling/importance_sampling_ratio/min': 0.24817818403244019, 'sampling/importance_sampling_ratio/mean': 1.0004949569702148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1420028284192085, 'clip_ratio/low_mean': 0.07078136503696442, 'clip_ratio/low_min': 0.07078136503696442, 'clip_ratio/high_mean': 0.049270967952907085, 'clip_ratio/high_max': 0.049270967952907085, 'clip_ratio/region_mean': 0.1200523329898715, 'reward_total_mean': 0.5605899691581726, 'reward_meter_mean': 0.9910345077514648, 'reward_meter_std': 0.006791371386498213, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9173648357391357, 'reward_repeat_soft_std': 0.03406515344977379, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5605899691581726, 'reward_total_composite_std': 0.010260870680212975, 'epoch': 0.11}

 33%|███▎      | 1103/3300 [2:45:10<7:51:54, 12.89s/it]INFO 04-13 10:06:08 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1104/3300 [2:45:17<6:50:43, 11.22s/it]2026-04-13 10:06:15,701 | INFO | train_grpo_train | metrics_logged mode=train step=1104

                                                       
{'loss': -0.0089, 'grad_norm': 12.141768455505371, 'learning_rate': 6.657575757575758e-06, 'num_tokens': 1952103.0, 'completions/mean_length': 43.625, 'completions/min_length': 37.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.47106683254241943, 'rewards/meter/std': 0.40478745102882385, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9712631702423096, 'rewards/repeat_soft/std': 0.017058251425623894, 'rewards/judge_quality/mean': 0.6025000214576721, 'rewards/judge_quality/std': 0.1976107507944107, 'rewards/total_composite/mean': 0.517354428768158, 'rewards/total_composite/std': 0.14876633882522583, 'reward': 0.517354428768158, 'reward_std': 0.14876633882522583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14809371531009674, 'sampling/sampling_logp_difference/max': 1.211209774017334, 'sampling/importance_sampling_ratio/min': 0.2978367507457733, 'sampling/importance_sampling_ratio/mean': 1.0351276397705078, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1666222214698792, 'clip_ratio/low_mean': 0.09970238246023655, 'clip_ratio/low_min': 0.09970238246023655, 'clip_ratio/high_mean': 0.06661933101713657, 'clip_ratio/high_max': 0.06661933101713657, 'clip_ratio/region_mean': 0.16632171347737312, 'reward_total_mean': 0.517354428768158, 'reward_meter_mean': 0.47106683254241943, 'reward_meter_std': 0.40478745102882385, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9712631702423096, 'reward_repeat_soft_std': 0.017058251425623894, 'reward_judge_quality_mean': 0.6025000214576721, 'reward_judge_quality_std': 0.1976107507944107, 'reward_total_composite_mean': 0.517354428768158, 'reward_total_composite_std': 0.14876633882522583, 'epoch': 0.11}

 33%|███▎      | 1104/3300 [2:45:17<6:50:43, 11.22s/it]INFO 04-13 10:06:15 [block_pool.py:378] Successfully reset prefix cache

 33%|███▎      | 1105/3300 [2:45:24<6:00:10,  9.85s/it]2026-04-13 10:06:22,335 | INFO | train_grpo_train | metrics_logged mode=train step=1105

                                                       
{'loss': 0.0742, 'grad_norm': 8.783102989196777, 'learning_rate': 6.654545454545455e-06, 'num_tokens': 1954012.0, 'completions/mean_length': 86.625, 'completions/min_length': 77.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.625, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9817788600921631, 'rewards/meter/std': 0.013821211643517017, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9379310607910156, 'rewards/repeat_soft/std': 0.04740970954298973, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6879007816314697, 'rewards/total_composite/std': 0.09658631682395935, 'reward': 0.6879007816314697, 'reward_std': 0.09658631682395935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1555800437927246, 'sampling/sampling_logp_difference/max': 1.3972806930541992, 'sampling/importance_sampling_ratio/min': 0.2472684532403946, 'sampling/importance_sampling_ratio/mean': 1.0163252353668213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1243841573596, 'clip_ratio/low_mean': 0.05159039422869682, 'clip_ratio/low_min': 0.05159039422869682, 'clip_ratio/high_mean': 0.09754106309264898, 'clip_ratio/high_max': 0.09754106309264898, 'clip_ratio/region_mean': 0.1491314573213458, 'reward_total_mean': 0.6879007816314697, 'reward_meter_mean': 0.9817788600921631, 'reward_meter_std': 0.013821211643517017, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9379310607910156, 'reward_repeat_soft_std': 0.04740970954298973, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6879007816314697, 'reward_total_composite_std': 0.09658631682395935, 'epoch': 0.11}

 33%|███▎      | 1105/3300 [2:45:24<6:00:10,  9.85s/it]INFO 04-13 10:06:22 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1106/3300 [2:45:30<5:14:48,  8.61s/it]2026-04-13 10:06:28,059 | INFO | train_grpo_train | metrics_logged mode=train step=1106

                                                       
{'loss': 0.1157, 'grad_norm': 20.9514217376709, 'learning_rate': 6.651515151515152e-06, 'num_tokens': 1955611.0, 'completions/mean_length': 25.875, 'completions/min_length': 20.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.9920638799667358, 'rewards/meter/std': 0.006432940252125263, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9376035928726196, 'rewards/repeat_soft/std': 0.06885240226984024, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.12117726355791092, 'rewards/total_composite/mean': 0.5561680793762207, 'rewards/total_composite/std': 0.08233477920293808, 'reward': 0.5561680793762207, 'reward_std': 0.08233474940061569, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1665658801794052, 'sampling/sampling_logp_difference/max': 1.387542724609375, 'sampling/importance_sampling_ratio/min': 0.24968811869621277, 'sampling/importance_sampling_ratio/mean': 1.0338551998138428, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3538757637143135, 'clip_ratio/low_mean': 0.10827161557972431, 'clip_ratio/low_min': 0.10827161557972431, 'clip_ratio/high_mean': 0.08063207007944584, 'clip_ratio/high_max': 0.08063207007944584, 'clip_ratio/region_mean': 0.18890368565917015, 'reward_total_mean': 0.5561680793762207, 'reward_meter_mean': 0.9920638799667358, 'reward_meter_std': 0.006432940252125263, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9376035928726196, 'reward_repeat_soft_std': 0.06885240226984024, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.12117726355791092, 'reward_total_composite_mean': 0.5561680793762207, 'reward_total_composite_std': 0.08233477920293808, 'epoch': 0.11}

 34%|███▎      | 1106/3300 [2:45:30<5:14:48,  8.61s/it]INFO 04-13 10:06:28 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1107/3300 [2:45:36<4:43:35,  7.76s/it]2026-04-13 10:06:33,839 | INFO | train_grpo_train | metrics_logged mode=train step=1107

                                                       
{'loss': 0.0631, 'grad_norm': 12.246903419494629, 'learning_rate': 6.6484848484848485e-06, 'num_tokens': 1957230.0, 'completions/mean_length': 38.375, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9808336496353149, 'rewards/meter/std': 0.0065889740362763405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8884369134902954, 'rewards/repeat_soft/std': 0.05048581212759018, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.6058225631713867, 'rewards/total_composite/std': 0.0140786562114954, 'reward': 0.6058225631713867, 'reward_std': 0.014078662730753422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12168633192777634, 'sampling/sampling_logp_difference/max': 1.398416519165039, 'sampling/importance_sampling_ratio/min': 0.24698776006698608, 'sampling/importance_sampling_ratio/mean': 1.0060867071151733, 'sampling/importance_sampling_ratio/max': 1.7908825874328613, 'entropy': 0.9118454605340958, 'clip_ratio/low_mean': 0.04358908720314503, 'clip_ratio/low_min': 0.04358908720314503, 'clip_ratio/high_mean': 0.06082251248881221, 'clip_ratio/high_max': 0.06082251248881221, 'clip_ratio/region_mean': 0.10441159969195724, 'reward_total_mean': 0.6058225631713867, 'reward_meter_mean': 0.9808336496353149, 'reward_meter_std': 0.0065889740362763405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8884369134902954, 'reward_repeat_soft_std': 0.05048581212759018, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.6058225631713867, 'reward_total_composite_std': 0.0140786562114954, 'epoch': 0.11}

 34%|███▎      | 1107/3300 [2:45:36<4:43:35,  7.76s/it]INFO 04-13 10:06:34 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1108/3300 [2:45:41<4:22:18,  7.18s/it]2026-04-13 10:06:39,666 | INFO | train_grpo_train | metrics_logged mode=train step=1108

                                                       
{'loss': 0.0263, 'grad_norm': 15.682332038879395, 'learning_rate': 6.645454545454546e-06, 'num_tokens': 1958885.0, 'completions/mean_length': 43.875, 'completions/min_length': 41.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5969927310943604, 'rewards/meter/std': 0.36882877349853516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8888386487960815, 'rewards/repeat_soft/std': 0.038945961743593216, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5122449398040771, 'rewards/total_composite/std': 0.11345456540584564, 'reward': 0.5122449398040771, 'reward_std': 0.11345456540584564, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14469872415065765, 'sampling/sampling_logp_difference/max': 4.074991226196289, 'sampling/importance_sampling_ratio/min': 0.016992364078760147, 'sampling/importance_sampling_ratio/mean': 1.0096375942230225, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7018393352627754, 'clip_ratio/low_mean': 0.019392292946577072, 'clip_ratio/low_min': 0.019392292946577072, 'clip_ratio/high_mean': 0.06644094968214631, 'clip_ratio/high_max': 0.06644094968214631, 'clip_ratio/region_mean': 0.08583324262872338, 'reward_total_mean': 0.5122449398040771, 'reward_meter_mean': 0.5969927310943604, 'reward_meter_std': 0.36882877349853516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8888386487960815, 'reward_repeat_soft_std': 0.038945961743593216, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5122449398040771, 'reward_total_composite_std': 0.11345456540584564, 'epoch': 0.11}

 34%|███▎      | 1108/3300 [2:45:41<4:22:18,  7.18s/it]INFO 04-13 10:06:39 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1109/3300 [2:45:48<4:15:30,  7.00s/it]2026-04-13 10:06:46,234 | INFO | train_grpo_train | metrics_logged mode=train step=1109

                                                       
{'loss': -0.0118, 'grad_norm': 11.471404075622559, 'learning_rate': 6.642424242424242e-06, 'num_tokens': 1960766.0, 'completions/mean_length': 56.125, 'completions/min_length': 52.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.6876039505004883, 'rewards/meter/std': 0.3078877031803131, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7964769601821899, 'rewards/repeat_soft/std': 0.05928468331694603, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.4661751687526703, 'rewards/total_composite/std': 0.13284611701965332, 'reward': 0.4661751687526703, 'reward_std': 0.13284608721733093, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11511439085006714, 'sampling/sampling_logp_difference/max': 1.5071403980255127, 'sampling/importance_sampling_ratio/min': 0.2215426117181778, 'sampling/importance_sampling_ratio/mean': 1.023069977760315, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8226094245910645, 'clip_ratio/low_mean': 0.06764401262626052, 'clip_ratio/low_min': 0.06764401262626052, 'clip_ratio/high_mean': 0.039518462494015694, 'clip_ratio/high_max': 0.039518462494015694, 'clip_ratio/region_mean': 0.10716247512027621, 'reward_total_mean': 0.4661751687526703, 'reward_meter_mean': 0.6876039505004883, 'reward_meter_std': 0.3078877031803131, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7964769601821899, 'reward_repeat_soft_std': 0.05928468331694603, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.4661751687526703, 'reward_total_composite_std': 0.13284611701965332, 'epoch': 0.11}

 34%|███▎      | 1109/3300 [2:45:48<4:15:30,  7.00s/it]INFO 04-13 10:06:46 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1110/3300 [2:45:54<4:09:40,  6.84s/it]2026-04-13 10:06:52,713 | INFO | train_grpo_train | metrics_logged mode=train step=1110

                                                       
{'loss': 0.0428, 'grad_norm': 9.730709075927734, 'learning_rate': 6.63939393939394e-06, 'num_tokens': 1962938.0, 'completions/mean_length': 77.5, 'completions/min_length': 64.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9428097009658813, 'rewards/meter/std': 0.12286613881587982, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8957542181015015, 'rewards/repeat_soft/std': 0.037643659859895706, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5306164026260376, 'rewards/total_composite/std': 0.05158373340964317, 'reward': 0.5306164026260376, 'reward_std': 0.05158371478319168, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1261560469865799, 'sampling/sampling_logp_difference/max': 1.6344184875488281, 'sampling/importance_sampling_ratio/min': 0.2225971221923828, 'sampling/importance_sampling_ratio/mean': 1.0119463205337524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8850665241479874, 'clip_ratio/low_mean': 0.012559335213154554, 'clip_ratio/low_min': 0.012559335213154554, 'clip_ratio/high_mean': 0.08973471214994788, 'clip_ratio/high_max': 0.08973471214994788, 'clip_ratio/region_mean': 0.10229404736310244, 'reward_total_mean': 0.5306164026260376, 'reward_meter_mean': 0.9428097009658813, 'reward_meter_std': 0.12286613881587982, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8957542181015015, 'reward_repeat_soft_std': 0.037643659859895706, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5306164026260376, 'reward_total_composite_std': 0.05158373340964317, 'epoch': 0.11}

 34%|███▎      | 1110/3300 [2:45:54<4:09:40,  6.84s/it]INFO 04-13 10:06:52 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1111/3300 [2:46:00<4:00:54,  6.60s/it]2026-04-13 10:06:58,759 | INFO | train_grpo_train | metrics_logged mode=train step=1111

                                                       
{'loss': 0.0351, 'grad_norm': 12.884098052978516, 'learning_rate': 6.6363636363636375e-06, 'num_tokens': 1964605.0, 'completions/mean_length': 47.375, 'completions/min_length': 44.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8017257452011108, 'rewards/meter/std': 0.30836212635040283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9168136119842529, 'rewards/repeat_soft/std': 0.041258662939071655, 'rewards/judge_quality/mean': 0.4437499940395355, 'rewards/judge_quality/std': 0.128834068775177, 'rewards/total_composite/mean': 0.5684049725532532, 'rewards/total_composite/std': 0.11981135606765747, 'reward': 0.5684049725532532, 'reward_std': 0.11981134861707687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11596706509590149, 'sampling/sampling_logp_difference/max': 2.4830503463745117, 'sampling/importance_sampling_ratio/min': 0.08348816633224487, 'sampling/importance_sampling_ratio/mean': 1.022456169128418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7652312889695168, 'clip_ratio/low_mean': 0.055207240860909224, 'clip_ratio/low_min': 0.055207240860909224, 'clip_ratio/high_mean': 0.05440958496183157, 'clip_ratio/high_max': 0.05440958496183157, 'clip_ratio/region_mean': 0.1096168258227408, 'reward_total_mean': 0.5684049725532532, 'reward_meter_mean': 0.8017257452011108, 'reward_meter_std': 0.30836212635040283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9168136119842529, 'reward_repeat_soft_std': 0.041258662939071655, 'reward_judge_quality_mean': 0.4437499940395355, 'reward_judge_quality_std': 0.128834068775177, 'reward_total_composite_mean': 0.5684049725532532, 'reward_total_composite_std': 0.11981135606765747, 'epoch': 0.11}

 34%|███▎      | 1111/3300 [2:46:00<4:00:54,  6.60s/it]INFO 04-13 10:06:58 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1112/3300 [2:46:07<3:57:30,  6.51s/it]2026-04-13 10:07:05,064 | INFO | train_grpo_train | metrics_logged mode=train step=1112

                                                       
{'loss': -0.0934, 'grad_norm': 12.25678825378418, 'learning_rate': 6.633333333333334e-06, 'num_tokens': 1966645.0, 'completions/mean_length': 67.0, 'completions/min_length': 53.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9791103601455688, 'rewards/meter/std': 0.010904289782047272, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8605831861495972, 'rewards/repeat_soft/std': 0.08659627288579941, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5227519869804382, 'rewards/total_composite/std': 0.049343012273311615, 'reward': 0.5227519869804382, 'reward_std': 0.049343008548021317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15082481503486633, 'sampling/sampling_logp_difference/max': 1.8279852867126465, 'sampling/importance_sampling_ratio/min': 0.160737082362175, 'sampling/importance_sampling_ratio/mean': 1.0180777311325073, 'sampling/importance_sampling_ratio/max': 1.8894423246383667, 'entropy': 1.0666228979825974, 'clip_ratio/low_mean': 0.05456842668354511, 'clip_ratio/low_min': 0.05456842668354511, 'clip_ratio/high_mean': 0.09077512472867966, 'clip_ratio/high_max': 0.09077512472867966, 'clip_ratio/region_mean': 0.14534355141222477, 'reward_total_mean': 0.5227519869804382, 'reward_meter_mean': 0.9791103601455688, 'reward_meter_std': 0.010904289782047272, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8605831861495972, 'reward_repeat_soft_std': 0.08659627288579941, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5227519869804382, 'reward_total_composite_std': 0.049343012273311615, 'epoch': 0.11}

 34%|███▎      | 1112/3300 [2:46:07<3:57:30,  6.51s/it]INFO 04-13 10:07:05 [block_pool.py:378] Successfully reset prefix cache

 34%|███▎      | 1113/3300 [2:46:18<4:48:50,  7.92s/it]2026-04-13 10:07:16,284 | INFO | train_grpo_train | metrics_logged mode=train step=1113

                                                       
{'loss': -0.1674, 'grad_norm': 3.0614638328552246, 'learning_rate': 6.630303030303031e-06, 'num_tokens': 1968481.0, 'completions/mean_length': 120.5, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.5714340209961, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.8282565474510193, 'rewards/meter/std': 0.32117608189582825, 'rewards/count_adherence/mean': 0.71875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9734363555908203, 'rewards/repeat_soft/std': 0.014392375946044922, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5041776895523071, 'rewards/total_composite/std': 0.22154462337493896, 'reward': 0.5041776895523071, 'reward_std': 0.22154460847377777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1511511206626892, 'sampling/sampling_logp_difference/max': 1.5392913818359375, 'sampling/importance_sampling_ratio/min': 0.21453307569026947, 'sampling/importance_sampling_ratio/mean': 1.0373806953430176, 'sampling/importance_sampling_ratio/max': 1.993828296661377, 'entropy': 1.2479336559772491, 'clip_ratio/low_mean': 0.025462962687015533, 'clip_ratio/low_min': 0.025462962687015533, 'clip_ratio/high_mean': 0.10990368854254484, 'clip_ratio/high_max': 0.10990368854254484, 'clip_ratio/region_mean': 0.13536665122956038, 'reward_total_mean': 0.5041776895523071, 'reward_meter_mean': 0.8282565474510193, 'reward_meter_std': 0.32117608189582825, 'reward_count_adherence_mean': 0.71875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9734363555908203, 'reward_repeat_soft_std': 0.014392375946044922, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5041776895523071, 'reward_total_composite_std': 0.22154462337493896, 'epoch': 0.11}

 34%|███▎      | 1113/3300 [2:46:18<4:48:50,  7.92s/it]INFO 04-13 10:07:16 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1114/3300 [2:46:25<4:35:07,  7.55s/it]2026-04-13 10:07:22,962 | INFO | train_grpo_train | metrics_logged mode=train step=1114

                                                       
{'loss': 0.0179, 'grad_norm': 13.786949157714844, 'learning_rate': 6.627272727272728e-06, 'num_tokens': 1970614.0, 'completions/mean_length': 69.625, 'completions/min_length': 54.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9920821189880371, 'rewards/meter/std': 0.005062623415142298, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8936679363250732, 'rewards/repeat_soft/std': 0.05867590755224228, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5490313172340393, 'rewards/total_composite/std': 0.07378941774368286, 'reward': 0.5490313172340393, 'reward_std': 0.07378942519426346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1403108388185501, 'sampling/sampling_logp_difference/max': 1.1537199020385742, 'sampling/importance_sampling_ratio/min': 0.31546109914779663, 'sampling/importance_sampling_ratio/mean': 1.0259828567504883, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0104362666606903, 'clip_ratio/low_mean': 0.11913142912089825, 'clip_ratio/low_min': 0.11913142912089825, 'clip_ratio/high_mean': 0.00844594556838274, 'clip_ratio/high_max': 0.00844594556838274, 'clip_ratio/region_mean': 0.127577374689281, 'reward_total_mean': 0.5490313172340393, 'reward_meter_mean': 0.9920821189880371, 'reward_meter_std': 0.005062623415142298, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8936679363250732, 'reward_repeat_soft_std': 0.05867590755224228, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5490313172340393, 'reward_total_composite_std': 0.07378941774368286, 'epoch': 0.11}

 34%|███▍      | 1114/3300 [2:46:25<4:35:07,  7.55s/it]INFO 04-13 10:07:23 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1115/3300 [2:46:31<4:19:26,  7.12s/it]2026-04-13 10:07:29,090 | INFO | train_grpo_train | metrics_logged mode=train step=1115

                                                       
{'loss': 0.0496, 'grad_norm': 10.43582534790039, 'learning_rate': 6.624242424242425e-06, 'num_tokens': 1972654.0, 'completions/mean_length': 69.0, 'completions/min_length': 58.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.6927752494812012, 'rewards/meter/std': 0.350618451833725, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9682995080947876, 'rewards/repeat_soft/std': 0.024794267490506172, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5085471868515015, 'rewards/total_composite/std': 0.1273602545261383, 'reward': 0.5085471868515015, 'reward_std': 0.1273602545261383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15537992119789124, 'sampling/sampling_logp_difference/max': 1.9183073043823242, 'sampling/importance_sampling_ratio/min': 0.14685533940792084, 'sampling/importance_sampling_ratio/mean': 1.0243021249771118, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1463446840643883, 'clip_ratio/low_mean': 0.0369216650724411, 'clip_ratio/low_min': 0.0369216650724411, 'clip_ratio/high_mean': 0.08321494609117508, 'clip_ratio/high_max': 0.08321494609117508, 'clip_ratio/region_mean': 0.12013661116361618, 'reward_total_mean': 0.5085471868515015, 'reward_meter_mean': 0.6927752494812012, 'reward_meter_std': 0.350618451833725, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9682995080947876, 'reward_repeat_soft_std': 0.024794267490506172, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5085471868515015, 'reward_total_composite_std': 0.1273602545261383, 'epoch': 0.11}

 34%|███▍      | 1115/3300 [2:46:31<4:19:26,  7.12s/it]INFO 04-13 10:07:29 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1116/3300 [2:46:37<4:06:55,  6.78s/it]2026-04-13 10:07:35,078 | INFO | train_grpo_train | metrics_logged mode=train step=1116

                                                       
{'loss': 0.0279, 'grad_norm': 12.833725929260254, 'learning_rate': 6.621212121212121e-06, 'num_tokens': 1974263.0, 'completions/mean_length': 43.125, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7856926918029785, 'rewards/meter/std': 0.21308495104312897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9201638102531433, 'rewards/repeat_soft/std': 0.08705054223537445, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6049907207489014, 'rewards/total_composite/std': 0.10036706924438477, 'reward': 0.6049907207489014, 'reward_std': 0.10036708414554596, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14687471091747284, 'sampling/sampling_logp_difference/max': 2.573352336883545, 'sampling/importance_sampling_ratio/min': 0.0762794017791748, 'sampling/importance_sampling_ratio/mean': 1.0123547315597534, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.838761143386364, 'clip_ratio/low_mean': 0.08220163360238075, 'clip_ratio/low_min': 0.08220163360238075, 'clip_ratio/high_mean': 0.04869186133146286, 'clip_ratio/high_max': 0.04869186133146286, 'clip_ratio/region_mean': 0.1308934949338436, 'reward_total_mean': 0.6049907207489014, 'reward_meter_mean': 0.7856926918029785, 'reward_meter_std': 0.21308495104312897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9201638102531433, 'reward_repeat_soft_std': 0.08705054223537445, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6049907207489014, 'reward_total_composite_std': 0.10036706924438477, 'epoch': 0.11}

 34%|███▍      | 1116/3300 [2:46:37<4:06:55,  6.78s/it]INFO 04-13 10:07:35 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1117/3300 [2:46:43<3:57:31,  6.53s/it]2026-04-13 10:07:41,009 | INFO | train_grpo_train | metrics_logged mode=train step=1117

                                                       
{'loss': 0.0666, 'grad_norm': 15.617884635925293, 'learning_rate': 6.618181818181819e-06, 'num_tokens': 1975941.0, 'completions/mean_length': 36.75, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9262579083442688, 'rewards/meter/std': 0.0824204757809639, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9577685594558716, 'rewards/repeat_soft/std': 0.050275254994630814, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012580931186676, 'rewards/total_composite/mean': 0.6820363998413086, 'rewards/total_composite/std': 0.12780505418777466, 'reward': 0.6820363998413086, 'reward_std': 0.12780506908893585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15699218213558197, 'sampling/sampling_logp_difference/max': 5.221889495849609, 'sampling/importance_sampling_ratio/min': 0.005397121887654066, 'sampling/importance_sampling_ratio/mean': 1.0000742673873901, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7281496822834015, 'clip_ratio/low_mean': 0.0680940211750567, 'clip_ratio/low_min': 0.0680940211750567, 'clip_ratio/high_mean': 0.03378378227353096, 'clip_ratio/high_max': 0.03378378227353096, 'clip_ratio/region_mean': 0.10187780344858766, 'reward_total_mean': 0.6820363998413086, 'reward_meter_mean': 0.9262579083442688, 'reward_meter_std': 0.0824204757809639, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9577685594558716, 'reward_repeat_soft_std': 0.050275254994630814, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012580931186676, 'reward_total_composite_mean': 0.6820363998413086, 'reward_total_composite_std': 0.12780505418777466, 'epoch': 0.11}

 34%|███▍      | 1117/3300 [2:46:43<3:57:31,  6.53s/it]INFO 04-13 10:07:41 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1118/3300 [2:46:50<4:00:42,  6.62s/it]2026-04-13 10:07:47,861 | INFO | train_grpo_train | metrics_logged mode=train step=1118

                                                       
{'loss': 0.0051, 'grad_norm': 9.490935325622559, 'learning_rate': 6.615151515151516e-06, 'num_tokens': 1978404.0, 'completions/mean_length': 94.875, 'completions/min_length': 78.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.875, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.9267277717590332, 'rewards/meter/std': 0.16889002919197083, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9018123149871826, 'rewards/repeat_soft/std': 0.04975356161594391, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5456591844558716, 'rewards/total_composite/std': 0.04382103309035301, 'reward': 0.5456591844558716, 'reward_std': 0.04382103309035301, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12349020689725876, 'sampling/sampling_logp_difference/max': 1.7906038761138916, 'sampling/importance_sampling_ratio/min': 0.17763397097587585, 'sampling/importance_sampling_ratio/mean': 1.005535364151001, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8234644457697868, 'clip_ratio/low_mean': 0.02817411068826914, 'clip_ratio/low_min': 0.02817411068826914, 'clip_ratio/high_mean': 0.09194431267678738, 'clip_ratio/high_max': 0.09194431267678738, 'clip_ratio/region_mean': 0.12011842336505651, 'reward_total_mean': 0.5456591844558716, 'reward_meter_mean': 0.9267277717590332, 'reward_meter_std': 0.16889002919197083, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9018123149871826, 'reward_repeat_soft_std': 0.04975356161594391, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5456591844558716, 'reward_total_composite_std': 0.04382103309035301, 'epoch': 0.11}

 34%|███▍      | 1118/3300 [2:46:50<4:00:42,  6.62s/it]INFO 04-13 10:07:48 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1119/3300 [2:46:57<4:06:18,  6.78s/it]2026-04-13 10:07:55,022 | INFO | train_grpo_train | metrics_logged mode=train step=1119

                                                       
{'loss': 0.0328, 'grad_norm': 12.819561004638672, 'learning_rate': 6.612121212121213e-06, 'num_tokens': 1980241.0, 'completions/mean_length': 71.625, 'completions/min_length': 60.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.625, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9107191562652588, 'rewards/meter/std': 0.2147468477487564, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8732104301452637, 'rewards/repeat_soft/std': 0.06761850416660309, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.16291432082653046, 'rewards/total_composite/mean': 0.5724388360977173, 'rewards/total_composite/std': 0.13450887799263, 'reward': 0.5724388360977173, 'reward_std': 0.1345088630914688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15446336567401886, 'sampling/sampling_logp_difference/max': 1.814864158630371, 'sampling/importance_sampling_ratio/min': 0.16286003589630127, 'sampling/importance_sampling_ratio/mean': 1.038291096687317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.326532170176506, 'clip_ratio/low_mean': 0.07853138726204634, 'clip_ratio/low_min': 0.07853138726204634, 'clip_ratio/high_mean': 0.028238224796950817, 'clip_ratio/high_max': 0.028238224796950817, 'clip_ratio/region_mean': 0.10676961205899715, 'reward_total_mean': 0.5724388360977173, 'reward_meter_mean': 0.9107191562652588, 'reward_meter_std': 0.2147468477487564, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8732104301452637, 'reward_repeat_soft_std': 0.06761850416660309, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.16291432082653046, 'reward_total_composite_mean': 0.5724388360977173, 'reward_total_composite_std': 0.13450887799263, 'epoch': 0.11}

 34%|███▍      | 1119/3300 [2:46:57<4:06:18,  6.78s/it]INFO 04-13 10:07:55 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1120/3300 [2:47:04<4:10:02,  6.88s/it]2026-04-13 10:08:02,117 | INFO | train_grpo_train | metrics_logged mode=train step=1120

                                                       
{'loss': 0.0482, 'grad_norm': 9.773730278015137, 'learning_rate': 6.609090909090909e-06, 'num_tokens': 1982294.0, 'completions/mean_length': 76.625, 'completions/min_length': 63.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.6077883243560791, 'rewards/meter/std': 0.3168320655822754, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.837867021560669, 'rewards/repeat_soft/std': 0.08495821058750153, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.46740472316741943, 'rewards/total_composite/std': 0.13453896343708038, 'reward': 0.46740472316741943, 'reward_std': 0.13453896343708038, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14241869747638702, 'sampling/sampling_logp_difference/max': 2.31327223777771, 'sampling/importance_sampling_ratio/min': 0.09893697500228882, 'sampling/importance_sampling_ratio/mean': 0.9904230833053589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1188526898622513, 'clip_ratio/low_mean': 0.06325502693653107, 'clip_ratio/low_min': 0.06325502693653107, 'clip_ratio/high_mean': 0.051562500186264515, 'clip_ratio/high_max': 0.051562500186264515, 'clip_ratio/region_mean': 0.11481752712279558, 'reward_total_mean': 0.46740472316741943, 'reward_meter_mean': 0.6077883243560791, 'reward_meter_std': 0.3168320655822754, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.837867021560669, 'reward_repeat_soft_std': 0.08495821058750153, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.46740472316741943, 'reward_total_composite_std': 0.13453896343708038, 'epoch': 0.11}

 34%|███▍      | 1120/3300 [2:47:04<4:10:02,  6.88s/it]INFO 04-13 10:08:02 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1121/3300 [2:47:10<4:05:22,  6.76s/it]2026-04-13 10:08:08,613 | INFO | train_grpo_train | metrics_logged mode=train step=1121

                                                       
{'loss': -0.0491, 'grad_norm': 10.90406608581543, 'learning_rate': 6.606060606060607e-06, 'num_tokens': 1984213.0, 'completions/mean_length': 60.875, 'completions/min_length': 47.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9739933013916016, 'rewards/meter/std': 0.03326532989740372, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9222980737686157, 'rewards/repeat_soft/std': 0.04028952121734619, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5566431879997253, 'rewards/total_composite/std': 0.01162195298820734, 'reward': 0.5566431879997253, 'reward_std': 0.011621958576142788, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1495954841375351, 'sampling/sampling_logp_difference/max': 1.8567464351654053, 'sampling/importance_sampling_ratio/min': 0.15617994964122772, 'sampling/importance_sampling_ratio/mean': 1.0220969915390015, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1848102807998657, 'clip_ratio/low_mean': 0.03970585763454437, 'clip_ratio/low_min': 0.03970585763454437, 'clip_ratio/high_mean': 0.08167519699782133, 'clip_ratio/high_max': 0.08167519699782133, 'clip_ratio/region_mean': 0.1213810546323657, 'reward_total_mean': 0.5566431879997253, 'reward_meter_mean': 0.9739933013916016, 'reward_meter_std': 0.03326532989740372, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9222980737686157, 'reward_repeat_soft_std': 0.04028952121734619, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5566431879997253, 'reward_total_composite_std': 0.01162195298820734, 'epoch': 0.11}

 34%|███▍      | 1121/3300 [2:47:10<4:05:22,  6.76s/it]INFO 04-13 10:08:08 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1122/3300 [2:47:17<4:04:19,  6.73s/it]2026-04-13 10:08:15,249 | INFO | train_grpo_train | metrics_logged mode=train step=1122

                                                       
{'loss': -0.0273, 'grad_norm': 8.987536430358887, 'learning_rate': 6.603030303030303e-06, 'num_tokens': 1986309.0, 'completions/mean_length': 77.0, 'completions/min_length': 68.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.8478034734725952, 'rewards/meter/std': 0.17993468046188354, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.94336998462677, 'rewards/repeat_soft/std': 0.024993643164634705, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.5010881423950195, 'rewards/total_composite/std': 0.06411530077457428, 'reward': 0.5010881423950195, 'reward_std': 0.06411530822515488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1549995243549347, 'sampling/sampling_logp_difference/max': 2.250798225402832, 'sampling/importance_sampling_ratio/min': 0.10531511902809143, 'sampling/importance_sampling_ratio/mean': 1.0179111957550049, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4146091043949127, 'clip_ratio/low_mean': 0.05337740574032068, 'clip_ratio/low_min': 0.05337740574032068, 'clip_ratio/high_mean': 0.06918917689472437, 'clip_ratio/high_max': 0.06918917689472437, 'clip_ratio/region_mean': 0.12256658263504505, 'reward_total_mean': 0.5010881423950195, 'reward_meter_mean': 0.8478034734725952, 'reward_meter_std': 0.17993468046188354, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.94336998462677, 'reward_repeat_soft_std': 0.024993643164634705, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.5010881423950195, 'reward_total_composite_std': 0.06411530077457428, 'epoch': 0.11}

 34%|███▍      | 1122/3300 [2:47:17<4:04:19,  6.73s/it]INFO 04-13 10:08:15 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1123/3300 [2:47:23<3:56:48,  6.53s/it]2026-04-13 10:08:21,297 | INFO | train_grpo_train | metrics_logged mode=train step=1123

                                                       
{'loss': 0.0172, 'grad_norm': 11.716697692871094, 'learning_rate': 6.600000000000001e-06, 'num_tokens': 1988086.0, 'completions/mean_length': 55.125, 'completions/min_length': 47.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9374167323112488, 'rewards/meter/std': 0.10473205149173737, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9076154232025146, 'rewards/repeat_soft/std': 0.08004806190729141, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.5522748231887817, 'rewards/total_composite/std': 0.0624866783618927, 'reward': 0.5522748231887817, 'reward_std': 0.0624866746366024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1449204832315445, 'sampling/sampling_logp_difference/max': 1.5615653991699219, 'sampling/importance_sampling_ratio/min': 0.2098073810338974, 'sampling/importance_sampling_ratio/mean': 1.012521505355835, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2464976161718369, 'clip_ratio/low_mean': 0.04081095149740577, 'clip_ratio/low_min': 0.04081095149740577, 'clip_ratio/high_mean': 0.07589972019195557, 'clip_ratio/high_max': 0.07589972019195557, 'clip_ratio/region_mean': 0.11671067168936133, 'reward_total_mean': 0.5522748231887817, 'reward_meter_mean': 0.9374167323112488, 'reward_meter_std': 0.10473205149173737, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9076154232025146, 'reward_repeat_soft_std': 0.08004806190729141, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.5522748231887817, 'reward_total_composite_std': 0.0624866783618927, 'epoch': 0.11}

 34%|███▍      | 1123/3300 [2:47:23<3:56:48,  6.53s/it]INFO 04-13 10:08:21 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1124/3300 [2:47:30<3:59:44,  6.61s/it]2026-04-13 10:08:28,108 | INFO | train_grpo_train | metrics_logged mode=train step=1124

                                                       
{'loss': 0.0262, 'grad_norm': 9.961459159851074, 'learning_rate': 6.596969696969698e-06, 'num_tokens': 1990035.0, 'completions/mean_length': 62.625, 'completions/min_length': 54.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.7475841045379639, 'rewards/meter/std': 0.19611549377441406, 'rewards/count_adherence/mean': 0.7708333134651184, 'rewards/count_adherence/std': 0.0862581729888916, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8471733927726746, 'rewards/repeat_soft/std': 0.07111194729804993, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.4932330250740051, 'rewards/total_composite/std': 0.08194217085838318, 'reward': 0.4932330250740051, 'reward_std': 0.08194217085838318, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10107006132602692, 'sampling/sampling_logp_difference/max': 3.2778851985931396, 'sampling/importance_sampling_ratio/min': 0.236572265625, 'sampling/importance_sampling_ratio/mean': 1.0227466821670532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5160208456218243, 'clip_ratio/low_mean': 0.03161199390888214, 'clip_ratio/low_min': 0.03161199390888214, 'clip_ratio/high_mean': 0.06855028681457043, 'clip_ratio/high_max': 0.06855028681457043, 'clip_ratio/region_mean': 0.10016228072345257, 'reward_total_mean': 0.4932330250740051, 'reward_meter_mean': 0.7475841045379639, 'reward_meter_std': 0.19611549377441406, 'reward_count_adherence_mean': 0.7708333134651184, 'reward_count_adherence_std': 0.0862581729888916, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8471733927726746, 'reward_repeat_soft_std': 0.07111194729804993, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.4932330250740051, 'reward_total_composite_std': 0.08194217085838318, 'epoch': 0.11}

 34%|███▍      | 1124/3300 [2:47:30<3:59:44,  6.61s/it]INFO 04-13 10:08:28 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1125/3300 [2:47:36<3:55:05,  6.49s/it]2026-04-13 10:08:34,298 | INFO | train_grpo_train | metrics_logged mode=train step=1125

                                                       
{'loss': -0.0262, 'grad_norm': 7.455492973327637, 'learning_rate': 6.593939393939395e-06, 'num_tokens': 1991683.0, 'completions/mean_length': 41.0, 'completions/min_length': 37.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8595741391181946, 'rewards/meter/std': 0.3456464409828186, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8770755529403687, 'rewards/repeat_soft/std': 0.11364620178937912, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5710352659225464, 'rewards/total_composite/std': 0.11277426779270172, 'reward': 0.5710352659225464, 'reward_std': 0.11277425289154053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12854094803333282, 'sampling/sampling_logp_difference/max': 1.1401591300964355, 'sampling/importance_sampling_ratio/min': 0.31976813077926636, 'sampling/importance_sampling_ratio/mean': 1.043609380722046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0445300117135048, 'clip_ratio/low_mean': 0.013513513840734959, 'clip_ratio/low_min': 0.013513513840734959, 'clip_ratio/high_mean': 0.12344709038734436, 'clip_ratio/high_max': 0.12344709038734436, 'clip_ratio/region_mean': 0.13696060422807932, 'reward_total_mean': 0.5710352659225464, 'reward_meter_mean': 0.8595741391181946, 'reward_meter_std': 0.3456464409828186, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8770755529403687, 'reward_repeat_soft_std': 0.11364620178937912, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5710352659225464, 'reward_total_composite_std': 0.11277426779270172, 'epoch': 0.11}

 34%|███▍      | 1125/3300 [2:47:36<3:55:05,  6.49s/it]INFO 04-13 10:08:34 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1126/3300 [2:47:42<3:51:19,  6.38s/it]2026-04-13 10:08:40,447 | INFO | train_grpo_train | metrics_logged mode=train step=1126

                                                       
{'loss': -0.0292, 'grad_norm': 14.455120086669922, 'learning_rate': 6.590909090909091e-06, 'num_tokens': 1993459.0, 'completions/mean_length': 42.0, 'completions/min_length': 34.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.9213700294494629, 'rewards/meter/std': 0.11122234910726547, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9119148850440979, 'rewards/repeat_soft/std': 0.0637424886226654, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6664919257164001, 'rewards/total_composite/std': 0.14875824749469757, 'reward': 0.6664919257164001, 'reward_std': 0.14875824749469757, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14410021901130676, 'sampling/sampling_logp_difference/max': 1.0909414291381836, 'sampling/importance_sampling_ratio/min': 0.33590012788772583, 'sampling/importance_sampling_ratio/mean': 1.0286000967025757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.042415238916874, 'clip_ratio/low_mean': 0.09365024883300066, 'clip_ratio/low_min': 0.09365024883300066, 'clip_ratio/high_mean': 0.022001934237778187, 'clip_ratio/high_max': 0.022001934237778187, 'clip_ratio/region_mean': 0.11565218307077885, 'reward_total_mean': 0.6664919257164001, 'reward_meter_mean': 0.9213700294494629, 'reward_meter_std': 0.11122234910726547, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9119148850440979, 'reward_repeat_soft_std': 0.0637424886226654, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6664919257164001, 'reward_total_composite_std': 0.14875824749469757, 'epoch': 0.11}

 34%|███▍      | 1126/3300 [2:47:42<3:51:19,  6.38s/it]INFO 04-13 10:08:40 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1127/3300 [2:47:53<4:40:16,  7.74s/it]2026-04-13 10:08:51,349 | INFO | train_grpo_train | metrics_logged mode=train step=1127

                                                       
{'loss': -0.0753, 'grad_norm': 2.7037055492401123, 'learning_rate': 6.5878787878787885e-06, 'num_tokens': 1994871.0, 'completions/mean_length': 84.5, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.428571701049805, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7526909112930298, 'rewards/meter/std': 0.36496224999427795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9083139300346375, 'rewards/repeat_soft/std': 0.07009585946798325, 'rewards/judge_quality/mean': 0.32249999046325684, 'rewards/judge_quality/std': 0.1528538018465042, 'rewards/total_composite/mean': 0.4741763472557068, 'rewards/total_composite/std': 0.2111952304840088, 'reward': 0.4741763472557068, 'reward_std': 0.2111952304840088, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13462936878204346, 'sampling/sampling_logp_difference/max': 0.811007022857666, 'sampling/importance_sampling_ratio/min': 0.45685768127441406, 'sampling/importance_sampling_ratio/mean': 1.0266838073730469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.125832237303257, 'clip_ratio/low_mean': 0.024038461968302727, 'clip_ratio/low_min': 0.024038461968302727, 'clip_ratio/high_mean': 0.11665728315711021, 'clip_ratio/high_max': 0.11665728315711021, 'clip_ratio/region_mean': 0.14069574512541294, 'reward_total_mean': 0.4741763472557068, 'reward_meter_mean': 0.7526909112930298, 'reward_meter_std': 0.36496224999427795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9083139300346375, 'reward_repeat_soft_std': 0.07009585946798325, 'reward_judge_quality_mean': 0.32249999046325684, 'reward_judge_quality_std': 0.1528538018465042, 'reward_total_composite_mean': 0.4741763472557068, 'reward_total_composite_std': 0.2111952304840088, 'epoch': 0.11}

 34%|███▍      | 1127/3300 [2:47:53<4:40:16,  7.74s/it]INFO 04-13 10:08:51 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1128/3300 [2:48:04<5:19:57,  8.84s/it]2026-04-13 10:09:02,753 | INFO | train_grpo_train | metrics_logged mode=train step=1128

                                                       
{'loss': -0.0627, 'grad_norm': 3.1919827461242676, 'learning_rate': 6.584848484848485e-06, 'num_tokens': 1996325.0, 'completions/mean_length': 91.75, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 31.71428680419922, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6264134049415588, 'rewards/meter/std': 0.4235175848007202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.998651385307312, 'rewards/repeat_soft/std': 0.003814458381384611, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.5159874558448792, 'rewards/total_composite/std': 0.2651650309562683, 'reward': 0.5159874558448792, 'reward_std': 0.2651650309562683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14543375372886658, 'sampling/sampling_logp_difference/max': 1.6496882438659668, 'sampling/importance_sampling_ratio/min': 0.19210979342460632, 'sampling/importance_sampling_ratio/mean': 1.036872386932373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6218914948403835, 'clip_ratio/low_mean': 0.036608658730983734, 'clip_ratio/low_min': 0.036608658730983734, 'clip_ratio/high_mean': 0.09086930099874735, 'clip_ratio/high_max': 0.09086930099874735, 'clip_ratio/region_mean': 0.12747795972973108, 'reward_total_mean': 0.5159874558448792, 'reward_meter_mean': 0.6264134049415588, 'reward_meter_std': 0.4235175848007202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.998651385307312, 'reward_repeat_soft_std': 0.003814458381384611, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.5159874558448792, 'reward_total_composite_std': 0.2651650309562683, 'epoch': 0.11}

 34%|███▍      | 1128/3300 [2:48:04<5:19:57,  8.84s/it]INFO 04-13 10:09:02 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1129/3300 [2:48:15<5:43:29,  9.49s/it]2026-04-13 10:09:13,772 | INFO | train_grpo_train | metrics_logged mode=train step=1129

                                                       
{'loss': -0.1115, 'grad_norm': 2.871786117553711, 'learning_rate': 6.581818181818182e-06, 'num_tokens': 1997814.0, 'completions/mean_length': 98.125, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9012106657028198, 'rewards/meter/std': 0.17986968159675598, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9321990609169006, 'rewards/repeat_soft/std': 0.03417094051837921, 'rewards/judge_quality/mean': 0.4012500047683716, 'rewards/judge_quality/std': 0.19773268699645996, 'rewards/total_composite/mean': 0.5356782078742981, 'rewards/total_composite/std': 0.24061591923236847, 'reward': 0.5356782078742981, 'reward_std': 0.24061590433120728, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1484362781047821, 'sampling/sampling_logp_difference/max': 1.0686688423156738, 'sampling/importance_sampling_ratio/min': 0.3434654176235199, 'sampling/importance_sampling_ratio/mean': 1.011346459388733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0026362836360931, 'clip_ratio/low_mean': 0.013888888992369175, 'clip_ratio/low_min': 0.013888888992369175, 'clip_ratio/high_mean': 0.08541359659284353, 'clip_ratio/high_max': 0.08541359659284353, 'clip_ratio/region_mean': 0.09930248558521271, 'reward_total_mean': 0.5356782078742981, 'reward_meter_mean': 0.9012106657028198, 'reward_meter_std': 0.17986968159675598, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9321990609169006, 'reward_repeat_soft_std': 0.03417094051837921, 'reward_judge_quality_mean': 0.4012500047683716, 'reward_judge_quality_std': 0.19773268699645996, 'reward_total_composite_mean': 0.5356782078742981, 'reward_total_composite_std': 0.24061591923236847, 'epoch': 0.11}

 34%|███▍      | 1129/3300 [2:48:16<5:43:29,  9.49s/it]INFO 04-13 10:09:13 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1130/3300 [2:48:21<5:04:28,  8.42s/it]2026-04-13 10:09:19,722 | INFO | train_grpo_train | metrics_logged mode=train step=1130

                                                       
{'loss': 0.0142, 'grad_norm': 9.468161582946777, 'learning_rate': 6.578787878787879e-06, 'num_tokens': 1999386.0, 'completions/mean_length': 44.5, 'completions/min_length': 43.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5640295147895813, 'rewards/meter/std': 0.34981825947761536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8968439102172852, 'rewards/repeat_soft/std': 0.054242637008428574, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5763897895812988, 'rewards/total_composite/std': 0.17131243646144867, 'reward': 0.5763897895812988, 'reward_std': 0.17131242156028748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10996052622795105, 'sampling/sampling_logp_difference/max': 2.1871838569641113, 'sampling/importance_sampling_ratio/min': 0.11223236471414566, 'sampling/importance_sampling_ratio/mean': 0.9965682625770569, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5740429349243641, 'clip_ratio/low_mean': 0.05572951911017299, 'clip_ratio/low_min': 0.05572951911017299, 'clip_ratio/high_mean': 0.025766385719180107, 'clip_ratio/high_max': 0.025766385719180107, 'clip_ratio/region_mean': 0.0814959048293531, 'reward_total_mean': 0.5763897895812988, 'reward_meter_mean': 0.5640295147895813, 'reward_meter_std': 0.34981825947761536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8968439102172852, 'reward_repeat_soft_std': 0.054242637008428574, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5763897895812988, 'reward_total_composite_std': 0.17131243646144867, 'epoch': 0.11}

 34%|███▍      | 1130/3300 [2:48:21<5:04:28,  8.42s/it]INFO 04-13 10:09:20 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1131/3300 [2:48:29<4:54:20,  8.14s/it]2026-04-13 10:09:27,219 | INFO | train_grpo_train | metrics_logged mode=train step=1131

                                                       
{'loss': -0.1178, 'grad_norm': 14.717326164245605, 'learning_rate': 6.575757575757577e-06, 'num_tokens': 2001222.0, 'completions/mean_length': 63.5, 'completions/min_length': 41.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8636382818222046, 'rewards/meter/std': 0.2630917727947235, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9741989374160767, 'rewards/repeat_soft/std': 0.04893767088651657, 'rewards/judge_quality/mean': 0.6575000286102295, 'rewards/judge_quality/std': 0.20658794045448303, 'rewards/total_composite/mean': 0.7093788385391235, 'rewards/total_composite/std': 0.18725824356079102, 'reward': 0.7093788385391235, 'reward_std': 0.18725824356079102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1483689695596695, 'sampling/sampling_logp_difference/max': 2.488309860229492, 'sampling/importance_sampling_ratio/min': 0.0830502137541771, 'sampling/importance_sampling_ratio/mean': 1.0160596370697021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6833672076463699, 'clip_ratio/low_mean': 0.05194827914237976, 'clip_ratio/low_min': 0.05194827914237976, 'clip_ratio/high_mean': 0.07146300747990608, 'clip_ratio/high_max': 0.07146300747990608, 'clip_ratio/region_mean': 0.12341128662228584, 'reward_total_mean': 0.7093788385391235, 'reward_meter_mean': 0.8636382818222046, 'reward_meter_std': 0.2630917727947235, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9741989374160767, 'reward_repeat_soft_std': 0.04893767088651657, 'reward_judge_quality_mean': 0.6575000286102295, 'reward_judge_quality_std': 0.20658794045448303, 'reward_total_composite_mean': 0.7093788385391235, 'reward_total_composite_std': 0.18725824356079102, 'epoch': 0.11}

 34%|███▍      | 1131/3300 [2:48:29<4:54:20,  8.14s/it]INFO 04-13 10:09:27 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1132/3300 [2:48:35<4:38:00,  7.69s/it]2026-04-13 10:09:33,865 | INFO | train_grpo_train | metrics_logged mode=train step=1132

                                                       
{'loss': 0.0658, 'grad_norm': 12.287823677062988, 'learning_rate': 6.572727272727273e-06, 'num_tokens': 2002814.0, 'completions/mean_length': 40.0, 'completions/min_length': 34.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9884093999862671, 'rewards/meter/std': 0.006536794826388359, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9248701333999634, 'rewards/repeat_soft/std': 0.06530743837356567, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.61822110414505, 'rewards/total_composite/std': 0.017233116552233696, 'reward': 0.61822110414505, 'reward_std': 0.01723311096429825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14291036128997803, 'sampling/sampling_logp_difference/max': 2.1756439208984375, 'sampling/importance_sampling_ratio/min': 0.11353502422571182, 'sampling/importance_sampling_ratio/mean': 1.0182663202285767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1009890586137772, 'clip_ratio/low_mean': 0.06933821551501751, 'clip_ratio/low_min': 0.06933821551501751, 'clip_ratio/high_mean': 0.07251258380711079, 'clip_ratio/high_max': 0.07251258380711079, 'clip_ratio/region_mean': 0.1418507993221283, 'reward_total_mean': 0.61822110414505, 'reward_meter_mean': 0.9884093999862671, 'reward_meter_std': 0.006536794826388359, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9248701333999634, 'reward_repeat_soft_std': 0.06530743837356567, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.61822110414505, 'reward_total_composite_std': 0.017233116552233696, 'epoch': 0.11}

 34%|███▍      | 1132/3300 [2:48:36<4:38:00,  7.69s/it]INFO 04-13 10:09:34 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1133/3300 [2:48:47<5:20:47,  8.88s/it]2026-04-13 10:09:45,487 | INFO | train_grpo_train | metrics_logged mode=train step=1133

                                                       
{'loss': -0.1135, 'grad_norm': 2.8444020748138428, 'learning_rate': 6.56969696969697e-06, 'num_tokens': 2004482.0, 'completions/mean_length': 102.5, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 44.000003814697266, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6006203293800354, 'rewards/meter/std': 0.4255194664001465, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9695420861244202, 'rewards/repeat_soft/std': 0.0406782403588295, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.4698367714881897, 'rewards/total_composite/std': 0.21603558957576752, 'reward': 0.4698367714881897, 'reward_std': 0.21603557467460632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1557750254869461, 'sampling/sampling_logp_difference/max': 1.7177009582519531, 'sampling/importance_sampling_ratio/min': 0.17947831749916077, 'sampling/importance_sampling_ratio/mean': 1.0362001657485962, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0773121491074562, 'clip_ratio/low_mean': 0.037367021664977074, 'clip_ratio/low_min': 0.037367021664977074, 'clip_ratio/high_mean': 0.07085578422993422, 'clip_ratio/high_max': 0.07085578422993422, 'clip_ratio/region_mean': 0.10822280589491129, 'reward_total_mean': 0.4698367714881897, 'reward_meter_mean': 0.6006203293800354, 'reward_meter_std': 0.4255194664001465, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9695420861244202, 'reward_repeat_soft_std': 0.0406782403588295, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.4698367714881897, 'reward_total_composite_std': 0.21603558957576752, 'epoch': 0.11}

 34%|███▍      | 1133/3300 [2:48:47<5:20:47,  8.88s/it]INFO 04-13 10:09:45 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1134/3300 [2:48:54<4:53:31,  8.13s/it]2026-04-13 10:09:51,860 | INFO | train_grpo_train | metrics_logged mode=train step=1134

                                                       
{'loss': 0.0101, 'grad_norm': 10.173377990722656, 'learning_rate': 6.566666666666667e-06, 'num_tokens': 2006521.0, 'completions/mean_length': 66.875, 'completions/min_length': 54.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.875, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9178117513656616, 'rewards/meter/std': 0.11717532575130463, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9686312675476074, 'rewards/repeat_soft/std': 0.019978869706392288, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.5446088910102844, 'rewards/total_composite/std': 0.05727742984890938, 'reward': 0.5446088910102844, 'reward_std': 0.05727742612361908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18602125346660614, 'sampling/sampling_logp_difference/max': 1.8618534803390503, 'sampling/importance_sampling_ratio/min': 0.155384361743927, 'sampling/importance_sampling_ratio/mean': 1.010296106338501, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.7691242843866348, 'clip_ratio/low_mean': 0.0799316605553031, 'clip_ratio/low_min': 0.0799316605553031, 'clip_ratio/high_mean': 0.06743836030364037, 'clip_ratio/high_max': 0.06743836030364037, 'clip_ratio/region_mean': 0.14737002085894346, 'reward_total_mean': 0.5446088910102844, 'reward_meter_mean': 0.9178117513656616, 'reward_meter_std': 0.11717532575130463, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9686312675476074, 'reward_repeat_soft_std': 0.019978869706392288, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.5446088910102844, 'reward_total_composite_std': 0.05727742984890938, 'epoch': 0.11}

 34%|███▍      | 1134/3300 [2:48:54<4:53:31,  8.13s/it]INFO 04-13 10:09:52 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1135/3300 [2:49:00<4:36:21,  7.66s/it]2026-04-13 10:09:58,426 | INFO | train_grpo_train | metrics_logged mode=train step=1135

                                                       
{'loss': 0.0051, 'grad_norm': 10.730149269104004, 'learning_rate': 6.563636363636364e-06, 'num_tokens': 2008640.0, 'completions/mean_length': 80.875, 'completions/min_length': 76.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.5640273690223694, 'rewards/meter/std': 0.45509153604507446, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9958585500717163, 'rewards/repeat_soft/std': 0.0028345701284706593, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1440981924533844, 'rewards/total_composite/mean': 0.50409996509552, 'rewards/total_composite/std': 0.19318099319934845, 'reward': 0.50409996509552, 'reward_std': 0.19318099319934845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15317434072494507, 'sampling/sampling_logp_difference/max': 1.7908809185028076, 'sampling/importance_sampling_ratio/min': 0.16681316494941711, 'sampling/importance_sampling_ratio/mean': 1.0111745595932007, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2982793375849724, 'clip_ratio/low_mean': 0.07364986091852188, 'clip_ratio/low_min': 0.07364986091852188, 'clip_ratio/high_mean': 0.07550835236907005, 'clip_ratio/high_max': 0.07550835236907005, 'clip_ratio/region_mean': 0.14915821328759193, 'reward_total_mean': 0.50409996509552, 'reward_meter_mean': 0.5640273690223694, 'reward_meter_std': 0.45509153604507446, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9958585500717163, 'reward_repeat_soft_std': 0.0028345701284706593, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1440981924533844, 'reward_total_composite_mean': 0.50409996509552, 'reward_total_composite_std': 0.19318099319934845, 'epoch': 0.11}

 34%|███▍      | 1135/3300 [2:49:00<4:36:21,  7.66s/it]INFO 04-13 10:09:58 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1136/3300 [2:49:06<4:18:33,  7.17s/it]2026-04-13 10:10:04,445 | INFO | train_grpo_train | metrics_logged mode=train step=1136

                                                       
{'loss': 0.0649, 'grad_norm': 14.582271575927734, 'learning_rate': 6.56060606060606e-06, 'num_tokens': 2010272.0, 'completions/mean_length': 46.0, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9082609415054321, 'rewards/meter/std': 0.18846864998340607, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9454102516174316, 'rewards/repeat_soft/std': 0.04497417062520981, 'rewards/judge_quality/mean': 0.690000057220459, 'rewards/judge_quality/std': 0.22315914928913116, 'rewards/total_composite/mean': 0.7473007440567017, 'rewards/total_composite/std': 0.1645580530166626, 'reward': 0.7473007440567017, 'reward_std': 0.1645580530166626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1381937712430954, 'sampling/sampling_logp_difference/max': 1.9469313621520996, 'sampling/importance_sampling_ratio/min': 0.1427113264799118, 'sampling/importance_sampling_ratio/mean': 1.019103765487671, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0522950664162636, 'clip_ratio/low_mean': 0.051837935112416744, 'clip_ratio/low_min': 0.051837935112416744, 'clip_ratio/high_mean': 0.08532747812569141, 'clip_ratio/high_max': 0.08532747812569141, 'clip_ratio/region_mean': 0.13716541323810816, 'reward_total_mean': 0.7473007440567017, 'reward_meter_mean': 0.9082609415054321, 'reward_meter_std': 0.18846864998340607, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9454102516174316, 'reward_repeat_soft_std': 0.04497417062520981, 'reward_judge_quality_mean': 0.690000057220459, 'reward_judge_quality_std': 0.22315914928913116, 'reward_total_composite_mean': 0.7473007440567017, 'reward_total_composite_std': 0.1645580530166626, 'epoch': 0.11}

 34%|███▍      | 1136/3300 [2:49:06<4:18:33,  7.17s/it]INFO 04-13 10:10:04 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1137/3300 [2:49:18<5:05:17,  8.47s/it]2026-04-13 10:10:15,946 | INFO | train_grpo_train | metrics_logged mode=train step=1137

                                                       
{'loss': -0.1119, 'grad_norm': 3.1062843799591064, 'learning_rate': 6.5575757575757585e-06, 'num_tokens': 2011803.0, 'completions/mean_length': 100.375, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.57143020629883, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8784385919570923, 'rewards/meter/std': 0.2813236713409424, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9840956926345825, 'rewards/repeat_soft/std': 0.015371386893093586, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.2665520906448364, 'rewards/total_composite/mean': 0.5879136323928833, 'rewards/total_composite/std': 0.2742556035518646, 'reward': 0.5879136323928833, 'reward_std': 0.2742556035518646, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13355565071105957, 'sampling/sampling_logp_difference/max': 1.1007308959960938, 'sampling/importance_sampling_ratio/min': 0.33262789249420166, 'sampling/importance_sampling_ratio/mean': 1.0302464962005615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0896887928247452, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.10518785053864121, 'clip_ratio/high_max': 0.10518785053864121, 'clip_ratio/region_mean': 0.11456285091117024, 'reward_total_mean': 0.5879136323928833, 'reward_meter_mean': 0.8784385919570923, 'reward_meter_std': 0.2813236713409424, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9840956926345825, 'reward_repeat_soft_std': 0.015371386893093586, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.2665520906448364, 'reward_total_composite_mean': 0.5879136323928833, 'reward_total_composite_std': 0.2742556035518646, 'epoch': 0.11}

 34%|███▍      | 1137/3300 [2:49:18<5:05:17,  8.47s/it]INFO 04-13 10:10:16 [block_pool.py:378] Successfully reset prefix cache

 34%|███▍      | 1138/3300 [2:49:24<4:37:46,  7.71s/it]2026-04-13 10:10:21,884 | INFO | train_grpo_train | metrics_logged mode=train step=1138

                                                       
{'loss': 0.0294, 'grad_norm': 14.06969928741455, 'learning_rate': 6.554545454545455e-06, 'num_tokens': 2013311.0, 'completions/mean_length': 31.5, 'completions/min_length': 28.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.4891081750392914, 'rewards/meter/std': 0.3816754221916199, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9245496988296509, 'rewards/repeat_soft/std': 0.07919017225503922, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5305708050727844, 'rewards/total_composite/std': 0.17777620255947113, 'reward': 0.5305708050727844, 'reward_std': 0.17777620255947113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14619868993759155, 'sampling/sampling_logp_difference/max': 1.7692971229553223, 'sampling/importance_sampling_ratio/min': 0.17045274376869202, 'sampling/importance_sampling_ratio/mean': 0.9935710430145264, 'sampling/importance_sampling_ratio/max': 1.962249755859375, 'entropy': 0.8535003513097763, 'clip_ratio/low_mean': 0.06597222480922937, 'clip_ratio/low_min': 0.06597222480922937, 'clip_ratio/high_mean': 0.06373106129467487, 'clip_ratio/high_max': 0.06373106129467487, 'clip_ratio/region_mean': 0.12970328610390425, 'reward_total_mean': 0.5305708050727844, 'reward_meter_mean': 0.4891081750392914, 'reward_meter_std': 0.3816754221916199, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9245496988296509, 'reward_repeat_soft_std': 0.07919017225503922, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5305708050727844, 'reward_total_composite_std': 0.17777620255947113, 'epoch': 0.11}

 34%|███▍      | 1138/3300 [2:49:24<4:37:46,  7.71s/it]INFO 04-13 10:10:22 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1139/3300 [2:49:34<5:12:41,  8.68s/it]2026-04-13 10:10:32,836 | INFO | train_grpo_train | metrics_logged mode=train step=1139

                                                       
{'loss': -0.1127, 'grad_norm': 1.582905650138855, 'learning_rate': 6.551515151515152e-06, 'num_tokens': 2014823.0, 'completions/mean_length': 95.0, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.42857360839844, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.909144401550293, 'rewards/meter/std': 0.22075851261615753, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9273666143417358, 'rewards/repeat_soft/std': 0.059534281492233276, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.53968346118927, 'rewards/total_composite/std': 0.21839000284671783, 'reward': 0.53968346118927, 'reward_std': 0.21838998794555664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11558738350868225, 'sampling/sampling_logp_difference/max': 1.6367340087890625, 'sampling/importance_sampling_ratio/min': 0.19461461901664734, 'sampling/importance_sampling_ratio/mean': 1.0194705724716187, 'sampling/importance_sampling_ratio/max': 1.8529249429702759, 'entropy': 0.665039174258709, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08882309589534998, 'clip_ratio/high_max': 0.08882309589534998, 'clip_ratio/region_mean': 0.08882309589534998, 'reward_total_mean': 0.53968346118927, 'reward_meter_mean': 0.909144401550293, 'reward_meter_std': 0.22075851261615753, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9273666143417358, 'reward_repeat_soft_std': 0.059534281492233276, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.53968346118927, 'reward_total_composite_std': 0.21839000284671783, 'epoch': 0.11}

 35%|███▍      | 1139/3300 [2:49:35<5:12:41,  8.68s/it]INFO 04-13 10:10:33 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1140/3300 [2:49:47<5:50:42,  9.74s/it]2026-04-13 10:10:45,051 | INFO | train_grpo_train | metrics_logged mode=train step=1140

                                                       
{'loss': -0.1527, 'grad_norm': 3.4708738327026367, 'learning_rate': 6.5484848484848494e-06, 'num_tokens': 2016805.0, 'completions/mean_length': 115.75, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.142860412597656, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.35854005813598633, 'rewards/meter/std': 0.371934711933136, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9448579549789429, 'rewards/repeat_soft/std': 0.04365341365337372, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.18074746429920197, 'rewards/total_composite/mean': 0.35877785086631775, 'rewards/total_composite/std': 0.17009501159191132, 'reward': 0.35877785086631775, 'reward_std': 0.17009499669075012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.133514866232872, 'sampling/sampling_logp_difference/max': 1.222212791442871, 'sampling/importance_sampling_ratio/min': 0.29457759857177734, 'sampling/importance_sampling_ratio/mean': 1.0175260305404663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8386684954166412, 'clip_ratio/low_mean': 0.05946615990251303, 'clip_ratio/low_min': 0.05946615990251303, 'clip_ratio/high_mean': 0.06686240807175636, 'clip_ratio/high_max': 0.06686240807175636, 'clip_ratio/region_mean': 0.1263285679742694, 'reward_total_mean': 0.35877785086631775, 'reward_meter_mean': 0.35854005813598633, 'reward_meter_std': 0.371934711933136, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9448579549789429, 'reward_repeat_soft_std': 0.04365341365337372, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.18074746429920197, 'reward_total_composite_mean': 0.35877785086631775, 'reward_total_composite_std': 0.17009501159191132, 'epoch': 0.11}

 35%|███▍      | 1140/3300 [2:49:47<5:50:42,  9.74s/it]INFO 04-13 10:10:45 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1141/3300 [2:49:53<5:09:29,  8.60s/it]2026-04-13 10:10:50,989 | INFO | train_grpo_train | metrics_logged mode=train step=1141

                                                       
{'loss': 0.039, 'grad_norm': 12.786428451538086, 'learning_rate': 6.545454545454546e-06, 'num_tokens': 2018367.0, 'completions/mean_length': 41.25, 'completions/min_length': 39.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6099901795387268, 'rewards/meter/std': 0.3610473871231079, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8980588912963867, 'rewards/repeat_soft/std': 0.08926127105951309, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.574455976486206, 'rewards/total_composite/std': 0.16856855154037476, 'reward': 0.574455976486206, 'reward_std': 0.16856855154037476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14846009016036987, 'sampling/sampling_logp_difference/max': 1.8245267868041992, 'sampling/importance_sampling_ratio/min': 0.16129395365715027, 'sampling/importance_sampling_ratio/mean': 1.027154564857483, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.111771211028099, 'clip_ratio/low_mean': 0.056686026975512505, 'clip_ratio/low_min': 0.056686026975512505, 'clip_ratio/high_mean': 0.06761363800615072, 'clip_ratio/high_max': 0.06761363800615072, 'clip_ratio/region_mean': 0.12429966498166323, 'reward_total_mean': 0.574455976486206, 'reward_meter_mean': 0.6099901795387268, 'reward_meter_std': 0.3610473871231079, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8980588912963867, 'reward_repeat_soft_std': 0.08926127105951309, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.574455976486206, 'reward_total_composite_std': 0.16856855154037476, 'epoch': 0.11}

 35%|███▍      | 1141/3300 [2:49:53<5:09:29,  8.60s/it]INFO 04-13 10:10:51 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1142/3300 [2:50:03<5:24:41,  9.03s/it]2026-04-13 10:11:01,012 | INFO | train_grpo_train | metrics_logged mode=train step=1142

                                                       
{'loss': -0.0552, 'grad_norm': 12.804361343383789, 'learning_rate': 6.542424242424243e-06, 'num_tokens': 2020081.0, 'completions/mean_length': 52.25, 'completions/min_length': 43.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8938835859298706, 'rewards/meter/std': 0.09182074666023254, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.895793080329895, 'rewards/repeat_soft/std': 0.048144880682229996, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.6230459213256836, 'rewards/total_composite/std': 0.13977524638175964, 'reward': 0.6230459213256836, 'reward_std': 0.13977523148059845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12277662754058838, 'sampling/sampling_logp_difference/max': 1.502573013305664, 'sampling/importance_sampling_ratio/min': 0.22255678474903107, 'sampling/importance_sampling_ratio/mean': 1.0013785362243652, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6223147809505463, 'clip_ratio/low_mean': 0.07965277880430222, 'clip_ratio/low_min': 0.07965277880430222, 'clip_ratio/high_mean': 0.02281746082007885, 'clip_ratio/high_max': 0.02281746082007885, 'clip_ratio/region_mean': 0.10247023962438107, 'reward_total_mean': 0.6230459213256836, 'reward_meter_mean': 0.8938835859298706, 'reward_meter_std': 0.09182074666023254, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.895793080329895, 'reward_repeat_soft_std': 0.048144880682229996, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.6230459213256836, 'reward_total_composite_std': 0.13977524638175964, 'epoch': 0.11}

 35%|███▍      | 1142/3300 [2:50:03<5:24:41,  9.03s/it]INFO 04-13 10:11:01 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1143/3300 [2:50:09<4:52:22,  8.13s/it]2026-04-13 10:11:07,054 | INFO | train_grpo_train | metrics_logged mode=train step=1143

                                                       
{'loss': -0.0002, 'grad_norm': 15.238978385925293, 'learning_rate': 6.5393939393939395e-06, 'num_tokens': 2021646.0, 'completions/mean_length': 41.625, 'completions/min_length': 33.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9243209958076477, 'rewards/meter/std': 0.10788122564554214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.989984393119812, 'rewards/repeat_soft/std': 0.01322022546082735, 'rewards/judge_quality/mean': 0.5987499952316284, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.7137107849121094, 'rewards/total_composite/std': 0.1404464691877365, 'reward': 0.7137107849121094, 'reward_std': 0.14044643938541412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15316607058048248, 'sampling/sampling_logp_difference/max': 1.1398444175720215, 'sampling/importance_sampling_ratio/min': 0.319868803024292, 'sampling/importance_sampling_ratio/mean': 1.0187184810638428, 'sampling/importance_sampling_ratio/max': 1.9264450073242188, 'entropy': 1.0886655524373055, 'clip_ratio/low_mean': 0.06184093654155731, 'clip_ratio/low_min': 0.06184093654155731, 'clip_ratio/high_mean': 0.0540414759889245, 'clip_ratio/high_max': 0.0540414759889245, 'clip_ratio/region_mean': 0.11588241253048182, 'reward_total_mean': 0.7137107849121094, 'reward_meter_mean': 0.9243209958076477, 'reward_meter_std': 0.10788122564554214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.989984393119812, 'reward_repeat_soft_std': 0.01322022546082735, 'reward_judge_quality_mean': 0.5987499952316284, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.7137107849121094, 'reward_total_composite_std': 0.1404464691877365, 'epoch': 0.11}

 35%|███▍      | 1143/3300 [2:50:09<4:52:22,  8.13s/it]INFO 04-13 10:11:07 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1144/3300 [2:50:15<4:35:01,  7.65s/it]2026-04-13 10:11:13,592 | INFO | train_grpo_train | metrics_logged mode=train step=1144

                                                       
{'loss': 0.1071, 'grad_norm': 9.581685066223145, 'learning_rate': 6.536363636363638e-06, 'num_tokens': 2023599.0, 'completions/mean_length': 71.125, 'completions/min_length': 52.0, 'completions/max_length': 85.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.9341791868209839, 'rewards/meter/std': 0.11174824833869934, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9516667127609253, 'rewards/repeat_soft/std': 0.028464632108807564, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/total_composite/mean': 0.6210188269615173, 'rewards/total_composite/std': 0.08535400032997131, 'reward': 0.6210188269615173, 'reward_std': 0.08535400032997131, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14601223170757294, 'sampling/sampling_logp_difference/max': 1.1267385482788086, 'sampling/importance_sampling_ratio/min': 0.32408854365348816, 'sampling/importance_sampling_ratio/mean': 1.020800232887268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2325196266174316, 'clip_ratio/low_mean': 0.08254536800086498, 'clip_ratio/low_min': 0.08254536800086498, 'clip_ratio/high_mean': 0.06810528226196766, 'clip_ratio/high_max': 0.06810528226196766, 'clip_ratio/region_mean': 0.15065065026283264, 'reward_total_mean': 0.6210188269615173, 'reward_meter_mean': 0.9341791868209839, 'reward_meter_std': 0.11174824833869934, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9516667127609253, 'reward_repeat_soft_std': 0.028464632108807564, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.6210188269615173, 'reward_total_composite_std': 0.08535400032997131, 'epoch': 0.11}

 35%|███▍      | 1144/3300 [2:50:15<4:35:01,  7.65s/it]INFO 04-13 10:11:13 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1145/3300 [2:50:22<4:28:22,  7.47s/it]2026-04-13 10:11:20,641 | INFO | train_grpo_train | metrics_logged mode=train step=1145

                                                       
{'loss': 0.0505, 'grad_norm': 10.74338436126709, 'learning_rate': 6.533333333333334e-06, 'num_tokens': 2025811.0, 'completions/mean_length': 77.5, 'completions/min_length': 68.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.8966401815414429, 'rewards/meter/std': 0.14012953639030457, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9699361324310303, 'rewards/repeat_soft/std': 0.021746700629591942, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.5656353831291199, 'rewards/total_composite/std': 0.10508348792791367, 'reward': 0.5656353831291199, 'reward_std': 0.10508348792791367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1363961398601532, 'sampling/sampling_logp_difference/max': 1.8495888710021973, 'sampling/importance_sampling_ratio/min': 0.15730181336402893, 'sampling/importance_sampling_ratio/mean': 1.0134751796722412, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9776269495487213, 'clip_ratio/low_mean': 0.0936512565240264, 'clip_ratio/low_min': 0.0936512565240264, 'clip_ratio/high_mean': 0.03307184763252735, 'clip_ratio/high_max': 0.03307184763252735, 'clip_ratio/region_mean': 0.12672310415655375, 'reward_total_mean': 0.5656353831291199, 'reward_meter_mean': 0.8966401815414429, 'reward_meter_std': 0.14012953639030457, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9699361324310303, 'reward_repeat_soft_std': 0.021746700629591942, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.5656353831291199, 'reward_total_composite_std': 0.10508348792791367, 'epoch': 0.12}

 35%|███▍      | 1145/3300 [2:50:22<4:28:22,  7.47s/it]INFO 04-13 10:11:20 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1146/3300 [2:50:28<4:13:08,  7.05s/it]2026-04-13 10:11:26,708 | INFO | train_grpo_train | metrics_logged mode=train step=1146

                                                       
{'loss': 0.0492, 'grad_norm': 15.230727195739746, 'learning_rate': 6.530303030303031e-06, 'num_tokens': 2027346.0, 'completions/mean_length': 36.875, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.875, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9046503305435181, 'rewards/meter/std': 0.15936368703842163, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9611223936080933, 'rewards/repeat_soft/std': 0.04192990064620972, 'rewards/judge_quality/mean': 0.627500057220459, 'rewards/judge_quality/std': 0.21952873468399048, 'rewards/total_composite/mean': 0.7060563564300537, 'rewards/total_composite/std': 0.12517179548740387, 'reward': 0.7060563564300537, 'reward_std': 0.12517178058624268, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14486348628997803, 'sampling/sampling_logp_difference/max': 0.9896438121795654, 'sampling/importance_sampling_ratio/min': 0.38649776577949524, 'sampling/importance_sampling_ratio/mean': 1.0297596454620361, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0335094183683395, 'clip_ratio/low_mean': 0.07998835388571024, 'clip_ratio/low_min': 0.07998835388571024, 'clip_ratio/high_mean': 0.04699411056935787, 'clip_ratio/high_max': 0.04699411056935787, 'clip_ratio/region_mean': 0.1269824644550681, 'reward_total_mean': 0.7060563564300537, 'reward_meter_mean': 0.9046503305435181, 'reward_meter_std': 0.15936368703842163, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9611223936080933, 'reward_repeat_soft_std': 0.04192990064620972, 'reward_judge_quality_mean': 0.627500057220459, 'reward_judge_quality_std': 0.21952873468399048, 'reward_total_composite_mean': 0.7060563564300537, 'reward_total_composite_std': 0.12517179548740387, 'epoch': 0.12}

 35%|███▍      | 1146/3300 [2:50:28<4:13:08,  7.05s/it]INFO 04-13 10:11:26 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1147/3300 [2:50:35<4:06:41,  6.87s/it]2026-04-13 10:11:33,174 | INFO | train_grpo_train | metrics_logged mode=train step=1147

                                                       
{'loss': 0.0052, 'grad_norm': 12.32846736907959, 'learning_rate': 6.527272727272728e-06, 'num_tokens': 2029195.0, 'completions/mean_length': 45.125, 'completions/min_length': 41.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.47214794158935547, 'rewards/meter/std': 0.3173607885837555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995770275592804, 'rewards/repeat_soft/std': 0.0094536691904068, 'rewards/judge_quality/mean': 0.7400000095367432, 'rewards/judge_quality/std': 0.24859607219696045, 'rewards/total_composite/mean': 0.56976318359375, 'rewards/total_composite/std': 0.16080376505851746, 'reward': 0.56976318359375, 'reward_std': 0.16080376505851746, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16579216718673706, 'sampling/sampling_logp_difference/max': 2.801460027694702, 'sampling/importance_sampling_ratio/min': 0.06072134152054787, 'sampling/importance_sampling_ratio/mean': 1.005444049835205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9021644480526447, 'clip_ratio/low_mean': 0.08514065854251385, 'clip_ratio/low_min': 0.08514065854251385, 'clip_ratio/high_mean': 0.04831978306174278, 'clip_ratio/high_max': 0.04831978306174278, 'clip_ratio/region_mean': 0.13346044160425663, 'reward_total_mean': 0.56976318359375, 'reward_meter_mean': 0.47214794158935547, 'reward_meter_std': 0.3173607885837555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995770275592804, 'reward_repeat_soft_std': 0.0094536691904068, 'reward_judge_quality_mean': 0.7400000095367432, 'reward_judge_quality_std': 0.24859607219696045, 'reward_total_composite_mean': 0.56976318359375, 'reward_total_composite_std': 0.16080376505851746, 'epoch': 0.12}

 35%|███▍      | 1147/3300 [2:50:35<4:06:41,  6.87s/it]INFO 04-13 10:11:33 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1148/3300 [2:50:41<3:55:13,  6.56s/it]2026-04-13 10:11:38,990 | INFO | train_grpo_train | metrics_logged mode=train step=1148

                                                       
{'loss': -0.0042, 'grad_norm': 12.439154624938965, 'learning_rate': 6.524242424242425e-06, 'num_tokens': 2030820.0, 'completions/mean_length': 48.125, 'completions/min_length': 42.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7631425261497498, 'rewards/meter/std': 0.33935314416885376, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9642658233642578, 'rewards/repeat_soft/std': 0.020984716713428497, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.13979578018188477, 'rewards/total_composite/mean': 0.5847371220588684, 'rewards/total_composite/std': 0.10088668018579483, 'reward': 0.5847371220588684, 'reward_std': 0.10088668018579483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13191036880016327, 'sampling/sampling_logp_difference/max': 1.2770562171936035, 'sampling/importance_sampling_ratio/min': 0.2788569927215576, 'sampling/importance_sampling_ratio/mean': 1.0432844161987305, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1150656566023827, 'clip_ratio/low_mean': 0.03852657042443752, 'clip_ratio/low_min': 0.03852657042443752, 'clip_ratio/high_mean': 0.08616977697238326, 'clip_ratio/high_max': 0.08616977697238326, 'clip_ratio/region_mean': 0.12469634739682078, 'reward_total_mean': 0.5847371220588684, 'reward_meter_mean': 0.7631425261497498, 'reward_meter_std': 0.33935314416885376, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9642658233642578, 'reward_repeat_soft_std': 0.020984716713428497, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.13979578018188477, 'reward_total_composite_mean': 0.5847371220588684, 'reward_total_composite_std': 0.10088668018579483, 'epoch': 0.12}

 35%|███▍      | 1148/3300 [2:50:41<3:55:13,  6.56s/it]INFO 04-13 10:11:39 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1149/3300 [2:50:47<3:52:00,  6.47s/it]2026-04-13 10:11:45,261 | INFO | train_grpo_train | metrics_logged mode=train step=1149

                                                       
{'loss': 0.0663, 'grad_norm': 9.508204460144043, 'learning_rate': 6.521212121212121e-06, 'num_tokens': 2032794.0, 'completions/mean_length': 72.75, 'completions/min_length': 59.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.75, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9851264953613281, 'rewards/meter/std': 0.006965228822082281, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9444782137870789, 'rewards/repeat_soft/std': 0.03553704917430878, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.5871252417564392, 'rewards/total_composite/std': 0.12597087025642395, 'reward': 0.5871252417564392, 'reward_std': 0.12597087025642395, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13400357961654663, 'sampling/sampling_logp_difference/max': 2.046328067779541, 'sampling/importance_sampling_ratio/min': 0.12920847535133362, 'sampling/importance_sampling_ratio/mean': 1.027452826499939, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1493511721491814, 'clip_ratio/low_mean': 0.09198104683309793, 'clip_ratio/low_min': 0.09198104683309793, 'clip_ratio/high_mean': 0.012711863964796066, 'clip_ratio/high_max': 0.012711863964796066, 'clip_ratio/region_mean': 0.104692910797894, 'reward_total_mean': 0.5871252417564392, 'reward_meter_mean': 0.9851264953613281, 'reward_meter_std': 0.006965228822082281, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9444782137870789, 'reward_repeat_soft_std': 0.03553704917430878, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.5871252417564392, 'reward_total_composite_std': 0.12597087025642395, 'epoch': 0.12}

 35%|███▍      | 1149/3300 [2:50:47<3:52:00,  6.47s/it]INFO 04-13 10:11:45 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1150/3300 [2:50:54<3:54:11,  6.54s/it]2026-04-13 10:11:51,949 | INFO | train_grpo_train | metrics_logged mode=train step=1150

                                                       
{'loss': 0.0055, 'grad_norm': 9.685317039489746, 'learning_rate': 6.5181818181818195e-06, 'num_tokens': 2034777.0, 'completions/mean_length': 76.875, 'completions/min_length': 64.0, 'completions/max_length': 89.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.34443455934524536, 'rewards/meter/std': 0.2653238773345947, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9492671489715576, 'rewards/repeat_soft/std': 0.03478346765041351, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.42426055669784546, 'rewards/total_composite/std': 0.14111271500587463, 'reward': 0.42426055669784546, 'reward_std': 0.14111271500587463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15354497730731964, 'sampling/sampling_logp_difference/max': 1.3122196197509766, 'sampling/importance_sampling_ratio/min': 0.26922181248664856, 'sampling/importance_sampling_ratio/mean': 1.0177932977676392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.103570282459259, 'clip_ratio/low_mean': 0.0882663344964385, 'clip_ratio/low_min': 0.0882663344964385, 'clip_ratio/high_mean': 0.03345371223986149, 'clip_ratio/high_max': 0.03345371223986149, 'clip_ratio/region_mean': 0.12172004673629999, 'reward_total_mean': 0.42426055669784546, 'reward_meter_mean': 0.34443455934524536, 'reward_meter_std': 0.2653238773345947, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.05892555043101311, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9492671489715576, 'reward_repeat_soft_std': 0.03478346765041351, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.42426055669784546, 'reward_total_composite_std': 0.14111271500587463, 'epoch': 0.12}

 35%|███▍      | 1150/3300 [2:50:54<3:54:11,  6.54s/it]INFO 04-13 10:11:52 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:15,  1.90s/it]

 30%|███       | 3/10 [00:12<00:33,  4.76s/it]

 40%|████      | 4/10 [00:15<00:24,  4.15s/it]

 50%|█████     | 5/10 [00:19<00:20,  4.15s/it]

 60%|██████    | 6/10 [00:23<00:15,  3.92s/it]

 70%|███████   | 7/10 [00:26<00:11,  3.81s/it]

 80%|████████  | 8/10 [00:30<00:07,  3.62s/it]

 90%|█████████ | 9/10 [00:38<00:05,  5.21s/it]

100%|██████████| 10/10 [00:42<00:00,  4.59s/it]2026-04-13 10:12:43,433 | INFO | train_grpo_train | metrics_logged mode=eval step=1150

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 51.4175, 'eval_samples_per_second': 1.556, 'eval_steps_per_second': 0.194, 'eval_num_tokens': 2034777.0, 'eval_completions/mean_length': 79.525, 'eval_completions/min_length': 30.5, 'eval_completions/max_length': 213.0, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 56.7702392578125, 'eval_completions/min_terminated_length': 30.5, 'eval_completions/max_terminated_length': 87.7, 'eval_rewards/meter/mean': 0.7342909038066864, 'eval_rewards/meter/std': 0.3166229337453842, 'eval_rewards/count_adherence/mean': 0.8370833396911621, 'eval_rewards/count_adherence/std': 0.1520685613155365, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.11700168251991272, 'eval_rewards/repeat_soft/mean': 0.9545246958732605, 'eval_rewards/repeat_soft/std': 0.04832870401442051, 'eval_rewards/judge_quality/mean': 0.4768750011920929, 'eval_rewards/judge_quality/std': 0.19408679232001305, 'eval_rewards/total_composite/mean': 0.5226333200931549, 'eval_rewards/total_composite/std': 0.16456383764743804, 'eval_reward': 0.5226333200931549, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08284016698598862, 'eval_sampling/sampling_logp_difference/max': 0.9862382411956787, 'eval_sampling/importance_sampling_ratio/min': 0.37612539529800415, 'eval_sampling/importance_sampling_ratio/mean': 1.022798728942871, 'eval_sampling/importance_sampling_ratio/max': 1.4848750710487366, 'eval_entropy': 0.9602272927761077, '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': 0.5226333200931549, 'eval_reward_meter_mean': 0.7342909038066864, 'eval_reward_meter_std': 0.3166229337453842, 'eval_reward_count_adherence_mean': 0.8370833396911621, 'eval_reward_count_adherence_std': 0.1520685613155365, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.11700168251991272, 'eval_reward_repeat_soft_mean': 0.9545246958732605, 'eval_reward_repeat_soft_std': 0.04832870401442051, 'eval_reward_judge_quality_mean': 0.4768750011920929, 'eval_reward_judge_quality_std': 0.19408679232001305, 'eval_reward_total_composite_mean': 0.5226333200931549, 'eval_reward_total_composite_std': 0.16456383764743804, 'epoch': 0.12}

 35%|███▍      | 1150/3300 [2:51:45<3:54:11,  6.54s/it]

100%|██████████| 10/10 [00:42<00:00,  4.59s/it]

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

 35%|███▍      | 1151/3300 [2:51:54<13:37:25, 22.82s/it]2026-04-13 10:12:52,786 | INFO | train_grpo_train | metrics_logged mode=train step=1151

                                                        
{'loss': 0.0478, 'grad_norm': 11.886629104614258, 'learning_rate': 6.515151515151516e-06, 'num_tokens': 2036449.0, 'completions/mean_length': 52.0, 'completions/min_length': 41.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8322956562042236, 'rewards/meter/std': 0.23459459841251373, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9498786926269531, 'rewards/repeat_soft/std': 0.06581722944974899, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6352229714393616, 'rewards/total_composite/std': 0.14727750420570374, 'reward': 0.6352229714393616, 'reward_std': 0.14727750420570374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10419005900621414, 'sampling/sampling_logp_difference/max': 1.838627815246582, 'sampling/importance_sampling_ratio/min': 0.15903550386428833, 'sampling/importance_sampling_ratio/mean': 1.0121780633926392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3742055706679821, 'clip_ratio/low_mean': 0.07162434328347445, 'clip_ratio/low_min': 0.07162434328347445, 'clip_ratio/high_mean': 0.03747049532830715, 'clip_ratio/high_max': 0.03747049532830715, 'clip_ratio/region_mean': 0.1090948386117816, 'reward_total_mean': 0.6352229714393616, 'reward_meter_mean': 0.8322956562042236, 'reward_meter_std': 0.23459459841251373, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9498786926269531, 'reward_repeat_soft_std': 0.06581722944974899, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6352229714393616, 'reward_total_composite_std': 0.14727750420570374, 'epoch': 0.12}

 35%|███▍      | 1151/3300 [2:51:55<13:37:25, 22.82s/it]INFO 04-13 10:12:53 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1152/3300 [2:52:01<10:43:28, 17.97s/it]2026-04-13 10:12:59,442 | INFO | train_grpo_train | metrics_logged mode=train step=1152

                                                        
{'loss': 0.1161, 'grad_norm': 15.425283432006836, 'learning_rate': 6.512121212121213e-06, 'num_tokens': 2037983.0, 'completions/mean_length': 41.75, 'completions/min_length': 34.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8951907753944397, 'rewards/meter/std': 0.2683439254760742, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.959107518196106, 'rewards/repeat_soft/std': 0.027892453595995903, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.11310552060604095, 'rewards/total_composite/mean': 0.6202410459518433, 'rewards/total_composite/std': 0.11311757564544678, 'reward': 0.6202410459518433, 'reward_std': 0.11311756819486618, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14502514898777008, 'sampling/sampling_logp_difference/max': 1.3253741264343262, 'sampling/importance_sampling_ratio/min': 0.26570355892181396, 'sampling/importance_sampling_ratio/mean': 1.032871961593628, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8751308396458626, 'clip_ratio/low_mean': 0.0844203990418464, 'clip_ratio/low_min': 0.0844203990418464, 'clip_ratio/high_mean': 0.043231177143752575, 'clip_ratio/high_max': 0.043231177143752575, 'clip_ratio/region_mean': 0.12765157618559897, 'reward_total_mean': 0.6202410459518433, 'reward_meter_mean': 0.8951907753944397, 'reward_meter_std': 0.2683439254760742, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.959107518196106, 'reward_repeat_soft_std': 0.027892453595995903, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.11310552060604095, 'reward_total_composite_mean': 0.6202410459518433, 'reward_total_composite_std': 0.11311757564544678, 'epoch': 0.12}

 35%|███▍      | 1152/3300 [2:52:01<10:43:28, 17.97s/it]INFO 04-13 10:12:59 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1153/3300 [2:52:08<8:39:12, 14.51s/it] 2026-04-13 10:13:05,873 | INFO | train_grpo_train | metrics_logged mode=train step=1153

                                                       
{'loss': 0.0392, 'grad_norm': 15.552083015441895, 'learning_rate': 6.5090909090909095e-06, 'num_tokens': 2039585.0, 'completions/mean_length': 44.25, 'completions/min_length': 37.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7080739736557007, 'rewards/meter/std': 0.23568610846996307, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9552506804466248, 'rewards/repeat_soft/std': 0.04192620888352394, 'rewards/judge_quality/mean': 0.7100000381469727, 'rewards/judge_quality/std': 0.2331768274307251, 'rewards/total_composite/mean': 0.6645849943161011, 'rewards/total_composite/std': 0.16087202727794647, 'reward': 0.6645849943161011, 'reward_std': 0.16087201237678528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11269795894622803, 'sampling/sampling_logp_difference/max': 1.5774636268615723, 'sampling/importance_sampling_ratio/min': 0.2064981907606125, 'sampling/importance_sampling_ratio/mean': 1.0016475915908813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6939624734222889, 'clip_ratio/low_mean': 0.04442748334258795, 'clip_ratio/low_min': 0.04442748334258795, 'clip_ratio/high_mean': 0.06681038066744804, 'clip_ratio/high_max': 0.06681038066744804, 'clip_ratio/region_mean': 0.11123786401003599, 'reward_total_mean': 0.6645849943161011, 'reward_meter_mean': 0.7080739736557007, 'reward_meter_std': 0.23568610846996307, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9552506804466248, 'reward_repeat_soft_std': 0.04192620888352394, 'reward_judge_quality_mean': 0.7100000381469727, 'reward_judge_quality_std': 0.2331768274307251, 'reward_total_composite_mean': 0.6645849943161011, 'reward_total_composite_std': 0.16087202727794647, 'epoch': 0.12}

 35%|███▍      | 1153/3300 [2:52:08<8:39:12, 14.51s/it]INFO 04-13 10:13:06 [block_pool.py:378] Successfully reset prefix cache

 35%|███▍      | 1154/3300 [2:52:19<8:09:27, 13.68s/it]2026-04-13 10:13:17,625 | INFO | train_grpo_train | metrics_logged mode=train step=1154

                                                       
{'loss': -0.1162, 'grad_norm': 2.4770925045013428, 'learning_rate': 6.506060606060607e-06, 'num_tokens': 2041124.0, 'completions/mean_length': 98.375, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.28571701049805, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9925278425216675, 'rewards/meter/std': 0.003449532203376293, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9869427680969238, 'rewards/repeat_soft/std': 0.02140633761882782, 'rewards/judge_quality/mean': 0.5087500214576721, 'rewards/judge_quality/std': 0.24002604186534882, 'rewards/total_composite/mean': 0.6286358833312988, 'rewards/total_composite/std': 0.2727552056312561, 'reward': 0.6286358833312988, 'reward_std': 0.2727552056312561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14411436021327972, 'sampling/sampling_logp_difference/max': 1.1456871032714844, 'sampling/importance_sampling_ratio/min': 0.3180053234100342, 'sampling/importance_sampling_ratio/mean': 1.0198239088058472, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8729628100991249, 'clip_ratio/low_mean': 0.03125, 'clip_ratio/low_min': 0.03125, 'clip_ratio/high_mean': 0.0861591063439846, 'clip_ratio/high_max': 0.0861591063439846, 'clip_ratio/region_mean': 0.1174091063439846, 'reward_total_mean': 0.6286358833312988, 'reward_meter_mean': 0.9925278425216675, 'reward_meter_std': 0.003449532203376293, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9869427680969238, 'reward_repeat_soft_std': 0.02140633761882782, 'reward_judge_quality_mean': 0.5087500214576721, 'reward_judge_quality_std': 0.24002604186534882, 'reward_total_composite_mean': 0.6286358833312988, 'reward_total_composite_std': 0.2727552056312561, 'epoch': 0.12}

 35%|███▍      | 1154/3300 [2:52:19<8:09:27, 13.68s/it]INFO 04-13 10:13:17 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1155/3300 [2:52:25<6:48:24, 11.42s/it]2026-04-13 10:13:23,775 | INFO | train_grpo_train | metrics_logged mode=train step=1155

                                                       
{'loss': 0.0446, 'grad_norm': 13.718782424926758, 'learning_rate': 6.503030303030303e-06, 'num_tokens': 2042771.0, 'completions/mean_length': 46.875, 'completions/min_length': 36.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8706246614456177, 'rewards/meter/std': 0.25902047753334045, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9424266815185547, 'rewards/repeat_soft/std': 0.040095433592796326, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.5893842577934265, 'rewards/total_composite/std': 0.05839797854423523, 'reward': 0.5893842577934265, 'reward_std': 0.058397967368364334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.143097922205925, 'sampling/sampling_logp_difference/max': 1.1727938652038574, 'sampling/importance_sampling_ratio/min': 0.3095010221004486, 'sampling/importance_sampling_ratio/mean': 1.0094670057296753, 'sampling/importance_sampling_ratio/max': 1.9955836534500122, 'entropy': 1.087383322417736, 'clip_ratio/low_mean': 0.022058824077248573, 'clip_ratio/low_min': 0.022058824077248573, 'clip_ratio/high_mean': 0.10677625681273639, 'clip_ratio/high_max': 0.10677625681273639, 'clip_ratio/region_mean': 0.12883508088998497, 'reward_total_mean': 0.5893842577934265, 'reward_meter_mean': 0.8706246614456177, 'reward_meter_std': 0.25902047753334045, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9424266815185547, 'reward_repeat_soft_std': 0.040095433592796326, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.5893842577934265, 'reward_total_composite_std': 0.05839797854423523, 'epoch': 0.12}

 35%|███▌      | 1155/3300 [2:52:26<6:48:24, 11.42s/it]INFO 04-13 10:13:24 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1156/3300 [2:52:33<6:05:17, 10.22s/it]2026-04-13 10:13:31,234 | INFO | train_grpo_train | metrics_logged mode=train step=1156

                                                       
{'loss': -0.0461, 'grad_norm': 12.073512077331543, 'learning_rate': 6.5000000000000004e-06, 'num_tokens': 2044454.0, 'completions/mean_length': 41.375, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9720487594604492, 'rewards/meter/std': 0.018122032284736633, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9285119771957397, 'rewards/repeat_soft/std': 0.011059910990297794, 'rewards/judge_quality/mean': 0.5049999952316284, 'rewards/judge_quality/std': 0.16801361739635468, 'rewards/total_composite/mean': 0.6586222648620605, 'rewards/total_composite/std': 0.10675440728664398, 'reward': 0.6586222648620605, 'reward_std': 0.10675440728664398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1238454058766365, 'sampling/sampling_logp_difference/max': 2.65266752243042, 'sampling/importance_sampling_ratio/min': 0.07046300172805786, 'sampling/importance_sampling_ratio/mean': 1.0210554599761963, 'sampling/importance_sampling_ratio/max': 1.9586304426193237, 'entropy': 1.0119702145457268, 'clip_ratio/low_mean': 0.10703061474487185, 'clip_ratio/low_min': 0.10703061474487185, 'clip_ratio/high_mean': 0.018617020919919014, 'clip_ratio/high_max': 0.018617020919919014, 'clip_ratio/region_mean': 0.12564763566479087, 'reward_total_mean': 0.6586222648620605, 'reward_meter_mean': 0.9720487594604492, 'reward_meter_std': 0.018122032284736633, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9285119771957397, 'reward_repeat_soft_std': 0.011059910990297794, 'reward_judge_quality_mean': 0.5049999952316284, 'reward_judge_quality_std': 0.16801361739635468, 'reward_total_composite_mean': 0.6586222648620605, 'reward_total_composite_std': 0.10675440728664398, 'epoch': 0.12}

 35%|███▌      | 1156/3300 [2:52:33<6:05:17, 10.22s/it]INFO 04-13 10:13:31 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1157/3300 [2:52:39<5:22:21,  9.03s/it]2026-04-13 10:13:37,422 | INFO | train_grpo_train | metrics_logged mode=train step=1157

                                                       
{'loss': -0.022, 'grad_norm': 10.855276107788086, 'learning_rate': 6.496969696969697e-06, 'num_tokens': 2046156.0, 'completions/mean_length': 43.75, 'completions/min_length': 38.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8936759233474731, 'rewards/meter/std': 0.1595841944217682, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.895618736743927, 'rewards/repeat_soft/std': 0.0682944655418396, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.6663017868995667, 'rewards/total_composite/std': 0.15666808187961578, 'reward': 0.6663017868995667, 'reward_std': 0.1566680669784546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11408718675374985, 'sampling/sampling_logp_difference/max': 1.7818901538848877, 'sampling/importance_sampling_ratio/min': 0.27114689350128174, 'sampling/importance_sampling_ratio/mean': 1.0291129350662231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.74884133040905, 'clip_ratio/low_mean': 0.0684937359765172, 'clip_ratio/low_min': 0.0684937359765172, 'clip_ratio/high_mean': 0.03380434773862362, 'clip_ratio/high_max': 0.03380434773862362, 'clip_ratio/region_mean': 0.10229808371514082, 'reward_total_mean': 0.6663017868995667, 'reward_meter_mean': 0.8936759233474731, 'reward_meter_std': 0.1595841944217682, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.895618736743927, 'reward_repeat_soft_std': 0.0682944655418396, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.6663017868995667, 'reward_total_composite_std': 0.15666808187961578, 'epoch': 0.12}

 35%|███▌      | 1157/3300 [2:52:39<5:22:21,  9.03s/it]INFO 04-13 10:13:37 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1158/3300 [2:52:50<5:45:44,  9.68s/it]2026-04-13 10:13:48,642 | INFO | train_grpo_train | metrics_logged mode=train step=1158

                                                       
{'loss': -0.1513, 'grad_norm': 3.5333056449890137, 'learning_rate': 6.493939393939395e-06, 'num_tokens': 2048145.0, 'completions/mean_length': 124.625, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.28572082519531, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.7109284996986389, 'rewards/meter/std': 0.3987520635128021, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9296675324440002, 'rewards/repeat_soft/std': 0.04415125399827957, 'rewards/judge_quality/mean': 0.4649999737739563, 'rewards/judge_quality/std': 0.3147788643836975, 'rewards/total_composite/mean': 0.5115648508071899, 'rewards/total_composite/std': 0.2610779404640198, 'reward': 0.5115648508071899, 'reward_std': 0.2610779404640198, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12852534651756287, 'sampling/sampling_logp_difference/max': 1.8264281749725342, 'sampling/importance_sampling_ratio/min': 0.16098757088184357, 'sampling/importance_sampling_ratio/mean': 1.0188608169555664, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7715061381459236, 'clip_ratio/low_mean': 0.03934571333229542, 'clip_ratio/low_min': 0.03934571333229542, 'clip_ratio/high_mean': 0.08331483043730259, 'clip_ratio/high_max': 0.08331483043730259, 'clip_ratio/region_mean': 0.12266054376959801, 'reward_total_mean': 0.5115648508071899, 'reward_meter_mean': 0.7109284996986389, 'reward_meter_std': 0.3987520635128021, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9296675324440002, 'reward_repeat_soft_std': 0.04415125399827957, 'reward_judge_quality_mean': 0.4649999737739563, 'reward_judge_quality_std': 0.3147788643836975, 'reward_total_composite_mean': 0.5115648508071899, 'reward_total_composite_std': 0.2610779404640198, 'epoch': 0.12}

 35%|███▌      | 1158/3300 [2:52:50<5:45:44,  9.68s/it]INFO 04-13 10:13:48 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1159/3300 [2:52:57<5:10:24,  8.70s/it]2026-04-13 10:13:55,079 | INFO | train_grpo_train | metrics_logged mode=train step=1159

                                                       
{'loss': 0.0341, 'grad_norm': 16.398767471313477, 'learning_rate': 6.490909090909091e-06, 'num_tokens': 2049736.0, 'completions/mean_length': 39.875, 'completions/min_length': 35.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9786190390586853, 'rewards/meter/std': 0.013541036285459995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9787521362304688, 'rewards/repeat_soft/std': 0.027974126860499382, 'rewards/judge_quality/mean': 0.7200000286102295, 'rewards/judge_quality/std': 0.20701968669891357, 'rewards/total_composite/mean': 0.8045221567153931, 'rewards/total_composite/std': 0.13010771572589874, 'reward': 0.8045221567153931, 'reward_std': 0.13010771572589874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1522352546453476, 'sampling/sampling_logp_difference/max': 1.1938486099243164, 'sampling/importance_sampling_ratio/min': 0.303052693605423, 'sampling/importance_sampling_ratio/mean': 1.0002986192703247, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0424583405256271, 'clip_ratio/low_mean': 0.04302912671118975, 'clip_ratio/low_min': 0.04302912671118975, 'clip_ratio/high_mean': 0.09753315430134535, 'clip_ratio/high_max': 0.09753315430134535, 'clip_ratio/region_mean': 0.1405622810125351, 'reward_total_mean': 0.8045221567153931, 'reward_meter_mean': 0.9786190390586853, 'reward_meter_std': 0.013541036285459995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9787521362304688, 'reward_repeat_soft_std': 0.027974126860499382, 'reward_judge_quality_mean': 0.7200000286102295, 'reward_judge_quality_std': 0.20701968669891357, 'reward_total_composite_mean': 0.8045221567153931, 'reward_total_composite_std': 0.13010771572589874, 'epoch': 0.12}

 35%|███▌      | 1159/3300 [2:52:57<5:10:24,  8.70s/it]INFO 04-13 10:13:55 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1160/3300 [2:53:03<4:49:09,  8.11s/it]2026-04-13 10:14:01,810 | INFO | train_grpo_train | metrics_logged mode=train step=1160

                                                       
{'loss': -0.0229, 'grad_norm': 16.45473861694336, 'learning_rate': 6.487878787878789e-06, 'num_tokens': 2050986.0, 'completions/mean_length': 19.25, 'completions/min_length': 16.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.33942797780036926, 'rewards/meter/std': 0.3242776095867157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9583057761192322, 'rewards/repeat_soft/std': 0.007864047773182392, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.3720680773258209, 'rewards/total_composite/std': 0.17117416858673096, 'reward': 0.3720680773258209, 'reward_std': 0.17117416858673096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15709532797336578, 'sampling/sampling_logp_difference/max': 1.241135597229004, 'sampling/importance_sampling_ratio/min': 0.2890557646751404, 'sampling/importance_sampling_ratio/mean': 1.0157854557037354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1322322264313698, 'clip_ratio/low_mean': 0.06802545534446836, 'clip_ratio/low_min': 0.06802545534446836, 'clip_ratio/high_mean': 0.09996639750897884, 'clip_ratio/high_max': 0.09996639750897884, 'clip_ratio/region_mean': 0.1679918528534472, 'reward_total_mean': 0.3720680773258209, 'reward_meter_mean': 0.33942797780036926, 'reward_meter_std': 0.3242776095867157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9583057761192322, 'reward_repeat_soft_std': 0.007864047773182392, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.3720680773258209, 'reward_total_composite_std': 0.17117416858673096, 'epoch': 0.12}

 35%|███▌      | 1160/3300 [2:53:04<4:49:09,  8.11s/it]INFO 04-13 10:14:02 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1161/3300 [2:53:09<4:25:58,  7.46s/it]2026-04-13 10:14:07,720 | INFO | train_grpo_train | metrics_logged mode=train step=1161

                                                       
{'loss': 0.0639, 'grad_norm': 19.172563552856445, 'learning_rate': 6.484848484848485e-06, 'num_tokens': 2052447.0, 'completions/mean_length': 18.625, 'completions/min_length': 15.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.7257708311080933, 'rewards/meter/std': 0.25202879309654236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9509689807891846, 'rewards/repeat_soft/std': 0.02166803367435932, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.540780782699585, 'rewards/total_composite/std': 0.06680676341056824, 'reward': 0.540780782699585, 'reward_std': 0.06680675595998764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18771231174468994, 'sampling/sampling_logp_difference/max': 1.4687366485595703, 'sampling/importance_sampling_ratio/min': 0.2302161604166031, 'sampling/importance_sampling_ratio/mean': 1.011730670928955, 'sampling/importance_sampling_ratio/max': 1.923264980316162, 'entropy': 1.036086767911911, 'clip_ratio/low_mean': 0.11173768062144518, 'clip_ratio/low_min': 0.11173768062144518, 'clip_ratio/high_mean': 0.09665388148277998, 'clip_ratio/high_max': 0.09665388148277998, 'clip_ratio/region_mean': 0.20839156210422516, 'reward_total_mean': 0.540780782699585, 'reward_meter_mean': 0.7257708311080933, 'reward_meter_std': 0.25202879309654236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9509689807891846, 'reward_repeat_soft_std': 0.02166803367435932, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.540780782699585, 'reward_total_composite_std': 0.06680676341056824, 'epoch': 0.12}

 35%|███▌      | 1161/3300 [2:53:09<4:25:58,  7.46s/it]INFO 04-13 10:14:07 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1162/3300 [2:53:16<4:15:13,  7.16s/it]2026-04-13 10:14:14,187 | INFO | train_grpo_train | metrics_logged mode=train step=1162

                                                       
{'loss': -0.0015, 'grad_norm': 9.042016983032227, 'learning_rate': 6.481818181818182e-06, 'num_tokens': 2054554.0, 'completions/mean_length': 80.375, 'completions/min_length': 69.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.9644762277603149, 'rewards/meter/std': 0.05867907777428627, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.938398003578186, 'rewards/repeat_soft/std': 0.02318570762872696, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5540616512298584, 'rewards/total_composite/std': 0.01620873436331749, 'reward': 0.5540616512298584, 'reward_std': 0.01620873622596264, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14428913593292236, 'sampling/sampling_logp_difference/max': 2.3007876873016357, 'sampling/importance_sampling_ratio/min': 0.10017991065979004, 'sampling/importance_sampling_ratio/mean': 1.0051047801971436, 'sampling/importance_sampling_ratio/max': 1.9472465515136719, 'entropy': 0.9574398621916771, 'clip_ratio/low_mean': 0.039586388505995274, 'clip_ratio/low_min': 0.039586388505995274, 'clip_ratio/high_mean': 0.07957096491008997, 'clip_ratio/high_max': 0.07957096491008997, 'clip_ratio/region_mean': 0.11915735341608524, 'reward_total_mean': 0.5540616512298584, 'reward_meter_mean': 0.9644762277603149, 'reward_meter_std': 0.05867907777428627, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.938398003578186, 'reward_repeat_soft_std': 0.02318570762872696, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5540616512298584, 'reward_total_composite_std': 0.01620873436331749, 'epoch': 0.12}

 35%|███▌      | 1162/3300 [2:53:16<4:15:13,  7.16s/it]INFO 04-13 10:14:14 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1163/3300 [2:53:22<4:07:19,  6.94s/it]2026-04-13 10:14:20,623 | INFO | train_grpo_train | metrics_logged mode=train step=1163

                                                       
{'loss': 0.0314, 'grad_norm': 16.131433486938477, 'learning_rate': 6.478787878787879e-06, 'num_tokens': 2056378.0, 'completions/mean_length': 52.0, 'completions/min_length': 49.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7054102420806885, 'rewards/meter/std': 0.28646501898765564, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9489527344703674, 'rewards/repeat_soft/std': 0.03885876014828682, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.45700687170028687, 'rewards/total_composite/std': 0.10430697351694107, 'reward': 0.45700687170028687, 'reward_std': 0.10430697351694107, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13044683635234833, 'sampling/sampling_logp_difference/max': 2.052244186401367, 'sampling/importance_sampling_ratio/min': 0.1284463107585907, 'sampling/importance_sampling_ratio/mean': 0.9911546111106873, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6575556993484497, 'clip_ratio/low_mean': 0.05782278999686241, 'clip_ratio/low_min': 0.05782278999686241, 'clip_ratio/high_mean': 0.04991029482334852, 'clip_ratio/high_max': 0.04991029482334852, 'clip_ratio/region_mean': 0.10773308482021093, 'reward_total_mean': 0.45700687170028687, 'reward_meter_mean': 0.7054102420806885, 'reward_meter_std': 0.28646501898765564, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9489527344703674, 'reward_repeat_soft_std': 0.03885876014828682, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.45700687170028687, 'reward_total_composite_std': 0.10430697351694107, 'epoch': 0.12}

 35%|███▌      | 1163/3300 [2:53:22<4:07:19,  6.94s/it]INFO 04-13 10:14:20 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1164/3300 [2:53:28<3:53:57,  6.57s/it]2026-04-13 10:14:26,325 | INFO | train_grpo_train | metrics_logged mode=train step=1164

                                                       
{'loss': -0.0325, 'grad_norm': 19.62907600402832, 'learning_rate': 6.475757575757576e-06, 'num_tokens': 2057812.0, 'completions/mean_length': 20.25, 'completions/min_length': 17.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9464387893676758, 'rewards/meter/std': 0.09658417850732803, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.961837112903595, 'rewards/repeat_soft/std': 0.0018749026348814368, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6116173267364502, 'rewards/total_composite/std': 0.0269559845328331, 'reward': 0.6116173267364502, 'reward_std': 0.026955988258123398, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15353576838970184, 'sampling/sampling_logp_difference/max': 1.5081539154052734, 'sampling/importance_sampling_ratio/min': 0.22131817042827606, 'sampling/importance_sampling_ratio/mean': 0.997308075428009, 'sampling/importance_sampling_ratio/max': 1.7216521501541138, 'entropy': 1.0296223536133766, 'clip_ratio/low_mean': 0.035153554286807775, 'clip_ratio/low_min': 0.035153554286807775, 'clip_ratio/high_mean': 0.06448688125237823, 'clip_ratio/high_max': 0.06448688125237823, 'clip_ratio/region_mean': 0.099640435539186, 'reward_total_mean': 0.6116173267364502, 'reward_meter_mean': 0.9464387893676758, 'reward_meter_std': 0.09658417850732803, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.961837112903595, 'reward_repeat_soft_std': 0.0018749026348814368, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6116173267364502, 'reward_total_composite_std': 0.0269559845328331, 'epoch': 0.12}

 35%|███▌      | 1164/3300 [2:53:28<3:53:57,  6.57s/it]INFO 04-13 10:14:26 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1165/3300 [2:53:35<3:58:15,  6.70s/it]2026-04-13 10:14:33,317 | INFO | train_grpo_train | metrics_logged mode=train step=1165

                                                       
{'loss': 0.0137, 'grad_norm': 8.567177772521973, 'learning_rate': 6.472727272727272e-06, 'num_tokens': 2060085.0, 'completions/mean_length': 103.125, 'completions/min_length': 95.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.125, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.9605042338371277, 'rewards/meter/std': 0.04659838229417801, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9363154768943787, 'rewards/repeat_soft/std': 0.03352692350745201, 'rewards/judge_quality/mean': 0.5399999618530273, 'rewards/judge_quality/std': 0.22245386242866516, 'rewards/total_composite/mean': 0.4727051556110382, 'rewards/total_composite/std': 0.3168168067932129, 'reward': 0.4727051556110382, 'reward_std': 0.3168168067932129, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14320671558380127, 'sampling/sampling_logp_difference/max': 1.9507179260253906, 'sampling/importance_sampling_ratio/min': 0.1421719640493393, 'sampling/importance_sampling_ratio/mean': 1.007926344871521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0821733325719833, 'clip_ratio/low_mean': 0.0317208468914032, 'clip_ratio/low_min': 0.0317208468914032, 'clip_ratio/high_mean': 0.09785328712314367, 'clip_ratio/high_max': 0.09785328712314367, 'clip_ratio/region_mean': 0.12957413401454687, 'reward_total_mean': 0.4727051556110382, 'reward_meter_mean': 0.9605042338371277, 'reward_meter_std': 0.04659838229417801, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9363154768943787, 'reward_repeat_soft_std': 0.03352692350745201, 'reward_judge_quality_mean': 0.5399999618530273, 'reward_judge_quality_std': 0.22245386242866516, 'reward_total_composite_mean': 0.4727051556110382, 'reward_total_composite_std': 0.3168168067932129, 'epoch': 0.12}

 35%|███▌      | 1165/3300 [2:53:35<3:58:15,  6.70s/it]INFO 04-13 10:14:33 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1166/3300 [2:53:42<3:58:15,  6.70s/it]2026-04-13 10:14:40,014 | INFO | train_grpo_train | metrics_logged mode=train step=1166

                                                       
{'loss': 0.0913, 'grad_norm': 9.98704719543457, 'learning_rate': 6.4696969696969705e-06, 'num_tokens': 2062101.0, 'completions/mean_length': 76.0, 'completions/min_length': 62.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9051650166511536, 'rewards/meter/std': 0.17993085086345673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9518389701843262, 'rewards/repeat_soft/std': 0.056906893849372864, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5898858904838562, 'rewards/total_composite/std': 0.04827408120036125, 'reward': 0.5898858904838562, 'reward_std': 0.048274096101522446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13661645352840424, 'sampling/sampling_logp_difference/max': 2.071254014968872, 'sampling/importance_sampling_ratio/min': 0.1260276436805725, 'sampling/importance_sampling_ratio/mean': 1.0345901250839233, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0170091539621353, 'clip_ratio/low_mean': 0.05436989106237888, 'clip_ratio/low_min': 0.05436989106237888, 'clip_ratio/high_mean': 0.07861310988664627, 'clip_ratio/high_max': 0.07861310988664627, 'clip_ratio/region_mean': 0.13298300094902515, 'reward_total_mean': 0.5898858904838562, 'reward_meter_mean': 0.9051650166511536, 'reward_meter_std': 0.17993085086345673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9518389701843262, 'reward_repeat_soft_std': 0.056906893849372864, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5898858904838562, 'reward_total_composite_std': 0.04827408120036125, 'epoch': 0.12}

 35%|███▌      | 1166/3300 [2:53:42<3:58:15,  6.70s/it]INFO 04-13 10:14:40 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1167/3300 [2:53:48<3:51:22,  6.51s/it]2026-04-13 10:14:46,082 | INFO | train_grpo_train | metrics_logged mode=train step=1167

                                                       
{'loss': 0.0675, 'grad_norm': 18.763057708740234, 'learning_rate': 6.466666666666667e-06, 'num_tokens': 2063765.0, 'completions/mean_length': 35.0, 'completions/min_length': 27.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.3749932050704956, 'rewards/meter/std': 0.3746505677700043, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9954110383987427, 'rewards/repeat_soft/std': 0.00569864921271801, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.4894949495792389, 'rewards/total_composite/std': 0.1945095956325531, 'reward': 0.4894949495792389, 'reward_std': 0.1945095956325531, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1903914213180542, 'sampling/sampling_logp_difference/max': 2.062476634979248, 'sampling/importance_sampling_ratio/min': 0.12713870406150818, 'sampling/importance_sampling_ratio/mean': 1.0226589441299438, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8588284626603127, 'clip_ratio/low_mean': 0.07798821665346622, 'clip_ratio/low_min': 0.07798821665346622, 'clip_ratio/high_mean': 0.06570973061025143, 'clip_ratio/high_max': 0.06570973061025143, 'clip_ratio/region_mean': 0.14369794726371765, 'reward_total_mean': 0.4894949495792389, 'reward_meter_mean': 0.3749932050704956, 'reward_meter_std': 0.3746505677700043, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9954110383987427, 'reward_repeat_soft_std': 0.00569864921271801, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.4894949495792389, 'reward_total_composite_std': 0.1945095956325531, 'epoch': 0.12}

 35%|███▌      | 1167/3300 [2:53:48<3:51:22,  6.51s/it]INFO 04-13 10:14:46 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1168/3300 [2:53:54<3:48:29,  6.43s/it]2026-04-13 10:14:52,330 | INFO | train_grpo_train | metrics_logged mode=train step=1168

                                                       
{'loss': 0.0471, 'grad_norm': 15.621354103088379, 'learning_rate': 6.463636363636364e-06, 'num_tokens': 2065411.0, 'completions/mean_length': 35.75, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7622805833816528, 'rewards/meter/std': 0.3560439348220825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.896129310131073, 'rewards/repeat_soft/std': 0.07596825063228607, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2353682667016983, 'rewards/total_composite/mean': 0.6478046178817749, 'rewards/total_composite/std': 0.21685576438903809, 'reward': 0.6478046178817749, 'reward_std': 0.21685576438903809, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13290895521640778, 'sampling/sampling_logp_difference/max': 1.4618310928344727, 'sampling/importance_sampling_ratio/min': 0.23181141912937164, 'sampling/importance_sampling_ratio/mean': 1.0047551393508911, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6507822498679161, 'clip_ratio/low_mean': 0.06357021653093398, 'clip_ratio/low_min': 0.06357021653093398, 'clip_ratio/high_mean': 0.06838235631585121, 'clip_ratio/high_max': 0.06838235631585121, 'clip_ratio/region_mean': 0.1319525728467852, 'reward_total_mean': 0.6478046178817749, 'reward_meter_mean': 0.7622805833816528, 'reward_meter_std': 0.3560439348220825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.896129310131073, 'reward_repeat_soft_std': 0.07596825063228607, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2353682667016983, 'reward_total_composite_mean': 0.6478046178817749, 'reward_total_composite_std': 0.21685576438903809, 'epoch': 0.12}

 35%|███▌      | 1168/3300 [2:53:54<3:48:29,  6.43s/it]INFO 04-13 10:14:52 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1169/3300 [2:54:00<3:46:57,  6.39s/it]2026-04-13 10:14:58,668 | INFO | train_grpo_train | metrics_logged mode=train step=1169

                                                       
{'loss': 0.0867, 'grad_norm': 14.500654220581055, 'learning_rate': 6.460606060606061e-06, 'num_tokens': 2067120.0, 'completions/mean_length': 50.625, 'completions/min_length': 39.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.791524350643158, 'rewards/meter/std': 0.24436300992965698, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9756710529327393, 'rewards/repeat_soft/std': 0.03941075876355171, 'rewards/judge_quality/mean': 0.5274999737739563, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.6255395412445068, 'rewards/total_composite/std': 0.14700382947921753, 'reward': 0.6255395412445068, 'reward_std': 0.14700381457805634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1730264574289322, 'sampling/sampling_logp_difference/max': 1.6852831840515137, 'sampling/importance_sampling_ratio/min': 0.18539191782474518, 'sampling/importance_sampling_ratio/mean': 1.0151746273040771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2209726721048355, 'clip_ratio/low_mean': 0.07802170515060425, 'clip_ratio/low_min': 0.07802170515060425, 'clip_ratio/high_mean': 0.059146737679839134, 'clip_ratio/high_max': 0.059146737679839134, 'clip_ratio/region_mean': 0.13716844283044338, 'reward_total_mean': 0.6255395412445068, 'reward_meter_mean': 0.791524350643158, 'reward_meter_std': 0.24436300992965698, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9756710529327393, 'reward_repeat_soft_std': 0.03941075876355171, 'reward_judge_quality_mean': 0.5274999737739563, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.6255395412445068, 'reward_total_composite_std': 0.14700382947921753, 'epoch': 0.12}

 35%|███▌      | 1169/3300 [2:54:00<3:46:57,  6.39s/it]INFO 04-13 10:14:59 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1170/3300 [2:54:07<3:49:46,  6.47s/it]2026-04-13 10:15:05,291 | INFO | train_grpo_train | metrics_logged mode=train step=1170

                                                       
{'loss': 0.0595, 'grad_norm': 13.179718017578125, 'learning_rate': 6.457575757575758e-06, 'num_tokens': 2068715.0, 'completions/mean_length': 43.375, 'completions/min_length': 35.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8890571594238281, 'rewards/meter/std': 0.28032368421554565, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9450587034225464, 'rewards/repeat_soft/std': 0.10344314575195312, 'rewards/judge_quality/mean': 0.3474999964237213, 'rewards/judge_quality/std': 0.11310551315546036, 'rewards/total_composite/mean': 0.5488234162330627, 'rewards/total_composite/std': 0.0971178188920021, 'reward': 0.5488234162330627, 'reward_std': 0.09711781144142151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12511882185935974, 'sampling/sampling_logp_difference/max': 1.342888355255127, 'sampling/importance_sampling_ratio/min': 0.2610904574394226, 'sampling/importance_sampling_ratio/mean': 1.0211719274520874, 'sampling/importance_sampling_ratio/max': 1.907301664352417, 'entropy': 1.0861545503139496, 'clip_ratio/low_mean': 0.05386972241103649, 'clip_ratio/low_min': 0.05386972241103649, 'clip_ratio/high_mean': 0.09366679657250643, 'clip_ratio/high_max': 0.09366679657250643, 'clip_ratio/region_mean': 0.14753651898354292, 'reward_total_mean': 0.5488234162330627, 'reward_meter_mean': 0.8890571594238281, 'reward_meter_std': 0.28032368421554565, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9450587034225464, 'reward_repeat_soft_std': 0.10344314575195312, 'reward_judge_quality_mean': 0.3474999964237213, 'reward_judge_quality_std': 0.11310551315546036, 'reward_total_composite_mean': 0.5488234162330627, 'reward_total_composite_std': 0.0971178188920021, 'epoch': 0.12}

 35%|███▌      | 1170/3300 [2:54:07<3:49:46,  6.47s/it]INFO 04-13 10:15:05 [block_pool.py:378] Successfully reset prefix cache

 35%|███▌      | 1171/3300 [2:54:18<4:41:57,  7.95s/it]2026-04-13 10:15:16,681 | INFO | train_grpo_train | metrics_logged mode=train step=1171

                                                       
{'loss': -0.1137, 'grad_norm': 2.412520408630371, 'learning_rate': 6.454545454545456e-06, 'num_tokens': 2070387.0, 'completions/mean_length': 106.0, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.000003814697266, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.8596898317337036, 'rewards/meter/std': 0.2784144878387451, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9502856135368347, 'rewards/repeat_soft/std': 0.050328876823186874, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.2876474857330322, 'rewards/total_composite/mean': 0.6112303137779236, 'rewards/total_composite/std': 0.28258681297302246, 'reward': 0.6112303137779236, 'reward_std': 0.28258681297302246, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13302764296531677, 'sampling/sampling_logp_difference/max': 4.004399299621582, 'sampling/importance_sampling_ratio/min': 0.018235240131616592, 'sampling/importance_sampling_ratio/mean': 1.0074357986450195, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7949735596776009, 'clip_ratio/low_mean': 0.034433962777256966, 'clip_ratio/low_min': 0.034433962777256966, 'clip_ratio/high_mean': 0.06723435246385634, 'clip_ratio/high_max': 0.06723435246385634, 'clip_ratio/region_mean': 0.1016683152411133, 'reward_total_mean': 0.6112303137779236, 'reward_meter_mean': 0.8596898317337036, 'reward_meter_std': 0.2784144878387451, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9502856135368347, 'reward_repeat_soft_std': 0.050328876823186874, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.2876474857330322, 'reward_total_composite_mean': 0.6112303137779236, 'reward_total_composite_std': 0.28258681297302246, 'epoch': 0.12}

 35%|███▌      | 1171/3300 [2:54:18<4:41:57,  7.95s/it]INFO 04-13 10:15:16 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1172/3300 [2:54:26<4:35:03,  7.76s/it]2026-04-13 10:15:24,024 | INFO | train_grpo_train | metrics_logged mode=train step=1172

                                                       
{'loss': 0.0459, 'grad_norm': 9.369461059570312, 'learning_rate': 6.451515151515152e-06, 'num_tokens': 2072506.0, 'completions/mean_length': 84.875, 'completions/min_length': 67.0, 'completions/max_length': 97.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.9398049712181091, 'rewards/meter/std': 0.09466666728258133, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8908748626708984, 'rewards/repeat_soft/std': 0.05745447054505348, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.517151415348053, 'rewards/total_composite/std': 0.06308780610561371, 'reward': 0.517151415348053, 'reward_std': 0.06308779865503311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14235681295394897, 'sampling/sampling_logp_difference/max': 1.635481834411621, 'sampling/importance_sampling_ratio/min': 0.19485846161842346, 'sampling/importance_sampling_ratio/mean': 1.026593565940857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.134580135345459, 'clip_ratio/low_mean': 0.03076496720314026, 'clip_ratio/low_min': 0.03076496720314026, 'clip_ratio/high_mean': 0.10033840034157038, 'clip_ratio/high_max': 0.10033840034157038, 'clip_ratio/region_mean': 0.13110336754471064, 'reward_total_mean': 0.517151415348053, 'reward_meter_mean': 0.9398049712181091, 'reward_meter_std': 0.09466666728258133, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8908748626708984, 'reward_repeat_soft_std': 0.05745447054505348, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.517151415348053, 'reward_total_composite_std': 0.06308780610561371, 'epoch': 0.12}

 36%|███▌      | 1172/3300 [2:54:26<4:35:03,  7.76s/it]INFO 04-13 10:15:24 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1173/3300 [2:54:32<4:21:49,  7.39s/it]2026-04-13 10:15:30,547 | INFO | train_grpo_train | metrics_logged mode=train step=1173

                                                       
{'loss': -0.031, 'grad_norm': 13.581130981445312, 'learning_rate': 6.4484848484848496e-06, 'num_tokens': 2074062.0, 'completions/mean_length': 30.5, 'completions/min_length': 24.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8466594815254211, 'rewards/meter/std': 0.31667131185531616, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9836012125015259, 'rewards/repeat_soft/std': 0.007643247488886118, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6549550294876099, 'rewards/total_composite/std': 0.17814242839813232, 'reward': 0.6549550294876099, 'reward_std': 0.17814242839813232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11450258642435074, 'sampling/sampling_logp_difference/max': 1.3845956325531006, 'sampling/importance_sampling_ratio/min': 0.2504250407218933, 'sampling/importance_sampling_ratio/mean': 0.9979570508003235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6376455649733543, 'clip_ratio/low_mean': 0.08072647964581847, 'clip_ratio/low_min': 0.08072647964581847, 'clip_ratio/high_mean': 0.026751894503831863, 'clip_ratio/high_max': 0.026751894503831863, 'clip_ratio/region_mean': 0.10747837414965034, 'reward_total_mean': 0.6549550294876099, 'reward_meter_mean': 0.8466594815254211, 'reward_meter_std': 0.31667131185531616, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9836012125015259, 'reward_repeat_soft_std': 0.007643247488886118, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6549550294876099, 'reward_total_composite_std': 0.17814242839813232, 'epoch': 0.12}

 36%|███▌      | 1173/3300 [2:54:32<4:21:49,  7.39s/it]INFO 04-13 10:15:30 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1174/3300 [2:54:39<4:14:02,  7.17s/it]2026-04-13 10:15:37,212 | INFO | train_grpo_train | metrics_logged mode=train step=1174

                                                       
{'loss': 0.0277, 'grad_norm': 15.254562377929688, 'learning_rate': 6.445454545454546e-06, 'num_tokens': 2075722.0, 'completions/mean_length': 43.5, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7098647356033325, 'rewards/meter/std': 0.26970329880714417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9789924621582031, 'rewards/repeat_soft/std': 0.03237193077802658, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5406419038772583, 'rewards/total_composite/std': 0.07226701825857162, 'reward': 0.5406419038772583, 'reward_std': 0.07226699590682983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16433273255825043, 'sampling/sampling_logp_difference/max': 2.1549975872039795, 'sampling/importance_sampling_ratio/min': 0.11590346693992615, 'sampling/importance_sampling_ratio/mean': 1.0102665424346924, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.253235936164856, 'clip_ratio/low_mean': 0.044484478421509266, 'clip_ratio/low_min': 0.044484478421509266, 'clip_ratio/high_mean': 0.10984905622899532, 'clip_ratio/high_max': 0.10984905622899532, 'clip_ratio/region_mean': 0.1543335346505046, 'reward_total_mean': 0.5406419038772583, 'reward_meter_mean': 0.7098647356033325, 'reward_meter_std': 0.26970329880714417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9789924621582031, 'reward_repeat_soft_std': 0.03237193077802658, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5406419038772583, 'reward_total_composite_std': 0.07226701825857162, 'epoch': 0.12}

 36%|███▌      | 1174/3300 [2:54:39<4:14:02,  7.17s/it]INFO 04-13 10:15:37 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1175/3300 [2:54:47<4:19:54,  7.34s/it]2026-04-13 10:15:44,947 | INFO | train_grpo_train | metrics_logged mode=train step=1175

                                                       
{'loss': 0.0175, 'grad_norm': 10.267904281616211, 'learning_rate': 6.442424242424243e-06, 'num_tokens': 2078073.0, 'completions/mean_length': 75.875, 'completions/min_length': 65.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.9792505502700806, 'rewards/meter/std': 0.010109245777130127, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9222643971443176, 'rewards/repeat_soft/std': 0.050244636833667755, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.539008378982544, 'rewards/total_composite/std': 0.007422552909702063, 'reward': 0.539008378982544, 'reward_std': 0.007422561291605234, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1581365019083023, 'sampling/sampling_logp_difference/max': 2.3960189819335938, 'sampling/importance_sampling_ratio/min': 0.09107982367277145, 'sampling/importance_sampling_ratio/mean': 1.0091606378555298, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7431683018803596, 'clip_ratio/low_mean': 0.09852162003517151, 'clip_ratio/low_min': 0.09852162003517151, 'clip_ratio/high_mean': 0.04452393390238285, 'clip_ratio/high_max': 0.04452393390238285, 'clip_ratio/region_mean': 0.14304555393755436, 'reward_total_mean': 0.539008378982544, 'reward_meter_mean': 0.9792505502700806, 'reward_meter_std': 0.010109245777130127, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9222643971443176, 'reward_repeat_soft_std': 0.050244636833667755, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.539008378982544, 'reward_total_composite_std': 0.007422552909702063, 'epoch': 0.12}

 36%|███▌      | 1175/3300 [2:54:47<4:19:54,  7.34s/it]INFO 04-13 10:15:45 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1176/3300 [2:54:53<4:08:42,  7.03s/it]2026-04-13 10:15:51,238 | INFO | train_grpo_train | metrics_logged mode=train step=1176

                                                       
{'loss': 0.0091, 'grad_norm': 12.990802764892578, 'learning_rate': 6.43939393939394e-06, 'num_tokens': 2079566.0, 'completions/mean_length': 27.625, 'completions/min_length': 26.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.7428567409515381, 'rewards/meter/std': 0.3768162429332733, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/total_composite/mean': 0.7134602665901184, 'rewards/total_composite/std': 0.21302269399166107, 'reward': 0.7134602665901184, 'reward_std': 0.21302267909049988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1083531305193901, 'sampling/sampling_logp_difference/max': 1.336371898651123, 'sampling/importance_sampling_ratio/min': 0.26279738545417786, 'sampling/importance_sampling_ratio/mean': 1.006940484046936, 'sampling/importance_sampling_ratio/max': 1.6972180604934692, 'entropy': 0.7061746343970299, 'clip_ratio/low_mean': 0.1073880223557353, 'clip_ratio/low_min': 0.1073880223557353, 'clip_ratio/high_mean': 0.03640110045671463, 'clip_ratio/high_max': 0.03640110045671463, 'clip_ratio/region_mean': 0.14378912281244993, 'reward_total_mean': 0.7134602665901184, 'reward_meter_mean': 0.7428567409515381, 'reward_meter_std': 0.3768162429332733, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.7134602665901184, 'reward_total_composite_std': 0.21302269399166107, 'epoch': 0.12}

 36%|███▌      | 1176/3300 [2:54:53<4:08:42,  7.03s/it]INFO 04-13 10:15:51 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1177/3300 [2:54:59<4:03:30,  6.88s/it]2026-04-13 10:15:57,751 | INFO | train_grpo_train | metrics_logged mode=train step=1177

                                                       
{'loss': -0.1547, 'grad_norm': 22.922889709472656, 'learning_rate': 6.436363636363637e-06, 'num_tokens': 2080972.0, 'completions/mean_length': 21.75, 'completions/min_length': 14.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.75, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.5803344249725342, 'rewards/meter/std': 0.4434331953525543, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9334905743598938, 'rewards/repeat_soft/std': 0.06022332236170769, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.493272066116333, 'rewards/total_composite/std': 0.11983183771371841, 'reward': 0.493272066116333, 'reward_std': 0.11983183771371841, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17677754163742065, 'sampling/sampling_logp_difference/max': 1.2432184219360352, 'sampling/importance_sampling_ratio/min': 0.2884543538093567, 'sampling/importance_sampling_ratio/mean': 1.046566367149353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0221305936574936, 'clip_ratio/low_mean': 0.10992063768208027, 'clip_ratio/low_min': 0.10992063768208027, 'clip_ratio/high_mean': 0.09235384874045849, 'clip_ratio/high_max': 0.09235384874045849, 'clip_ratio/region_mean': 0.20227448642253876, 'reward_total_mean': 0.493272066116333, 'reward_meter_mean': 0.5803344249725342, 'reward_meter_std': 0.4434331953525543, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9334905743598938, 'reward_repeat_soft_std': 0.06022332236170769, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.493272066116333, 'reward_total_composite_std': 0.11983183771371841, 'epoch': 0.12}

 36%|███▌      | 1177/3300 [2:54:59<4:03:30,  6.88s/it]INFO 04-13 10:15:57 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1178/3300 [2:55:06<4:00:44,  6.81s/it]2026-04-13 10:16:04,383 | INFO | train_grpo_train | metrics_logged mode=train step=1178

                                                       
{'loss': 0.027, 'grad_norm': 11.848631858825684, 'learning_rate': 6.433333333333333e-06, 'num_tokens': 2082710.0, 'completions/mean_length': 48.25, 'completions/min_length': 42.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7625800967216492, 'rewards/meter/std': 0.2791503965854645, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9629368185997009, 'rewards/repeat_soft/std': 0.04091007634997368, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.6047937870025635, 'rewards/total_composite/std': 0.1743733137845993, 'reward': 0.6047937870025635, 'reward_std': 0.1743733137845993, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12717106938362122, 'sampling/sampling_logp_difference/max': 2.013184070587158, 'sampling/importance_sampling_ratio/min': 0.13356272876262665, 'sampling/importance_sampling_ratio/mean': 1.0019422769546509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.581928551197052, 'clip_ratio/low_mean': 0.06915176194161177, 'clip_ratio/low_min': 0.06915176194161177, 'clip_ratio/high_mean': 0.03293771017342806, 'clip_ratio/high_max': 0.03293771017342806, 'clip_ratio/region_mean': 0.10208947211503983, 'reward_total_mean': 0.6047937870025635, 'reward_meter_mean': 0.7625800967216492, 'reward_meter_std': 0.2791503965854645, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9629368185997009, 'reward_repeat_soft_std': 0.04091007634997368, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.6047937870025635, 'reward_total_composite_std': 0.1743733137845993, 'epoch': 0.12}

 36%|███▌      | 1178/3300 [2:55:06<4:00:44,  6.81s/it]INFO 04-13 10:16:04 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1179/3300 [2:55:14<4:15:16,  7.22s/it]2026-04-13 10:16:12,569 | INFO | train_grpo_train | metrics_logged mode=train step=1179

                                                       
{'loss': 0.0509, 'grad_norm': 13.395605087280273, 'learning_rate': 6.430303030303031e-06, 'num_tokens': 2084447.0, 'completions/mean_length': 49.125, 'completions/min_length': 42.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8613986968994141, 'rewards/meter/std': 0.1988326907157898, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9677962064743042, 'rewards/repeat_soft/std': 0.05451434105634689, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.11310552060604095, 'rewards/total_composite/mean': 0.605843722820282, 'rewards/total_composite/std': 0.07325293868780136, 'reward': 0.605843722820282, 'reward_std': 0.07325293123722076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1554005891084671, 'sampling/sampling_logp_difference/max': 1.7847270965576172, 'sampling/importance_sampling_ratio/min': 0.16784286499023438, 'sampling/importance_sampling_ratio/mean': 1.0024949312210083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1742741093039513, 'clip_ratio/low_mean': 0.023413379210978746, 'clip_ratio/low_min': 0.023413379210978746, 'clip_ratio/high_mean': 0.12985734874382615, 'clip_ratio/high_max': 0.12985734874382615, 'clip_ratio/region_mean': 0.1532707279548049, 'reward_total_mean': 0.605843722820282, 'reward_meter_mean': 0.8613986968994141, 'reward_meter_std': 0.1988326907157898, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9677962064743042, 'reward_repeat_soft_std': 0.05451434105634689, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.11310552060604095, 'reward_total_composite_mean': 0.605843722820282, 'reward_total_composite_std': 0.07325293868780136, 'epoch': 0.12}

 36%|███▌      | 1179/3300 [2:55:14<4:15:16,  7.22s/it]INFO 04-13 10:16:12 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1180/3300 [2:55:20<4:02:35,  6.87s/it]2026-04-13 10:16:18,606 | INFO | train_grpo_train | metrics_logged mode=train step=1180

                                                       
{'loss': 0.065, 'grad_norm': 12.635881423950195, 'learning_rate': 6.427272727272728e-06, 'num_tokens': 2086088.0, 'completions/mean_length': 42.125, 'completions/min_length': 31.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8736544847488403, 'rewards/meter/std': 0.129952535033226, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9728291630744934, 'rewards/repeat_soft/std': 0.022369252517819405, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.12739142775535583, 'rewards/total_composite/mean': 0.6076606512069702, 'rewards/total_composite/std': 0.09539171308279037, 'reward': 0.6076606512069702, 'reward_std': 0.09539171308279037, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17623835802078247, 'sampling/sampling_logp_difference/max': 1.3247051239013672, 'sampling/importance_sampling_ratio/min': 0.26588135957717896, 'sampling/importance_sampling_ratio/mean': 1.0084788799285889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.268315128982067, 'clip_ratio/low_mean': 0.06542681530117989, 'clip_ratio/low_min': 0.06542681530117989, 'clip_ratio/high_mean': 0.08177647786214948, 'clip_ratio/high_max': 0.08177647786214948, 'clip_ratio/region_mean': 0.14720329316332936, 'reward_total_mean': 0.6076606512069702, 'reward_meter_mean': 0.8736544847488403, 'reward_meter_std': 0.129952535033226, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9728291630744934, 'reward_repeat_soft_std': 0.022369252517819405, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.12739142775535583, 'reward_total_composite_mean': 0.6076606512069702, 'reward_total_composite_std': 0.09539171308279037, 'epoch': 0.12}

 36%|███▌      | 1180/3300 [2:55:20<4:02:35,  6.87s/it]INFO 04-13 10:16:18 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1181/3300 [2:55:27<4:01:35,  6.84s/it]2026-04-13 10:16:25,390 | INFO | train_grpo_train | metrics_logged mode=train step=1181

                                                       
{'loss': 0.014, 'grad_norm': 10.029298782348633, 'learning_rate': 6.424242424242425e-06, 'num_tokens': 2088503.0, 'completions/mean_length': 79.875, 'completions/min_length': 76.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.9892352819442749, 'rewards/meter/std': 0.002484049880877137, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9803063869476318, 'rewards/repeat_soft/std': 0.014201012440025806, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.5421372652053833, 'rewards/total_composite/std': 0.05115868151187897, 'reward': 0.5421372652053833, 'reward_std': 0.05115867778658867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15114262700080872, 'sampling/sampling_logp_difference/max': 1.9406929016113281, 'sampling/importance_sampling_ratio/min': 0.14360441267490387, 'sampling/importance_sampling_ratio/mean': 1.02043616771698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8896564543247223, 'clip_ratio/low_mean': 0.032595694065093994, 'clip_ratio/low_min': 0.032595694065093994, 'clip_ratio/high_mean': 0.09571088943630457, 'clip_ratio/high_max': 0.09571088943630457, 'clip_ratio/region_mean': 0.12830658350139856, 'reward_total_mean': 0.5421372652053833, 'reward_meter_mean': 0.9892352819442749, 'reward_meter_std': 0.002484049880877137, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9803063869476318, 'reward_repeat_soft_std': 0.014201012440025806, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.5421372652053833, 'reward_total_composite_std': 0.05115868151187897, 'epoch': 0.12}

 36%|███▌      | 1181/3300 [2:55:27<4:01:35,  6.84s/it]INFO 04-13 10:16:25 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1182/3300 [2:55:34<3:57:55,  6.74s/it]2026-04-13 10:16:31,937 | INFO | train_grpo_train | metrics_logged mode=train step=1182

                                                       
{'loss': 0.0771, 'grad_norm': 60.17795944213867, 'learning_rate': 6.4212121212121215e-06, 'num_tokens': 2089810.0, 'completions/mean_length': 18.375, 'completions/min_length': 15.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.375, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9784665107727051, 'rewards/meter/std': 0.019904328510165215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9497883319854736, 'rewards/repeat_soft/std': 0.018914852291345596, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6262871026992798, 'rewards/total_composite/std': 0.01159062422811985, 'reward': 0.6262871026992798, 'reward_std': 0.011590617708861828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12480399757623672, 'sampling/sampling_logp_difference/max': 1.5895189046859741, 'sampling/importance_sampling_ratio/min': 0.20402373373508453, 'sampling/importance_sampling_ratio/mean': 1.0191082954406738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7400579452514648, 'clip_ratio/low_mean': 0.04062500037252903, 'clip_ratio/low_min': 0.04062500037252903, 'clip_ratio/high_mean': 0.037738095968961716, 'clip_ratio/high_max': 0.037738095968961716, 'clip_ratio/region_mean': 0.07836309634149075, 'reward_total_mean': 0.6262871026992798, 'reward_meter_mean': 0.9784665107727051, 'reward_meter_std': 0.019904328510165215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9497883319854736, 'reward_repeat_soft_std': 0.018914852291345596, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6262871026992798, 'reward_total_composite_std': 0.01159062422811985, 'epoch': 0.12}

 36%|███▌      | 1182/3300 [2:55:34<3:57:55,  6.74s/it]INFO 04-13 10:16:32 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1183/3300 [2:55:40<3:59:20,  6.78s/it]2026-04-13 10:16:38,818 | INFO | train_grpo_train | metrics_logged mode=train step=1183

                                                       
{'loss': 0.0269, 'grad_norm': 15.561697006225586, 'learning_rate': 6.418181818181819e-06, 'num_tokens': 2091487.0, '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': 0.920923113822937, 'rewards/meter/std': 0.044823601841926575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9764682054519653, 'rewards/repeat_soft/std': 0.02194131538271904, 'rewards/judge_quality/mean': 0.44749999046325684, 'rewards/judge_quality/std': 0.12848013639450073, 'rewards/total_composite/mean': 0.6128696799278259, 'rewards/total_composite/std': 0.06621865183115005, 'reward': 0.6128696799278259, 'reward_std': 0.06621863692998886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12935635447502136, 'sampling/sampling_logp_difference/max': 1.3288724422454834, 'sampling/importance_sampling_ratio/min': 0.2647756338119507, 'sampling/importance_sampling_ratio/mean': 1.0072948932647705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7739732190966606, 'clip_ratio/low_mean': 0.07058823620900512, 'clip_ratio/low_min': 0.07058823620900512, 'clip_ratio/high_mean': 0.03732493054121733, 'clip_ratio/high_max': 0.03732493054121733, 'clip_ratio/region_mean': 0.10791316675022244, 'reward_total_mean': 0.6128696799278259, 'reward_meter_mean': 0.920923113822937, 'reward_meter_std': 0.044823601841926575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9764682054519653, 'reward_repeat_soft_std': 0.02194131538271904, 'reward_judge_quality_mean': 0.44749999046325684, 'reward_judge_quality_std': 0.12848013639450073, 'reward_total_composite_mean': 0.6128696799278259, 'reward_total_composite_std': 0.06621865183115005, 'epoch': 0.12}

 36%|███▌      | 1183/3300 [2:55:41<3:59:20,  6.78s/it]INFO 04-13 10:16:39 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1184/3300 [2:55:47<3:58:25,  6.76s/it]2026-04-13 10:16:45,523 | INFO | train_grpo_train | metrics_logged mode=train step=1184

                                                       
{'loss': 0.0227, 'grad_norm': 14.760201454162598, 'learning_rate': 6.415151515151515e-06, 'num_tokens': 2093134.0, 'completions/mean_length': 40.875, 'completions/min_length': 37.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9717105627059937, 'rewards/meter/std': 0.03349750488996506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9868008494377136, 'rewards/repeat_soft/std': 0.010895688086748123, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.22696760296821594, 'rewards/total_composite/mean': 0.7206075191497803, 'rewards/total_composite/std': 0.14404785633087158, 'reward': 0.7206075191497803, 'reward_std': 0.14404785633087158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15891246497631073, 'sampling/sampling_logp_difference/max': 1.3936738967895508, 'sampling/importance_sampling_ratio/min': 0.2481618970632553, 'sampling/importance_sampling_ratio/mean': 1.0071918964385986, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1980241239070892, 'clip_ratio/low_mean': 0.12017614394426346, 'clip_ratio/low_min': 0.12017614394426346, 'clip_ratio/high_mean': 0.05986295826733112, 'clip_ratio/high_max': 0.05986295826733112, 'clip_ratio/region_mean': 0.18003910221159458, 'reward_total_mean': 0.7206075191497803, 'reward_meter_mean': 0.9717105627059937, 'reward_meter_std': 0.03349750488996506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9868008494377136, 'reward_repeat_soft_std': 0.010895688086748123, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.22696760296821594, 'reward_total_composite_mean': 0.7206075191497803, 'reward_total_composite_std': 0.14404785633087158, 'epoch': 0.12}

 36%|███▌      | 1184/3300 [2:55:47<3:58:25,  6.76s/it]INFO 04-13 10:16:45 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1185/3300 [2:55:54<3:56:05,  6.70s/it]2026-04-13 10:16:52,079 | INFO | train_grpo_train | metrics_logged mode=train step=1185

                                                       
{'loss': -0.0049, 'grad_norm': 16.361007690429688, 'learning_rate': 6.412121212121213e-06, 'num_tokens': 2094701.0, 'completions/mean_length': 43.875, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7773550152778625, 'rewards/meter/std': 0.3657516539096832, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9184046983718872, 'rewards/repeat_soft/std': 0.06233862787485123, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.25150617957115173, 'rewards/total_composite/mean': 0.6161110401153564, 'rewards/total_composite/std': 0.20731280744075775, 'reward': 0.6161110401153564, 'reward_std': 0.20731279253959656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14475561678409576, 'sampling/sampling_logp_difference/max': 2.3401975631713867, 'sampling/importance_sampling_ratio/min': 0.09630860388278961, 'sampling/importance_sampling_ratio/mean': 1.0052721500396729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0590978786349297, 'clip_ratio/low_mean': 0.06890421081334352, 'clip_ratio/low_min': 0.06890421081334352, 'clip_ratio/high_mean': 0.031746032647788525, 'clip_ratio/high_max': 0.031746032647788525, 'clip_ratio/region_mean': 0.10065024346113205, 'reward_total_mean': 0.6161110401153564, 'reward_meter_mean': 0.7773550152778625, 'reward_meter_std': 0.3657516539096832, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9184046983718872, 'reward_repeat_soft_std': 0.06233862787485123, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.25150617957115173, 'reward_total_composite_mean': 0.6161110401153564, 'reward_total_composite_std': 0.20731280744075775, 'epoch': 0.12}

 36%|███▌      | 1185/3300 [2:55:54<3:56:05,  6.70s/it]INFO 04-13 10:16:52 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1186/3300 [2:56:05<4:45:35,  8.11s/it]2026-04-13 10:17:03,426 | INFO | train_grpo_train | metrics_logged mode=train step=1186

                                                       
{'loss': -0.1406, 'grad_norm': 2.0124919414520264, 'learning_rate': 6.40909090909091e-06, 'num_tokens': 2096340.0, 'completions/mean_length': 106.875, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9718071222305298, 'rewards/meter/std': 0.026004765182733536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9764590263366699, 'rewards/repeat_soft/std': 0.006248929537832737, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.14029940962791443, 'rewards/total_composite/mean': 0.5499769449234009, 'rewards/total_composite/std': 0.22244948148727417, 'reward': 0.5499769449234009, 'reward_std': 0.22244949638843536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1452147215604782, 'sampling/sampling_logp_difference/max': 1.3545036315917969, 'sampling/importance_sampling_ratio/min': 0.2580753564834595, 'sampling/importance_sampling_ratio/mean': 1.0315886735916138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2035367637872696, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.12365831341594458, 'clip_ratio/high_max': 0.12365831341594458, 'clip_ratio/region_mean': 0.12365831341594458, 'reward_total_mean': 0.5499769449234009, 'reward_meter_mean': 0.9718071222305298, 'reward_meter_std': 0.026004765182733536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9764590263366699, 'reward_repeat_soft_std': 0.006248929537832737, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.14029940962791443, 'reward_total_composite_mean': 0.5499769449234009, 'reward_total_composite_std': 0.22244948148727417, 'epoch': 0.12}

 36%|███▌      | 1186/3300 [2:56:05<4:45:35,  8.11s/it]INFO 04-13 10:17:03 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1187/3300 [2:56:17<5:26:01,  9.26s/it]2026-04-13 10:17:15,373 | INFO | train_grpo_train | metrics_logged mode=train step=1187

                                                       
{'loss': -0.1604, 'grad_norm': 4.2426886558532715, 'learning_rate': 6.406060606060607e-06, 'num_tokens': 2098159.0, 'completions/mean_length': 124.375, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.44092947244644165, 'rewards/meter/std': 0.3489566147327423, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9642821550369263, 'rewards/repeat_soft/std': 0.04394963011145592, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.22947145998477936, 'rewards/total_composite/mean': 0.45709294080734253, 'rewards/total_composite/std': 0.2502937316894531, 'reward': 0.45709294080734253, 'reward_std': 0.25029370188713074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13997969031333923, 'sampling/sampling_logp_difference/max': 1.2993731498718262, 'sampling/importance_sampling_ratio/min': 0.272702693939209, 'sampling/importance_sampling_ratio/mean': 1.0237239599227905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8705766573548317, 'clip_ratio/low_mean': 0.07386473566293716, 'clip_ratio/low_min': 0.07386473566293716, 'clip_ratio/high_mean': 0.035914814099669456, 'clip_ratio/high_max': 0.035914814099669456, 'clip_ratio/region_mean': 0.10977954976260662, 'reward_total_mean': 0.45709294080734253, 'reward_meter_mean': 0.44092947244644165, 'reward_meter_std': 0.3489566147327423, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9642821550369263, 'reward_repeat_soft_std': 0.04394963011145592, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.22947145998477936, 'reward_total_composite_mean': 0.45709294080734253, 'reward_total_composite_std': 0.2502937316894531, 'epoch': 0.12}

 36%|███▌      | 1187/3300 [2:56:17<5:26:01,  9.26s/it]INFO 04-13 10:17:15 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1188/3300 [2:56:23<4:54:49,  8.38s/it]2026-04-13 10:17:21,694 | INFO | train_grpo_train | metrics_logged mode=train step=1188

                                                       
{'loss': 0.0424, 'grad_norm': 12.961089134216309, 'learning_rate': 6.403030303030303e-06, 'num_tokens': 2100051.0, 'completions/mean_length': 55.5, 'completions/min_length': 50.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.5, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9832470417022705, 'rewards/meter/std': 0.01328179519623518, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9672013521194458, 'rewards/repeat_soft/std': 0.020161520689725876, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.6988184452056885, 'rewards/total_composite/std': 0.1485285609960556, 'reward': 0.6988184452056885, 'reward_std': 0.1485285609960556, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14231687784194946, 'sampling/sampling_logp_difference/max': 1.3428674936294556, 'sampling/importance_sampling_ratio/min': 0.26109591126441956, 'sampling/importance_sampling_ratio/mean': 1.02568781375885, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.107128344476223, 'clip_ratio/low_mean': 0.10461588110774755, 'clip_ratio/low_min': 0.10461588110774755, 'clip_ratio/high_mean': 0.02727272640913725, 'clip_ratio/high_max': 0.02727272640913725, 'clip_ratio/region_mean': 0.1318886075168848, 'reward_total_mean': 0.6988184452056885, 'reward_meter_mean': 0.9832470417022705, 'reward_meter_std': 0.01328179519623518, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9672013521194458, 'reward_repeat_soft_std': 0.020161520689725876, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.6988184452056885, 'reward_total_composite_std': 0.1485285609960556, 'epoch': 0.12}

 36%|███▌      | 1188/3300 [2:56:23<4:54:49,  8.38s/it]INFO 04-13 10:17:21 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1189/3300 [2:56:30<4:34:47,  7.81s/it]2026-04-13 10:17:28,187 | INFO | train_grpo_train | metrics_logged mode=train step=1189

                                                       
{'loss': -0.0018, 'grad_norm': 10.991680145263672, 'learning_rate': 6.4000000000000006e-06, 'num_tokens': 2101752.0, 'completions/mean_length': 55.625, 'completions/min_length': 44.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9614124298095703, 'rewards/meter/std': 0.06390176713466644, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9865628480911255, 'rewards/repeat_soft/std': 0.017839282751083374, 'rewards/judge_quality/mean': 0.8062500357627869, 'rewards/judge_quality/std': 0.16860245168209076, 'rewards/total_composite/mean': 0.8503421545028687, 'rewards/total_composite/std': 0.10541792213916779, 'reward': 0.8503421545028687, 'reward_std': 0.10541790723800659, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13659781217575073, 'sampling/sampling_logp_difference/max': 1.2765960693359375, 'sampling/importance_sampling_ratio/min': 0.27898532152175903, 'sampling/importance_sampling_ratio/mean': 1.0041850805282593, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9991523623466492, 'clip_ratio/low_mean': 0.08090193755924702, 'clip_ratio/low_min': 0.08090193755924702, 'clip_ratio/high_mean': 0.06871948949992657, 'clip_ratio/high_max': 0.06871948949992657, 'clip_ratio/region_mean': 0.14962142705917358, 'reward_total_mean': 0.8503421545028687, 'reward_meter_mean': 0.9614124298095703, 'reward_meter_std': 0.06390176713466644, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9865628480911255, 'reward_repeat_soft_std': 0.017839282751083374, 'reward_judge_quality_mean': 0.8062500357627869, 'reward_judge_quality_std': 0.16860245168209076, 'reward_total_composite_mean': 0.8503421545028687, 'reward_total_composite_std': 0.10541792213916779, 'epoch': 0.12}

 36%|███▌      | 1189/3300 [2:56:30<4:34:47,  7.81s/it]INFO 04-13 10:17:28 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1190/3300 [2:56:36<4:18:25,  7.35s/it]2026-04-13 10:17:34,453 | INFO | train_grpo_train | metrics_logged mode=train step=1190

                                                       
{'loss': 0.0464, 'grad_norm': 14.893120765686035, 'learning_rate': 6.396969696969697e-06, 'num_tokens': 2103394.0, 'completions/mean_length': 51.25, 'completions/min_length': 47.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.983229398727417, 'rewards/meter/std': 0.011052844114601612, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9670378565788269, 'rewards/repeat_soft/std': 0.033053942024707794, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906256318092346, 'rewards/total_composite/mean': 0.6262995004653931, 'rewards/total_composite/std': 0.08399740606546402, 'reward': 0.6262995004653931, 'reward_std': 0.08399741351604462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15653058886528015, 'sampling/sampling_logp_difference/max': 1.4083642959594727, 'sampling/importance_sampling_ratio/min': 0.2445429563522339, 'sampling/importance_sampling_ratio/mean': 1.0138646364212036, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1422750800848007, 'clip_ratio/low_mean': 0.09773599030449986, 'clip_ratio/low_min': 0.09773599030449986, 'clip_ratio/high_mean': 0.039455581456422806, 'clip_ratio/high_max': 0.039455581456422806, 'clip_ratio/region_mean': 0.13719157176092267, 'reward_total_mean': 0.6262995004653931, 'reward_meter_mean': 0.983229398727417, 'reward_meter_std': 0.011052844114601612, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9670378565788269, 'reward_repeat_soft_std': 0.033053942024707794, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906256318092346, 'reward_total_composite_mean': 0.6262995004653931, 'reward_total_composite_std': 0.08399740606546402, 'epoch': 0.12}

 36%|███▌      | 1190/3300 [2:56:36<4:18:25,  7.35s/it]INFO 04-13 10:17:34 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1191/3300 [2:56:43<4:15:54,  7.28s/it]2026-04-13 10:17:41,582 | INFO | train_grpo_train | metrics_logged mode=train step=1191

                                                       
{'loss': 0.051, 'grad_norm': 10.140250205993652, 'learning_rate': 6.393939393939394e-06, 'num_tokens': 2105598.0, 'completions/mean_length': 88.5, 'completions/min_length': 73.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9199206829071045, 'rewards/meter/std': 0.12989304959774017, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9780960083007812, 'rewards/repeat_soft/std': 0.016376817598938942, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5732858180999756, 'rewards/total_composite/std': 0.07956836372613907, 'reward': 0.5732858180999756, 'reward_std': 0.07956835627555847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14494368433952332, 'sampling/sampling_logp_difference/max': 1.548975944519043, 'sampling/importance_sampling_ratio/min': 0.21246543526649475, 'sampling/importance_sampling_ratio/mean': 1.0270488262176514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1746059358119965, 'clip_ratio/low_mean': 0.12986954487860203, 'clip_ratio/low_min': 0.12986954487860203, 'clip_ratio/high_mean': 0.0150602413341403, 'clip_ratio/high_max': 0.0150602413341403, 'clip_ratio/region_mean': 0.14492978621274233, 'reward_total_mean': 0.5732858180999756, 'reward_meter_mean': 0.9199206829071045, 'reward_meter_std': 0.12989304959774017, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9780960083007812, 'reward_repeat_soft_std': 0.016376817598938942, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5732858180999756, 'reward_total_composite_std': 0.07956836372613907, 'epoch': 0.12}

 36%|███▌      | 1191/3300 [2:56:43<4:15:54,  7.28s/it]INFO 04-13 10:17:41 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1192/3300 [2:56:50<4:08:16,  7.07s/it]2026-04-13 10:17:48,150 | INFO | train_grpo_train | metrics_logged mode=train step=1192

                                                       
{'loss': 0.0775, 'grad_norm': 12.860295295715332, 'learning_rate': 6.390909090909091e-06, 'num_tokens': 2107245.0, 'completions/mean_length': 51.875, 'completions/min_length': 39.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7144586443901062, 'rewards/meter/std': 0.37646904587745667, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9928861260414124, 'rewards/repeat_soft/std': 0.014112521894276142, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5860370993614197, 'rewards/total_composite/std': 0.17091825604438782, 'reward': 0.5860370993614197, 'reward_std': 0.17091824114322662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14491435885429382, 'sampling/sampling_logp_difference/max': 1.269066333770752, 'sampling/importance_sampling_ratio/min': 0.281093955039978, 'sampling/importance_sampling_ratio/mean': 1.0024237632751465, 'sampling/importance_sampling_ratio/max': 1.8214935064315796, 'entropy': 1.239496372640133, 'clip_ratio/low_mean': 0.05775345675647259, 'clip_ratio/low_min': 0.05775345675647259, 'clip_ratio/high_mean': 0.0726571804843843, 'clip_ratio/high_max': 0.0726571804843843, 'clip_ratio/region_mean': 0.13041063724085689, 'reward_total_mean': 0.5860370993614197, 'reward_meter_mean': 0.7144586443901062, 'reward_meter_std': 0.37646904587745667, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9928861260414124, 'reward_repeat_soft_std': 0.014112521894276142, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5860370993614197, 'reward_total_composite_std': 0.17091825604438782, 'epoch': 0.12}

 36%|███▌      | 1192/3300 [2:56:50<4:08:16,  7.07s/it]INFO 04-13 10:17:48 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1193/3300 [2:56:56<4:01:20,  6.87s/it]2026-04-13 10:17:54,564 | INFO | train_grpo_train | metrics_logged mode=train step=1193

                                                       
{'loss': 0.1202, 'grad_norm': 13.604833602905273, 'learning_rate': 6.387878787878789e-06, 'num_tokens': 2108993.0, 'completions/mean_length': 44.5, 'completions/min_length': 26.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6406293511390686, 'rewards/meter/std': 0.36187151074409485, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9930717945098877, 'rewards/repeat_soft/std': 0.006495763082057238, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.4684458076953888, 'rewards/total_composite/std': 0.07827360928058624, 'reward': 0.4684458076953888, 'reward_std': 0.07827360183000565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17882013320922852, 'sampling/sampling_logp_difference/max': 2.479620933532715, 'sampling/importance_sampling_ratio/min': 0.08377497643232346, 'sampling/importance_sampling_ratio/mean': 0.98419189453125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.017261803150177, 'clip_ratio/low_mean': 0.049816531129181385, 'clip_ratio/low_min': 0.049816531129181385, 'clip_ratio/high_mean': 0.1255369170103222, 'clip_ratio/high_max': 0.1255369170103222, 'clip_ratio/region_mean': 0.1753534481395036, 'reward_total_mean': 0.4684458076953888, 'reward_meter_mean': 0.6406293511390686, 'reward_meter_std': 0.36187151074409485, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9930717945098877, 'reward_repeat_soft_std': 0.006495763082057238, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.4684458076953888, 'reward_total_composite_std': 0.07827360928058624, 'epoch': 0.12}

 36%|███▌      | 1193/3300 [2:56:56<4:01:20,  6.87s/it]INFO 04-13 10:17:54 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1194/3300 [2:57:02<3:50:05,  6.56s/it]2026-04-13 10:18:00,377 | INFO | train_grpo_train | metrics_logged mode=train step=1194

                                                       
{'loss': 0.0689, 'grad_norm': 19.21489143371582, 'learning_rate': 6.384848484848485e-06, 'num_tokens': 2110251.0, 'completions/mean_length': 21.25, 'completions/min_length': 17.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.91082763671875, 'rewards/meter/std': 0.15774142742156982, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9571585655212402, 'rewards/repeat_soft/std': 0.01005425676703453, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.2775370180606842, 'rewards/total_composite/mean': 0.6985938549041748, 'rewards/total_composite/std': 0.1734265834093094, 'reward': 0.6985938549041748, 'reward_std': 0.1734265685081482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15081371366977692, 'sampling/sampling_logp_difference/max': 1.801047921180725, 'sampling/importance_sampling_ratio/min': 0.165125772356987, 'sampling/importance_sampling_ratio/mean': 1.0244626998901367, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1096824705600739, 'clip_ratio/low_mean': 0.07454019039869308, 'clip_ratio/low_min': 0.07454019039869308, 'clip_ratio/high_mean': 0.07072027772665024, 'clip_ratio/high_max': 0.07072027772665024, 'clip_ratio/region_mean': 0.14526046812534332, 'reward_total_mean': 0.6985938549041748, 'reward_meter_mean': 0.91082763671875, 'reward_meter_std': 0.15774142742156982, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9571585655212402, 'reward_repeat_soft_std': 0.01005425676703453, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.2775370180606842, 'reward_total_composite_mean': 0.6985938549041748, 'reward_total_composite_std': 0.1734265834093094, 'epoch': 0.12}

 36%|███▌      | 1194/3300 [2:57:02<3:50:05,  6.56s/it]INFO 04-13 10:18:00 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1195/3300 [2:57:08<3:45:00,  6.41s/it]2026-04-13 10:18:06,458 | INFO | train_grpo_train | metrics_logged mode=train step=1195

                                                       
{'loss': 0.0729, 'grad_norm': 21.43175506591797, 'learning_rate': 6.381818181818182e-06, 'num_tokens': 2111668.0, 'completions/mean_length': 27.125, 'completions/min_length': 22.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.7170892357826233, 'rewards/meter/std': 0.4193773567676544, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9456349611282349, 'rewards/repeat_soft/std': 0.02970712073147297, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5472277402877808, 'rewards/total_composite/std': 0.1188732162117958, 'reward': 0.5472277402877808, 'reward_std': 0.1188732162117958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19483132660388947, 'sampling/sampling_logp_difference/max': 1.9253969192504883, 'sampling/importance_sampling_ratio/min': 0.1458178609609604, 'sampling/importance_sampling_ratio/mean': 1.0215253829956055, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1635321155190468, 'clip_ratio/low_mean': 0.03362068999558687, 'clip_ratio/low_min': 0.03362068999558687, 'clip_ratio/high_mean': 0.13018791005015373, 'clip_ratio/high_max': 0.13018791005015373, 'clip_ratio/region_mean': 0.1638086000457406, 'reward_total_mean': 0.5472277402877808, 'reward_meter_mean': 0.7170892357826233, 'reward_meter_std': 0.4193773567676544, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9456349611282349, 'reward_repeat_soft_std': 0.02970712073147297, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5472277402877808, 'reward_total_composite_std': 0.1188732162117958, 'epoch': 0.12}

 36%|███▌      | 1195/3300 [2:57:08<3:45:00,  6.41s/it]INFO 04-13 10:18:06 [block_pool.py:378] Successfully reset prefix cache

 36%|███▌      | 1196/3300 [2:57:14<3:43:32,  6.37s/it]2026-04-13 10:18:12,745 | INFO | train_grpo_train | metrics_logged mode=train step=1196

                                                       
{'loss': 0.0023, 'grad_norm': 22.521121978759766, 'learning_rate': 6.37878787878788e-06, 'num_tokens': 2113405.0, 'completions/mean_length': 41.125, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8774996995925903, 'rewards/meter/std': 0.20249055325984955, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9252956509590149, 'rewards/repeat_soft/std': 0.041638702154159546, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5887066721916199, 'rewards/total_composite/std': 0.054475292563438416, 'reward': 0.5887066721916199, 'reward_std': 0.054475292563438416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1631079614162445, 'sampling/sampling_logp_difference/max': 1.6264400482177734, 'sampling/importance_sampling_ratio/min': 0.19662831723690033, 'sampling/importance_sampling_ratio/mean': 1.0326478481292725, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0275902897119522, 'clip_ratio/low_mean': 0.04046817682683468, 'clip_ratio/low_min': 0.04046817682683468, 'clip_ratio/high_mean': 0.10343743953853846, 'clip_ratio/high_max': 0.10343743953853846, 'clip_ratio/region_mean': 0.14390561636537313, 'reward_total_mean': 0.5887066721916199, 'reward_meter_mean': 0.8774996995925903, 'reward_meter_std': 0.20249055325984955, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9252956509590149, 'reward_repeat_soft_std': 0.041638702154159546, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5887066721916199, 'reward_total_composite_std': 0.054475292563438416, 'epoch': 0.12}

 36%|███▌      | 1196/3300 [2:57:14<3:43:32,  6.37s/it]INFO 04-13 10:18:12 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1197/3300 [2:57:21<3:43:58,  6.39s/it]2026-04-13 10:18:19,178 | INFO | train_grpo_train | metrics_logged mode=train step=1197

                                                       
{'loss': 0.108, 'grad_norm': 14.894396781921387, 'learning_rate': 6.375757575757576e-06, 'num_tokens': 2115159.0, 'completions/mean_length': 44.25, 'completions/min_length': 34.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8855868577957153, 'rewards/meter/std': 0.16638562083244324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9769302010536194, 'rewards/repeat_soft/std': 0.028060367330908775, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5948513746261597, 'rewards/total_composite/std': 0.049289148300886154, 'reward': 0.5948513746261597, 'reward_std': 0.04928916320204735, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1516123265028, 'sampling/sampling_logp_difference/max': 2.5276780128479004, 'sampling/importance_sampling_ratio/min': 0.07984420657157898, 'sampling/importance_sampling_ratio/mean': 1.0203876495361328, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8630895391106606, 'clip_ratio/low_mean': 0.030555556528270245, 'clip_ratio/low_min': 0.030555556528270245, 'clip_ratio/high_mean': 0.07832845905795693, 'clip_ratio/high_max': 0.07832845905795693, 'clip_ratio/region_mean': 0.10888401558622718, 'reward_total_mean': 0.5948513746261597, 'reward_meter_mean': 0.8855868577957153, 'reward_meter_std': 0.16638562083244324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9769302010536194, 'reward_repeat_soft_std': 0.028060367330908775, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5948513746261597, 'reward_total_composite_std': 0.049289148300886154, 'epoch': 0.12}

 36%|███▋      | 1197/3300 [2:57:21<3:43:58,  6.39s/it]INFO 04-13 10:18:19 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1198/3300 [2:57:27<3:43:24,  6.38s/it]2026-04-13 10:18:25,519 | INFO | train_grpo_train | metrics_logged mode=train step=1198

                                                       
{'loss': -0.0627, 'grad_norm': 12.346944808959961, 'learning_rate': 6.372727272727274e-06, 'num_tokens': 2117026.0, 'completions/mean_length': 59.375, 'completions/min_length': 50.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8909122943878174, 'rewards/meter/std': 0.14564402401447296, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9201802611351013, 'rewards/repeat_soft/std': 0.05797836557030678, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5553241968154907, 'rewards/total_composite/std': 0.08962288498878479, 'reward': 0.5553241968154907, 'reward_std': 0.089622862637043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15406766533851624, 'sampling/sampling_logp_difference/max': 2.4812746047973633, 'sampling/importance_sampling_ratio/min': 0.08363655209541321, 'sampling/importance_sampling_ratio/mean': 1.0228735208511353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7829513102769852, 'clip_ratio/low_mean': 0.11219498421996832, 'clip_ratio/low_min': 0.11219498421996832, 'clip_ratio/high_mean': 0.04119047708809376, 'clip_ratio/high_max': 0.04119047708809376, 'clip_ratio/region_mean': 0.15338546130806208, 'reward_total_mean': 0.5553241968154907, 'reward_meter_mean': 0.8909122943878174, 'reward_meter_std': 0.14564402401447296, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9201802611351013, 'reward_repeat_soft_std': 0.05797836557030678, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5553241968154907, 'reward_total_composite_std': 0.08962288498878479, 'epoch': 0.12}

 36%|███▋      | 1198/3300 [2:57:27<3:43:24,  6.38s/it]INFO 04-13 10:18:25 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1199/3300 [2:57:33<3:39:42,  6.27s/it]2026-04-13 10:18:31,552 | INFO | train_grpo_train | metrics_logged mode=train step=1199

                                                       
{'loss': -0.0039, 'grad_norm': 12.607721328735352, 'learning_rate': 6.3696969696969706e-06, 'num_tokens': 2118740.0, 'completions/mean_length': 45.25, 'completions/min_length': 36.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9817730784416199, 'rewards/meter/std': 0.014380814507603645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8610401153564453, 'rewards/repeat_soft/std': 0.04926241561770439, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7600692510604858, 'rewards/total_composite/std': 0.17443729937076569, 'reward': 0.7600692510604858, 'reward_std': 0.1744372844696045, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12888361513614655, 'sampling/sampling_logp_difference/max': 1.837851643562317, 'sampling/importance_sampling_ratio/min': 0.15915898978710175, 'sampling/importance_sampling_ratio/mean': 1.0331822633743286, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9980748891830444, 'clip_ratio/low_mean': 0.0693011898547411, 'clip_ratio/low_min': 0.0693011898547411, 'clip_ratio/high_mean': 0.05967627977952361, 'clip_ratio/high_max': 0.05967627977952361, 'clip_ratio/region_mean': 0.1289774696342647, 'reward_total_mean': 0.7600692510604858, 'reward_meter_mean': 0.9817730784416199, 'reward_meter_std': 0.014380814507603645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8610401153564453, 'reward_repeat_soft_std': 0.04926241561770439, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7600692510604858, 'reward_total_composite_std': 0.17443729937076569, 'epoch': 0.12}

 36%|███▋      | 1199/3300 [2:57:33<3:39:42,  6.27s/it]INFO 04-13 10:18:31 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1200/3300 [2:57:39<3:38:16,  6.24s/it]2026-04-13 10:18:37,701 | INFO | train_grpo_train | metrics_logged mode=train step=1200

                                                       
{'loss': 0.0559, 'grad_norm': 12.950010299682617, 'learning_rate': 6.366666666666668e-06, 'num_tokens': 2120393.0, 'completions/mean_length': 46.625, 'completions/min_length': 41.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8502044677734375, 'rewards/meter/std': 0.1859995424747467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9613869190216064, 'rewards/repeat_soft/std': 0.028637342154979706, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.6189001798629761, 'rewards/total_composite/std': 0.14047762751579285, 'reward': 0.6189001798629761, 'reward_std': 0.14047759771347046, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1475149244070053, 'sampling/sampling_logp_difference/max': 1.9337434768676758, 'sampling/importance_sampling_ratio/min': 0.1446058601140976, 'sampling/importance_sampling_ratio/mean': 1.035650610923767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.14615136384964, 'clip_ratio/low_mean': 0.08140896027907729, 'clip_ratio/low_min': 0.08140896027907729, 'clip_ratio/high_mean': 0.0376881156116724, 'clip_ratio/high_max': 0.0376881156116724, 'clip_ratio/region_mean': 0.11909707589074969, 'reward_total_mean': 0.6189001798629761, 'reward_meter_mean': 0.8502044677734375, 'reward_meter_std': 0.1859995424747467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9613869190216064, 'reward_repeat_soft_std': 0.028637342154979706, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.6189001798629761, 'reward_total_composite_std': 0.14047762751579285, 'epoch': 0.12}

 36%|███▋      | 1200/3300 [2:57:39<3:38:16,  6.24s/it]INFO 04-13 10:18:37 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:08<00:34,  4.30s/it]

 30%|███       | 3/10 [00:17<00:43,  6.28s/it]

 40%|████      | 4/10 [00:26<00:44,  7.36s/it]

 50%|█████     | 5/10 [00:31<00:31,  6.27s/it]

 60%|██████    | 6/10 [00:34<00:21,  5.47s/it]

 70%|███████   | 7/10 [00:43<00:19,  6.59s/it]

 80%|████████  | 8/10 [00:52<00:14,  7.27s/it]

 90%|█████████ | 9/10 [00:56<00:06,  6.13s/it]

100%|██████████| 10/10 [00:59<00:00,  5.32s/it]2026-04-13 10:19:46,369 | INFO | train_grpo_train | metrics_logged mode=eval step=1200

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 68.5981, 'eval_samples_per_second': 1.166, 'eval_steps_per_second': 0.146, 'eval_num_tokens': 2120393.0, 'eval_completions/mean_length': 98.275, 'eval_completions/min_length': 30.7, 'eval_completions/max_length': 349.6, 'eval_completions/clipped_ratio': 0.0875, 'eval_completions/mean_terminated_length': 58.62321586608887, 'eval_completions/min_terminated_length': 30.7, 'eval_completions/max_terminated_length': 92.1, 'eval_rewards/meter/mean': 0.6437475919723511, 'eval_rewards/meter/std': 0.378156253695488, 'eval_rewards/count_adherence/mean': 0.8241666734218598, 'eval_rewards/count_adherence/std': 0.15415611043572425, 'eval_rewards/hard_gate/mean': 0.9125, 'eval_rewards/hard_gate/std': 0.2230676978826523, 'eval_rewards/repeat_soft/mean': 0.9744892477989197, 'eval_rewards/repeat_soft/std': 0.030129980575293303, 'eval_rewards/judge_quality/mean': 0.45325000286102296, 'eval_rewards/judge_quality/std': 0.19670253470540047, 'eval_rewards/total_composite/mean': 0.4832571804523468, 'eval_rewards/total_composite/std': 0.19506984874606131, 'eval_reward': 0.4832571804523468, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08069026917219162, 'eval_sampling/sampling_logp_difference/max': 0.9795767307281494, 'eval_sampling/importance_sampling_ratio/min': 0.38638010025024416, 'eval_sampling/importance_sampling_ratio/mean': 1.0223345279693603, 'eval_sampling/importance_sampling_ratio/max': 1.4525879859924316, 'eval_entropy': 0.9961097061634063, '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': 0.4832571804523468, 'eval_reward_meter_mean': 0.6437475919723511, 'eval_reward_meter_std': 0.378156253695488, 'eval_reward_count_adherence_mean': 0.8241666734218598, 'eval_reward_count_adherence_std': 0.15415611043572425, 'eval_reward_hard_gate_mean': 0.9125, 'eval_reward_hard_gate_std': 0.2230676978826523, 'eval_reward_repeat_soft_mean': 0.9744892477989197, 'eval_reward_repeat_soft_std': 0.030129980575293303, 'eval_reward_judge_quality_mean': 0.45325000286102296, 'eval_reward_judge_quality_std': 0.19670253470540047, 'eval_reward_total_composite_mean': 0.4832571804523468, 'eval_reward_total_composite_std': 0.19506984874606131, 'epoch': 0.12}

 36%|███▋      | 1200/3300 [2:58:48<3:38:16,  6.24s/it]

100%|██████████| 10/10 [00:59<00:00,  5.32s/it]

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

 36%|███▋      | 1201/3300 [2:59:03<17:11:53, 29.50s/it]2026-04-13 10:20:01,473 | INFO | train_grpo_train | metrics_logged mode=train step=1201

                                                        
{'loss': -0.1373, 'grad_norm': 3.051819086074829, 'learning_rate': 6.363636363636364e-06, 'num_tokens': 2122638.0, 'completions/mean_length': 134.625, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.71428680419922, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7561625242233276, 'rewards/meter/std': 0.3270168900489807, 'rewards/count_adherence/mean': 0.6500000357627869, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9848114252090454, 'rewards/repeat_soft/std': 0.00961005873978138, 'rewards/judge_quality/mean': 0.3824999928474426, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.4379623830318451, 'rewards/total_composite/std': 0.19056583940982819, 'reward': 0.4379623830318451, 'reward_std': 0.190565824508667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15711110830307007, 'sampling/sampling_logp_difference/max': 1.6434390544891357, 'sampling/importance_sampling_ratio/min': 0.1933140754699707, 'sampling/importance_sampling_ratio/mean': 1.002091407775879, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8832895681262016, 'clip_ratio/low_mean': 0.024122806265950203, 'clip_ratio/low_min': 0.024122806265950203, 'clip_ratio/high_mean': 0.114609788171947, 'clip_ratio/high_max': 0.114609788171947, 'clip_ratio/region_mean': 0.1387325944378972, 'reward_total_mean': 0.4379623830318451, 'reward_meter_mean': 0.7561625242233276, 'reward_meter_std': 0.3270168900489807, 'reward_count_adherence_mean': 0.6500000357627869, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9848114252090454, 'reward_repeat_soft_std': 0.00961005873978138, 'reward_judge_quality_mean': 0.3824999928474426, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.4379623830318451, 'reward_total_composite_std': 0.19056583940982819, 'epoch': 0.12}

 36%|███▋      | 1201/3300 [2:59:03<17:11:53, 29.50s/it]INFO 04-13 10:20:01 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1202/3300 [2:59:09<13:03:36, 22.41s/it]2026-04-13 10:20:07,351 | INFO | train_grpo_train | metrics_logged mode=train step=1202

                                                        
{'loss': 0.1207, 'grad_norm': 33.19580841064453, 'learning_rate': 6.3606060606060615e-06, 'num_tokens': 2124021.0, 'completions/mean_length': 19.875, 'completions/min_length': 15.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8972835540771484, 'rewards/meter/std': 0.1867736577987671, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9558625221252441, 'rewards/repeat_soft/std': 0.018056858330965042, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5894105434417725, 'rewards/total_composite/std': 0.047464992851018906, 'reward': 0.5894105434417725, 'reward_std': 0.047464992851018906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16092872619628906, 'sampling/sampling_logp_difference/max': 1.2862911224365234, 'sampling/importance_sampling_ratio/min': 0.27629363536834717, 'sampling/importance_sampling_ratio/mean': 1.0243053436279297, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0373342111706734, 'clip_ratio/low_mean': 0.02604166604578495, 'clip_ratio/low_min': 0.02604166604578495, 'clip_ratio/high_mean': 0.12085863808169961, 'clip_ratio/high_max': 0.12085863808169961, 'clip_ratio/region_mean': 0.14690030412748456, 'reward_total_mean': 0.5894105434417725, 'reward_meter_mean': 0.8972835540771484, 'reward_meter_std': 0.1867736577987671, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9558625221252441, 'reward_repeat_soft_std': 0.018056858330965042, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5894105434417725, 'reward_total_composite_std': 0.047464992851018906, 'epoch': 0.12}

 36%|███▋      | 1202/3300 [2:59:09<13:03:36, 22.41s/it]INFO 04-13 10:20:07 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1203/3300 [2:59:15<10:11:25, 17.49s/it]2026-04-13 10:20:13,370 | INFO | train_grpo_train | metrics_logged mode=train step=1203

                                                        
{'loss': -0.0348, 'grad_norm': 27.95676040649414, 'learning_rate': 6.357575757575758e-06, 'num_tokens': 2125408.0, 'completions/mean_length': 19.375, 'completions/min_length': 14.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.375, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.6801613569259644, 'rewards/meter/std': 0.3801872134208679, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9488004446029663, 'rewards/repeat_soft/std': 0.03874802216887474, 'rewards/judge_quality/mean': 0.36374998092651367, 'rewards/judge_quality/std': 0.09500939399003983, 'rewards/total_composite/mean': 0.5100680589675903, 'rewards/total_composite/std': 0.1137004867196083, 'reward': 0.5100680589675903, 'reward_std': 0.11370047926902771, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.2133399397134781, 'sampling/sampling_logp_difference/max': 1.2526464462280273, 'sampling/importance_sampling_ratio/min': 0.28574758768081665, 'sampling/importance_sampling_ratio/mean': 1.0028443336486816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 2.0662638172507286, 'clip_ratio/low_mean': 0.0587418326176703, 'clip_ratio/low_min': 0.0587418326176703, 'clip_ratio/high_mean': 0.09089940413832664, 'clip_ratio/high_max': 0.09089940413832664, 'clip_ratio/region_mean': 0.14964123675599694, 'reward_total_mean': 0.5100680589675903, 'reward_meter_mean': 0.6801613569259644, 'reward_meter_std': 0.3801872134208679, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9488004446029663, 'reward_repeat_soft_std': 0.03874802216887474, 'reward_judge_quality_mean': 0.36374998092651367, 'reward_judge_quality_std': 0.09500939399003983, 'reward_total_composite_mean': 0.5100680589675903, 'reward_total_composite_std': 0.1137004867196083, 'epoch': 0.12}

 36%|███▋      | 1203/3300 [2:59:15<10:11:25, 17.49s/it]INFO 04-13 10:20:13 [block_pool.py:378] Successfully reset prefix cache

 36%|███▋      | 1204/3300 [2:59:26<9:03:57, 15.57s/it] 2026-04-13 10:20:24,456 | INFO | train_grpo_train | metrics_logged mode=train step=1204

                                                       
{'loss': -0.1478, 'grad_norm': 2.8766109943389893, 'learning_rate': 6.354545454545455e-06, 'num_tokens': 2127018.0, 'completions/mean_length': 120.25, 'completions/min_length': 54.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.28572082519531, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.8735039234161377, 'rewards/meter/std': 0.2158215194940567, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9839870929718018, 'rewards/repeat_soft/std': 0.011796806007623672, 'rewards/judge_quality/mean': 0.3100000023841858, 'rewards/judge_quality/std': 0.1345893144607544, 'rewards/total_composite/mean': 0.4356952905654907, 'rewards/total_composite/std': 0.19193559885025024, 'reward': 0.4356952905654907, 'reward_std': 0.19193559885025024, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15613046288490295, 'sampling/sampling_logp_difference/max': 1.7565383911132812, 'sampling/importance_sampling_ratio/min': 0.17264144122600555, 'sampling/importance_sampling_ratio/mean': 1.0262750387191772, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2565559148788452, 'clip_ratio/low_mean': 0.02321428619325161, 'clip_ratio/low_min': 0.02321428619325161, 'clip_ratio/high_mean': 0.12299090530723333, 'clip_ratio/high_max': 0.12299090530723333, 'clip_ratio/region_mean': 0.14620519150048494, 'reward_total_mean': 0.4356952905654907, 'reward_meter_mean': 0.8735039234161377, 'reward_meter_std': 0.2158215194940567, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9839870929718018, 'reward_repeat_soft_std': 0.011796806007623672, 'reward_judge_quality_mean': 0.3100000023841858, 'reward_judge_quality_std': 0.1345893144607544, 'reward_total_composite_mean': 0.4356952905654907, 'reward_total_composite_std': 0.19193559885025024, 'epoch': 0.12}

 36%|███▋      | 1204/3300 [2:59:26<9:03:57, 15.57s/it]INFO 04-13 10:20:24 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1205/3300 [2:59:38<8:20:32, 14.34s/it]2026-04-13 10:20:35,906 | INFO | train_grpo_train | metrics_logged mode=train step=1205

                                                       
{'loss': 0.0408, 'grad_norm': 16.52765655517578, 'learning_rate': 6.3515151515151516e-06, 'num_tokens': 2128647.0, 'completions/mean_length': 43.625, 'completions/min_length': 41.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6900322437286377, 'rewards/meter/std': 0.3865191638469696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9763786792755127, 'rewards/repeat_soft/std': 0.022137310355901718, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.2887411117553711, 'rewards/total_composite/mean': 0.649908721446991, 'rewards/total_composite/std': 0.2352641224861145, 'reward': 0.649908721446991, 'reward_std': 0.2352641224861145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.151315838098526, 'sampling/sampling_logp_difference/max': 1.98641037940979, 'sampling/importance_sampling_ratio/min': 0.13718698918819427, 'sampling/importance_sampling_ratio/mean': 1.007492184638977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5421228930354118, 'clip_ratio/low_mean': 0.06591028068214655, 'clip_ratio/low_min': 0.06591028068214655, 'clip_ratio/high_mean': 0.04702952038496733, 'clip_ratio/high_max': 0.04702952038496733, 'clip_ratio/region_mean': 0.11293980106711388, 'reward_total_mean': 0.649908721446991, 'reward_meter_mean': 0.6900322437286377, 'reward_meter_std': 0.3865191638469696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9763786792755127, 'reward_repeat_soft_std': 0.022137310355901718, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.2887411117553711, 'reward_total_composite_mean': 0.649908721446991, 'reward_total_composite_std': 0.2352641224861145, 'epoch': 0.12}

 37%|███▋      | 1205/3300 [2:59:38<8:20:32, 14.34s/it]INFO 04-13 10:20:36 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1206/3300 [2:59:44<7:00:31, 12.05s/it]2026-04-13 10:20:42,628 | INFO | train_grpo_train | metrics_logged mode=train step=1206

                                                       
{'loss': 0.0696, 'grad_norm': 9.657299995422363, 'learning_rate': 6.34848484848485e-06, 'num_tokens': 2130488.0, 'completions/mean_length': 52.125, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9881328344345093, 'rewards/meter/std': 0.006366685964167118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8367646932601929, 'rewards/repeat_soft/std': 0.07589803636074066, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.19334924221038818, 'rewards/total_composite/mean': 0.6268348097801208, 'rewards/total_composite/std': 0.1335621327161789, 'reward': 0.6268348097801208, 'reward_std': 0.1335621327161789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12683773040771484, 'sampling/sampling_logp_difference/max': 1.4757649898529053, 'sampling/importance_sampling_ratio/min': 0.33859530091285706, 'sampling/importance_sampling_ratio/mean': 1.0261094570159912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0749919340014458, 'clip_ratio/low_mean': 0.09005030803382397, 'clip_ratio/low_min': 0.09005030803382397, 'clip_ratio/high_mean': 0.02380952425301075, 'clip_ratio/high_max': 0.02380952425301075, 'clip_ratio/region_mean': 0.11385983228683472, 'reward_total_mean': 0.6268348097801208, 'reward_meter_mean': 0.9881328344345093, 'reward_meter_std': 0.006366685964167118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8367646932601929, 'reward_repeat_soft_std': 0.07589803636074066, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.19334924221038818, 'reward_total_composite_mean': 0.6268348097801208, 'reward_total_composite_std': 0.1335621327161789, 'epoch': 0.12}

 37%|███▋      | 1206/3300 [2:59:44<7:00:31, 12.05s/it]INFO 04-13 10:20:42 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1207/3300 [2:59:50<5:57:09, 10.24s/it]2026-04-13 10:20:48,633 | INFO | train_grpo_train | metrics_logged mode=train step=1207

                                                       
{'loss': 0.0543, 'grad_norm': 29.118677139282227, 'learning_rate': 6.345454545454546e-06, 'num_tokens': 2132026.0, 'completions/mean_length': 25.25, 'completions/min_length': 22.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.25, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.8704559803009033, 'rewards/meter/std': 0.26886969804763794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.948982834815979, 'rewards/repeat_soft/std': 0.03707736358046532, 'rewards/judge_quality/mean': 0.3762499988079071, 'rewards/judge_quality/std': 0.11287634819746017, 'rewards/total_composite/mean': 0.5548781156539917, 'rewards/total_composite/std': 0.1002330482006073, 'reward': 0.5548781156539917, 'reward_std': 0.1002330407500267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1961924135684967, 'sampling/sampling_logp_difference/max': 1.9085369110107422, 'sampling/importance_sampling_ratio/min': 0.14829720556735992, 'sampling/importance_sampling_ratio/mean': 1.0134321451187134, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2682722434401512, 'clip_ratio/low_mean': 0.029230769723653793, 'clip_ratio/low_min': 0.029230769723653793, 'clip_ratio/high_mean': 0.11837542243301868, 'clip_ratio/high_max': 0.11837542243301868, 'clip_ratio/region_mean': 0.14760619215667248, 'reward_total_mean': 0.5548781156539917, 'reward_meter_mean': 0.8704559803009033, 'reward_meter_std': 0.26886969804763794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.948982834815979, 'reward_repeat_soft_std': 0.03707736358046532, 'reward_judge_quality_mean': 0.3762499988079071, 'reward_judge_quality_std': 0.11287634819746017, 'reward_total_composite_mean': 0.5548781156539917, 'reward_total_composite_std': 0.1002330482006073, 'epoch': 0.12}

 37%|███▋      | 1207/3300 [2:59:50<5:57:09, 10.24s/it]INFO 04-13 10:20:48 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1208/3300 [3:00:02<6:12:56, 10.70s/it]2026-04-13 10:21:00,401 | INFO | train_grpo_train | metrics_logged mode=train step=1208

                                                       
{'loss': -0.1173, 'grad_norm': 2.0062859058380127, 'learning_rate': 6.342424242424243e-06, 'num_tokens': 2133776.0, 'completions/mean_length': 167.75, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.740172803401947, 'rewards/meter/std': 0.41915446519851685, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9885010719299316, 'rewards/repeat_soft/std': 0.01072147861123085, 'rewards/judge_quality/mean': 0.5712499618530273, 'rewards/judge_quality/std': 0.4018328785896301, 'rewards/total_composite/mean': 0.6020362973213196, 'rewards/total_composite/std': 0.40308788418769836, 'reward': 0.6020362973213196, 'reward_std': 0.40308788418769836, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13050059974193573, 'sampling/sampling_logp_difference/max': 1.5308656692504883, 'sampling/importance_sampling_ratio/min': 0.2163483202457428, 'sampling/importance_sampling_ratio/mean': 1.0152531862258911, 'sampling/importance_sampling_ratio/max': 1.7760437726974487, 'entropy': 0.87591902166605, 'clip_ratio/low_mean': 0.028764205053448677, 'clip_ratio/low_min': 0.028764205053448677, 'clip_ratio/high_mean': 0.047134652733802795, 'clip_ratio/high_max': 0.047134652733802795, 'clip_ratio/region_mean': 0.07589885778725147, 'reward_total_mean': 0.6020362973213196, 'reward_meter_mean': 0.740172803401947, 'reward_meter_std': 0.41915446519851685, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9885010719299316, 'reward_repeat_soft_std': 0.01072147861123085, 'reward_judge_quality_mean': 0.5712499618530273, 'reward_judge_quality_std': 0.4018328785896301, 'reward_total_composite_mean': 0.6020362973213196, 'reward_total_composite_std': 0.40308788418769836, 'epoch': 0.12}

 37%|███▋      | 1208/3300 [3:00:02<6:12:56, 10.70s/it]INFO 04-13 10:21:00 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1209/3300 [3:00:14<6:25:49, 11.07s/it]2026-04-13 10:21:12,345 | INFO | train_grpo_train | metrics_logged mode=train step=1209

                                                       
{'loss': -0.124, 'grad_norm': 3.3813726902008057, 'learning_rate': 6.33939393939394e-06, 'num_tokens': 2135597.0, 'completions/mean_length': 115.625, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.000003814697266, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.6542345285415649, 'rewards/meter/std': 0.41005656123161316, 'rewards/count_adherence/mean': 0.7000000476837158, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9741815328598022, 'rewards/repeat_soft/std': 0.022501269355416298, 'rewards/judge_quality/mean': 0.3499999940395355, 'rewards/judge_quality/std': 0.15445756912231445, 'rewards/total_composite/mean': 0.41418832540512085, 'rewards/total_composite/std': 0.19170942902565002, 'reward': 0.41418832540512085, 'reward_std': 0.19170941412448883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1486818790435791, 'sampling/sampling_logp_difference/max': 1.7528820037841797, 'sampling/importance_sampling_ratio/min': 0.17327384650707245, 'sampling/importance_sampling_ratio/mean': 1.0278178453445435, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.713676430284977, 'clip_ratio/low_mean': 0.05007684417068958, 'clip_ratio/low_min': 0.05007684417068958, 'clip_ratio/high_mean': 0.08644945221021771, 'clip_ratio/high_max': 0.08644945221021771, 'clip_ratio/region_mean': 0.1365262963809073, 'reward_total_mean': 0.41418832540512085, 'reward_meter_mean': 0.6542345285415649, 'reward_meter_std': 0.41005656123161316, 'reward_count_adherence_mean': 0.7000000476837158, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9741815328598022, 'reward_repeat_soft_std': 0.022501269355416298, 'reward_judge_quality_mean': 0.3499999940395355, 'reward_judge_quality_std': 0.15445756912231445, 'reward_total_composite_mean': 0.41418832540512085, 'reward_total_composite_std': 0.19170942902565002, 'epoch': 0.12}

 37%|███▋      | 1209/3300 [3:00:14<6:25:49, 11.07s/it]INFO 04-13 10:21:12 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1210/3300 [3:00:20<5:30:50,  9.50s/it]2026-04-13 10:21:18,175 | INFO | train_grpo_train | metrics_logged mode=train step=1210

                                                       
{'loss': 0.1422, 'grad_norm': 18.450532913208008, 'learning_rate': 6.336363636363637e-06, 'num_tokens': 2136956.0, 'completions/mean_length': 23.875, 'completions/min_length': 14.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.7502397298812866, 'rewards/meter/std': 0.3737407624721527, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9594065546989441, 'rewards/repeat_soft/std': 0.008749538101255894, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465452551841736, 'rewards/total_composite/mean': 0.7429962158203125, 'rewards/total_composite/std': 0.237124502658844, 'reward': 0.7429962158203125, 'reward_std': 0.2371244877576828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1377735584974289, 'sampling/sampling_logp_difference/max': 1.2020378112792969, 'sampling/importance_sampling_ratio/min': 0.30058106780052185, 'sampling/importance_sampling_ratio/mean': 1.015731692314148, 'sampling/importance_sampling_ratio/max': 1.8095163106918335, 'entropy': 0.9119813740253448, 'clip_ratio/low_mean': 0.030828773509711027, 'clip_ratio/low_min': 0.030828773509711027, 'clip_ratio/high_mean': 0.08687951043248177, 'clip_ratio/high_max': 0.08687951043248177, 'clip_ratio/region_mean': 0.11770828394219279, 'reward_total_mean': 0.7429962158203125, 'reward_meter_mean': 0.7502397298812866, 'reward_meter_std': 0.3737407624721527, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9594065546989441, 'reward_repeat_soft_std': 0.008749538101255894, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465452551841736, 'reward_total_composite_mean': 0.7429962158203125, 'reward_total_composite_std': 0.237124502658844, 'epoch': 0.12}

 37%|███▋      | 1210/3300 [3:00:20<5:30:50,  9.50s/it]INFO 04-13 10:21:18 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1211/3300 [3:00:31<5:49:32, 10.04s/it]2026-04-13 10:21:29,475 | INFO | train_grpo_train | metrics_logged mode=train step=1211

                                                       
{'loss': -0.0765, 'grad_norm': 2.047220468521118, 'learning_rate': 6.333333333333333e-06, 'num_tokens': 2138377.0, 'completions/mean_length': 152.625, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 32.833335876464844, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.3923335075378418, 'rewards/meter/std': 0.39527276158332825, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9937474727630615, 'rewards/repeat_soft/std': 0.011514980345964432, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.33065953850746155, 'rewards/total_composite/mean': 0.39616629481315613, 'rewards/total_composite/std': 0.2613082528114319, 'reward': 0.39616629481315613, 'reward_std': 0.2613082528114319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14189504086971283, 'sampling/sampling_logp_difference/max': 1.5270648002624512, 'sampling/importance_sampling_ratio/min': 0.21717219054698944, 'sampling/importance_sampling_ratio/mean': 1.041178584098816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7455131411552429, 'clip_ratio/low_mean': 0.010135134682059288, 'clip_ratio/low_min': 0.010135134682059288, 'clip_ratio/high_mean': 0.06774892006069422, 'clip_ratio/high_max': 0.06774892006069422, 'clip_ratio/region_mean': 0.0778840547427535, 'reward_total_mean': 0.39616629481315613, 'reward_meter_mean': 0.3923335075378418, 'reward_meter_std': 0.39527276158332825, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9937474727630615, 'reward_repeat_soft_std': 0.011514980345964432, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.33065953850746155, 'reward_total_composite_mean': 0.39616629481315613, 'reward_total_composite_std': 0.2613082528114319, 'epoch': 0.12}

 37%|███▋      | 1211/3300 [3:00:31<5:49:32, 10.04s/it]INFO 04-13 10:21:29 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1212/3300 [3:00:38<5:15:46,  9.07s/it]2026-04-13 10:21:36,297 | INFO | train_grpo_train | metrics_logged mode=train step=1212

                                                       
{'loss': 0.0196, 'grad_norm': 9.696742057800293, 'learning_rate': 6.330303030303031e-06, 'num_tokens': 2140550.0, 'completions/mean_length': 77.625, 'completions/min_length': 70.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.625, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9477736353874207, 'rewards/meter/std': 0.08169277012348175, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9340115785598755, 'rewards/repeat_soft/std': 0.04480970278382301, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5658751726150513, 'rewards/total_composite/std': 0.12052886933088303, 'reward': 0.5658751726150513, 'reward_std': 0.12052884697914124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14696049690246582, 'sampling/sampling_logp_difference/max': 1.8601083755493164, 'sampling/importance_sampling_ratio/min': 0.15565575659275055, 'sampling/importance_sampling_ratio/mean': 1.0185673236846924, 'sampling/importance_sampling_ratio/max': 1.997831106185913, 'entropy': 1.1255382373929024, 'clip_ratio/low_mean': 0.11282453499734402, 'clip_ratio/low_min': 0.11282453499734402, 'clip_ratio/high_mean': 0.017123287543654442, 'clip_ratio/high_max': 0.017123287543654442, 'clip_ratio/region_mean': 0.12994782254099846, 'reward_total_mean': 0.5658751726150513, 'reward_meter_mean': 0.9477736353874207, 'reward_meter_std': 0.08169277012348175, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9340115785598755, 'reward_repeat_soft_std': 0.04480970278382301, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5658751726150513, 'reward_total_composite_std': 0.12052886933088303, 'epoch': 0.12}

 37%|███▋      | 1212/3300 [3:00:38<5:15:46,  9.07s/it]INFO 04-13 10:21:36 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1213/3300 [3:00:44<4:43:09,  8.14s/it]2026-04-13 10:21:42,266 | INFO | train_grpo_train | metrics_logged mode=train step=1213

                                                       
{'loss': 0.0752, 'grad_norm': 24.94639778137207, 'learning_rate': 6.327272727272727e-06, 'num_tokens': 2142167.0, 'completions/mean_length': 36.125, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8696717023849487, 'rewards/meter/std': 0.14692988991737366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9133477210998535, 'rewards/repeat_soft/std': 0.03822728246450424, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.5706425905227661, 'rewards/total_composite/std': 0.05210505425930023, 'reward': 0.5706425905227661, 'reward_std': 0.05210505798459053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12412320077419281, 'sampling/sampling_logp_difference/max': 1.1363520622253418, 'sampling/importance_sampling_ratio/min': 0.3209878206253052, 'sampling/importance_sampling_ratio/mean': 1.0149773359298706, 'sampling/importance_sampling_ratio/max': 1.9755158424377441, 'entropy': 0.82471963763237, 'clip_ratio/low_mean': 0.06491832248866558, 'clip_ratio/low_min': 0.06491832248866558, 'clip_ratio/high_mean': 0.07793703209608793, 'clip_ratio/high_max': 0.07793703209608793, 'clip_ratio/region_mean': 0.1428553545847535, 'reward_total_mean': 0.5706425905227661, 'reward_meter_mean': 0.8696717023849487, 'reward_meter_std': 0.14692988991737366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9133477210998535, 'reward_repeat_soft_std': 0.03822728246450424, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.5706425905227661, 'reward_total_composite_std': 0.05210505425930023, 'epoch': 0.12}

 37%|███▋      | 1213/3300 [3:00:44<4:43:09,  8.14s/it]INFO 04-13 10:21:42 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1214/3300 [3:00:55<5:15:44,  9.08s/it]2026-04-13 10:21:53,584 | INFO | train_grpo_train | metrics_logged mode=train step=1214

                                                       
{'loss': -0.0485, 'grad_norm': 6.261666774749756, 'learning_rate': 6.324242424242425e-06, 'num_tokens': 2143780.0, 'completions/mean_length': 100.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.85714340209961, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.23466385900974274, 'rewards/meter/std': 0.17924705147743225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9908305406570435, 'rewards/repeat_soft/std': 0.009137889370322227, 'rewards/judge_quality/mean': 0.45499998331069946, 'rewards/judge_quality/std': 0.2334829568862915, 'rewards/total_composite/mean': 0.41426029801368713, 'rewards/total_composite/std': 0.054228782653808594, 'reward': 0.41426029801368713, 'reward_std': 0.05422878637909889, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1751175969839096, 'sampling/sampling_logp_difference/max': 1.8739943504333496, 'sampling/importance_sampling_ratio/min': 0.153509259223938, 'sampling/importance_sampling_ratio/mean': 0.9763612151145935, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7590757980942726, 'clip_ratio/low_mean': 0.029545455239713192, 'clip_ratio/low_min': 0.029545455239713192, 'clip_ratio/high_mean': 0.1381261870265007, 'clip_ratio/high_max': 0.1381261870265007, 'clip_ratio/region_mean': 0.1676716422662139, 'reward_total_mean': 0.41426029801368713, 'reward_meter_mean': 0.23466385900974274, 'reward_meter_std': 0.17924705147743225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9908305406570435, 'reward_repeat_soft_std': 0.009137889370322227, 'reward_judge_quality_mean': 0.45499998331069946, 'reward_judge_quality_std': 0.2334829568862915, 'reward_total_composite_mean': 0.41426029801368713, 'reward_total_composite_std': 0.054228782653808594, 'epoch': 0.12}

 37%|███▋      | 1214/3300 [3:00:55<5:15:44,  9.08s/it]INFO 04-13 10:21:53 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1215/3300 [3:01:03<4:58:42,  8.60s/it]2026-04-13 10:22:00,999 | INFO | train_grpo_train | metrics_logged mode=train step=1215

                                                       
{'loss': 0.0459, 'grad_norm': 12.655740737915039, 'learning_rate': 6.3212121212121216e-06, 'num_tokens': 2145739.0, 'completions/mean_length': 63.875, 'completions/min_length': 53.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6042683124542236, 'rewards/meter/std': 0.23978233337402344, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9636364579200745, 'rewards/repeat_soft/std': 0.01974622532725334, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.10541586577892303, 'rewards/total_composite/mean': 0.5332971811294556, 'rewards/total_composite/std': 0.1182984784245491, 'reward': 0.5332971811294556, 'reward_std': 0.1182984858751297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1559729278087616, 'sampling/sampling_logp_difference/max': 1.490548849105835, 'sampling/importance_sampling_ratio/min': 0.22524899244308472, 'sampling/importance_sampling_ratio/mean': 1.002070665359497, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1542490348219872, 'clip_ratio/low_mean': 0.05461149103939533, 'clip_ratio/low_min': 0.05461149103939533, 'clip_ratio/high_mean': 0.08680233173072338, 'clip_ratio/high_max': 0.08680233173072338, 'clip_ratio/region_mean': 0.1414138227701187, 'reward_total_mean': 0.5332971811294556, 'reward_meter_mean': 0.6042683124542236, 'reward_meter_std': 0.23978233337402344, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9636364579200745, 'reward_repeat_soft_std': 0.01974622532725334, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.10541586577892303, 'reward_total_composite_mean': 0.5332971811294556, 'reward_total_composite_std': 0.1182984784245491, 'epoch': 0.12}

 37%|███▋      | 1215/3300 [3:01:03<4:58:42,  8.60s/it]INFO 04-13 10:22:01 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1216/3300 [3:01:09<4:34:29,  7.90s/it]2026-04-13 10:22:07,288 | INFO | train_grpo_train | metrics_logged mode=train step=1216

                                                       
{'loss': -0.0055, 'grad_norm': 16.260692596435547, 'learning_rate': 6.318181818181819e-06, 'num_tokens': 2147291.0, 'completions/mean_length': 42.0, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.985662579536438, 'rewards/meter/std': 0.008431307040154934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9779547452926636, 'rewards/repeat_soft/std': 0.015956100076436996, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.21084441244602203, 'rewards/total_composite/mean': 0.6705451011657715, 'rewards/total_composite/std': 0.1335761845111847, 'reward': 0.6705451011657715, 'reward_std': 0.1335761994123459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13670815527439117, 'sampling/sampling_logp_difference/max': 1.4380803108215332, 'sampling/importance_sampling_ratio/min': 0.23738302290439606, 'sampling/importance_sampling_ratio/mean': 1.0070143938064575, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9411209970712662, 'clip_ratio/low_mean': 0.07499094074591994, 'clip_ratio/low_min': 0.07499094074591994, 'clip_ratio/high_mean': 0.01648550760000944, 'clip_ratio/high_max': 0.01648550760000944, 'clip_ratio/region_mean': 0.09147644834592938, 'reward_total_mean': 0.6705451011657715, 'reward_meter_mean': 0.985662579536438, 'reward_meter_std': 0.008431307040154934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9779547452926636, 'reward_repeat_soft_std': 0.015956100076436996, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.21084441244602203, 'reward_total_composite_mean': 0.6705451011657715, 'reward_total_composite_std': 0.1335761845111847, 'epoch': 0.12}

 37%|███▋      | 1216/3300 [3:01:09<4:34:29,  7.90s/it]INFO 04-13 10:22:07 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1217/3300 [3:01:15<4:14:29,  7.33s/it]2026-04-13 10:22:13,280 | INFO | train_grpo_train | metrics_logged mode=train step=1217

                                                       
{'loss': 0.0067, 'grad_norm': 13.174397468566895, 'learning_rate': 6.315151515151515e-06, 'num_tokens': 2148780.0, 'completions/mean_length': 44.125, 'completions/min_length': 32.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.568111777305603, 'rewards/meter/std': 0.25833097100257874, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.968593955039978, 'rewards/repeat_soft/std': 0.024218551814556122, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.48826876282691956, 'rewards/total_composite/std': 0.09525391459465027, 'reward': 0.48826876282691956, 'reward_std': 0.09525391459465027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17072643339633942, 'sampling/sampling_logp_difference/max': 1.8180158138275146, 'sampling/importance_sampling_ratio/min': 0.16234755516052246, 'sampling/importance_sampling_ratio/mean': 1.0049854516983032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.038679413497448, 'clip_ratio/low_mean': 0.06209849356673658, 'clip_ratio/low_min': 0.06209849356673658, 'clip_ratio/high_mean': 0.08722630236297846, 'clip_ratio/high_max': 0.08722630236297846, 'clip_ratio/region_mean': 0.14932479592971504, 'reward_total_mean': 0.48826876282691956, 'reward_meter_mean': 0.568111777305603, 'reward_meter_std': 0.25833097100257874, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.968593955039978, 'reward_repeat_soft_std': 0.024218551814556122, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.48826876282691956, 'reward_total_composite_std': 0.09525391459465027, 'epoch': 0.12}

 37%|███▋      | 1217/3300 [3:01:15<4:14:29,  7.33s/it]INFO 04-13 10:22:13 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1218/3300 [3:01:22<4:07:18,  7.13s/it]2026-04-13 10:22:19,932 | INFO | train_grpo_train | metrics_logged mode=train step=1218

                                                       
{'loss': 0.098, 'grad_norm': 11.035683631896973, 'learning_rate': 6.3121212121212125e-06, 'num_tokens': 2150629.0, 'completions/mean_length': 73.125, 'completions/min_length': 59.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.7245287895202637, 'rewards/meter/std': 0.37711793184280396, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9799288511276245, 'rewards/repeat_soft/std': 0.01665422134101391, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.49478572607040405, 'rewards/total_composite/std': 0.10128509998321533, 'reward': 0.49478572607040405, 'reward_std': 0.10128511488437653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13519148528575897, 'sampling/sampling_logp_difference/max': 2.9129977226257324, 'sampling/importance_sampling_ratio/min': 0.05431266874074936, 'sampling/importance_sampling_ratio/mean': 1.0262887477874756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8518487140536308, 'clip_ratio/low_mean': 0.04737659823149443, 'clip_ratio/low_min': 0.04737659823149443, 'clip_ratio/high_mean': 0.06279222015291452, 'clip_ratio/high_max': 0.06279222015291452, 'clip_ratio/region_mean': 0.11016881838440895, 'reward_total_mean': 0.49478572607040405, 'reward_meter_mean': 0.7245287895202637, 'reward_meter_std': 0.37711793184280396, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9799288511276245, 'reward_repeat_soft_std': 0.01665422134101391, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.49478572607040405, 'reward_total_composite_std': 0.10128509998321533, 'epoch': 0.12}

 37%|███▋      | 1218/3300 [3:01:22<4:07:18,  7.13s/it]INFO 04-13 10:22:20 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1219/3300 [3:01:28<3:59:33,  6.91s/it]2026-04-13 10:22:26,326 | INFO | train_grpo_train | metrics_logged mode=train step=1219

                                                       
{'loss': -0.1254, 'grad_norm': 14.57003116607666, 'learning_rate': 6.309090909090909e-06, 'num_tokens': 2152274.0, 'completions/mean_length': 45.625, 'completions/min_length': 34.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.03117642179131508, 'rewards/meter/std': 0.05061213672161102, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9878250360488892, 'rewards/repeat_soft/std': 0.009674848057329655, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.0975411981344223, 'rewards/total_composite/mean': 0.318015456199646, 'rewards/total_composite/std': 0.022586451843380928, 'reward': 0.318015456199646, 'reward_std': 0.022586463019251823, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16530752182006836, 'sampling/sampling_logp_difference/max': 1.4637900590896606, 'sampling/importance_sampling_ratio/min': 0.23135775327682495, 'sampling/importance_sampling_ratio/mean': 1.0113688707351685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9115501157939434, 'clip_ratio/low_mean': 0.08679158054292202, 'clip_ratio/low_min': 0.08679158054292202, 'clip_ratio/high_mean': 0.060084665194153786, 'clip_ratio/high_max': 0.060084665194153786, 'clip_ratio/region_mean': 0.1468762457370758, 'reward_total_mean': 0.318015456199646, 'reward_meter_mean': 0.03117642179131508, 'reward_meter_std': 0.05061213672161102, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9878250360488892, 'reward_repeat_soft_std': 0.009674848057329655, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.0975411981344223, 'reward_total_composite_mean': 0.318015456199646, 'reward_total_composite_std': 0.022586451843380928, 'epoch': 0.12}

 37%|███▋      | 1219/3300 [3:01:28<3:59:33,  6.91s/it]INFO 04-13 10:22:26 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1220/3300 [3:01:35<3:59:40,  6.91s/it]2026-04-13 10:22:33,255 | INFO | train_grpo_train | metrics_logged mode=train step=1220

                                                       
{'loss': -0.0312, 'grad_norm': 11.163514137268066, 'learning_rate': 6.306060606060607e-06, 'num_tokens': 2154227.0, 'completions/mean_length': 71.125, 'completions/min_length': 51.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.9727790355682373, 'rewards/meter/std': 0.024574285373091698, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9593303203582764, 'rewards/repeat_soft/std': 0.05558227747678757, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.6099128127098083, 'rewards/total_composite/std': 0.09227675199508667, 'reward': 0.6099128127098083, 'reward_std': 0.09227675199508667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14367783069610596, 'sampling/sampling_logp_difference/max': 1.4354729652404785, 'sampling/importance_sampling_ratio/min': 0.23800277709960938, 'sampling/importance_sampling_ratio/mean': 1.013994812965393, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9198137000203133, 'clip_ratio/low_mean': 0.10493800975382328, 'clip_ratio/low_min': 0.10493800975382328, 'clip_ratio/high_mean': 0.0327856270596385, 'clip_ratio/high_max': 0.0327856270596385, 'clip_ratio/region_mean': 0.13772363681346178, 'reward_total_mean': 0.6099128127098083, 'reward_meter_mean': 0.9727790355682373, 'reward_meter_std': 0.024574285373091698, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9593303203582764, 'reward_repeat_soft_std': 0.05558227747678757, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.6099128127098083, 'reward_total_composite_std': 0.09227675199508667, 'epoch': 0.12}

 37%|███▋      | 1220/3300 [3:01:35<3:59:40,  6.91s/it]INFO 04-13 10:22:33 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1221/3300 [3:01:41<3:52:53,  6.72s/it]2026-04-13 10:22:39,529 | INFO | train_grpo_train | metrics_logged mode=train step=1221

                                                       
{'loss': 0.0271, 'grad_norm': 12.287282943725586, 'learning_rate': 6.303030303030303e-06, 'num_tokens': 2155872.0, 'completions/mean_length': 51.625, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.625, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5839359164237976, 'rewards/meter/std': 0.3622641861438751, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9617362022399902, 'rewards/repeat_soft/std': 0.05584303289651871, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5050953030586243, 'rewards/total_composite/std': 0.10229077190160751, 'reward': 0.5050953030586243, 'reward_std': 0.10229074954986572, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15814530849456787, 'sampling/sampling_logp_difference/max': 1.9131996631622314, 'sampling/importance_sampling_ratio/min': 0.14760734140872955, 'sampling/importance_sampling_ratio/mean': 1.007646918296814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0089810118079185, 'clip_ratio/low_mean': 0.06683951243758202, 'clip_ratio/low_min': 0.06683951243758202, 'clip_ratio/high_mean': 0.0798441655933857, 'clip_ratio/high_max': 0.0798441655933857, 'clip_ratio/region_mean': 0.1466836780309677, 'reward_total_mean': 0.5050953030586243, 'reward_meter_mean': 0.5839359164237976, 'reward_meter_std': 0.3622641861438751, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9617362022399902, 'reward_repeat_soft_std': 0.05584303289651871, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5050953030586243, 'reward_total_composite_std': 0.10229077190160751, 'epoch': 0.12}

 37%|███▋      | 1221/3300 [3:01:41<3:52:53,  6.72s/it]INFO 04-13 10:22:39 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1222/3300 [3:01:48<3:52:52,  6.72s/it]2026-04-13 10:22:46,257 | INFO | train_grpo_train | metrics_logged mode=train step=1222

                                                       
{'loss': 0.0436, 'grad_norm': 15.896768569946289, 'learning_rate': 6.300000000000001e-06, 'num_tokens': 2157397.0, 'completions/mean_length': 40.625, 'completions/min_length': 30.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6093494296073914, 'rewards/meter/std': 0.3976246118545532, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9590468406677246, 'rewards/repeat_soft/std': 0.048096902668476105, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.49619966745376587, 'rewards/total_composite/std': 0.17820341885089874, 'reward': 0.49619966745376587, 'reward_std': 0.17820340394973755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15029707551002502, 'sampling/sampling_logp_difference/max': 1.7684967517852783, 'sampling/importance_sampling_ratio/min': 0.17058923840522766, 'sampling/importance_sampling_ratio/mean': 1.0053178071975708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9463603720068932, 'clip_ratio/low_mean': 0.07149485219269991, 'clip_ratio/low_min': 0.07149485219269991, 'clip_ratio/high_mean': 0.06642336864024401, 'clip_ratio/high_max': 0.06642336864024401, 'clip_ratio/region_mean': 0.13791822083294392, 'reward_total_mean': 0.49619966745376587, 'reward_meter_mean': 0.6093494296073914, 'reward_meter_std': 0.3976246118545532, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9590468406677246, 'reward_repeat_soft_std': 0.048096902668476105, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.49619966745376587, 'reward_total_composite_std': 0.17820341885089874, 'epoch': 0.12}

 37%|███▋      | 1222/3300 [3:01:48<3:52:52,  6.72s/it]INFO 04-13 10:22:46 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1223/3300 [3:01:54<3:51:07,  6.68s/it]2026-04-13 10:22:52,823 | INFO | train_grpo_train | metrics_logged mode=train step=1223

                                                       
{'loss': 0.0731, 'grad_norm': 11.490691184997559, 'learning_rate': 6.296969696969697e-06, 'num_tokens': 2159255.0, 'completions/mean_length': 58.25, 'completions/min_length': 47.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.25, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.7078536748886108, 'rewards/meter/std': 0.2628180682659149, 'rewards/count_adherence/mean': 0.65625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9808167219161987, 'rewards/repeat_soft/std': 0.016820814460515976, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.1911618709564209, 'rewards/total_composite/mean': 0.531480073928833, 'rewards/total_composite/std': 0.13504460453987122, 'reward': 0.531480073928833, 'reward_std': 0.1350446194410324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13397450745105743, 'sampling/sampling_logp_difference/max': 2.532672643661499, 'sampling/importance_sampling_ratio/min': 0.07944640517234802, 'sampling/importance_sampling_ratio/mean': 1.0112334489822388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8477783501148224, 'clip_ratio/low_mean': 0.0765253221616149, 'clip_ratio/low_min': 0.0765253221616149, 'clip_ratio/high_mean': 0.050978997722268105, 'clip_ratio/high_max': 0.050978997722268105, 'clip_ratio/region_mean': 0.127504319883883, 'reward_total_mean': 0.531480073928833, 'reward_meter_mean': 0.7078536748886108, 'reward_meter_std': 0.2628180682659149, 'reward_count_adherence_mean': 0.65625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9808167219161987, 'reward_repeat_soft_std': 0.016820814460515976, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.1911618709564209, 'reward_total_composite_mean': 0.531480073928833, 'reward_total_composite_std': 0.13504460453987122, 'epoch': 0.12}

 37%|███▋      | 1223/3300 [3:01:55<3:51:07,  6.68s/it]INFO 04-13 10:22:53 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1224/3300 [3:02:01<3:50:05,  6.65s/it]2026-04-13 10:22:59,412 | INFO | train_grpo_train | metrics_logged mode=train step=1224

                                                       
{'loss': 0.0102, 'grad_norm': 11.508916854858398, 'learning_rate': 6.293939393939394e-06, 'num_tokens': 2161308.0, 'completions/mean_length': 69.625, 'completions/min_length': 55.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9514361023902893, 'rewards/meter/std': 0.07141122221946716, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9387154579162598, 'rewards/repeat_soft/std': 0.04537630081176758, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5205493569374084, 'rewards/total_composite/std': 0.020482094958424568, 'reward': 0.5205493569374084, 'reward_std': 0.020482098683714867, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1273832470178604, 'sampling/sampling_logp_difference/max': 1.6701726913452148, 'sampling/importance_sampling_ratio/min': 0.18821455538272858, 'sampling/importance_sampling_ratio/mean': 1.0133082866668701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8839424699544907, 'clip_ratio/low_mean': 0.03731842152774334, 'clip_ratio/low_min': 0.03731842152774334, 'clip_ratio/high_mean': 0.07454585377126932, 'clip_ratio/high_max': 0.07454585377126932, 'clip_ratio/region_mean': 0.11186427529901266, 'reward_total_mean': 0.5205493569374084, 'reward_meter_mean': 0.9514361023902893, 'reward_meter_std': 0.07141122221946716, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9387154579162598, 'reward_repeat_soft_std': 0.04537630081176758, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5205493569374084, 'reward_total_composite_std': 0.020482094958424568, 'epoch': 0.12}

 37%|███▋      | 1224/3300 [3:02:01<3:50:05,  6.65s/it]INFO 04-13 10:22:59 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1225/3300 [3:02:07<3:45:11,  6.51s/it]2026-04-13 10:23:05,603 | INFO | train_grpo_train | metrics_logged mode=train step=1225

                                                       
{'loss': 0.0963, 'grad_norm': 13.007549285888672, 'learning_rate': 6.290909090909092e-06, 'num_tokens': 2162785.0, 'completions/mean_length': 34.625, 'completions/min_length': 30.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7389113306999207, 'rewards/meter/std': 0.3665609359741211, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.933692216873169, 'rewards/repeat_soft/std': 0.06272569298744202, 'rewards/judge_quality/mean': 0.5137500166893005, 'rewards/judge_quality/std': 0.20777307450771332, 'rewards/total_composite/mean': 0.6081284284591675, 'rewards/total_composite/std': 0.18631310760974884, 'reward': 0.6081284284591675, 'reward_std': 0.18631310760974884, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13696837425231934, 'sampling/sampling_logp_difference/max': 1.710554599761963, 'sampling/importance_sampling_ratio/min': 0.18076550960540771, 'sampling/importance_sampling_ratio/mean': 1.015040397644043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7119800746440887, 'clip_ratio/low_mean': 0.07738095428794622, 'clip_ratio/low_min': 0.07738095428794622, 'clip_ratio/high_mean': 0.04546847240999341, 'clip_ratio/high_max': 0.04546847240999341, 'clip_ratio/region_mean': 0.12284942669793963, 'reward_total_mean': 0.6081284284591675, 'reward_meter_mean': 0.7389113306999207, 'reward_meter_std': 0.3665609359741211, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.933692216873169, 'reward_repeat_soft_std': 0.06272569298744202, 'reward_judge_quality_mean': 0.5137500166893005, 'reward_judge_quality_std': 0.20777307450771332, 'reward_total_composite_mean': 0.6081284284591675, 'reward_total_composite_std': 0.18631310760974884, 'epoch': 0.12}

 37%|███▋      | 1225/3300 [3:02:07<3:45:11,  6.51s/it]INFO 04-13 10:23:05 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1226/3300 [3:02:13<3:41:09,  6.40s/it]2026-04-13 10:23:11,735 | INFO | train_grpo_train | metrics_logged mode=train step=1226

                                                       
{'loss': 0.0561, 'grad_norm': 11.328813552856445, 'learning_rate': 6.287878787878788e-06, 'num_tokens': 2164366.0, 'completions/mean_length': 38.625, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9663524627685547, 'rewards/meter/std': 0.048049844801425934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9799480438232422, 'rewards/repeat_soft/std': 0.018092192709445953, 'rewards/judge_quality/mean': 0.4937499761581421, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.657935619354248, 'rewards/total_composite/std': 0.11399483680725098, 'reward': 0.657935619354248, 'reward_std': 0.11399482190608978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13801391422748566, 'sampling/sampling_logp_difference/max': 1.7289395332336426, 'sampling/importance_sampling_ratio/min': 0.17747251689434052, 'sampling/importance_sampling_ratio/mean': 0.9950167536735535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.901582732796669, 'clip_ratio/low_mean': 0.10738228587433696, 'clip_ratio/low_min': 0.10738228587433696, 'clip_ratio/high_mean': 0.02302631549537182, 'clip_ratio/high_max': 0.02302631549537182, 'clip_ratio/region_mean': 0.13040860136970878, 'reward_total_mean': 0.657935619354248, 'reward_meter_mean': 0.9663524627685547, 'reward_meter_std': 0.048049844801425934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9799480438232422, 'reward_repeat_soft_std': 0.018092192709445953, 'reward_judge_quality_mean': 0.4937499761581421, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.657935619354248, 'reward_total_composite_std': 0.11399483680725098, 'epoch': 0.12}

 37%|███▋      | 1226/3300 [3:02:13<3:41:09,  6.40s/it]INFO 04-13 10:23:11 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1227/3300 [3:02:24<4:28:41,  7.78s/it]2026-04-13 10:23:22,730 | INFO | train_grpo_train | metrics_logged mode=train step=1227

                                                       
{'loss': -0.0594, 'grad_norm': 3.061837911605835, 'learning_rate': 6.284848484848486e-06, 'num_tokens': 2165732.0, 'completions/mean_length': 83.75, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.571430206298828, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.4164186120033264, 'rewards/meter/std': 0.4045465290546417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9296796321868896, 'rewards/repeat_soft/std': 0.0432172529399395, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.40845227241516113, 'rewards/total_composite/std': 0.1992538571357727, 'reward': 0.40845227241516113, 'reward_std': 0.1992538571357727, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15326885879039764, 'sampling/sampling_logp_difference/max': 1.05519437789917, 'sampling/importance_sampling_ratio/min': 0.34812474250793457, 'sampling/importance_sampling_ratio/mean': 1.0181257724761963, 'sampling/importance_sampling_ratio/max': 1.7947677373886108, 'entropy': 0.9504754319787025, 'clip_ratio/low_mean': 0.034852758049964905, 'clip_ratio/low_min': 0.034852758049964905, 'clip_ratio/high_mean': 0.047062008175998926, 'clip_ratio/high_max': 0.047062008175998926, 'clip_ratio/region_mean': 0.08191476622596383, 'reward_total_mean': 0.40845227241516113, 'reward_meter_mean': 0.4164186120033264, 'reward_meter_std': 0.4045465290546417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9296796321868896, 'reward_repeat_soft_std': 0.0432172529399395, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.40845227241516113, 'reward_total_composite_std': 0.1992538571357727, 'epoch': 0.12}

 37%|███▋      | 1227/3300 [3:02:24<4:28:41,  7.78s/it]INFO 04-13 10:23:22 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1228/3300 [3:02:30<4:10:58,  7.27s/it]2026-04-13 10:23:28,807 | INFO | train_grpo_train | metrics_logged mode=train step=1228

                                                       
{'loss': 0.0354, 'grad_norm': 13.09066390991211, 'learning_rate': 6.2818181818181825e-06, 'num_tokens': 2167365.0, 'completions/mean_length': 48.125, 'completions/min_length': 40.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6213930249214172, 'rewards/meter/std': 0.372549831867218, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9939090013504028, 'rewards/repeat_soft/std': 0.003991083241999149, 'rewards/judge_quality/mean': 0.6525000333786011, 'rewards/judge_quality/std': 0.2418234497308731, 'rewards/total_composite/mean': 0.5973623394966125, 'rewards/total_composite/std': 0.18629467487335205, 'reward': 0.5973623394966125, 'reward_std': 0.18629464507102966, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17509528994560242, 'sampling/sampling_logp_difference/max': 1.982809066772461, 'sampling/importance_sampling_ratio/min': 0.13768193125724792, 'sampling/importance_sampling_ratio/mean': 1.0242944955825806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2939858064055443, 'clip_ratio/low_mean': 0.08759469725191593, 'clip_ratio/low_min': 0.08759469725191593, 'clip_ratio/high_mean': 0.0906465258449316, 'clip_ratio/high_max': 0.0906465258449316, 'clip_ratio/region_mean': 0.17824122309684753, 'reward_total_mean': 0.5973623394966125, 'reward_meter_mean': 0.6213930249214172, 'reward_meter_std': 0.372549831867218, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9939090013504028, 'reward_repeat_soft_std': 0.003991083241999149, 'reward_judge_quality_mean': 0.6525000333786011, 'reward_judge_quality_std': 0.2418234497308731, 'reward_total_composite_mean': 0.5973623394966125, 'reward_total_composite_std': 0.18629467487335205, 'epoch': 0.12}

 37%|███▋      | 1228/3300 [3:02:31<4:10:58,  7.27s/it]INFO 04-13 10:23:29 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1229/3300 [3:02:37<3:58:30,  6.91s/it]2026-04-13 10:23:34,888 | INFO | train_grpo_train | metrics_logged mode=train step=1229

                                                       
{'loss': 0.1696, 'grad_norm': 27.195316314697266, 'learning_rate': 6.27878787878788e-06, 'num_tokens': 2168751.0, 'completions/mean_length': 26.25, 'completions/min_length': 19.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.47008028626441956, 'rewards/meter/std': 0.3700377345085144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9698743224143982, 'rewards/repeat_soft/std': 0.030791107565164566, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.5009973049163818, 'rewards/total_composite/std': 0.11817308515310287, 'reward': 0.5009973049163818, 'reward_std': 0.11817307770252228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1846369206905365, 'sampling/sampling_logp_difference/max': 1.955707311630249, 'sampling/importance_sampling_ratio/min': 0.14146436750888824, 'sampling/importance_sampling_ratio/mean': 0.993920087814331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9901880025863647, 'clip_ratio/low_mean': 0.08829365111887455, 'clip_ratio/low_min': 0.08829365111887455, 'clip_ratio/high_mean': 0.08385943993926048, 'clip_ratio/high_max': 0.08385943993926048, 'clip_ratio/region_mean': 0.17215309105813503, 'reward_total_mean': 0.5009973049163818, 'reward_meter_mean': 0.47008028626441956, 'reward_meter_std': 0.3700377345085144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9698743224143982, 'reward_repeat_soft_std': 0.030791107565164566, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.5009973049163818, 'reward_total_composite_std': 0.11817308515310287, 'epoch': 0.12}

 37%|███▋      | 1229/3300 [3:02:37<3:58:30,  6.91s/it]INFO 04-13 10:23:35 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1230/3300 [3:02:48<4:45:53,  8.29s/it]2026-04-13 10:23:46,385 | INFO | train_grpo_train | metrics_logged mode=train step=1230

                                                       
{'loss': -0.1841, 'grad_norm': 2.2863810062408447, 'learning_rate': 6.275757575757576e-06, 'num_tokens': 2171009.0, 'completions/mean_length': 147.25, 'completions/min_length': 75.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 95.14286041259766, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.8865006566047668, 'rewards/meter/std': 0.1508045494556427, 'rewards/count_adherence/mean': 0.5208333730697632, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9836091995239258, 'rewards/repeat_soft/std': 0.013653808273375034, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.4220849275588989, 'rewards/total_composite/std': 0.17547102272510529, 'reward': 0.4220849275588989, 'reward_std': 0.17547102272510529, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1561456173658371, 'sampling/sampling_logp_difference/max': 2.5082197189331055, 'sampling/importance_sampling_ratio/min': 0.08141304552555084, 'sampling/importance_sampling_ratio/mean': 1.0233533382415771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.137905977666378, 'clip_ratio/low_mean': 0.017326733097434044, 'clip_ratio/low_min': 0.017326733097434044, 'clip_ratio/high_mean': 0.11770072672516108, 'clip_ratio/high_max': 0.11770072672516108, 'clip_ratio/region_mean': 0.13502745982259512, 'reward_total_mean': 0.4220849275588989, 'reward_meter_mean': 0.8865006566047668, 'reward_meter_std': 0.1508045494556427, 'reward_count_adherence_mean': 0.5208333730697632, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9836091995239258, 'reward_repeat_soft_std': 0.013653808273375034, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.4220849275588989, 'reward_total_composite_std': 0.17547102272510529, 'epoch': 0.12}

 37%|███▋      | 1230/3300 [3:02:48<4:45:53,  8.29s/it]INFO 04-13 10:23:46 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1231/3300 [3:02:55<4:29:55,  7.83s/it]2026-04-13 10:23:53,140 | INFO | train_grpo_train | metrics_logged mode=train step=1231

                                                       
{'loss': 0.06, 'grad_norm': 9.06653881072998, 'learning_rate': 6.2727272727272734e-06, 'num_tokens': 2173136.0, 'completions/mean_length': 83.875, 'completions/min_length': 60.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.949677586555481, 'rewards/meter/std': 0.0947742909193039, 'rewards/count_adherence/mean': 0.5416666865348816, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9752757549285889, 'rewards/repeat_soft/std': 0.015269976109266281, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5621194839477539, 'rewards/total_composite/std': 0.09587524831295013, 'reward': 0.5621194839477539, 'reward_std': 0.09587525576353073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14602883160114288, 'sampling/sampling_logp_difference/max': 2.093897819519043, 'sampling/importance_sampling_ratio/min': 0.12320596724748611, 'sampling/importance_sampling_ratio/mean': 1.019765853881836, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1949023455381393, 'clip_ratio/low_mean': 0.09645771980285645, 'clip_ratio/low_min': 0.09645771980285645, 'clip_ratio/high_mean': 0.03787878900766373, 'clip_ratio/high_max': 0.03787878900766373, 'clip_ratio/region_mean': 0.13433650881052017, 'reward_total_mean': 0.5621194839477539, 'reward_meter_mean': 0.949677586555481, 'reward_meter_std': 0.0947742909193039, 'reward_count_adherence_mean': 0.5416666865348816, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9752757549285889, 'reward_repeat_soft_std': 0.015269976109266281, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5621194839477539, 'reward_total_composite_std': 0.09587524831295013, 'epoch': 0.12}

 37%|███▋      | 1231/3300 [3:02:55<4:29:55,  7.83s/it]INFO 04-13 10:23:53 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1232/3300 [3:03:06<5:06:16,  8.89s/it]2026-04-13 10:24:04,541 | INFO | train_grpo_train | metrics_logged mode=train step=1232

                                                       
{'loss': -0.1636, 'grad_norm': 3.878553867340088, 'learning_rate': 6.26969696969697e-06, 'num_tokens': 2174893.0, 'completions/mean_length': 115.625, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.000003814697266, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6648431420326233, 'rewards/meter/std': 0.3916427791118622, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9351570010185242, 'rewards/repeat_soft/std': 0.07089769840240479, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.24224768579006195, 'rewards/total_composite/mean': 0.5185441970825195, 'rewards/total_composite/std': 0.26280826330184937, 'reward': 0.5185441970825195, 'reward_std': 0.26280826330184937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1412878930568695, 'sampling/sampling_logp_difference/max': 2.252089023590088, 'sampling/importance_sampling_ratio/min': 0.10517927259206772, 'sampling/importance_sampling_ratio/mean': 1.015083909034729, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8335136622190475, 'clip_ratio/low_mean': 0.02855603490024805, 'clip_ratio/low_min': 0.02855603490024805, 'clip_ratio/high_mean': 0.08079096302390099, 'clip_ratio/high_max': 0.08079096302390099, 'clip_ratio/region_mean': 0.10934699792414904, 'reward_total_mean': 0.5185441970825195, 'reward_meter_mean': 0.6648431420326233, 'reward_meter_std': 0.3916427791118622, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.17817415297031403, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9351570010185242, 'reward_repeat_soft_std': 0.07089769840240479, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.24224768579006195, 'reward_total_composite_mean': 0.5185441970825195, 'reward_total_composite_std': 0.26280826330184937, 'epoch': 0.12}

 37%|███▋      | 1232/3300 [3:03:06<5:06:16,  8.89s/it]INFO 04-13 10:24:04 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1233/3300 [3:03:18<5:33:01,  9.67s/it]2026-04-13 10:24:15,988 | INFO | train_grpo_train | metrics_logged mode=train step=1233

                                                       
{'loss': -0.1058, 'grad_norm': 2.93407940864563, 'learning_rate': 6.266666666666668e-06, 'num_tokens': 2176433.0, 'completions/mean_length': 97.5, 'completions/min_length': 28.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.28571701049805, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.6622532606124878, 'rewards/meter/std': 0.3364217281341553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9898523092269897, 'rewards/repeat_soft/std': 0.019080858677625656, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.28965190052986145, 'rewards/total_composite/mean': 0.5197000503540039, 'rewards/total_composite/std': 0.2545000910758972, 'reward': 0.5197000503540039, 'reward_std': 0.2545000910758972, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13664236664772034, 'sampling/sampling_logp_difference/max': 1.1050457954406738, 'sampling/importance_sampling_ratio/min': 0.3311957120895386, 'sampling/importance_sampling_ratio/mean': 1.0266621112823486, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.853924110531807, 'clip_ratio/low_mean': 0.04910714365541935, 'clip_ratio/low_min': 0.04910714365541935, 'clip_ratio/high_mean': 0.10055717173963785, 'clip_ratio/high_max': 0.10055717173963785, 'clip_ratio/region_mean': 0.1496643153950572, 'reward_total_mean': 0.5197000503540039, 'reward_meter_mean': 0.6622532606124878, 'reward_meter_std': 0.3364217281341553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9898523092269897, 'reward_repeat_soft_std': 0.019080858677625656, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.28965190052986145, 'reward_total_composite_mean': 0.5197000503540039, 'reward_total_composite_std': 0.2545000910758972, 'epoch': 0.12}

 37%|███▋      | 1233/3300 [3:03:18<5:33:01,  9.67s/it]INFO 04-13 10:24:16 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1234/3300 [3:03:24<4:55:49,  8.59s/it]2026-04-13 10:24:22,105 | INFO | train_grpo_train | metrics_logged mode=train step=1234

                                                       
{'loss': -0.0091, 'grad_norm': 12.671320915222168, 'learning_rate': 6.263636363636364e-06, 'num_tokens': 2178029.0, 'completions/mean_length': 32.5, 'completions/min_length': 28.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.40019360184669495, 'rewards/meter/std': 0.38638606667518616, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9779560565948486, 'rewards/repeat_soft/std': 0.021496865898370743, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4300210475921631, 'rewards/total_composite/std': 0.08451351523399353, 'reward': 0.4300210475921631, 'reward_std': 0.08451351523399353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.136244535446167, 'sampling/sampling_logp_difference/max': 1.4354257583618164, 'sampling/importance_sampling_ratio/min': 0.23801401257514954, 'sampling/importance_sampling_ratio/mean': 1.0039359331130981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9081961289048195, 'clip_ratio/low_mean': 0.0855260866228491, 'clip_ratio/low_min': 0.0855260866228491, 'clip_ratio/high_mean': 0.04243326187133789, 'clip_ratio/high_max': 0.04243326187133789, 'clip_ratio/region_mean': 0.127959348494187, 'reward_total_mean': 0.4300210475921631, 'reward_meter_mean': 0.40019360184669495, 'reward_meter_std': 0.38638606667518616, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9779560565948486, 'reward_repeat_soft_std': 0.021496865898370743, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4300210475921631, 'reward_total_composite_std': 0.08451351523399353, 'epoch': 0.12}

 37%|███▋      | 1234/3300 [3:03:24<4:55:49,  8.59s/it]INFO 04-13 10:24:22 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1235/3300 [3:03:30<4:36:38,  8.04s/it]2026-04-13 10:24:28,858 | INFO | train_grpo_train | metrics_logged mode=train step=1235

                                                       
{'loss': 0.0277, 'grad_norm': 10.910974502563477, 'learning_rate': 6.260606060606062e-06, 'num_tokens': 2179750.0, 'completions/mean_length': 47.125, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6514851450920105, 'rewards/meter/std': 0.457214891910553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9426051378250122, 'rewards/repeat_soft/std': 0.08489826321601868, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.22461079061031342, 'rewards/total_composite/mean': 0.5894964933395386, 'rewards/total_composite/std': 0.19443294405937195, 'reward': 0.5894964933395386, 'reward_std': 0.19443294405937195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15471337735652924, 'sampling/sampling_logp_difference/max': 1.7690258026123047, 'sampling/importance_sampling_ratio/min': 0.1704990118741989, 'sampling/importance_sampling_ratio/mean': 1.0226340293884277, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.998061791062355, 'clip_ratio/low_mean': 0.058662546798586845, 'clip_ratio/low_min': 0.058662546798586845, 'clip_ratio/high_mean': 0.06591849401593208, 'clip_ratio/high_max': 0.06591849401593208, 'clip_ratio/region_mean': 0.12458104081451893, 'reward_total_mean': 0.5894964933395386, 'reward_meter_mean': 0.6514851450920105, 'reward_meter_std': 0.457214891910553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9426051378250122, 'reward_repeat_soft_std': 0.08489826321601868, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.22461079061031342, 'reward_total_composite_mean': 0.5894964933395386, 'reward_total_composite_std': 0.19443294405937195, 'epoch': 0.12}

 37%|███▋      | 1235/3300 [3:03:31<4:36:38,  8.04s/it]INFO 04-13 10:24:29 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1236/3300 [3:03:37<4:22:41,  7.64s/it]2026-04-13 10:24:35,557 | INFO | train_grpo_train | metrics_logged mode=train step=1236

                                                       
{'loss': -0.0314, 'grad_norm': 18.540569305419922, 'learning_rate': 6.257575757575758e-06, 'num_tokens': 2181435.0, 'completions/mean_length': 31.625, 'completions/min_length': 26.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5110172033309937, 'rewards/meter/std': 0.3751196563243866, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9423044323921204, 'rewards/repeat_soft/std': 0.05376303568482399, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.42815661430358887, 'rewards/total_composite/std': 0.11032947897911072, 'reward': 0.42815661430358887, 'reward_std': 0.11032947152853012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1298311948776245, 'sampling/sampling_logp_difference/max': 1.2114909887313843, 'sampling/importance_sampling_ratio/min': 0.2977529764175415, 'sampling/importance_sampling_ratio/mean': 1.0086385011672974, 'sampling/importance_sampling_ratio/max': 1.7687610387802124, 'entropy': 0.8549127578735352, 'clip_ratio/low_mean': 0.07775939162820578, 'clip_ratio/low_min': 0.07775939162820578, 'clip_ratio/high_mean': 0.07042957190424204, 'clip_ratio/high_max': 0.07042957190424204, 'clip_ratio/region_mean': 0.14818896353244781, 'reward_total_mean': 0.42815661430358887, 'reward_meter_mean': 0.5110172033309937, 'reward_meter_std': 0.3751196563243866, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9423044323921204, 'reward_repeat_soft_std': 0.05376303568482399, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.42815661430358887, 'reward_total_composite_std': 0.11032947897911072, 'epoch': 0.12}

 37%|███▋      | 1236/3300 [3:03:37<4:22:41,  7.64s/it]INFO 04-13 10:24:35 [block_pool.py:378] Successfully reset prefix cache

 37%|███▋      | 1237/3300 [3:03:45<4:25:37,  7.73s/it]2026-04-13 10:24:43,486 | INFO | train_grpo_train | metrics_logged mode=train step=1237

                                                       
{'loss': -0.0864, 'grad_norm': 13.851546287536621, 'learning_rate': 6.254545454545455e-06, 'num_tokens': 2183394.0, 'completions/mean_length': 58.875, 'completions/min_length': 48.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.7594131231307983, 'rewards/meter/std': 0.28806328773498535, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.985525369644165, 'rewards/repeat_soft/std': 0.0346342995762825, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.49545133113861084, 'rewards/total_composite/std': 0.0897204726934433, 'reward': 0.49545133113861084, 'reward_std': 0.0897204726934433, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17434343695640564, 'sampling/sampling_logp_difference/max': 2.0161962509155273, 'sampling/importance_sampling_ratio/min': 0.1331610232591629, 'sampling/importance_sampling_ratio/mean': 1.0120218992233276, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1033561006188393, 'clip_ratio/low_mean': 0.07087666913866997, 'clip_ratio/low_min': 0.07087666913866997, 'clip_ratio/high_mean': 0.07048771437257528, 'clip_ratio/high_max': 0.07048771437257528, 'clip_ratio/region_mean': 0.14136438351124525, 'reward_total_mean': 0.49545133113861084, 'reward_meter_mean': 0.7594131231307983, 'reward_meter_std': 0.28806328773498535, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.985525369644165, 'reward_repeat_soft_std': 0.0346342995762825, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.49545133113861084, 'reward_total_composite_std': 0.0897204726934433, 'epoch': 0.12}

 37%|███▋      | 1237/3300 [3:03:45<4:25:37,  7.73s/it]INFO 04-13 10:24:43 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1238/3300 [3:03:52<4:13:28,  7.38s/it]2026-04-13 10:24:50,002 | INFO | train_grpo_train | metrics_logged mode=train step=1238

                                                       
{'loss': 0.018, 'grad_norm': 11.206317901611328, 'learning_rate': 6.251515151515152e-06, 'num_tokens': 2185182.0, 'completions/mean_length': 53.5, 'completions/min_length': 39.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9882075786590576, 'rewards/meter/std': 0.007610471919178963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9839343428611755, 'rewards/repeat_soft/std': 0.015634674578905106, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.6221863627433777, 'rewards/total_composite/std': 0.009470634162425995, 'reward': 0.6221863627433777, 'reward_std': 0.009470636025071144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1383993923664093, 'sampling/sampling_logp_difference/max': 1.1137313842773438, 'sampling/importance_sampling_ratio/min': 0.32833153009414673, 'sampling/importance_sampling_ratio/mean': 1.0288126468658447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1552726477384567, 'clip_ratio/low_mean': 0.06292570754885674, 'clip_ratio/low_min': 0.06292570754885674, 'clip_ratio/high_mean': 0.06277895253151655, 'clip_ratio/high_max': 0.06277895253151655, 'clip_ratio/region_mean': 0.1257046600803733, 'reward_total_mean': 0.6221863627433777, 'reward_meter_mean': 0.9882075786590576, 'reward_meter_std': 0.007610471919178963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9839343428611755, 'reward_repeat_soft_std': 0.015634674578905106, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.6221863627433777, 'reward_total_composite_std': 0.009470634162425995, 'epoch': 0.12}

 38%|███▊      | 1238/3300 [3:03:52<4:13:28,  7.38s/it]INFO 04-13 10:24:50 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1239/3300 [3:03:59<4:10:24,  7.29s/it]2026-04-13 10:24:57,100 | INFO | train_grpo_train | metrics_logged mode=train step=1239

                                                       
{'loss': 0.0055, 'grad_norm': 9.425826072692871, 'learning_rate': 6.248484848484849e-06, 'num_tokens': 2187544.0, 'completions/mean_length': 96.25, 'completions/min_length': 75.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.25, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.8532570600509644, 'rewards/meter/std': 0.23092512786388397, 'rewards/count_adherence/mean': 0.6458333730697632, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9596713781356812, 'rewards/repeat_soft/std': 0.02742459625005722, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.5543482303619385, 'rewards/total_composite/std': 0.1316426694393158, 'reward': 0.5543482303619385, 'reward_std': 0.1316426545381546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14521437883377075, 'sampling/sampling_logp_difference/max': 1.6073007583618164, 'sampling/importance_sampling_ratio/min': 0.2171570062637329, 'sampling/importance_sampling_ratio/mean': 1.0121631622314453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0436527580022812, 'clip_ratio/low_mean': 0.10271982569247484, 'clip_ratio/low_min': 0.10271982569247484, 'clip_ratio/high_mean': 0.030097167938947678, 'clip_ratio/high_max': 0.030097167938947678, 'clip_ratio/region_mean': 0.13281699363142252, 'reward_total_mean': 0.5543482303619385, 'reward_meter_mean': 0.8532570600509644, 'reward_meter_std': 0.23092512786388397, 'reward_count_adherence_mean': 0.6458333730697632, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9596713781356812, 'reward_repeat_soft_std': 0.02742459625005722, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.5543482303619385, 'reward_total_composite_std': 0.1316426694393158, 'epoch': 0.12}

 38%|███▊      | 1239/3300 [3:03:59<4:10:24,  7.29s/it]INFO 04-13 10:24:57 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1240/3300 [3:04:05<3:59:24,  6.97s/it]2026-04-13 10:25:03,325 | INFO | train_grpo_train | metrics_logged mode=train step=1240

                                                       
{'loss': -0.0157, 'grad_norm': 12.375225067138672, 'learning_rate': 6.245454545454545e-06, 'num_tokens': 2189245.0, 'completions/mean_length': 52.625, 'completions/min_length': 44.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7396665215492249, 'rewards/meter/std': 0.3439599275588989, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9697411060333252, 'rewards/repeat_soft/std': 0.02785804308950901, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5908524990081787, 'rewards/total_composite/std': 0.15924030542373657, 'reward': 0.5908524990081787, 'reward_std': 0.15924030542373657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14979474246501923, 'sampling/sampling_logp_difference/max': 2.096376419067383, 'sampling/importance_sampling_ratio/min': 0.12290095537900925, 'sampling/importance_sampling_ratio/mean': 1.023373007774353, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0163159668445587, 'clip_ratio/low_mean': 0.047969188541173935, 'clip_ratio/low_min': 0.047969188541173935, 'clip_ratio/high_mean': 0.07672291994094849, 'clip_ratio/high_max': 0.07672291994094849, 'clip_ratio/region_mean': 0.12469210848212242, 'reward_total_mean': 0.5908524990081787, 'reward_meter_mean': 0.7396665215492249, 'reward_meter_std': 0.3439599275588989, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9697411060333252, 'reward_repeat_soft_std': 0.02785804308950901, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5908524990081787, 'reward_total_composite_std': 0.15924030542373657, 'epoch': 0.12}

 38%|███▊      | 1240/3300 [3:04:05<3:59:24,  6.97s/it]INFO 04-13 10:25:03 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1241/3300 [3:04:11<3:51:07,  6.73s/it]2026-04-13 10:25:09,506 | INFO | train_grpo_train | metrics_logged mode=train step=1241

                                                       
{'loss': 0.033, 'grad_norm': 13.497233390808105, 'learning_rate': 6.2424242424242434e-06, 'num_tokens': 2190866.0, 'completions/mean_length': 50.625, 'completions/min_length': 40.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9617208242416382, 'rewards/meter/std': 0.07272892445325851, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.968368411064148, 'rewards/repeat_soft/std': 0.028197133913636208, 'rewards/judge_quality/mean': 0.5862499475479126, 'rewards/judge_quality/std': 0.19558978080749512, 'rewards/total_composite/mean': 0.714325487613678, 'rewards/total_composite/std': 0.1335224211215973, 'reward': 0.714325487613678, 'reward_std': 0.1335224211215973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1718737632036209, 'sampling/sampling_logp_difference/max': 1.2398936748504639, 'sampling/importance_sampling_ratio/min': 0.29771336913108826, 'sampling/importance_sampling_ratio/mean': 1.018097996711731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3330445513129234, 'clip_ratio/low_mean': 0.0491032125428319, 'clip_ratio/low_min': 0.0491032125428319, 'clip_ratio/high_mean': 0.07631996087729931, 'clip_ratio/high_max': 0.07631996087729931, 'clip_ratio/region_mean': 0.1254231734201312, 'reward_total_mean': 0.714325487613678, 'reward_meter_mean': 0.9617208242416382, 'reward_meter_std': 0.07272892445325851, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.968368411064148, 'reward_repeat_soft_std': 0.028197133913636208, 'reward_judge_quality_mean': 0.5862499475479126, 'reward_judge_quality_std': 0.19558978080749512, 'reward_total_composite_mean': 0.714325487613678, 'reward_total_composite_std': 0.1335224211215973, 'epoch': 0.12}

 38%|███▊      | 1241/3300 [3:04:11<3:51:07,  6.73s/it]INFO 04-13 10:25:09 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1242/3300 [3:04:22<4:34:21,  8.00s/it]2026-04-13 10:25:20,456 | INFO | train_grpo_train | metrics_logged mode=train step=1242

                                                       
{'loss': -0.1331, 'grad_norm': 2.4201745986938477, 'learning_rate': 6.23939393939394e-06, 'num_tokens': 2192596.0, 'completions/mean_length': 109.25, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.71428680419922, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8386589288711548, 'rewards/meter/std': 0.26981183886528015, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9880883693695068, 'rewards/repeat_soft/std': 0.0190111193805933, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.15982133150100708, 'rewards/total_composite/mean': 0.5404645800590515, 'rewards/total_composite/std': 0.2306319624185562, 'reward': 0.5404645800590515, 'reward_std': 0.2306319624185562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1571856141090393, 'sampling/sampling_logp_difference/max': 1.2705446481704712, 'sampling/importance_sampling_ratio/min': 0.28067871928215027, 'sampling/importance_sampling_ratio/mean': 1.0289143323898315, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0195204690098763, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.10808935575187206, 'clip_ratio/high_max': 0.10808935575187206, 'clip_ratio/region_mean': 0.1259464994072914, 'reward_total_mean': 0.5404645800590515, 'reward_meter_mean': 0.8386589288711548, 'reward_meter_std': 0.26981183886528015, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9880883693695068, 'reward_repeat_soft_std': 0.0190111193805933, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.15982133150100708, 'reward_total_composite_mean': 0.5404645800590515, 'reward_total_composite_std': 0.2306319624185562, 'epoch': 0.12}

 38%|███▊      | 1242/3300 [3:04:22<4:34:21,  8.00s/it]INFO 04-13 10:25:20 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1243/3300 [3:04:29<4:23:10,  7.68s/it]2026-04-13 10:25:27,381 | INFO | train_grpo_train | metrics_logged mode=train step=1243

                                                       
{'loss': 0.0496, 'grad_norm': 10.702670097351074, 'learning_rate': 6.236363636363637e-06, 'num_tokens': 2194803.0, 'completions/mean_length': 76.875, 'completions/min_length': 60.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.46705207228660583, 'rewards/meter/std': 0.3572114109992981, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9794241189956665, 'rewards/repeat_soft/std': 0.013340122997760773, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3944188356399536, 'rewards/total_composite/std': 0.09902052581310272, 'reward': 0.3944188356399536, 'reward_std': 0.09902053326368332, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15322375297546387, 'sampling/sampling_logp_difference/max': 2.087918281555176, 'sampling/importance_sampling_ratio/min': 0.12394487857818604, 'sampling/importance_sampling_ratio/mean': 1.0137479305267334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9834109246730804, 'clip_ratio/low_mean': 0.06600053561851382, 'clip_ratio/low_min': 0.06600053561851382, 'clip_ratio/high_mean': 0.05654762126505375, 'clip_ratio/high_max': 0.05654762126505375, 'clip_ratio/region_mean': 0.12254815688356757, 'reward_total_mean': 0.3944188356399536, 'reward_meter_mean': 0.46705207228660583, 'reward_meter_std': 0.3572114109992981, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9794241189956665, 'reward_repeat_soft_std': 0.013340122997760773, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3944188356399536, 'reward_total_composite_std': 0.09902052581310272, 'epoch': 0.12}

 38%|███▊      | 1243/3300 [3:04:29<4:23:10,  7.68s/it]INFO 04-13 10:25:27 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1244/3300 [3:04:41<5:02:02,  8.81s/it]2026-04-13 10:25:38,848 | INFO | train_grpo_train | metrics_logged mode=train step=1244

                                                       
{'loss': -0.1396, 'grad_norm': 3.0950686931610107, 'learning_rate': 6.2333333333333335e-06, 'num_tokens': 2196963.0, 'completions/mean_length': 135.0, 'completions/min_length': 71.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 81.14286041259766, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.8606762886047363, 'rewards/meter/std': 0.34499090909957886, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9489689469337463, 'rewards/repeat_soft/std': 0.04853681102395058, 'rewards/judge_quality/mean': 0.6899999976158142, 'rewards/judge_quality/std': 0.32035693526268005, 'rewards/total_composite/mean': 0.685646116733551, 'rewards/total_composite/std': 0.30944472551345825, 'reward': 0.685646116733551, 'reward_std': 0.30944475531578064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14565177261829376, 'sampling/sampling_logp_difference/max': 1.917027235031128, 'sampling/importance_sampling_ratio/min': 0.14704343676567078, 'sampling/importance_sampling_ratio/mean': 1.0036393404006958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9305947348475456, 'clip_ratio/low_mean': 0.0375093650072813, 'clip_ratio/low_min': 0.0375093650072813, 'clip_ratio/high_mean': 0.06487716641277075, 'clip_ratio/high_max': 0.06487716641277075, 'clip_ratio/region_mean': 0.10238653142005205, 'reward_total_mean': 0.685646116733551, 'reward_meter_mean': 0.8606762886047363, 'reward_meter_std': 0.34499090909957886, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9489689469337463, 'reward_repeat_soft_std': 0.04853681102395058, 'reward_judge_quality_mean': 0.6899999976158142, 'reward_judge_quality_std': 0.32035693526268005, 'reward_total_composite_mean': 0.685646116733551, 'reward_total_composite_std': 0.30944472551345825, 'epoch': 0.12}

 38%|███▊      | 1244/3300 [3:04:41<5:02:02,  8.81s/it]INFO 04-13 10:25:39 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1245/3300 [3:04:47<4:33:13,  7.98s/it]2026-04-13 10:25:44,877 | INFO | train_grpo_train | metrics_logged mode=train step=1245

                                                       
{'loss': 0.0049, 'grad_norm': 23.15663719177246, 'learning_rate': 6.230303030303031e-06, 'num_tokens': 2198306.0, 'completions/mean_length': 22.875, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9307630062103271, 'rewards/meter/std': 0.1662379801273346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9394959211349487, 'rewards/repeat_soft/std': 0.06506536155939102, 'rewards/judge_quality/mean': 0.4649999737739563, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.6307128667831421, 'rewards/total_composite/std': 0.14116933941841125, 'reward': 0.6307128667831421, 'reward_std': 0.14116933941841125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16882959008216858, 'sampling/sampling_logp_difference/max': 1.842193365097046, 'sampling/importance_sampling_ratio/min': 0.15846946835517883, 'sampling/importance_sampling_ratio/mean': 1.0106979608535767, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7975779548287392, 'clip_ratio/low_mean': 0.0898249763995409, 'clip_ratio/low_min': 0.0898249763995409, 'clip_ratio/high_mean': 0.03217391204088926, 'clip_ratio/high_max': 0.03217391204088926, 'clip_ratio/region_mean': 0.12199888844043016, 'reward_total_mean': 0.6307128667831421, 'reward_meter_mean': 0.9307630062103271, 'reward_meter_std': 0.1662379801273346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9394959211349487, 'reward_repeat_soft_std': 0.06506536155939102, 'reward_judge_quality_mean': 0.4649999737739563, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.6307128667831421, 'reward_total_composite_std': 0.14116933941841125, 'epoch': 0.13}

 38%|███▊      | 1245/3300 [3:04:47<4:33:13,  7.98s/it]INFO 04-13 10:25:45 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1246/3300 [3:04:52<4:10:16,  7.31s/it]2026-04-13 10:25:50,640 | INFO | train_grpo_train | metrics_logged mode=train step=1246

                                                       
{'loss': 0.0269, 'grad_norm': 13.141584396362305, 'learning_rate': 6.227272727272727e-06, 'num_tokens': 2199643.0, 'completions/mean_length': 23.125, 'completions/min_length': 18.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.986322283744812, 'rewards/meter/std': 0.006948336027562618, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9170810580253601, 'rewards/repeat_soft/std': 0.06967313587665558, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.06998724490404129, 'rewards/total_composite/mean': 0.6075693368911743, 'rewards/total_composite/std': 0.05287746340036392, 'reward': 0.6075693368911743, 'reward_std': 0.05287746340036392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15039460361003876, 'sampling/sampling_logp_difference/max': 1.7972307205200195, 'sampling/importance_sampling_ratio/min': 0.16575726866722107, 'sampling/importance_sampling_ratio/mean': 1.0215646028518677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0975348427891731, 'clip_ratio/low_mean': 0.009999999776482582, 'clip_ratio/low_min': 0.009999999776482582, 'clip_ratio/high_mean': 0.09229386271908879, 'clip_ratio/high_max': 0.09229386271908879, 'clip_ratio/region_mean': 0.10229386249557137, 'reward_total_mean': 0.6075693368911743, 'reward_meter_mean': 0.986322283744812, 'reward_meter_std': 0.006948336027562618, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9170810580253601, 'reward_repeat_soft_std': 0.06967313587665558, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.06998724490404129, 'reward_total_composite_mean': 0.6075693368911743, 'reward_total_composite_std': 0.05287746340036392, 'epoch': 0.13}

 38%|███▊      | 1246/3300 [3:04:52<4:10:16,  7.31s/it]INFO 04-13 10:25:50 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1247/3300 [3:05:04<4:55:01,  8.62s/it]2026-04-13 10:26:02,353 | INFO | train_grpo_train | metrics_logged mode=train step=1247

                                                       
{'loss': -0.1776, 'grad_norm': 2.893418550491333, 'learning_rate': 6.224242424242425e-06, 'num_tokens': 2201604.0, 'completions/mean_length': 134.125, 'completions/min_length': 65.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 80.14286041259766, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.709868311882019, 'rewards/meter/std': 0.27947330474853516, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9677321314811707, 'rewards/repeat_soft/std': 0.021217970177531242, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.45401251316070557, 'rewards/total_composite/std': 0.19332006573677063, 'reward': 0.45401251316070557, 'reward_std': 0.19332006573677063, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15028485655784607, 'sampling/sampling_logp_difference/max': 2.978888988494873, 'sampling/importance_sampling_ratio/min': 0.050849296152591705, 'sampling/importance_sampling_ratio/mean': 1.010841965675354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8377620056271553, 'clip_ratio/low_mean': 0.03322988003492355, 'clip_ratio/low_min': 0.03322988003492355, 'clip_ratio/high_mean': 0.09654430113732815, 'clip_ratio/high_max': 0.09654430113732815, 'clip_ratio/region_mean': 0.1297741811722517, 'reward_total_mean': 0.45401251316070557, 'reward_meter_mean': 0.709868311882019, 'reward_meter_std': 0.27947330474853516, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9677321314811707, 'reward_repeat_soft_std': 0.021217970177531242, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.45401251316070557, 'reward_total_composite_std': 0.19332006573677063, 'epoch': 0.13}

 38%|███▊      | 1247/3300 [3:05:04<4:55:01,  8.62s/it]INFO 04-13 10:26:02 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1248/3300 [3:05:11<4:41:32,  8.23s/it]2026-04-13 10:26:09,678 | INFO | train_grpo_train | metrics_logged mode=train step=1248

                                                       
{'loss': 0.0259, 'grad_norm': 10.667177200317383, 'learning_rate': 6.221212121212121e-06, 'num_tokens': 2203609.0, 'completions/mean_length': 80.625, 'completions/min_length': 74.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.625, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.7917189598083496, 'rewards/meter/std': 0.26461970806121826, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.973153293132782, 'rewards/repeat_soft/std': 0.030183982104063034, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5621123313903809, 'rewards/total_composite/std': 0.07222156971693039, 'reward': 0.5621123313903809, 'reward_std': 0.07222156226634979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1624266654253006, 'sampling/sampling_logp_difference/max': 2.63492751121521, 'sampling/importance_sampling_ratio/min': 0.07172417640686035, 'sampling/importance_sampling_ratio/mean': 1.023848056793213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1514958143234253, 'clip_ratio/low_mean': 0.06772187165915966, 'clip_ratio/low_min': 0.06772187165915966, 'clip_ratio/high_mean': 0.08548036124557257, 'clip_ratio/high_max': 0.08548036124557257, 'clip_ratio/region_mean': 0.15320223290473223, 'reward_total_mean': 0.5621123313903809, 'reward_meter_mean': 0.7917189598083496, 'reward_meter_std': 0.26461970806121826, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.973153293132782, 'reward_repeat_soft_std': 0.030183982104063034, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5621123313903809, 'reward_total_composite_std': 0.07222156971693039, 'epoch': 0.13}

 38%|███▊      | 1248/3300 [3:05:11<4:41:32,  8.23s/it]INFO 04-13 10:26:10 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1249/3300 [3:05:18<4:29:10,  7.87s/it]2026-04-13 10:26:16,718 | INFO | train_grpo_train | metrics_logged mode=train step=1249

                                                       
{'loss': -0.0403, 'grad_norm': 14.570401191711426, 'learning_rate': 6.218181818181819e-06, 'num_tokens': 2205280.0, 'completions/mean_length': 43.875, 'completions/min_length': 27.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6555452942848206, 'rewards/meter/std': 0.30486875772476196, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9634662866592407, 'rewards/repeat_soft/std': 0.03781601041555405, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5708943605422974, 'rewards/total_composite/std': 0.16406582295894623, 'reward': 0.5708943605422974, 'reward_std': 0.16406580805778503, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17109832167625427, 'sampling/sampling_logp_difference/max': 1.7995681762695312, 'sampling/importance_sampling_ratio/min': 0.16537028551101685, 'sampling/importance_sampling_ratio/mean': 1.0338964462280273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2643868625164032, 'clip_ratio/low_mean': 0.09901054576039314, 'clip_ratio/low_min': 0.09901054576039314, 'clip_ratio/high_mean': 0.09047590382397175, 'clip_ratio/high_max': 0.09047590382397175, 'clip_ratio/region_mean': 0.1894864495843649, 'reward_total_mean': 0.5708943605422974, 'reward_meter_mean': 0.6555452942848206, 'reward_meter_std': 0.30486875772476196, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9634662866592407, 'reward_repeat_soft_std': 0.03781601041555405, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5708943605422974, 'reward_total_composite_std': 0.16406582295894623, 'epoch': 0.13}

 38%|███▊      | 1249/3300 [3:05:18<4:29:10,  7.87s/it]INFO 04-13 10:26:17 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1250/3300 [3:05:25<4:19:41,  7.60s/it]2026-04-13 10:26:23,679 | INFO | train_grpo_train | metrics_logged mode=train step=1250

                                                       
{'loss': 0.0766, 'grad_norm': 17.4160213470459, 'learning_rate': 6.215151515151515e-06, 'num_tokens': 2206889.0, 'completions/mean_length': 42.125, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8886697888374329, 'rewards/meter/std': 0.22303172945976257, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9861406087875366, 'rewards/repeat_soft/std': 0.016790125519037247, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.8407308459281921, 'rewards/total_composite/std': 0.16276921331882477, 'reward': 0.8407308459281921, 'reward_std': 0.16276919841766357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15097838640213013, 'sampling/sampling_logp_difference/max': 1.4429893493652344, 'sampling/importance_sampling_ratio/min': 0.2362205684185028, 'sampling/importance_sampling_ratio/mean': 1.0155298709869385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9842656850814819, 'clip_ratio/low_mean': 0.029223228804767132, 'clip_ratio/low_min': 0.029223228804767132, 'clip_ratio/high_mean': 0.13385965023189783, 'clip_ratio/high_max': 0.13385965023189783, 'clip_ratio/region_mean': 0.16308287903666496, 'reward_total_mean': 0.8407308459281921, 'reward_meter_mean': 0.8886697888374329, 'reward_meter_std': 0.22303172945976257, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9861406087875366, 'reward_repeat_soft_std': 0.016790125519037247, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.8407308459281921, 'reward_total_composite_std': 0.16276921331882477, 'epoch': 0.13}

 38%|███▊      | 1250/3300 [3:05:25<4:19:41,  7.60s/it]INFO 04-13 10:26:24 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.78s/it]

 30%|███       | 3/10 [00:12<00:32,  4.68s/it]

 40%|████      | 4/10 [00:21<00:37,  6.29s/it]

 50%|█████     | 5/10 [00:25<00:27,  5.58s/it]

 60%|██████    | 6/10 [00:29<00:20,  5.17s/it]

 70%|███████   | 7/10 [00:33<00:14,  4.70s/it]

 80%|████████  | 8/10 [00:42<00:12,  6.04s/it]

 90%|█████████ | 9/10 [00:45<00:05,  5.22s/it]

100%|██████████| 10/10 [00:49<00:00,  4.62s/it]2026-04-13 10:27:16,781 | INFO | train_grpo_train | metrics_logged mode=eval step=1250

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 53.0271, 'eval_samples_per_second': 1.509, 'eval_steps_per_second': 0.189, 'eval_num_tokens': 2206889.0, 'eval_completions/mean_length': 79.4625, 'eval_completions/min_length': 30.6, 'eval_completions/max_length': 223.3, 'eval_completions/clipped_ratio': 0.0375, 'eval_completions/mean_terminated_length': 62.56071548461914, 'eval_completions/min_terminated_length': 30.6, 'eval_completions/max_terminated_length': 97.8, 'eval_rewards/meter/mean': 0.7589913964271545, 'eval_rewards/meter/std': 0.2847945909947157, 'eval_rewards/count_adherence/mean': 0.8977083384990692, 'eval_rewards/count_adherence/std': 0.12962410226464272, 'eval_rewards/hard_gate/mean': 0.9625, 'eval_rewards/hard_gate/std': 0.10606601536273956, 'eval_rewards/repeat_soft/mean': 0.955734246969223, 'eval_rewards/repeat_soft/std': 0.04282893724739552, 'eval_rewards/judge_quality/mean': 0.47662500441074374, 'eval_rewards/judge_quality/std': 0.16404289174824954, 'eval_rewards/total_composite/mean': 0.5555821835994721, 'eval_rewards/total_composite/std': 0.16687288582324983, 'eval_reward': 0.5555821835994721, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.08312602117657661, 'eval_sampling/sampling_logp_difference/max': 0.983290719985962, 'eval_sampling/importance_sampling_ratio/min': 0.38558162450790406, 'eval_sampling/importance_sampling_ratio/mean': 1.0188080310821532, 'eval_sampling/importance_sampling_ratio/max': 1.5668889284133911, 'eval_entropy': 0.9307186782360077, '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': 0.5555821835994721, 'eval_reward_meter_mean': 0.7589913964271545, 'eval_reward_meter_std': 0.2847945909947157, 'eval_reward_count_adherence_mean': 0.8977083384990692, 'eval_reward_count_adherence_std': 0.12962410226464272, 'eval_reward_hard_gate_mean': 0.9625, 'eval_reward_hard_gate_std': 0.10606601536273956, 'eval_reward_repeat_soft_mean': 0.955734246969223, 'eval_reward_repeat_soft_std': 0.04282893724739552, 'eval_reward_judge_quality_mean': 0.47662500441074374, 'eval_reward_judge_quality_std': 0.16404289174824954, 'eval_reward_total_composite_mean': 0.5555821835994721, 'eval_reward_total_composite_std': 0.16687288582324983, 'epoch': 0.13}

 38%|███▊      | 1250/3300 [3:06:19<4:19:41,  7.60s/it]

100%|██████████| 10/10 [00:49<00:00,  4.62s/it]

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

 38%|███▊      | 1251/3300 [3:06:33<14:36:23, 25.66s/it]2026-04-13 10:27:31,446 | INFO | train_grpo_train | metrics_logged mode=train step=1251

                                                        
{'loss': -0.1316, 'grad_norm': 3.2266314029693604, 'learning_rate': 6.212121212121213e-06, 'num_tokens': 2208761.0, 'completions/mean_length': 129.0, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.28572082519531, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.47363999485969543, 'rewards/meter/std': 0.3648098111152649, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9949181079864502, 'rewards/repeat_soft/std': 0.003754963865503669, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.23046152293682098, 'rewards/total_composite/mean': 0.4828604459762573, 'rewards/total_composite/std': 0.24279141426086426, 'reward': 0.4828604459762573, 'reward_std': 0.24279142916202545, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.155914306640625, 'sampling/sampling_logp_difference/max': 1.3624074459075928, 'sampling/importance_sampling_ratio/min': 0.25604361295700073, 'sampling/importance_sampling_ratio/mean': 1.018334984779358, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0389277786016464, 'clip_ratio/low_mean': 0.05251501500606537, 'clip_ratio/low_min': 0.05251501500606537, 'clip_ratio/high_mean': 0.08952989056706429, 'clip_ratio/high_max': 0.08952989056706429, 'clip_ratio/region_mean': 0.14204490557312965, 'reward_total_mean': 0.4828604459762573, 'reward_meter_mean': 0.47363999485969543, 'reward_meter_std': 0.3648098111152649, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9949181079864502, 'reward_repeat_soft_std': 0.003754963865503669, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.23046152293682098, 'reward_total_composite_mean': 0.4828604459762573, 'reward_total_composite_std': 0.24279141426086426, 'epoch': 0.13}

 38%|███▊      | 1251/3300 [3:06:33<14:36:23, 25.66s/it]INFO 04-13 10:27:31 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1252/3300 [3:06:45<12:11:57, 21.44s/it]2026-04-13 10:27:43,056 | INFO | train_grpo_train | metrics_logged mode=train step=1252

                                                        
{'loss': -0.0997, 'grad_norm': 3.2271807193756104, 'learning_rate': 6.209090909090909e-06, 'num_tokens': 2210316.0, 'completions/mean_length': 96.375, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7558550834655762, 'rewards/meter/std': 0.3269907534122467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9275693893432617, 'rewards/repeat_soft/std': 0.05068346485495567, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.15665589272975922, 'rewards/total_composite/mean': 0.4778561294078827, 'rewards/total_composite/std': 0.21120095252990723, 'reward': 0.4778561294078827, 'reward_std': 0.21120095252990723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13341522216796875, 'sampling/sampling_logp_difference/max': 1.2374773025512695, 'sampling/importance_sampling_ratio/min': 0.2901151776313782, 'sampling/importance_sampling_ratio/mean': 1.0314818620681763, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7292150855064392, 'clip_ratio/low_mean': 0.0399305559694767, 'clip_ratio/low_min': 0.0399305559694767, 'clip_ratio/high_mean': 0.055104373721405864, 'clip_ratio/high_max': 0.055104373721405864, 'clip_ratio/region_mean': 0.09503492969088256, 'reward_total_mean': 0.4778561294078827, 'reward_meter_mean': 0.7558550834655762, 'reward_meter_std': 0.3269907534122467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9275693893432617, 'reward_repeat_soft_std': 0.05068346485495567, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.15665589272975922, 'reward_total_composite_mean': 0.4778561294078827, 'reward_total_composite_std': 0.21120095252990723, 'epoch': 0.13}

 38%|███▊      | 1252/3300 [3:06:45<12:11:57, 21.44s/it]INFO 04-13 10:27:43 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1253/3300 [3:06:52<9:45:12, 17.15s/it] 2026-04-13 10:27:50,192 | INFO | train_grpo_train | metrics_logged mode=train step=1253

                                                       
{'loss': 0.0138, 'grad_norm': 10.025697708129883, 'learning_rate': 6.206060606060606e-06, 'num_tokens': 2212242.0, 'completions/mean_length': 65.75, 'completions/min_length': 58.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.75, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.6598318815231323, 'rewards/meter/std': 0.27221354842185974, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9810633659362793, 'rewards/repeat_soft/std': 0.01688402332365513, 'rewards/judge_quality/mean': 0.6825000047683716, 'rewards/judge_quality/std': 0.23260945081710815, 'rewards/total_composite/mean': 0.6451053619384766, 'rewards/total_composite/std': 0.15650221705436707, 'reward': 0.6451053619384766, 'reward_std': 0.15650221705436707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1306808590888977, 'sampling/sampling_logp_difference/max': 2.628945827484131, 'sampling/importance_sampling_ratio/min': 0.07215448468923569, 'sampling/importance_sampling_ratio/mean': 0.9933652877807617, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5045273192226887, 'clip_ratio/low_mean': 0.05152179766446352, 'clip_ratio/low_min': 0.05152179766446352, 'clip_ratio/high_mean': 0.06639338470995426, 'clip_ratio/high_max': 0.06639338470995426, 'clip_ratio/region_mean': 0.11791518237441778, 'reward_total_mean': 0.6451053619384766, 'reward_meter_mean': 0.6598318815231323, 'reward_meter_std': 0.27221354842185974, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9810633659362793, 'reward_repeat_soft_std': 0.01688402332365513, 'reward_judge_quality_mean': 0.6825000047683716, 'reward_judge_quality_std': 0.23260945081710815, 'reward_total_composite_mean': 0.6451053619384766, 'reward_total_composite_std': 0.15650221705436707, 'epoch': 0.13}

 38%|███▊      | 1253/3300 [3:06:52<9:45:12, 17.15s/it]INFO 04-13 10:27:50 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1254/3300 [3:07:03<8:47:20, 15.46s/it]2026-04-13 10:28:01,722 | INFO | train_grpo_train | metrics_logged mode=train step=1254

                                                       
{'loss': -0.1015, 'grad_norm': 3.218287706375122, 'learning_rate': 6.203030303030304e-06, 'num_tokens': 2213875.0, 'completions/mean_length': 111.125, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.85714340209961, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7664738297462463, 'rewards/meter/std': 0.2953992187976837, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9914422035217285, 'rewards/repeat_soft/std': 0.011019350029528141, 'rewards/judge_quality/mean': 0.5400000214576721, 'rewards/judge_quality/std': 0.2958281636238098, 'rewards/total_composite/mean': 0.5712922811508179, 'rewards/total_composite/std': 0.283235102891922, 'reward': 0.5712922811508179, 'reward_std': 0.283235102891922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.155284583568573, 'sampling/sampling_logp_difference/max': 1.6692895889282227, 'sampling/importance_sampling_ratio/min': 0.18838085234165192, 'sampling/importance_sampling_ratio/mean': 1.0308421850204468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1521489322185516, 'clip_ratio/low_mean': 0.035098522901535034, 'clip_ratio/low_min': 0.035098522901535034, 'clip_ratio/high_mean': 0.0953163430094719, 'clip_ratio/high_max': 0.0953163430094719, 'clip_ratio/region_mean': 0.13041486591100693, 'reward_total_mean': 0.5712922811508179, 'reward_meter_mean': 0.7664738297462463, 'reward_meter_std': 0.2953992187976837, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9914422035217285, 'reward_repeat_soft_std': 0.011019350029528141, 'reward_judge_quality_mean': 0.5400000214576721, 'reward_judge_quality_std': 0.2958281636238098, 'reward_total_composite_mean': 0.5712922811508179, 'reward_total_composite_std': 0.283235102891922, 'epoch': 0.13}

 38%|███▊      | 1254/3300 [3:07:03<8:47:20, 15.46s/it]INFO 04-13 10:28:01 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1255/3300 [3:07:10<7:17:41, 12.84s/it]2026-04-13 10:28:08,440 | INFO | train_grpo_train | metrics_logged mode=train step=1255

                                                       
{'loss': 0.0348, 'grad_norm': 13.712302207946777, 'learning_rate': 6.200000000000001e-06, 'num_tokens': 2215651.0, 'completions/mean_length': 50.0, 'completions/min_length': 44.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7190017700195312, 'rewards/meter/std': 0.4052329659461975, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9594361186027527, 'rewards/repeat_soft/std': 0.03879115730524063, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.11667262762784958, 'rewards/total_composite/mean': 0.5490782856941223, 'rewards/total_composite/std': 0.14828507602214813, 'reward': 0.5490782856941223, 'reward_std': 0.14828507602214813, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15267081558704376, 'sampling/sampling_logp_difference/max': 1.351280927658081, 'sampling/importance_sampling_ratio/min': 0.25890839099884033, 'sampling/importance_sampling_ratio/mean': 1.0056182146072388, 'sampling/importance_sampling_ratio/max': 1.9183979034423828, 'entropy': 0.8657406121492386, 'clip_ratio/low_mean': 0.06455399002879858, 'clip_ratio/low_min': 0.06455399002879858, 'clip_ratio/high_mean': 0.1000808821991086, 'clip_ratio/high_max': 0.1000808821991086, 'clip_ratio/region_mean': 0.16463487222790718, 'reward_total_mean': 0.5490782856941223, 'reward_meter_mean': 0.7190017700195312, 'reward_meter_std': 0.4052329659461975, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9594361186027527, 'reward_repeat_soft_std': 0.03879115730524063, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.11667262762784958, 'reward_total_composite_mean': 0.5490782856941223, 'reward_total_composite_std': 0.14828507602214813, 'epoch': 0.13}

 38%|███▊      | 1255/3300 [3:07:10<7:17:41, 12.84s/it]INFO 04-13 10:28:08 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1256/3300 [3:07:17<6:13:23, 10.96s/it]2026-04-13 10:28:15,009 | INFO | train_grpo_train | metrics_logged mode=train step=1256

                                                       
{'loss': 0.0037, 'grad_norm': 11.209565162658691, 'learning_rate': 6.196969696969698e-06, 'num_tokens': 2217223.0, 'completions/mean_length': 39.5, 'completions/min_length': 31.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6701686382293701, 'rewards/meter/std': 0.3509586751461029, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9642779231071472, 'rewards/repeat_soft/std': 0.04290106147527695, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.49856626987457275, 'rewards/total_composite/std': 0.2385425567626953, 'reward': 0.49856626987457275, 'reward_std': 0.2385425567626953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15506722033023834, 'sampling/sampling_logp_difference/max': 1.6867170333862305, 'sampling/importance_sampling_ratio/min': 0.18512627482414246, 'sampling/importance_sampling_ratio/mean': 1.0095295906066895, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0518463253974915, 'clip_ratio/low_mean': 0.07911140471696854, 'clip_ratio/low_min': 0.07911140471696854, 'clip_ratio/high_mean': 0.08315983507782221, 'clip_ratio/high_max': 0.08315983507782221, 'clip_ratio/region_mean': 0.16227123979479074, 'reward_total_mean': 0.49856626987457275, 'reward_meter_mean': 0.6701686382293701, 'reward_meter_std': 0.3509586751461029, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9642779231071472, 'reward_repeat_soft_std': 0.04290106147527695, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.49856626987457275, 'reward_total_composite_std': 0.2385425567626953, 'epoch': 0.13}

 38%|███▊      | 1256/3300 [3:07:17<6:13:23, 10.96s/it]INFO 04-13 10:28:15 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1257/3300 [3:07:23<5:26:43,  9.60s/it]2026-04-13 10:28:21,414 | INFO | train_grpo_train | metrics_logged mode=train step=1257

                                                       
{'loss': 0.1837, 'grad_norm': 17.26266860961914, 'learning_rate': 6.1939393939393944e-06, 'num_tokens': 2218717.0, 'completions/mean_length': 36.75, 'completions/min_length': 26.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7167240977287292, 'rewards/meter/std': 0.43885913491249084, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9467388391494751, 'rewards/repeat_soft/std': 0.0382780097424984, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4993710517883301, 'rewards/total_composite/std': 0.14655928313732147, 'reward': 0.4993710517883301, 'reward_std': 0.14655928313732147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1434856504201889, 'sampling/sampling_logp_difference/max': 1.4102238416671753, 'sampling/importance_sampling_ratio/min': 0.24408864974975586, 'sampling/importance_sampling_ratio/mean': 1.023720383644104, 'sampling/importance_sampling_ratio/max': 1.8100582361221313, 'entropy': 1.0456532463431358, 'clip_ratio/low_mean': 0.06550586502999067, 'clip_ratio/low_min': 0.06550586502999067, 'clip_ratio/high_mean': 0.07707366440445185, 'clip_ratio/high_max': 0.07707366440445185, 'clip_ratio/region_mean': 0.14257952943444252, 'reward_total_mean': 0.4993710517883301, 'reward_meter_mean': 0.7167240977287292, 'reward_meter_std': 0.43885913491249084, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9467388391494751, 'reward_repeat_soft_std': 0.0382780097424984, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4993710517883301, 'reward_total_composite_std': 0.14655928313732147, 'epoch': 0.13}

 38%|███▊      | 1257/3300 [3:07:23<5:26:43,  9.60s/it]INFO 04-13 10:28:21 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1258/3300 [3:07:29<4:52:37,  8.60s/it]2026-04-13 10:28:27,726 | INFO | train_grpo_train | metrics_logged mode=train step=1258

                                                       
{'loss': 0.0344, 'grad_norm': 27.869922637939453, 'learning_rate': 6.190909090909092e-06, 'num_tokens': 2220012.0, 'completions/mean_length': 14.875, 'completions/min_length': 14.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 14.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.7957995533943176, 'rewards/meter/std': 0.22487275302410126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5616282820701599, 'rewards/total_composite/std': 0.061390265822410583, 'reward': 0.5616282820701599, 'reward_std': 0.06139027327299118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07037394493818283, 'sampling/sampling_logp_difference/max': 1.1815903186798096, 'sampling/importance_sampling_ratio/min': 0.30679044127464294, 'sampling/importance_sampling_ratio/mean': 0.9939095973968506, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3817711900919676, 'clip_ratio/low_mean': 0.056197479367256165, 'clip_ratio/low_min': 0.056197479367256165, 'clip_ratio/high_mean': 0.03511904925107956, 'clip_ratio/high_max': 0.03511904925107956, 'clip_ratio/region_mean': 0.09131652861833572, 'reward_total_mean': 0.5616282820701599, 'reward_meter_mean': 0.7957995533943176, 'reward_meter_std': 0.22487275302410126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5616282820701599, 'reward_total_composite_std': 0.061390265822410583, 'epoch': 0.13}

 38%|███▊      | 1258/3300 [3:07:29<4:52:37,  8.60s/it]INFO 04-13 10:28:28 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1259/3300 [3:07:42<5:35:43,  9.87s/it]2026-04-13 10:28:40,562 | INFO | train_grpo_train | metrics_logged mode=train step=1259

                                                       
{'loss': 0.0345, 'grad_norm': 10.268722534179688, 'learning_rate': 6.187878787878788e-06, 'num_tokens': 2222222.0, 'completions/mean_length': 96.25, 'completions/min_length': 78.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.25, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.8196989893913269, 'rewards/meter/std': 0.2490110844373703, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9899951219558716, 'rewards/repeat_soft/std': 0.008460160344839096, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5472770929336548, 'rewards/total_composite/std': 0.05546480044722557, 'reward': 0.5472770929336548, 'reward_std': 0.05546480044722557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16437067091464996, 'sampling/sampling_logp_difference/max': 1.327692985534668, 'sampling/importance_sampling_ratio/min': 0.26508811116218567, 'sampling/importance_sampling_ratio/mean': 1.0197168588638306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1947279572486877, 'clip_ratio/low_mean': 0.05987168662250042, 'clip_ratio/low_min': 0.05987168662250042, 'clip_ratio/high_mean': 0.09439416974782944, 'clip_ratio/high_max': 0.09439416974782944, 'clip_ratio/region_mean': 0.15426585637032986, 'reward_total_mean': 0.5472770929336548, 'reward_meter_mean': 0.8196989893913269, 'reward_meter_std': 0.2490110844373703, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9899951219558716, 'reward_repeat_soft_std': 0.008460160344839096, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5472770929336548, 'reward_total_composite_std': 0.05546480044722557, 'epoch': 0.13}

 38%|███▊      | 1259/3300 [3:07:42<5:35:43,  9.87s/it]INFO 04-13 10:28:40 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1260/3300 [3:07:50<5:12:44,  9.20s/it]2026-04-13 10:28:48,195 | INFO | train_grpo_train | metrics_logged mode=train step=1260

                                                       
{'loss': 0.008, 'grad_norm': 8.31602954864502, 'learning_rate': 6.184848484848485e-06, 'num_tokens': 2224439.0, 'completions/mean_length': 113.125, 'completions/min_length': 94.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.125, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.8391289710998535, 'rewards/meter/std': 0.24904930591583252, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9839187264442444, 'rewards/repeat_soft/std': 0.014312672428786755, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6189167499542236, 'rewards/total_composite/std': 0.1338718831539154, 'reward': 0.6189167499542236, 'reward_std': 0.1338718980550766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1532590389251709, 'sampling/sampling_logp_difference/max': 1.7896296977996826, 'sampling/importance_sampling_ratio/min': 0.1670220047235489, 'sampling/importance_sampling_ratio/mean': 1.0085344314575195, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1714848205447197, 'clip_ratio/low_mean': 0.07738458458334208, 'clip_ratio/low_min': 0.07738458458334208, 'clip_ratio/high_mean': 0.052012067288160324, 'clip_ratio/high_max': 0.052012067288160324, 'clip_ratio/region_mean': 0.1293966518715024, 'reward_total_mean': 0.6189167499542236, 'reward_meter_mean': 0.8391289710998535, 'reward_meter_std': 0.24904930591583252, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9839187264442444, 'reward_repeat_soft_std': 0.014312672428786755, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6189167499542236, 'reward_total_composite_std': 0.1338718831539154, 'epoch': 0.13}

 38%|███▊      | 1260/3300 [3:07:50<5:12:44,  9.20s/it]INFO 04-13 10:28:48 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1261/3300 [3:08:02<5:41:46, 10.06s/it]2026-04-13 10:29:00,215 | INFO | train_grpo_train | metrics_logged mode=train step=1261

                                                       
{'loss': -0.1167, 'grad_norm': 2.189189910888672, 'learning_rate': 6.181818181818182e-06, 'num_tokens': 2226139.0, 'completions/mean_length': 97.5, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.28571701049805, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8494099378585815, 'rewards/meter/std': 0.3407638669013977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9858989715576172, 'rewards/repeat_soft/std': 0.008430222980678082, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.5380397439002991, 'rewards/total_composite/std': 0.21747693419456482, 'reward': 0.5380397439002991, 'reward_std': 0.21747693419456482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19998954236507416, 'sampling/sampling_logp_difference/max': 3.6249895095825195, 'sampling/importance_sampling_ratio/min': 0.02664937637746334, 'sampling/importance_sampling_ratio/mean': 1.004363775253296, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8866943642497063, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.153715162537992, 'clip_ratio/high_max': 0.153715162537992, 'clip_ratio/region_mean': 0.153715162537992, 'reward_total_mean': 0.5380397439002991, 'reward_meter_mean': 0.8494099378585815, 'reward_meter_std': 0.3407638669013977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9858989715576172, 'reward_repeat_soft_std': 0.008430222980678082, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.5380397439002991, 'reward_total_composite_std': 0.21747693419456482, 'epoch': 0.13}

 38%|███▊      | 1261/3300 [3:08:02<5:41:46, 10.06s/it]INFO 04-13 10:29:00 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1262/3300 [3:08:13<5:56:07, 10.48s/it]2026-04-13 10:29:11,697 | INFO | train_grpo_train | metrics_logged mode=train step=1262

                                                       
{'loss': 0.0258, 'grad_norm': 5.586556434631348, 'learning_rate': 6.17878787878788e-06, 'num_tokens': 2227557.0, 'completions/mean_length': 89.25, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 28.85714340209961, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.7877958416938782, 'rewards/meter/std': 0.31315621733665466, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9601781368255615, 'rewards/repeat_soft/std': 0.019703209400177002, 'rewards/judge_quality/mean': 0.7487499713897705, 'rewards/judge_quality/std': 0.332154780626297, 'rewards/total_composite/mean': 0.7074991464614868, 'rewards/total_composite/std': 0.34219247102737427, 'reward': 0.7074991464614868, 'reward_std': 0.34219247102737427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19931429624557495, 'sampling/sampling_logp_difference/max': 3.077780246734619, 'sampling/importance_sampling_ratio/min': 0.04606138914823532, 'sampling/importance_sampling_ratio/mean': 1.0298024415969849, 'sampling/importance_sampling_ratio/max': 1.999465823173523, 'entropy': 0.8831368088722229, 'clip_ratio/low_mean': 0.023333333432674408, 'clip_ratio/low_min': 0.023333333432674408, 'clip_ratio/high_mean': 0.10578703787177801, 'clip_ratio/high_max': 0.10578703787177801, 'clip_ratio/region_mean': 0.12912037130445242, 'reward_total_mean': 0.7074991464614868, 'reward_meter_mean': 0.7877958416938782, 'reward_meter_std': 0.31315621733665466, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9601781368255615, 'reward_repeat_soft_std': 0.019703209400177002, 'reward_judge_quality_mean': 0.7487499713897705, 'reward_judge_quality_std': 0.332154780626297, 'reward_total_composite_mean': 0.7074991464614868, 'reward_total_composite_std': 0.34219247102737427, 'epoch': 0.13}

 38%|███▊      | 1262/3300 [3:08:13<5:56:07, 10.48s/it]INFO 04-13 10:29:11 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1263/3300 [3:08:25<6:03:50, 10.72s/it]2026-04-13 10:29:22,955 | INFO | train_grpo_train | metrics_logged mode=train step=1263

                                                       
{'loss': -0.1757, 'grad_norm': 3.078822612762451, 'learning_rate': 6.175757575757576e-06, 'num_tokens': 2229732.0, 'completions/mean_length': 152.875, 'completions/min_length': 73.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 101.5714340209961, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.8817800283432007, 'rewards/meter/std': 0.19807694852352142, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9752034544944763, 'rewards/repeat_soft/std': 0.01988898403942585, 'rewards/judge_quality/mean': 0.47749999165534973, 'rewards/judge_quality/std': 0.22839504480361938, 'rewards/total_composite/mean': 0.5431589484214783, 'rewards/total_composite/std': 0.2479376643896103, 'reward': 0.5431589484214783, 'reward_std': 0.2479376643896103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1447705328464508, 'sampling/sampling_logp_difference/max': 1.4552173614501953, 'sampling/importance_sampling_ratio/min': 0.23334965109825134, 'sampling/importance_sampling_ratio/mean': 1.0285652875900269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.032411813735962, 'clip_ratio/low_mean': 0.015287769958376884, 'clip_ratio/low_min': 0.015287769958376884, 'clip_ratio/high_mean': 0.09570078644901514, 'clip_ratio/high_max': 0.09570078644901514, 'clip_ratio/region_mean': 0.11098855640739202, 'reward_total_mean': 0.5431589484214783, 'reward_meter_mean': 0.8817800283432007, 'reward_meter_std': 0.19807694852352142, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9752034544944763, 'reward_repeat_soft_std': 0.01988898403942585, 'reward_judge_quality_mean': 0.47749999165534973, 'reward_judge_quality_std': 0.22839504480361938, 'reward_total_composite_mean': 0.5431589484214783, 'reward_total_composite_std': 0.2479376643896103, 'epoch': 0.13}

 38%|███▊      | 1263/3300 [3:08:25<6:03:50, 10.72s/it]INFO 04-13 10:29:23 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1264/3300 [3:08:32<5:25:37,  9.60s/it]2026-04-13 10:29:29,975 | INFO | train_grpo_train | metrics_logged mode=train step=1264

                                                       
{'loss': 0.0077, 'grad_norm': 7.699588298797607, 'learning_rate': 6.1727272727272735e-06, 'num_tokens': 2231561.0, 'completions/mean_length': 69.625, 'completions/min_length': 63.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9858421087265015, 'rewards/meter/std': 0.016840020194649696, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8454391360282898, 'rewards/repeat_soft/std': 0.07404352724552155, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6362375020980835, 'rewards/total_composite/std': 0.11040402948856354, 'reward': 0.6362375020980835, 'reward_std': 0.11040402948856354, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11907019466161728, 'sampling/sampling_logp_difference/max': 1.2024345397949219, 'sampling/importance_sampling_ratio/min': 0.3004618287086487, 'sampling/importance_sampling_ratio/mean': 1.007966160774231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8413858339190483, 'clip_ratio/low_mean': 0.08165121171623468, 'clip_ratio/low_min': 0.08165121171623468, 'clip_ratio/high_mean': 0.005281690042465925, 'clip_ratio/high_max': 0.005281690042465925, 'clip_ratio/region_mean': 0.08693290175870061, 'reward_total_mean': 0.6362375020980835, 'reward_meter_mean': 0.9858421087265015, 'reward_meter_std': 0.016840020194649696, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8454391360282898, 'reward_repeat_soft_std': 0.07404352724552155, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6362375020980835, 'reward_total_composite_std': 0.11040402948856354, 'epoch': 0.13}

 38%|███▊      | 1264/3300 [3:08:32<5:25:37,  9.60s/it]INFO 04-13 10:29:30 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1265/3300 [3:08:45<6:02:30, 10.69s/it]2026-04-13 10:29:43,172 | INFO | train_grpo_train | metrics_logged mode=train step=1265

                                                       
{'loss': 0.042, 'grad_norm': 8.433825492858887, 'learning_rate': 6.16969696969697e-06, 'num_tokens': 2233847.0, 'completions/mean_length': 114.75, 'completions/min_length': 95.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.75, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.8529812097549438, 'rewards/meter/std': 0.3162272274494171, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877867698669434, 'rewards/repeat_soft/std': 0.011356890201568604, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5834782123565674, 'rewards/total_composite/std': 0.10660649836063385, 'reward': 0.5834782123565674, 'reward_std': 0.10660647600889206, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1506912261247635, 'sampling/sampling_logp_difference/max': 2.9099059104919434, 'sampling/importance_sampling_ratio/min': 0.05448085442185402, 'sampling/importance_sampling_ratio/mean': 1.014310359954834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1703742891550064, 'clip_ratio/low_mean': 0.05421730037778616, 'clip_ratio/low_min': 0.05421730037778616, 'clip_ratio/high_mean': 0.08516057766973972, 'clip_ratio/high_max': 0.08516057766973972, 'clip_ratio/region_mean': 0.13937787804752588, 'reward_total_mean': 0.5834782123565674, 'reward_meter_mean': 0.8529812097549438, 'reward_meter_std': 0.3162272274494171, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877867698669434, 'reward_repeat_soft_std': 0.011356890201568604, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5834782123565674, 'reward_total_composite_std': 0.10660649836063385, 'epoch': 0.13}

 38%|███▊      | 1265/3300 [3:08:45<6:02:30, 10.69s/it]INFO 04-13 10:29:43 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1266/3300 [3:08:56<6:06:52, 10.82s/it]2026-04-13 10:29:54,307 | INFO | train_grpo_train | metrics_logged mode=train step=1266

                                                       
{'loss': -0.1018, 'grad_norm': 3.282893419265747, 'learning_rate': 6.166666666666667e-06, 'num_tokens': 2235270.0, 'completions/mean_length': 97.875, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 38.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.5459794402122498, 'rewards/meter/std': 0.4117375612258911, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9514325857162476, 'rewards/repeat_soft/std': 0.07520611584186554, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4297533631324768, 'rewards/total_composite/std': 0.2113325297832489, 'reward': 0.4297533631324768, 'reward_std': 0.2113325297832489, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13173063099384308, 'sampling/sampling_logp_difference/max': 1.712148904800415, 'sampling/importance_sampling_ratio/min': 0.1804775446653366, 'sampling/importance_sampling_ratio/mean': 1.0067899227142334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.718297965824604, 'clip_ratio/low_mean': 0.03996442258358002, 'clip_ratio/low_min': 0.03996442258358002, 'clip_ratio/high_mean': 0.07270388770848513, 'clip_ratio/high_max': 0.07270388770848513, 'clip_ratio/region_mean': 0.11266831029206514, 'reward_total_mean': 0.4297533631324768, 'reward_meter_mean': 0.5459794402122498, 'reward_meter_std': 0.4117375612258911, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9514325857162476, 'reward_repeat_soft_std': 0.07520611584186554, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4297533631324768, 'reward_total_composite_std': 0.2113325297832489, 'epoch': 0.13}

 38%|███▊      | 1266/3300 [3:08:56<6:06:52, 10.82s/it]INFO 04-13 10:29:54 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1267/3300 [3:09:02<5:21:18,  9.48s/it]2026-04-13 10:30:00,706 | INFO | train_grpo_train | metrics_logged mode=train step=1267

                                                       
{'loss': 0.0667, 'grad_norm': 11.786508560180664, 'learning_rate': 6.163636363636364e-06, 'num_tokens': 2236947.0, 'completions/mean_length': 41.625, 'completions/min_length': 20.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8319039344787598, 'rewards/meter/std': 0.31794923543930054, 'rewards/count_adherence/mean': 0.125, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9469939470291138, 'rewards/repeat_soft/std': 0.03675156831741333, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4016599953174591, 'rewards/total_composite/std': 0.1257942020893097, 'reward': 0.4016599953174591, 'reward_std': 0.1257941871881485, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14323072135448456, 'sampling/sampling_logp_difference/max': 1.8149604797363281, 'sampling/importance_sampling_ratio/min': 0.16284434497356415, 'sampling/importance_sampling_ratio/mean': 1.0372194051742554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1073968783020973, 'clip_ratio/low_mean': 0.03768939524888992, 'clip_ratio/low_min': 0.03768939524888992, 'clip_ratio/high_mean': 0.08446543430909514, 'clip_ratio/high_max': 0.08446543430909514, 'clip_ratio/region_mean': 0.12215482955798507, 'reward_total_mean': 0.4016599953174591, 'reward_meter_mean': 0.8319039344787598, 'reward_meter_std': 0.31794923543930054, 'reward_count_adherence_mean': 0.125, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9469939470291138, 'reward_repeat_soft_std': 0.03675156831741333, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4016599953174591, 'reward_total_composite_std': 0.1257942020893097, 'epoch': 0.13}

 38%|███▊      | 1267/3300 [3:09:02<5:21:18,  9.48s/it]INFO 04-13 10:30:01 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1268/3300 [3:09:14<5:45:51, 10.21s/it]2026-04-13 10:30:12,586 | INFO | train_grpo_train | metrics_logged mode=train step=1268

                                                       
{'loss': -0.0833, 'grad_norm': 3.330807685852051, 'learning_rate': 6.160606060606062e-06, 'num_tokens': 2238493.0, 'completions/mean_length': 95.25, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.71428680419922, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.16060248017311096, 'rewards/meter/std': 0.28669473528862, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9736111164093018, 'rewards/repeat_soft/std': 0.04457922652363777, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.27994900941848755, 'rewards/total_composite/mean': 0.36672818660736084, 'rewards/total_composite/std': 0.11939622461795807, 'reward': 0.36672818660736084, 'reward_std': 0.11939621716737747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15467126667499542, 'sampling/sampling_logp_difference/max': 1.3939218521118164, 'sampling/importance_sampling_ratio/min': 0.2481003850698471, 'sampling/importance_sampling_ratio/mean': 0.9885116815567017, 'sampling/importance_sampling_ratio/max': 1.71488618850708, 'entropy': 0.6161579042673111, 'clip_ratio/low_mean': 0.07780447881668806, 'clip_ratio/low_min': 0.07780447881668806, 'clip_ratio/high_mean': 0.06103704310953617, 'clip_ratio/high_max': 0.06103704310953617, 'clip_ratio/region_mean': 0.13884152192622423, 'reward_total_mean': 0.36672818660736084, 'reward_meter_mean': 0.16060248017311096, 'reward_meter_std': 0.28669473528862, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9736111164093018, 'reward_repeat_soft_std': 0.04457922652363777, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.27994900941848755, 'reward_total_composite_mean': 0.36672818660736084, 'reward_total_composite_std': 0.11939622461795807, 'epoch': 0.13}

 38%|███▊      | 1268/3300 [3:09:14<5:45:51, 10.21s/it]INFO 04-13 10:30:12 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1269/3300 [3:09:20<5:01:00,  8.89s/it]2026-04-13 10:30:18,390 | INFO | train_grpo_train | metrics_logged mode=train step=1269

                                                       
{'loss': -0.0118, 'grad_norm': 13.987906455993652, 'learning_rate': 6.157575757575758e-06, 'num_tokens': 2239957.0, 'completions/mean_length': 37.0, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.879828155040741, 'rewards/meter/std': 0.22661611437797546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.999053955078125, 'rewards/repeat_soft/std': 0.0014005512930452824, 'rewards/judge_quality/mean': 0.9237500429153442, 'rewards/judge_quality/std': 0.01060659158974886, 'rewards/total_composite/mean': 0.8783953189849854, 'rewards/total_composite/std': 0.13688848912715912, 'reward': 0.8783953189849854, 'reward_std': 0.13688848912715912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14671094715595245, 'sampling/sampling_logp_difference/max': 3.0388565063476562, 'sampling/importance_sampling_ratio/min': 0.04788962006568909, 'sampling/importance_sampling_ratio/mean': 0.988682746887207, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5730679966509342, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.12588980235159397, 'clip_ratio/high_max': 0.12588980235159397, 'clip_ratio/region_mean': 0.13346556015312672, 'reward_total_mean': 0.8783953189849854, 'reward_meter_mean': 0.879828155040741, 'reward_meter_std': 0.22661611437797546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.999053955078125, 'reward_repeat_soft_std': 0.0014005512930452824, 'reward_judge_quality_mean': 0.9237500429153442, 'reward_judge_quality_std': 0.01060659158974886, 'reward_total_composite_mean': 0.8783953189849854, 'reward_total_composite_std': 0.13688848912715912, 'epoch': 0.13}

 38%|███▊      | 1269/3300 [3:09:20<5:01:00,  8.89s/it]INFO 04-13 10:30:18 [block_pool.py:378] Successfully reset prefix cache

 38%|███▊      | 1270/3300 [3:09:27<4:39:26,  8.26s/it]2026-04-13 10:30:25,172 | INFO | train_grpo_train | metrics_logged mode=train step=1270

                                                       
{'loss': 0.0392, 'grad_norm': 11.309529304504395, 'learning_rate': 6.154545454545455e-06, 'num_tokens': 2241818.0, 'completions/mean_length': 52.625, 'completions/min_length': 43.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9154290556907654, 'rewards/meter/std': 0.06503310799598694, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9116422533988953, 'rewards/repeat_soft/std': 0.07046398520469666, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6622281074523926, 'rewards/total_composite/std': 0.13527481257915497, 'reward': 0.6622281074523926, 'reward_std': 0.13527479767799377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11792018264532089, 'sampling/sampling_logp_difference/max': 1.9887970685958862, 'sampling/importance_sampling_ratio/min': 0.1368599534034729, 'sampling/importance_sampling_ratio/mean': 1.0024830102920532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6192738823592663, 'clip_ratio/low_mean': 0.06327737215906382, 'clip_ratio/low_min': 0.06327737215906382, 'clip_ratio/high_mean': 0.025917832739651203, 'clip_ratio/high_max': 0.025917832739651203, 'clip_ratio/region_mean': 0.08919520489871502, 'reward_total_mean': 0.6622281074523926, 'reward_meter_mean': 0.9154290556907654, 'reward_meter_std': 0.06503310799598694, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9116422533988953, 'reward_repeat_soft_std': 0.07046398520469666, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6622281074523926, 'reward_total_composite_std': 0.13527481257915497, 'epoch': 0.13}

 38%|███▊      | 1270/3300 [3:09:27<4:39:26,  8.26s/it]INFO 04-13 10:30:25 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1271/3300 [3:09:33<4:23:01,  7.78s/it]2026-04-13 10:30:31,832 | INFO | train_grpo_train | metrics_logged mode=train step=1271

                                                       
{'loss': -0.0079, 'grad_norm': 10.42291259765625, 'learning_rate': 6.151515151515152e-06, 'num_tokens': 2244176.0, 'completions/mean_length': 92.75, 'completions/min_length': 84.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.9904111623764038, 'rewards/meter/std': 0.008908153511583805, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9086899757385254, 'rewards/repeat_soft/std': 0.045052893459796906, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.4309046268463135, 'rewards/total_composite/std': 0.18315014243125916, 'reward': 0.4309046268463135, 'reward_std': 0.18315015733242035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14469754695892334, 'sampling/sampling_logp_difference/max': 3.127166748046875, 'sampling/importance_sampling_ratio/min': 0.04384183511137962, 'sampling/importance_sampling_ratio/mean': 1.0195159912109375, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0048793256282806, 'clip_ratio/low_mean': 0.03726926166564226, 'clip_ratio/low_min': 0.03726926166564226, 'clip_ratio/high_mean': 0.07609590562060475, 'clip_ratio/high_max': 0.07609590562060475, 'clip_ratio/region_mean': 0.11336516728624701, 'reward_total_mean': 0.4309046268463135, 'reward_meter_mean': 0.9904111623764038, 'reward_meter_std': 0.008908153511583805, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9086899757385254, 'reward_repeat_soft_std': 0.045052893459796906, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.4309046268463135, 'reward_total_composite_std': 0.18315014243125916, 'epoch': 0.13}

 39%|███▊      | 1271/3300 [3:09:34<4:23:01,  7.78s/it]INFO 04-13 10:30:32 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1272/3300 [3:09:40<4:12:32,  7.47s/it]2026-04-13 10:30:38,590 | INFO | train_grpo_train | metrics_logged mode=train step=1272

                                                       
{'loss': 0.0736, 'grad_norm': 9.260095596313477, 'learning_rate': 6.148484848484849e-06, 'num_tokens': 2246422.0, 'completions/mean_length': 97.75, 'completions/min_length': 78.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 97.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.978448748588562, 'rewards/meter/std': 0.01660177670419216, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9162514209747314, 'rewards/repeat_soft/std': 0.032950934022665024, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.6151977777481079, 'rewards/total_composite/std': 0.08692056685686111, 'reward': 0.6151977777481079, 'reward_std': 0.08692057430744171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14708460867404938, 'sampling/sampling_logp_difference/max': 2.8121001720428467, 'sampling/importance_sampling_ratio/min': 0.060078684240579605, 'sampling/importance_sampling_ratio/mean': 1.0057787895202637, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8702584803104401, 'clip_ratio/low_mean': 0.10920719988644123, 'clip_ratio/low_min': 0.10920719988644123, 'clip_ratio/high_mean': 0.012500000186264515, 'clip_ratio/high_max': 0.012500000186264515, 'clip_ratio/region_mean': 0.12170720007270575, 'reward_total_mean': 0.6151977777481079, 'reward_meter_mean': 0.978448748588562, 'reward_meter_std': 0.01660177670419216, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9162514209747314, 'reward_repeat_soft_std': 0.032950934022665024, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.6151977777481079, 'reward_total_composite_std': 0.08692056685686111, 'epoch': 0.13}

 39%|███▊      | 1272/3300 [3:09:40<4:12:32,  7.47s/it]INFO 04-13 10:30:38 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1273/3300 [3:09:52<4:55:52,  8.76s/it]2026-04-13 10:30:50,386 | INFO | train_grpo_train | metrics_logged mode=train step=1273

                                                       
{'loss': -0.1782, 'grad_norm': 2.169473648071289, 'learning_rate': 6.1454545454545454e-06, 'num_tokens': 2248367.0, 'completions/mean_length': 199.125, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 94.83333587646484, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8148660063743591, 'rewards/meter/std': 0.23851215839385986, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9603255987167358, 'rewards/repeat_soft/std': 0.05261612311005592, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.23802761733531952, 'rewards/total_composite/mean': 0.4444544315338135, 'rewards/total_composite/std': 0.283040851354599, 'reward': 0.4444544315338135, 'reward_std': 0.283040851354599, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1279352456331253, 'sampling/sampling_logp_difference/max': 1.4324660301208496, 'sampling/importance_sampling_ratio/min': 0.23871950805187225, 'sampling/importance_sampling_ratio/mean': 1.016228437423706, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6070486083626747, 'clip_ratio/low_mean': 0.02227722853422165, 'clip_ratio/low_min': 0.02227722853422165, 'clip_ratio/high_mean': 0.06470211688429117, 'clip_ratio/high_max': 0.06470211688429117, 'clip_ratio/region_mean': 0.08697934541851282, 'reward_total_mean': 0.4444544315338135, 'reward_meter_mean': 0.8148660063743591, 'reward_meter_std': 0.23851215839385986, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9603255987167358, 'reward_repeat_soft_std': 0.05261612311005592, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.23802761733531952, 'reward_total_composite_mean': 0.4444544315338135, 'reward_total_composite_std': 0.283040851354599, 'epoch': 0.13}

 39%|███▊      | 1273/3300 [3:09:52<4:55:52,  8.76s/it]INFO 04-13 10:30:50 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1274/3300 [3:10:01<4:53:43,  8.70s/it]2026-04-13 10:30:58,952 | INFO | train_grpo_train | metrics_logged mode=train step=1274

                                                       
{'loss': 0.0791, 'grad_norm': 21.68623161315918, 'learning_rate': 6.142424242424243e-06, 'num_tokens': 2249843.0, 'completions/mean_length': 24.5, 'completions/min_length': 18.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7573328018188477, 'rewards/meter/std': 0.32801494002342224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9540283679962158, 'rewards/repeat_soft/std': 0.01626886986196041, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7008538246154785, 'rewards/total_composite/std': 0.22855205833911896, 'reward': 0.7008538246154785, 'reward_std': 0.22855207324028015, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16179020702838898, 'sampling/sampling_logp_difference/max': 1.4722857475280762, 'sampling/importance_sampling_ratio/min': 0.22940054535865784, 'sampling/importance_sampling_ratio/mean': 1.0096404552459717, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9626705348491669, 'clip_ratio/low_mean': 0.05670033674687147, 'clip_ratio/low_min': 0.05670033674687147, 'clip_ratio/high_mean': 0.09861111082136631, 'clip_ratio/high_max': 0.09861111082136631, 'clip_ratio/region_mean': 0.15531144756823778, 'reward_total_mean': 0.7008538246154785, 'reward_meter_mean': 0.7573328018188477, 'reward_meter_std': 0.32801494002342224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9540283679962158, 'reward_repeat_soft_std': 0.01626886986196041, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7008538246154785, 'reward_total_composite_std': 0.22855205833911896, 'epoch': 0.13}

 39%|███▊      | 1274/3300 [3:10:01<4:53:43,  8.70s/it]INFO 04-13 10:30:59 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1275/3300 [3:10:12<5:23:56,  9.60s/it]2026-04-13 10:31:10,650 | INFO | train_grpo_train | metrics_logged mode=train step=1275

                                                       
{'loss': -0.1136, 'grad_norm': 3.550029754638672, 'learning_rate': 6.139393939393939e-06, 'num_tokens': 2251640.0, 'completions/mean_length': 108.625, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.000003814697266, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8806706070899963, 'rewards/meter/std': 0.31563690304756165, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8966113924980164, 'rewards/repeat_soft/std': 0.06710071116685867, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.3508535623550415, 'rewards/total_composite/mean': 0.6288752555847168, 'rewards/total_composite/std': 0.32005971670150757, 'reward': 0.6288752555847168, 'reward_std': 0.32005971670150757, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1061878502368927, 'sampling/sampling_logp_difference/max': 1.4472990036010742, 'sampling/importance_sampling_ratio/min': 0.23520472645759583, 'sampling/importance_sampling_ratio/mean': 1.0176953077316284, 'sampling/importance_sampling_ratio/max': 1.9982917308807373, 'entropy': 0.7043255418539047, 'clip_ratio/low_mean': 0.04471670696511865, 'clip_ratio/low_min': 0.04471670696511865, 'clip_ratio/high_mean': 0.047743055038154125, 'clip_ratio/high_max': 0.047743055038154125, 'clip_ratio/region_mean': 0.09245976200327277, 'reward_total_mean': 0.6288752555847168, 'reward_meter_mean': 0.8806706070899963, 'reward_meter_std': 0.31563690304756165, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8966113924980164, 'reward_repeat_soft_std': 0.06710071116685867, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.3508535623550415, 'reward_total_composite_mean': 0.6288752555847168, 'reward_total_composite_std': 0.32005971670150757, 'epoch': 0.13}

 39%|███▊      | 1275/3300 [3:10:12<5:23:56,  9.60s/it]INFO 04-13 10:31:11 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1276/3300 [3:10:20<5:03:05,  8.98s/it]2026-04-13 10:31:18,202 | INFO | train_grpo_train | metrics_logged mode=train step=1276

                                                       
{'loss': -0.037, 'grad_norm': 8.634065628051758, 'learning_rate': 6.136363636363637e-06, 'num_tokens': 2254032.0, 'completions/mean_length': 91.0, 'completions/min_length': 77.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.0, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9490756392478943, 'rewards/meter/std': 0.09576056152582169, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9341107606887817, 'rewards/repeat_soft/std': 0.04371308907866478, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5992085933685303, 'rewards/total_composite/std': 0.1208617240190506, 'reward': 0.5992085933685303, 'reward_std': 0.1208617240190506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15283799171447754, 'sampling/sampling_logp_difference/max': 2.512065887451172, 'sampling/importance_sampling_ratio/min': 0.08110052347183228, 'sampling/importance_sampling_ratio/mean': 1.011644959449768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9222359731793404, 'clip_ratio/low_mean': 0.10171039402484894, 'clip_ratio/low_min': 0.10171039402484894, 'clip_ratio/high_mean': 0.01752336509525776, 'clip_ratio/high_max': 0.01752336509525776, 'clip_ratio/region_mean': 0.1192337591201067, 'reward_total_mean': 0.5992085933685303, 'reward_meter_mean': 0.9490756392478943, 'reward_meter_std': 0.09576056152582169, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9341107606887817, 'reward_repeat_soft_std': 0.04371308907866478, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5992085933685303, 'reward_total_composite_std': 0.1208617240190506, 'epoch': 0.13}

 39%|███▊      | 1276/3300 [3:10:20<5:03:05,  8.98s/it]INFO 04-13 10:31:18 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1277/3300 [3:10:26<4:35:21,  8.17s/it]2026-04-13 10:31:24,415 | INFO | train_grpo_train | metrics_logged mode=train step=1277

                                                       
{'loss': 0.0143, 'grad_norm': 18.64078712463379, 'learning_rate': 6.133333333333334e-06, 'num_tokens': 2255456.0, 'completions/mean_length': 29.0, 'completions/min_length': 24.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.13576379418373108, 'rewards/meter/std': 0.1197720319032669, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9961292743682861, 'rewards/repeat_soft/std': 0.0050612385384738445, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.16974246501922607, 'rewards/total_composite/mean': 0.40248921513557434, 'rewards/total_composite/std': 0.07096210867166519, 'reward': 0.40248921513557434, 'reward_std': 0.07096210867166519, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14276927709579468, 'sampling/sampling_logp_difference/max': 2.2207584381103516, 'sampling/importance_sampling_ratio/min': 0.10852676630020142, 'sampling/importance_sampling_ratio/mean': 0.9993550777435303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5751420333981514, 'clip_ratio/low_mean': 0.07176368683576584, 'clip_ratio/low_min': 0.07176368683576584, 'clip_ratio/high_mean': 0.034050178714096546, 'clip_ratio/high_max': 0.034050178714096546, 'clip_ratio/region_mean': 0.10581386554986238, 'reward_total_mean': 0.40248921513557434, 'reward_meter_mean': 0.13576379418373108, 'reward_meter_std': 0.1197720319032669, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9961292743682861, 'reward_repeat_soft_std': 0.0050612385384738445, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.16974246501922607, 'reward_total_composite_mean': 0.40248921513557434, 'reward_total_composite_std': 0.07096210867166519, 'epoch': 0.13}

 39%|███▊      | 1277/3300 [3:10:26<4:35:21,  8.17s/it]INFO 04-13 10:31:24 [block_pool.py:378] Successfully reset prefix cache

 39%|███▊      | 1278/3300 [3:10:32<4:15:20,  7.58s/it]2026-04-13 10:31:30,613 | INFO | train_grpo_train | metrics_logged mode=train step=1278

                                                       
{'loss': 0.0742, 'grad_norm': 16.209163665771484, 'learning_rate': 6.130303030303031e-06, 'num_tokens': 2257071.0, 'completions/mean_length': 46.875, 'completions/min_length': 37.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.9211243987083435, 'rewards/meter/std': 0.12842069566249847, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9725692272186279, 'rewards/repeat_soft/std': 0.03515557572245598, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.24348658323287964, 'rewards/total_composite/mean': 0.7098895907402039, 'rewards/total_composite/std': 0.1747393012046814, 'reward': 0.7098895907402039, 'reward_std': 0.1747393161058426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14765222370624542, 'sampling/sampling_logp_difference/max': 2.2170071601867676, 'sampling/importance_sampling_ratio/min': 0.10893464833498001, 'sampling/importance_sampling_ratio/mean': 1.0145690441131592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7762776911258698, 'clip_ratio/low_mean': 0.06682059494778514, 'clip_ratio/low_min': 0.06682059494778514, 'clip_ratio/high_mean': 0.055385487619787455, 'clip_ratio/high_max': 0.055385487619787455, 'clip_ratio/region_mean': 0.1222060825675726, 'reward_total_mean': 0.7098895907402039, 'reward_meter_mean': 0.9211243987083435, 'reward_meter_std': 0.12842069566249847, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9725692272186279, 'reward_repeat_soft_std': 0.03515557572245598, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.24348658323287964, 'reward_total_composite_mean': 0.7098895907402039, 'reward_total_composite_std': 0.1747393012046814, 'epoch': 0.13}

 39%|███▊      | 1278/3300 [3:10:32<4:15:20,  7.58s/it]INFO 04-13 10:31:30 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1279/3300 [3:10:39<4:01:57,  7.18s/it]2026-04-13 10:31:36,881 | INFO | train_grpo_train | metrics_logged mode=train step=1279

                                                       
{'loss': 0.0277, 'grad_norm': 14.2674560546875, 'learning_rate': 6.127272727272727e-06, 'num_tokens': 2258797.0, 'completions/mean_length': 43.75, 'completions/min_length': 40.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.75, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5102277398109436, 'rewards/meter/std': 0.2549363672733307, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9901503324508667, 'rewards/repeat_soft/std': 0.01757005974650383, 'rewards/judge_quality/mean': 0.5099999904632568, 'rewards/judge_quality/std': 0.13979578018188477, 'rewards/total_composite/mean': 0.5142813920974731, 'rewards/total_composite/std': 0.07932232320308685, 'reward': 0.5142813920974731, 'reward_std': 0.07932231575250626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11101723462343216, 'sampling/sampling_logp_difference/max': 1.4886724948883057, 'sampling/importance_sampling_ratio/min': 0.22567205131053925, 'sampling/importance_sampling_ratio/mean': 0.9959089159965515, 'sampling/importance_sampling_ratio/max': 1.619319200515747, 'entropy': 0.5416922643780708, 'clip_ratio/low_mean': 0.03985507180914283, 'clip_ratio/low_min': 0.03985507180914283, 'clip_ratio/high_mean': 0.0694272646214813, 'clip_ratio/high_max': 0.0694272646214813, 'clip_ratio/region_mean': 0.10928233643062413, 'reward_total_mean': 0.5142813920974731, 'reward_meter_mean': 0.5102277398109436, 'reward_meter_std': 0.2549363672733307, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9901503324508667, 'reward_repeat_soft_std': 0.01757005974650383, 'reward_judge_quality_mean': 0.5099999904632568, 'reward_judge_quality_std': 0.13979578018188477, 'reward_total_composite_mean': 0.5142813920974731, 'reward_total_composite_std': 0.07932232320308685, 'epoch': 0.13}

 39%|███▉      | 1279/3300 [3:10:39<4:01:57,  7.18s/it]INFO 04-13 10:31:37 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1280/3300 [3:10:45<3:49:57,  6.83s/it]2026-04-13 10:31:42,886 | INFO | train_grpo_train | metrics_logged mode=train step=1280

                                                       
{'loss': -0.1251, 'grad_norm': 17.665630340576172, 'learning_rate': 6.1242424242424245e-06, 'num_tokens': 2260445.0, 'completions/mean_length': 30.0, 'completions/min_length': 22.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8284243941307068, 'rewards/meter/std': 0.318401575088501, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9651641845703125, 'rewards/repeat_soft/std': 0.036734238266944885, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.22696760296821594, 'rewards/total_composite/mean': 0.6755133867263794, 'rewards/total_composite/std': 0.1895093470811844, 'reward': 0.6755133867263794, 'reward_std': 0.1895093470811844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1389528065919876, 'sampling/sampling_logp_difference/max': 1.9500927925109863, 'sampling/importance_sampling_ratio/min': 0.1422608643770218, 'sampling/importance_sampling_ratio/mean': 1.0018794536590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7707121223211288, 'clip_ratio/low_mean': 0.09540211223065853, 'clip_ratio/low_min': 0.09540211223065853, 'clip_ratio/high_mean': 0.052087913267314434, 'clip_ratio/high_max': 0.052087913267314434, 'clip_ratio/region_mean': 0.14749002549797297, 'reward_total_mean': 0.6755133867263794, 'reward_meter_mean': 0.8284243941307068, 'reward_meter_std': 0.318401575088501, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9651641845703125, 'reward_repeat_soft_std': 0.036734238266944885, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.22696760296821594, 'reward_total_composite_mean': 0.6755133867263794, 'reward_total_composite_std': 0.1895093470811844, 'epoch': 0.13}

 39%|███▉      | 1280/3300 [3:10:45<3:49:57,  6.83s/it]INFO 04-13 10:31:43 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1281/3300 [3:10:56<4:32:30,  8.10s/it]2026-04-13 10:31:53,943 | INFO | train_grpo_train | metrics_logged mode=train step=1281

                                                       
{'loss': -0.1069, 'grad_norm': 3.375562906265259, 'learning_rate': 6.121212121212121e-06, 'num_tokens': 2262374.0, 'completions/mean_length': 148.125, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 96.14286041259766, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.5884270071983337, 'rewards/meter/std': 0.38525184988975525, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9659799337387085, 'rewards/repeat_soft/std': 0.054748762398958206, 'rewards/judge_quality/mean': 0.36000001430511475, 'rewards/judge_quality/std': 0.140813946723938, 'rewards/total_composite/mean': 0.41940566897392273, 'rewards/total_composite/std': 0.1943313330411911, 'reward': 0.41940566897392273, 'reward_std': 0.1943313330411911, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15925894677639008, 'sampling/sampling_logp_difference/max': 1.474484920501709, 'sampling/importance_sampling_ratio/min': 0.2288966029882431, 'sampling/importance_sampling_ratio/mean': 1.011333703994751, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8784201517701149, 'clip_ratio/low_mean': 0.04367189295589924, 'clip_ratio/low_min': 0.04367189295589924, 'clip_ratio/high_mean': 0.08718924224376678, 'clip_ratio/high_max': 0.08718924224376678, 'clip_ratio/region_mean': 0.13086113519966602, 'reward_total_mean': 0.41940566897392273, 'reward_meter_mean': 0.5884270071983337, 'reward_meter_std': 0.38525184988975525, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9659799337387085, 'reward_repeat_soft_std': 0.054748762398958206, 'reward_judge_quality_mean': 0.36000001430511475, 'reward_judge_quality_std': 0.140813946723938, 'reward_total_composite_mean': 0.41940566897392273, 'reward_total_composite_std': 0.1943313330411911, 'epoch': 0.13}

 39%|███▉      | 1281/3300 [3:10:56<4:32:30,  8.10s/it]INFO 04-13 10:31:54 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1282/3300 [3:11:03<4:20:44,  7.75s/it]2026-04-13 10:32:00,895 | INFO | train_grpo_train | metrics_logged mode=train step=1282

                                                       
{'loss': 0.3394, 'grad_norm': 14.450453758239746, 'learning_rate': 6.118181818181819e-06, 'num_tokens': 2263872.0, 'completions/mean_length': 35.25, 'completions/min_length': 14.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9769443273544312, 'rewards/meter/std': 0.023884175345301628, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624143242835999, 'rewards/repeat_soft/std': 0.013508067466318607, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.6212908029556274, 'rewards/total_composite/std': 0.21466147899627686, 'reward': 0.6212908029556274, 'reward_std': 0.21466147899627686, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17344297468662262, 'sampling/sampling_logp_difference/max': 2.619584321975708, 'sampling/importance_sampling_ratio/min': 0.23441235721111298, 'sampling/importance_sampling_ratio/mean': 0.9908496737480164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8512249290943146, 'clip_ratio/low_mean': 0.08545100968331099, 'clip_ratio/low_min': 0.08545100968331099, 'clip_ratio/high_mean': 0.05839285720139742, 'clip_ratio/high_max': 0.05839285720139742, 'clip_ratio/region_mean': 0.1438438668847084, 'reward_total_mean': 0.6212908029556274, 'reward_meter_mean': 0.9769443273544312, 'reward_meter_std': 0.023884175345301628, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624143242835999, 'reward_repeat_soft_std': 0.013508067466318607, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.6212908029556274, 'reward_total_composite_std': 0.21466147899627686, 'epoch': 0.13}

 39%|███▉      | 1282/3300 [3:11:03<4:20:44,  7.75s/it]INFO 04-13 10:32:01 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1283/3300 [3:11:09<4:04:16,  7.27s/it]2026-04-13 10:32:07,048 | INFO | train_grpo_train | metrics_logged mode=train step=1283

                                                       
{'loss': -0.0548, 'grad_norm': 12.311016082763672, 'learning_rate': 6.115151515151516e-06, 'num_tokens': 2265598.0, 'completions/mean_length': 47.75, 'completions/min_length': 33.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.907586932182312, 'rewards/meter/std': 0.226583793759346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9689907431602478, 'rewards/repeat_soft/std': 0.03646272420883179, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6331464052200317, 'rewards/total_composite/std': 0.13213784992694855, 'reward': 0.6331464052200317, 'reward_std': 0.13213784992694855, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18488241732120514, 'sampling/sampling_logp_difference/max': 1.675678014755249, 'sampling/importance_sampling_ratio/min': 0.18718121945858002, 'sampling/importance_sampling_ratio/mean': 1.0197899341583252, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.362459436058998, 'clip_ratio/low_mean': 0.162873312830925, 'clip_ratio/low_min': 0.162873312830925, 'clip_ratio/high_mean': 0.01844262331724167, 'clip_ratio/high_max': 0.01844262331724167, 'clip_ratio/region_mean': 0.18131593614816666, 'reward_total_mean': 0.6331464052200317, 'reward_meter_mean': 0.907586932182312, 'reward_meter_std': 0.226583793759346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9689907431602478, 'reward_repeat_soft_std': 0.03646272420883179, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6331464052200317, 'reward_total_composite_std': 0.13213784992694855, 'epoch': 0.13}

 39%|███▉      | 1283/3300 [3:11:09<4:04:16,  7.27s/it]INFO 04-13 10:32:07 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1284/3300 [3:11:15<3:58:41,  7.10s/it]2026-04-13 10:32:13,749 | INFO | train_grpo_train | metrics_logged mode=train step=1284

                                                       
{'loss': 0.0621, 'grad_norm': 10.086344718933105, 'learning_rate': 6.112121212121213e-06, 'num_tokens': 2267807.0, 'completions/mean_length': 90.125, 'completions/min_length': 80.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.125, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.883400022983551, 'rewards/meter/std': 0.2530997395515442, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9838672280311584, 'rewards/repeat_soft/std': 0.01193454209715128, 'rewards/judge_quality/mean': 0.6074999570846558, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.7078959941864014, 'rewards/total_composite/std': 0.1462317556142807, 'reward': 0.7078959941864014, 'reward_std': 0.1462317407131195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16071321070194244, 'sampling/sampling_logp_difference/max': 2.1969804763793945, 'sampling/importance_sampling_ratio/min': 0.1111382395029068, 'sampling/importance_sampling_ratio/mean': 0.9974961280822754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9173566922545433, 'clip_ratio/low_mean': 0.05083981156349182, 'clip_ratio/low_min': 0.05083981156349182, 'clip_ratio/high_mean': 0.08948590233922005, 'clip_ratio/high_max': 0.08948590233922005, 'clip_ratio/region_mean': 0.14032571390271187, 'reward_total_mean': 0.7078959941864014, 'reward_meter_mean': 0.883400022983551, 'reward_meter_std': 0.2530997395515442, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9838672280311584, 'reward_repeat_soft_std': 0.01193454209715128, 'reward_judge_quality_mean': 0.6074999570846558, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.7078959941864014, 'reward_total_composite_std': 0.1462317556142807, 'epoch': 0.13}

 39%|███▉      | 1284/3300 [3:11:15<3:58:41,  7.10s/it]INFO 04-13 10:32:14 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1285/3300 [3:11:22<3:53:32,  6.95s/it]2026-04-13 10:32:20,352 | INFO | train_grpo_train | metrics_logged mode=train step=1285

                                                       
{'loss': 0.0938, 'grad_norm': 9.451906204223633, 'learning_rate': 6.10909090909091e-06, 'num_tokens': 2270093.0, 'completions/mean_length': 86.75, 'completions/min_length': 78.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.75, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.6015616059303284, 'rewards/meter/std': 0.412428617477417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9905730485916138, 'rewards/repeat_soft/std': 0.007027474697679281, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5160260200500488, 'rewards/total_composite/std': 0.1082032173871994, 'reward': 0.5160260200500488, 'reward_std': 0.1082032099366188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1459343284368515, 'sampling/sampling_logp_difference/max': 2.7698051929473877, 'sampling/importance_sampling_ratio/min': 0.06267420947551727, 'sampling/importance_sampling_ratio/mean': 1.020920991897583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8987893760204315, 'clip_ratio/low_mean': 0.056004864163696766, 'clip_ratio/low_min': 0.056004864163696766, 'clip_ratio/high_mean': 0.06314123515039682, 'clip_ratio/high_max': 0.06314123515039682, 'clip_ratio/region_mean': 0.11914609931409359, 'reward_total_mean': 0.5160260200500488, 'reward_meter_mean': 0.6015616059303284, 'reward_meter_std': 0.412428617477417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9905730485916138, 'reward_repeat_soft_std': 0.007027474697679281, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5160260200500488, 'reward_total_composite_std': 0.1082032173871994, 'epoch': 0.13}

 39%|███▉      | 1285/3300 [3:11:22<3:53:32,  6.95s/it]INFO 04-13 10:32:20 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1286/3300 [3:11:28<3:46:13,  6.74s/it]2026-04-13 10:32:26,589 | INFO | train_grpo_train | metrics_logged mode=train step=1286

                                                       
{'loss': 0.0092, 'grad_norm': 16.7330379486084, 'learning_rate': 6.106060606060606e-06, 'num_tokens': 2271688.0, 'completions/mean_length': 41.375, 'completions/min_length': 36.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5061711072921753, 'rewards/meter/std': 0.3633720874786377, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.991203784942627, 'rewards/repeat_soft/std': 0.008159817196428776, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5223467350006104, 'rewards/total_composite/std': 0.09643376618623734, 'reward': 0.5223467350006104, 'reward_std': 0.09643375873565674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14730842411518097, 'sampling/sampling_logp_difference/max': 2.860617160797119, 'sampling/importance_sampling_ratio/min': 0.057233426719903946, 'sampling/importance_sampling_ratio/mean': 1.0176209211349487, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.883740559220314, 'clip_ratio/low_mean': 0.0753450570628047, 'clip_ratio/low_min': 0.0753450570628047, 'clip_ratio/high_mean': 0.08628917392343283, 'clip_ratio/high_max': 0.08628917392343283, 'clip_ratio/region_mean': 0.16163423098623753, 'reward_total_mean': 0.5223467350006104, 'reward_meter_mean': 0.5061711072921753, 'reward_meter_std': 0.3633720874786377, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.991203784942627, 'reward_repeat_soft_std': 0.008159817196428776, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5223467350006104, 'reward_total_composite_std': 0.09643376618623734, 'epoch': 0.13}

 39%|███▉      | 1286/3300 [3:11:28<3:46:13,  6.74s/it]INFO 04-13 10:32:26 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1287/3300 [3:11:40<4:33:16,  8.15s/it]2026-04-13 10:32:38,015 | INFO | train_grpo_train | metrics_logged mode=train step=1287

                                                       
{'loss': -0.111, 'grad_norm': 3.8122456073760986, 'learning_rate': 6.103030303030304e-06, 'num_tokens': 2273309.0, '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': 67.0, 'rewards/meter/mean': 0.7934470772743225, 'rewards/meter/std': 0.28795894980430603, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9797966480255127, 'rewards/repeat_soft/std': 0.02857290394604206, 'rewards/judge_quality/mean': 0.7237499952316284, 'rewards/judge_quality/std': 0.32487085461616516, 'rewards/total_composite/mean': 0.6774677038192749, 'rewards/total_composite/std': 0.3324925899505615, 'reward': 0.6774677038192749, 'reward_std': 0.3324925899505615, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16111302375793457, 'sampling/sampling_logp_difference/max': 1.9014368057250977, 'sampling/importance_sampling_ratio/min': 0.14935387670993805, 'sampling/importance_sampling_ratio/mean': 0.9883249998092651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5149192363023758, 'clip_ratio/low_mean': 0.03316326625645161, 'clip_ratio/low_min': 0.03316326625645161, 'clip_ratio/high_mean': 0.09763261023908854, 'clip_ratio/high_max': 0.09763261023908854, 'clip_ratio/region_mean': 0.13079587649554014, 'reward_total_mean': 0.6774677038192749, 'reward_meter_mean': 0.7934470772743225, 'reward_meter_std': 0.28795894980430603, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9797966480255127, 'reward_repeat_soft_std': 0.02857290394604206, 'reward_judge_quality_mean': 0.7237499952316284, 'reward_judge_quality_std': 0.32487085461616516, 'reward_total_composite_mean': 0.6774677038192749, 'reward_total_composite_std': 0.3324925899505615, 'epoch': 0.13}

 39%|███▉      | 1287/3300 [3:11:40<4:33:16,  8.15s/it]INFO 04-13 10:32:38 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1288/3300 [3:11:51<5:05:25,  9.11s/it]2026-04-13 10:32:49,369 | INFO | train_grpo_train | metrics_logged mode=train step=1288

                                                       
{'loss': -0.0769, 'grad_norm': 2.889350652694702, 'learning_rate': 6.1e-06, 'num_tokens': 2274900.0, 'completions/mean_length': 89.875, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 29.571430206298828, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6511855125427246, 'rewards/meter/std': 0.36623069643974304, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.4337500035762787, 'rewards/judge_quality/std': 0.24663087725639343, 'rewards/total_composite/mean': 0.5051541924476624, 'rewards/total_composite/std': 0.24218609929084778, 'reward': 0.5051541924476624, 'reward_std': 0.24218609929084778, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1646328568458557, 'sampling/sampling_logp_difference/max': 1.674455165863037, 'sampling/importance_sampling_ratio/min': 0.18741025030612946, 'sampling/importance_sampling_ratio/mean': 0.9861435294151306, 'sampling/importance_sampling_ratio/max': 1.7221283912658691, 'entropy': 1.0403777062892914, 'clip_ratio/low_mean': 0.03987455181777477, 'clip_ratio/low_min': 0.03987455181777477, 'clip_ratio/high_mean': 0.1197044001892209, 'clip_ratio/high_max': 0.1197044001892209, 'clip_ratio/region_mean': 0.15957895200699568, 'reward_total_mean': 0.5051541924476624, 'reward_meter_mean': 0.6511855125427246, 'reward_meter_std': 0.36623069643974304, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.4337500035762787, 'reward_judge_quality_std': 0.24663087725639343, 'reward_total_composite_mean': 0.5051541924476624, 'reward_total_composite_std': 0.24218609929084778, 'epoch': 0.13}

 39%|███▉      | 1288/3300 [3:11:51<5:05:25,  9.11s/it]INFO 04-13 10:32:49 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1289/3300 [3:11:58<4:41:22,  8.40s/it]2026-04-13 10:32:56,103 | INFO | train_grpo_train | metrics_logged mode=train step=1289

                                                       
{'loss': 0.0438, 'grad_norm': 13.759574890136719, 'learning_rate': 6.096969696969698e-06, 'num_tokens': 2276788.0, 'completions/mean_length': 52.0, 'completions/min_length': 48.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.247971773147583, 'rewards/meter/std': 0.3090267777442932, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975694417953491, 'rewards/repeat_soft/std': 0.004525233991444111, 'rewards/judge_quality/mean': 0.690000057220459, 'rewards/judge_quality/std': 0.2343989461660385, 'rewards/total_composite/mean': 0.44359105825424194, 'rewards/total_composite/std': 0.10209126770496368, 'reward': 0.44359105825424194, 'reward_std': 0.10209126770496368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13923315703868866, 'sampling/sampling_logp_difference/max': 2.5892133712768555, 'sampling/importance_sampling_ratio/min': 0.07507907599210739, 'sampling/importance_sampling_ratio/mean': 0.9784464836120605, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5176977664232254, 'clip_ratio/low_mean': 0.057573966681957245, 'clip_ratio/low_min': 0.057573966681957245, 'clip_ratio/high_mean': 0.056528182700276375, 'clip_ratio/high_max': 0.056528182700276375, 'clip_ratio/region_mean': 0.11410214938223362, 'reward_total_mean': 0.44359105825424194, 'reward_meter_mean': 0.247971773147583, 'reward_meter_std': 0.3090267777442932, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975694417953491, 'reward_repeat_soft_std': 0.004525233991444111, 'reward_judge_quality_mean': 0.690000057220459, 'reward_judge_quality_std': 0.2343989461660385, 'reward_total_composite_mean': 0.44359105825424194, 'reward_total_composite_std': 0.10209126770496368, 'epoch': 0.13}

 39%|███▉      | 1289/3300 [3:11:58<4:41:22,  8.40s/it]INFO 04-13 10:32:56 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1290/3300 [3:12:09<5:07:55,  9.19s/it]2026-04-13 10:33:07,155 | INFO | train_grpo_train | metrics_logged mode=train step=1290

                                                       
{'loss': -0.1377, 'grad_norm': 2.128577470779419, 'learning_rate': 6.0939393939393946e-06, 'num_tokens': 2278437.0, 'completions/mean_length': 174.125, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8028546571731567, 'rewards/meter/std': 0.3164840340614319, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9554349184036255, 'rewards/repeat_soft/std': 0.03757408261299133, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.20694634318351746, 'rewards/total_composite/mean': 0.47944241762161255, 'rewards/total_composite/std': 0.30393680930137634, 'reward': 0.47944241762161255, 'reward_std': 0.30393680930137634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14244621992111206, 'sampling/sampling_logp_difference/max': 1.4288115501403809, 'sampling/importance_sampling_ratio/min': 0.239593505859375, 'sampling/importance_sampling_ratio/mean': 1.0086246728897095, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5359006002545357, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11664778739213943, 'clip_ratio/high_max': 0.11664778739213943, 'clip_ratio/region_mean': 0.11664778739213943, 'reward_total_mean': 0.47944241762161255, 'reward_meter_mean': 0.8028546571731567, 'reward_meter_std': 0.3164840340614319, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9554349184036255, 'reward_repeat_soft_std': 0.03757408261299133, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.20694634318351746, 'reward_total_composite_mean': 0.47944241762161255, 'reward_total_composite_std': 0.30393680930137634, 'epoch': 0.13}

 39%|███▉      | 1290/3300 [3:12:09<5:07:55,  9.19s/it]INFO 04-13 10:33:07 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1291/3300 [3:12:15<4:39:30,  8.35s/it]2026-04-13 10:33:13,529 | INFO | train_grpo_train | metrics_logged mode=train step=1291

                                                       
{'loss': 0.0015, 'grad_norm': 12.229418754577637, 'learning_rate': 6.090909090909092e-06, 'num_tokens': 2280580.0, 'completions/mean_length': 74.875, 'completions/min_length': 67.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.7657289505004883, 'rewards/meter/std': 0.3279200792312622, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.920647382736206, 'rewards/repeat_soft/std': 0.052993398159742355, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.517281174659729, 'rewards/total_composite/std': 0.07782205939292908, 'reward': 0.517281174659729, 'reward_std': 0.07782205939292908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14785484969615936, 'sampling/sampling_logp_difference/max': 1.4468560218811035, 'sampling/importance_sampling_ratio/min': 0.2353089302778244, 'sampling/importance_sampling_ratio/mean': 0.9972316026687622, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9556903690099716, 'clip_ratio/low_mean': 0.05760151706635952, 'clip_ratio/low_min': 0.05760151706635952, 'clip_ratio/high_mean': 0.0969934118911624, 'clip_ratio/high_max': 0.0969934118911624, 'clip_ratio/region_mean': 0.15459492895752192, 'reward_total_mean': 0.517281174659729, 'reward_meter_mean': 0.7657289505004883, 'reward_meter_std': 0.3279200792312622, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.920647382736206, 'reward_repeat_soft_std': 0.052993398159742355, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.517281174659729, 'reward_total_composite_std': 0.07782205939292908, 'epoch': 0.13}

 39%|███▉      | 1291/3300 [3:12:15<4:39:30,  8.35s/it]INFO 04-13 10:33:13 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1292/3300 [3:12:26<5:07:08,  9.18s/it]2026-04-13 10:33:24,640 | INFO | train_grpo_train | metrics_logged mode=train step=1292

                                                       
{'loss': -0.1627, 'grad_norm': 3.7194430828094482, 'learning_rate': 6.087878787878788e-06, 'num_tokens': 2282395.0, 'completions/mean_length': 124.875, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.5714340209961, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.35180729627609253, 'rewards/meter/std': 0.2963906526565552, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9926522374153137, 'rewards/repeat_soft/std': 0.010722406208515167, 'rewards/judge_quality/mean': 0.5987499952316284, 'rewards/judge_quality/std': 0.3210001289844513, 'rewards/total_composite/mean': 0.449861079454422, 'rewards/total_composite/std': 0.22524167597293854, 'reward': 0.449861079454422, 'reward_std': 0.22524167597293854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14167234301567078, 'sampling/sampling_logp_difference/max': 2.0441176891326904, 'sampling/importance_sampling_ratio/min': 0.12949439883232117, 'sampling/importance_sampling_ratio/mean': 1.0059019327163696, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6685599088668823, 'clip_ratio/low_mean': 0.04187853168696165, 'clip_ratio/low_min': 0.04187853168696165, 'clip_ratio/high_mean': 0.08393612876534462, 'clip_ratio/high_max': 0.08393612876534462, 'clip_ratio/region_mean': 0.12581466045230627, 'reward_total_mean': 0.449861079454422, 'reward_meter_mean': 0.35180729627609253, 'reward_meter_std': 0.2963906526565552, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9926522374153137, 'reward_repeat_soft_std': 0.010722406208515167, 'reward_judge_quality_mean': 0.5987499952316284, 'reward_judge_quality_std': 0.3210001289844513, 'reward_total_composite_mean': 0.449861079454422, 'reward_total_composite_std': 0.22524167597293854, 'epoch': 0.13}

 39%|███▉      | 1292/3300 [3:12:26<5:07:08,  9.18s/it]INFO 04-13 10:33:24 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1293/3300 [3:12:33<4:37:23,  8.29s/it]2026-04-13 10:33:30,875 | INFO | train_grpo_train | metrics_logged mode=train step=1293

                                                       
{'loss': 0.0077, 'grad_norm': 11.292625427246094, 'learning_rate': 6.0848484848484855e-06, 'num_tokens': 2284209.0, 'completions/mean_length': 63.75, 'completions/min_length': 56.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.9202028512954712, 'rewards/meter/std': 0.1847878098487854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9812027812004089, 'rewards/repeat_soft/std': 0.022314518690109253, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6703628301620483, 'rewards/total_composite/std': 0.12549921870231628, 'reward': 0.6703628301620483, 'reward_std': 0.12549921870231628, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15040865540504456, 'sampling/sampling_logp_difference/max': 1.460923194885254, 'sampling/importance_sampling_ratio/min': 0.2320219874382019, 'sampling/importance_sampling_ratio/mean': 1.0040225982666016, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8783659860491753, 'clip_ratio/low_mean': 0.08028423693031073, 'clip_ratio/low_min': 0.08028423693031073, 'clip_ratio/high_mean': 0.05312481801956892, 'clip_ratio/high_max': 0.05312481801956892, 'clip_ratio/region_mean': 0.13340905494987965, 'reward_total_mean': 0.6703628301620483, 'reward_meter_mean': 0.9202028512954712, 'reward_meter_std': 0.1847878098487854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9812027812004089, 'reward_repeat_soft_std': 0.022314518690109253, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6703628301620483, 'reward_total_composite_std': 0.12549921870231628, 'epoch': 0.13}

 39%|███▉      | 1293/3300 [3:12:33<4:37:23,  8.29s/it]INFO 04-13 10:33:31 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1294/3300 [3:12:39<4:15:44,  7.65s/it]2026-04-13 10:33:37,019 | INFO | train_grpo_train | metrics_logged mode=train step=1294

                                                       
{'loss': 0.3254, 'grad_norm': 20.19084358215332, 'learning_rate': 6.081818181818182e-06, 'num_tokens': 2285629.0, 'completions/mean_length': 27.5, 'completions/min_length': 20.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.684787929058075, 'rewards/meter/std': 0.3552640378475189, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.953275203704834, 'rewards/repeat_soft/std': 0.024769343435764313, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.6545860767364502, 'rewards/total_composite/std': 0.26531553268432617, 'reward': 0.6545860767364502, 'reward_std': 0.26531553268432617, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15634866058826447, 'sampling/sampling_logp_difference/max': 1.5979595184326172, 'sampling/importance_sampling_ratio/min': 0.20230890810489655, 'sampling/importance_sampling_ratio/mean': 1.0324431657791138, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9607767537236214, 'clip_ratio/low_mean': 0.10757675394415855, 'clip_ratio/low_min': 0.10757675394415855, 'clip_ratio/high_mean': 0.0832880437374115, 'clip_ratio/high_max': 0.0832880437374115, 'clip_ratio/region_mean': 0.19086479768157005, 'reward_total_mean': 0.6545860767364502, 'reward_meter_mean': 0.684787929058075, 'reward_meter_std': 0.3552640378475189, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.953275203704834, 'reward_repeat_soft_std': 0.024769343435764313, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.6545860767364502, 'reward_total_composite_std': 0.26531553268432617, 'epoch': 0.13}

 39%|███▉      | 1294/3300 [3:12:39<4:15:44,  7.65s/it]INFO 04-13 10:33:37 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1295/3300 [3:12:46<4:08:22,  7.43s/it]2026-04-13 10:33:43,948 | INFO | train_grpo_train | metrics_logged mode=train step=1295

                                                       
{'loss': 0.0639, 'grad_norm': 8.543797492980957, 'learning_rate': 6.07878787878788e-06, 'num_tokens': 2287947.0, 'completions/mean_length': 102.75, 'completions/min_length': 85.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.75, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.8280737996101379, 'rewards/meter/std': 0.25559964776039124, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.979012668132782, 'rewards/repeat_soft/std': 0.009831268340349197, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5636317729949951, 'rewards/total_composite/std': 0.10877915471792221, 'reward': 0.5636317729949951, 'reward_std': 0.10877914726734161, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15261870622634888, 'sampling/sampling_logp_difference/max': 1.567774772644043, 'sampling/importance_sampling_ratio/min': 0.20850864052772522, 'sampling/importance_sampling_ratio/mean': 1.0170245170593262, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9373430162668228, 'clip_ratio/low_mean': 0.03791691828519106, 'clip_ratio/low_min': 0.03791691828519106, 'clip_ratio/high_mean': 0.0900578573346138, 'clip_ratio/high_max': 0.0900578573346138, 'clip_ratio/region_mean': 0.12797477561980486, 'reward_total_mean': 0.5636317729949951, 'reward_meter_mean': 0.8280737996101379, 'reward_meter_std': 0.25559964776039124, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.979012668132782, 'reward_repeat_soft_std': 0.009831268340349197, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5636317729949951, 'reward_total_composite_std': 0.10877915471792221, 'epoch': 0.13}

 39%|███▉      | 1295/3300 [3:12:46<4:08:22,  7.43s/it]INFO 04-13 10:33:44 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1296/3300 [3:12:52<3:53:15,  6.98s/it]2026-04-13 10:33:49,883 | INFO | train_grpo_train | metrics_logged mode=train step=1296

                                                       
{'loss': 0.0749, 'grad_norm': 19.56597900390625, 'learning_rate': 6.0757575757575755e-06, 'num_tokens': 2289501.0, 'completions/mean_length': 41.25, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8564885854721069, 'rewards/meter/std': 0.26120027899742126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9801363348960876, 'rewards/repeat_soft/std': 0.010971256531774998, 'rewards/judge_quality/mean': 0.48875001072883606, 'rewards/judge_quality/std': 0.16591200232505798, 'rewards/total_composite/mean': 0.624244213104248, 'rewards/total_composite/std': 0.14069931209087372, 'reward': 0.624244213104248, 'reward_std': 0.14069931209087372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18532513082027435, 'sampling/sampling_logp_difference/max': 2.3976306915283203, 'sampling/importance_sampling_ratio/min': 0.09093314409255981, 'sampling/importance_sampling_ratio/mean': 1.0053044557571411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0849749743938446, 'clip_ratio/low_mean': 0.10320191085338593, 'clip_ratio/low_min': 0.10320191085338593, 'clip_ratio/high_mean': 0.07297545112669468, 'clip_ratio/high_max': 0.07297545112669468, 'clip_ratio/region_mean': 0.1761773619800806, 'reward_total_mean': 0.624244213104248, 'reward_meter_mean': 0.8564885854721069, 'reward_meter_std': 0.26120027899742126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9801363348960876, 'reward_repeat_soft_std': 0.010971256531774998, 'reward_judge_quality_mean': 0.48875001072883606, 'reward_judge_quality_std': 0.16591200232505798, 'reward_total_composite_mean': 0.624244213104248, 'reward_total_composite_std': 0.14069931209087372, 'epoch': 0.13}

 39%|███▉      | 1296/3300 [3:12:52<3:53:15,  6.98s/it]INFO 04-13 10:33:50 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1297/3300 [3:13:02<4:32:39,  8.17s/it]2026-04-13 10:34:00,815 | INFO | train_grpo_train | metrics_logged mode=train step=1297

                                                       
{'loss': -0.0931, 'grad_norm': 2.960124969482422, 'learning_rate': 6.072727272727274e-06, 'num_tokens': 2290954.0, 'completions/mean_length': 96.625, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.28571701049805, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.5484809875488281, 'rewards/meter/std': 0.34423375129699707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9816007614135742, 'rewards/repeat_soft/std': 0.02931581251323223, 'rewards/judge_quality/mean': 0.5762499570846558, 'rewards/judge_quality/std': 0.31513887643814087, 'rewards/total_composite/mean': 0.48849669098854065, 'rewards/total_composite/std': 0.21631859242916107, 'reward': 0.48849669098854065, 'reward_std': 0.21631857752799988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1759599894285202, 'sampling/sampling_logp_difference/max': 2.608419895172119, 'sampling/importance_sampling_ratio/min': 0.07365082949399948, 'sampling/importance_sampling_ratio/mean': 1.0088098049163818, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0217162147164345, 'clip_ratio/low_mean': 0.01944444514811039, 'clip_ratio/low_min': 0.01944444514811039, 'clip_ratio/high_mean': 0.13144625257700682, 'clip_ratio/high_max': 0.13144625257700682, 'clip_ratio/region_mean': 0.1508906977251172, 'reward_total_mean': 0.48849669098854065, 'reward_meter_mean': 0.5484809875488281, 'reward_meter_std': 0.34423375129699707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9816007614135742, 'reward_repeat_soft_std': 0.02931581251323223, 'reward_judge_quality_mean': 0.5762499570846558, 'reward_judge_quality_std': 0.31513887643814087, 'reward_total_composite_mean': 0.48849669098854065, 'reward_total_composite_std': 0.21631859242916107, 'epoch': 0.13}

 39%|███▉      | 1297/3300 [3:13:03<4:32:39,  8.17s/it]INFO 04-13 10:34:01 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1298/3300 [3:13:14<5:04:22,  9.12s/it]2026-04-13 10:34:12,163 | INFO | train_grpo_train | metrics_logged mode=train step=1298

                                                       
{'loss': -0.1556, 'grad_norm': 2.7274155616760254, 'learning_rate': 6.06969696969697e-06, 'num_tokens': 2292935.0, 'completions/mean_length': 137.625, 'completions/min_length': 72.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 84.14286041259766, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.7151336073875427, 'rewards/meter/std': 0.3311758041381836, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.981122612953186, 'rewards/repeat_soft/std': 0.020318450406193733, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.47950494289398193, 'rewards/total_composite/std': 0.20593856275081635, 'reward': 0.47950494289398193, 'reward_std': 0.20593856275081635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16555652022361755, 'sampling/sampling_logp_difference/max': 2.035820484161377, 'sampling/importance_sampling_ratio/min': 0.1305733025074005, 'sampling/importance_sampling_ratio/mean': 1.0415858030319214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0952675566077232, 'clip_ratio/low_mean': 0.025537634268403053, 'clip_ratio/low_min': 0.025537634268403053, 'clip_ratio/high_mean': 0.11700581107288599, 'clip_ratio/high_max': 0.11700581107288599, 'clip_ratio/region_mean': 0.14254344534128904, 'reward_total_mean': 0.47950494289398193, 'reward_meter_mean': 0.7151336073875427, 'reward_meter_std': 0.3311758041381836, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.981122612953186, 'reward_repeat_soft_std': 0.020318450406193733, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.47950494289398193, 'reward_total_composite_std': 0.20593856275081635, 'epoch': 0.13}

 39%|███▉      | 1298/3300 [3:13:14<5:04:22,  9.12s/it]INFO 04-13 10:34:12 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1299/3300 [3:13:20<4:32:03,  8.16s/it]2026-04-13 10:34:18,069 | INFO | train_grpo_train | metrics_logged mode=train step=1299

                                                       
{'loss': 0.056, 'grad_norm': 13.681201934814453, 'learning_rate': 6.066666666666667e-06, 'num_tokens': 2294597.0, 'completions/mean_length': 36.75, 'completions/min_length': 31.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.33265426754951477, 'rewards/meter/std': 0.28129392862319946, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9925545454025269, 'rewards/repeat_soft/std': 0.01810423843562603, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.1865811049938202, 'rewards/total_composite/mean': 0.45487838983535767, 'rewards/total_composite/std': 0.07872791588306427, 'reward': 0.45487838983535767, 'reward_std': 0.07872791588306427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14265796542167664, 'sampling/sampling_logp_difference/max': 2.450737953186035, 'sampling/importance_sampling_ratio/min': 0.08622992783784866, 'sampling/importance_sampling_ratio/mean': 1.0143132209777832, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7138647735118866, 'clip_ratio/low_mean': 0.060220772633329034, 'clip_ratio/low_min': 0.060220772633329034, 'clip_ratio/high_mean': 0.07355498988181353, 'clip_ratio/high_max': 0.07355498988181353, 'clip_ratio/region_mean': 0.13377576251514256, 'reward_total_mean': 0.45487838983535767, 'reward_meter_mean': 0.33265426754951477, 'reward_meter_std': 0.28129392862319946, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9925545454025269, 'reward_repeat_soft_std': 0.01810423843562603, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.1865811049938202, 'reward_total_composite_mean': 0.45487838983535767, 'reward_total_composite_std': 0.07872791588306427, 'epoch': 0.13}

 39%|███▉      | 1299/3300 [3:13:20<4:32:03,  8.16s/it]INFO 04-13 10:34:18 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1300/3300 [3:13:31<5:01:00,  9.03s/it]2026-04-13 10:34:29,135 | INFO | train_grpo_train | metrics_logged mode=train step=1300

                                                       
{'loss': -0.1511, 'grad_norm': 3.13342022895813, 'learning_rate': 6.063636363636364e-06, 'num_tokens': 2296608.0, 'completions/mean_length': 124.375, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.3995330035686493, 'rewards/meter/std': 0.4067206382751465, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9935516119003296, 'rewards/repeat_soft/std': 0.008096016943454742, 'rewards/judge_quality/mean': 0.4449999928474426, 'rewards/judge_quality/std': 0.21876277029514313, 'rewards/total_composite/mean': 0.417496919631958, 'rewards/total_composite/std': 0.19607169926166534, 'reward': 0.417496919631958, 'reward_std': 0.19607168436050415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1340249627828598, 'sampling/sampling_logp_difference/max': 1.2487568855285645, 'sampling/importance_sampling_ratio/min': 0.2868611812591553, 'sampling/importance_sampling_ratio/mean': 0.996375560760498, 'sampling/importance_sampling_ratio/max': 1.8476996421813965, 'entropy': 0.6958075687289238, 'clip_ratio/low_mean': 0.04703459283336997, 'clip_ratio/low_min': 0.04703459283336997, 'clip_ratio/high_mean': 0.0712701603770256, 'clip_ratio/high_max': 0.0712701603770256, 'clip_ratio/region_mean': 0.11830475321039557, 'reward_total_mean': 0.417496919631958, 'reward_meter_mean': 0.3995330035686493, 'reward_meter_std': 0.4067206382751465, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9935516119003296, 'reward_repeat_soft_std': 0.008096016943454742, 'reward_judge_quality_mean': 0.4449999928474426, 'reward_judge_quality_std': 0.21876277029514313, 'reward_total_composite_mean': 0.417496919631958, 'reward_total_composite_std': 0.19607169926166534, 'epoch': 0.13}

 39%|███▉      | 1300/3300 [3:13:31<5:01:00,  9.03s/it]INFO 04-13 10:34:29 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.76s/it]

 30%|███       | 3/10 [00:08<00:20,  2.93s/it]

 40%|████      | 4/10 [00:16<00:30,  5.04s/it]

 50%|█████     | 5/10 [00:20<00:23,  4.66s/it]

 60%|██████    | 6/10 [00:24<00:17,  4.39s/it]

 70%|███████   | 7/10 [00:33<00:17,  5.80s/it]

 80%|████████  | 8/10 [00:37<00:10,  5.21s/it]

 90%|█████████ | 9/10 [00:45<00:06,  6.26s/it]

100%|██████████| 10/10 [00:49<00:00,  5.34s/it]2026-04-13 10:35:21,967 | INFO | train_grpo_train | metrics_logged mode=eval step=1300

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 52.7541, 'eval_samples_per_second': 1.516, 'eval_steps_per_second': 0.19, 'eval_num_tokens': 2296608.0, 'eval_completions/mean_length': 83.8875, 'eval_completions/min_length': 27.0, 'eval_completions/max_length': 224.1, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 61.05892906188965, 'eval_completions/min_terminated_length': 27.0, 'eval_completions/max_terminated_length': 99.4, 'eval_rewards/meter/mean': 0.6011181086301803, 'eval_rewards/meter/std': 0.36246598362922666, 'eval_rewards/count_adherence/mean': 0.9479166507720947, 'eval_rewards/count_adherence/std': 0.08784548602998257, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.11700168251991272, 'eval_rewards/repeat_soft/mean': 0.951271939277649, 'eval_rewards/repeat_soft/std': 0.05197468902915716, 'eval_rewards/judge_quality/mean': 0.44524999260902404, 'eval_rewards/judge_quality/std': 0.15283784195780753, 'eval_rewards/total_composite/mean': 0.49333526790142057, 'eval_rewards/total_composite/std': 0.17335330545902253, 'eval_reward': 0.49333526790142057, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.07236771360039711, 'eval_sampling/sampling_logp_difference/max': 1.009927797317505, 'eval_sampling/importance_sampling_ratio/min': 0.38720138370990753, 'eval_sampling/importance_sampling_ratio/mean': 1.0154764890670775, 'eval_sampling/importance_sampling_ratio/max': 1.4079676747322083, 'eval_entropy': 0.8205519914627075, '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': 0.49333526790142057, 'eval_reward_meter_mean': 0.6011181086301803, 'eval_reward_meter_std': 0.36246598362922666, 'eval_reward_count_adherence_mean': 0.9479166507720947, 'eval_reward_count_adherence_std': 0.08784548602998257, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.11700168251991272, 'eval_reward_repeat_soft_mean': 0.951271939277649, 'eval_reward_repeat_soft_std': 0.05197468902915716, 'eval_reward_judge_quality_mean': 0.44524999260902404, 'eval_reward_judge_quality_std': 0.15283784195780753, 'eval_reward_total_composite_mean': 0.49333526790142057, 'eval_reward_total_composite_std': 0.17335330545902253, 'epoch': 0.13}

 39%|███▉      | 1300/3300 [3:14:24<5:01:00,  9.03s/it]

100%|██████████| 10/10 [00:49<00:00,  5.34s/it]

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

 39%|███▉      | 1301/3300 [3:14:39<14:53:02, 26.80s/it]2026-04-13 10:35:37,458 | INFO | train_grpo_train | metrics_logged mode=train step=1301

                                                        
{'loss': -0.1863, 'grad_norm': 2.081382989883423, 'learning_rate': 6.060606060606061e-06, 'num_tokens': 2298683.0, 'completions/mean_length': 202.375, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 99.16667175292969, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.4714367687702179, 'rewards/meter/std': 0.26599934697151184, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.18898223340511322, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9643048048019409, 'rewards/repeat_soft/std': 0.03838948532938957, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.3530154228210449, 'rewards/total_composite/std': 0.22604341804981232, 'reward': 0.3530154228210449, 'reward_std': 0.22604341804981232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12762080132961273, 'sampling/sampling_logp_difference/max': 1.3520245552062988, 'sampling/importance_sampling_ratio/min': 0.258715957403183, 'sampling/importance_sampling_ratio/mean': 1.0024654865264893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7316770181059837, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11028255522251129, 'clip_ratio/high_max': 0.11028255522251129, 'clip_ratio/region_mean': 0.11028255522251129, 'reward_total_mean': 0.3530154228210449, 'reward_meter_mean': 0.4714367687702179, 'reward_meter_std': 0.26599934697151184, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.18898223340511322, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9643048048019409, 'reward_repeat_soft_std': 0.03838948532938957, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.3530154228210449, 'reward_total_composite_std': 0.22604341804981232, 'epoch': 0.13}

 39%|███▉      | 1301/3300 [3:14:39<14:53:02, 26.80s/it]INFO 04-13 10:35:37 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1302/3300 [3:14:46<11:34:40, 20.86s/it]2026-04-13 10:35:44,459 | INFO | train_grpo_train | metrics_logged mode=train step=1302

                                                        
{'loss': 0.0757, 'grad_norm': 13.769412994384766, 'learning_rate': 6.057575757575757e-06, 'num_tokens': 2300221.0, 'completions/mean_length': 37.25, 'completions/min_length': 28.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6589182615280151, 'rewards/meter/std': 0.40660297870635986, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9563091993331909, 'rewards/repeat_soft/std': 0.03997744992375374, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.5515223145484924, 'rewards/total_composite/std': 0.11672757565975189, 'reward': 0.5515223145484924, 'reward_std': 0.11672757565975189, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11146819591522217, 'sampling/sampling_logp_difference/max': 0.8875398635864258, 'sampling/importance_sampling_ratio/min': 0.41166725754737854, 'sampling/importance_sampling_ratio/mean': 1.0080918073654175, 'sampling/importance_sampling_ratio/max': 1.6126405000686646, 'entropy': 0.8147349208593369, 'clip_ratio/low_mean': 0.020604395773261786, 'clip_ratio/low_min': 0.020604395773261786, 'clip_ratio/high_mean': 0.07930322573520243, 'clip_ratio/high_max': 0.07930322573520243, 'clip_ratio/region_mean': 0.09990762150846422, 'reward_total_mean': 0.5515223145484924, 'reward_meter_mean': 0.6589182615280151, 'reward_meter_std': 0.40660297870635986, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9563091993331909, 'reward_repeat_soft_std': 0.03997744992375374, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.5515223145484924, 'reward_total_composite_std': 0.11672757565975189, 'epoch': 0.13}

 39%|███▉      | 1302/3300 [3:14:46<11:34:40, 20.86s/it]INFO 04-13 10:35:44 [block_pool.py:378] Successfully reset prefix cache

 39%|███▉      | 1303/3300 [3:14:58<10:08:05, 18.27s/it]2026-04-13 10:35:56,632 | INFO | train_grpo_train | metrics_logged mode=train step=1303

                                                        
{'loss': -0.1772, 'grad_norm': 3.8782999515533447, 'learning_rate': 6.0545454545454555e-06, 'num_tokens': 2302401.0, 'completions/mean_length': 146.5, 'completions/min_length': 86.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 94.28572082519531, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.5795238018035889, 'rewards/meter/std': 0.248986154794693, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9817588329315186, 'rewards/repeat_soft/std': 0.017542414367198944, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.1524970829486847, 'rewards/total_composite/mean': 0.42917606234550476, 'rewards/total_composite/std': 0.1993083655834198, 'reward': 0.42917606234550476, 'reward_std': 0.1993083357810974, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1815846562385559, 'sampling/sampling_logp_difference/max': 2.7994613647460938, 'sampling/importance_sampling_ratio/min': 0.06084282696247101, 'sampling/importance_sampling_ratio/mean': 1.001242756843567, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9739576280117035, 'clip_ratio/low_mean': 0.06201563123613596, 'clip_ratio/low_min': 0.06201563123613596, 'clip_ratio/high_mean': 0.08822392486035824, 'clip_ratio/high_max': 0.08822392486035824, 'clip_ratio/region_mean': 0.1502395560964942, 'reward_total_mean': 0.42917606234550476, 'reward_meter_mean': 0.5795238018035889, 'reward_meter_std': 0.248986154794693, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9817588329315186, 'reward_repeat_soft_std': 0.017542414367198944, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.1524970829486847, 'reward_total_composite_mean': 0.42917606234550476, 'reward_total_composite_std': 0.1993083655834198, 'epoch': 0.13}

 39%|███▉      | 1303/3300 [3:14:58<10:08:05, 18.27s/it]INFO 04-13 10:35:56 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1304/3300 [3:15:05<8:17:07, 14.94s/it] 2026-04-13 10:36:03,815 | INFO | train_grpo_train | metrics_logged mode=train step=1304

                                                       
{'loss': 0.0499, 'grad_norm': 14.70734691619873, 'learning_rate': 6.051515151515152e-06, 'num_tokens': 2304378.0, 'completions/mean_length': 66.125, 'completions/min_length': 52.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.7551811933517456, 'rewards/meter/std': 0.2874099314212799, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9010825157165527, 'rewards/repeat_soft/std': 0.05824727937579155, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5380387306213379, 'rewards/total_composite/std': 0.09783925861120224, 'reward': 0.5380387306213379, 'reward_std': 0.09783925861120224, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15492960810661316, 'sampling/sampling_logp_difference/max': 2.2266108989715576, 'sampling/importance_sampling_ratio/min': 0.10789347440004349, 'sampling/importance_sampling_ratio/mean': 1.0085541009902954, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7587547898292542, 'clip_ratio/low_mean': 0.07222104258835316, 'clip_ratio/low_min': 0.07222104258835316, 'clip_ratio/high_mean': 0.07400674372911453, 'clip_ratio/high_max': 0.07400674372911453, 'clip_ratio/region_mean': 0.1462277863174677, 'reward_total_mean': 0.5380387306213379, 'reward_meter_mean': 0.7551811933517456, 'reward_meter_std': 0.2874099314212799, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9010825157165527, 'reward_repeat_soft_std': 0.05824727937579155, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5380387306213379, 'reward_total_composite_std': 0.09783925861120224, 'epoch': 0.13}

 40%|███▉      | 1304/3300 [3:15:06<8:17:07, 14.94s/it]INFO 04-13 10:36:04 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1305/3300 [3:15:12<6:52:31, 12.41s/it]2026-04-13 10:36:10,307 | INFO | train_grpo_train | metrics_logged mode=train step=1305

                                                       
{'loss': 0.0479, 'grad_norm': 8.494245529174805, 'learning_rate': 6.048484848484849e-06, 'num_tokens': 2306503.0, 'completions/mean_length': 83.625, 'completions/min_length': 78.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.625, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.8273429870605469, 'rewards/meter/std': 0.21092940866947174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.940427839756012, 'rewards/repeat_soft/std': 0.06156879663467407, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6176351308822632, 'rewards/total_composite/std': 0.08360213041305542, 'reward': 0.6176351308822632, 'reward_std': 0.08360213786363602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1396074891090393, 'sampling/sampling_logp_difference/max': 2.808736801147461, 'sampling/importance_sampling_ratio/min': 0.060281090438365936, 'sampling/importance_sampling_ratio/mean': 1.0032211542129517, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7733359336853027, 'clip_ratio/low_mean': 0.07828882802277803, 'clip_ratio/low_min': 0.07828882802277803, 'clip_ratio/high_mean': 0.05440238770097494, 'clip_ratio/high_max': 0.05440238770097494, 'clip_ratio/region_mean': 0.13269121572375298, 'reward_total_mean': 0.6176351308822632, 'reward_meter_mean': 0.8273429870605469, 'reward_meter_std': 0.21092940866947174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.940427839756012, 'reward_repeat_soft_std': 0.06156879663467407, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6176351308822632, 'reward_total_composite_std': 0.08360213041305542, 'epoch': 0.13}

 40%|███▉      | 1305/3300 [3:15:12<6:52:31, 12.41s/it]INFO 04-13 10:36:10 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1306/3300 [3:15:19<5:59:43, 10.82s/it]2026-04-13 10:36:17,443 | INFO | train_grpo_train | metrics_logged mode=train step=1306

                                                       
{'loss': 0.0345, 'grad_norm': 10.0457181930542, 'learning_rate': 6.0454545454545456e-06, 'num_tokens': 2308856.0, 'completions/mean_length': 93.125, 'completions/min_length': 76.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.125, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.23517444729804993, 'rewards/meter/std': 0.1635940819978714, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9481331706047058, 'rewards/repeat_soft/std': 0.035619623959064484, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.39940470457077026, 'rewards/total_composite/std': 0.051205825060606, 'reward': 0.39940470457077026, 'reward_std': 0.051205825060606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1481638103723526, 'sampling/sampling_logp_difference/max': 2.566396951675415, 'sampling/importance_sampling_ratio/min': 0.07681180536746979, 'sampling/importance_sampling_ratio/mean': 1.0056487321853638, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7814144939184189, 'clip_ratio/low_mean': 0.05869795382022858, 'clip_ratio/low_min': 0.05869795382022858, 'clip_ratio/high_mean': 0.0744721395894885, 'clip_ratio/high_max': 0.0744721395894885, 'clip_ratio/region_mean': 0.13317009340971708, 'reward_total_mean': 0.39940470457077026, 'reward_meter_mean': 0.23517444729804993, 'reward_meter_std': 0.1635940819978714, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9481331706047058, 'reward_repeat_soft_std': 0.035619623959064484, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.39940470457077026, 'reward_total_composite_std': 0.051205825060606, 'epoch': 0.13}

 40%|███▉      | 1306/3300 [3:15:19<5:59:43, 10.82s/it]INFO 04-13 10:36:17 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1307/3300 [3:15:25<5:11:28,  9.38s/it]2026-04-13 10:36:23,435 | INFO | train_grpo_train | metrics_logged mode=train step=1307

                                                       
{'loss': 0.1057, 'grad_norm': 18.926837921142578, 'learning_rate': 6.042424242424243e-06, 'num_tokens': 2310360.0, 'completions/mean_length': 33.0, 'completions/min_length': 29.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.613053560256958, 'rewards/meter/std': 0.41506704688072205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9923344850540161, 'rewards/repeat_soft/std': 0.012143817730247974, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6067043542861938, 'rewards/total_composite/std': 0.20889125764369965, 'reward': 0.6067043542861938, 'reward_std': 0.20889124274253845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12716905772686005, 'sampling/sampling_logp_difference/max': 1.055168867111206, 'sampling/importance_sampling_ratio/min': 0.45095694065093994, 'sampling/importance_sampling_ratio/mean': 1.0153416395187378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9925641939043999, 'clip_ratio/low_mean': 0.046996163204312325, 'clip_ratio/low_min': 0.046996163204312325, 'clip_ratio/high_mean': 0.0647506876848638, 'clip_ratio/high_max': 0.0647506876848638, 'clip_ratio/region_mean': 0.11174685088917613, 'reward_total_mean': 0.6067043542861938, 'reward_meter_mean': 0.613053560256958, 'reward_meter_std': 0.41506704688072205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9923344850540161, 'reward_repeat_soft_std': 0.012143817730247974, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6067043542861938, 'reward_total_composite_std': 0.20889125764369965, 'epoch': 0.13}

 40%|███▉      | 1307/3300 [3:15:25<5:11:28,  9.38s/it]INFO 04-13 10:36:23 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1308/3300 [3:15:32<4:42:51,  8.52s/it]2026-04-13 10:36:29,954 | INFO | train_grpo_train | metrics_logged mode=train step=1308

                                                       
{'loss': 0.0313, 'grad_norm': 11.330538749694824, 'learning_rate': 6.039393939393939e-06, 'num_tokens': 2312017.0, 'completions/mean_length': 60.125, 'completions/min_length': 53.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.125, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.46473997831344604, 'rewards/meter/std': 0.3321572542190552, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9457338452339172, 'rewards/repeat_soft/std': 0.03787406533956528, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.1524970978498459, 'rewards/total_composite/mean': 0.4909104108810425, 'rewards/total_composite/std': 0.0880972370505333, 'reward': 0.4909104108810425, 'reward_std': 0.0880972146987915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15379510819911957, 'sampling/sampling_logp_difference/max': 2.13840913772583, 'sampling/importance_sampling_ratio/min': 0.1178421676158905, 'sampling/importance_sampling_ratio/mean': 1.0030930042266846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8838731646537781, 'clip_ratio/low_mean': 0.08330471720546484, 'clip_ratio/low_min': 0.08330471720546484, 'clip_ratio/high_mean': 0.05535276886075735, 'clip_ratio/high_max': 0.05535276886075735, 'clip_ratio/region_mean': 0.1386574860662222, 'reward_total_mean': 0.4909104108810425, 'reward_meter_mean': 0.46473997831344604, 'reward_meter_std': 0.3321572542190552, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9457338452339172, 'reward_repeat_soft_std': 0.03787406533956528, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.1524970978498459, 'reward_total_composite_mean': 0.4909104108810425, 'reward_total_composite_std': 0.0880972370505333, 'epoch': 0.13}

 40%|███▉      | 1308/3300 [3:15:32<4:42:51,  8.52s/it]INFO 04-13 10:36:30 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1309/3300 [3:15:38<4:26:10,  8.02s/it]2026-04-13 10:36:36,813 | INFO | train_grpo_train | metrics_logged mode=train step=1309

                                                       
{'loss': -0.0157, 'grad_norm': 12.610530853271484, 'learning_rate': 6.0363636363636365e-06, 'num_tokens': 2313604.0, 'completions/mean_length': 36.375, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5330278873443604, 'rewards/meter/std': 0.3682415783405304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.939886212348938, 'rewards/repeat_soft/std': 0.05095674470067024, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.48956868052482605, 'rewards/total_composite/std': 0.10176670551300049, 'reward': 0.48956868052482605, 'reward_std': 0.10176671296358109, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12218648195266724, 'sampling/sampling_logp_difference/max': 1.1730799674987793, 'sampling/importance_sampling_ratio/min': 0.30941247940063477, 'sampling/importance_sampling_ratio/mean': 1.0113162994384766, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7353175655007362, 'clip_ratio/low_mean': 0.08282269071787596, 'clip_ratio/low_min': 0.08282269071787596, 'clip_ratio/high_mean': 0.06039342097938061, 'clip_ratio/high_max': 0.06039342097938061, 'clip_ratio/region_mean': 0.14321611169725657, 'reward_total_mean': 0.48956868052482605, 'reward_meter_mean': 0.5330278873443604, 'reward_meter_std': 0.3682415783405304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.939886212348938, 'reward_repeat_soft_std': 0.05095674470067024, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.48956868052482605, 'reward_total_composite_std': 0.10176670551300049, 'epoch': 0.13}

 40%|███▉      | 1309/3300 [3:15:39<4:26:10,  8.02s/it]INFO 04-13 10:36:37 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1310/3300 [3:15:50<4:57:31,  8.97s/it]2026-04-13 10:36:47,996 | INFO | train_grpo_train | metrics_logged mode=train step=1310

                                                       
{'loss': -0.1008, 'grad_norm': 4.485108852386475, 'learning_rate': 6.033333333333335e-06, 'num_tokens': 2315242.0, 'completions/mean_length': 101.75, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.18399572372436523, 'rewards/meter/std': 0.15946964919567108, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9867174625396729, 'rewards/repeat_soft/std': 0.011840468272566795, 'rewards/judge_quality/mean': 0.48124998807907104, 'rewards/judge_quality/std': 0.2531762421131134, 'rewards/total_composite/mean': 0.3581715226173401, 'rewards/total_composite/std': 0.17232245206832886, 'reward': 0.3581715226173401, 'reward_std': 0.17232243716716766, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1770992875099182, 'sampling/sampling_logp_difference/max': 1.5108189582824707, 'sampling/importance_sampling_ratio/min': 0.22072914242744446, 'sampling/importance_sampling_ratio/mean': 1.0185973644256592, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7373760268092155, 'clip_ratio/low_mean': 0.057486264035105705, 'clip_ratio/low_min': 0.057486264035105705, 'clip_ratio/high_mean': 0.08373156748712063, 'clip_ratio/high_max': 0.08373156748712063, 'clip_ratio/region_mean': 0.14121783152222633, 'reward_total_mean': 0.3581715226173401, 'reward_meter_mean': 0.18399572372436523, 'reward_meter_std': 0.15946964919567108, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9867174625396729, 'reward_repeat_soft_std': 0.011840468272566795, 'reward_judge_quality_mean': 0.48124998807907104, 'reward_judge_quality_std': 0.2531762421131134, 'reward_total_composite_mean': 0.3581715226173401, 'reward_total_composite_std': 0.17232245206832886, 'epoch': 0.13}

 40%|███▉      | 1310/3300 [3:15:50<4:57:31,  8.97s/it]INFO 04-13 10:36:48 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1311/3300 [3:16:01<5:22:05,  9.72s/it]2026-04-13 10:36:59,455 | INFO | train_grpo_train | metrics_logged mode=train step=1311

                                                       
{'loss': -0.0929, 'grad_norm': 2.16886830329895, 'learning_rate': 6.030303030303031e-06, 'num_tokens': 2317169.0, 'completions/mean_length': 226.875, 'completions/min_length': 52.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 55.79999923706055, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.5467915534973145, 'rewards/meter/std': 0.3904469907283783, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6797391176223755, 'rewards/repeat_soft/std': 0.3173426687717438, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.2834985554218292, 'rewards/total_composite/mean': 0.4485192596912384, 'rewards/total_composite/std': 0.33092156052589417, 'reward': 0.4485192596912384, 'reward_std': 0.33092156052589417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09218810498714447, 'sampling/sampling_logp_difference/max': 1.2387669086456299, 'sampling/importance_sampling_ratio/min': 0.28974127769470215, 'sampling/importance_sampling_ratio/mean': 0.9867146611213684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.22532642632722855, 'clip_ratio/low_mean': 0.021189458668231964, 'clip_ratio/low_min': 0.021189458668231964, 'clip_ratio/high_mean': 0.02835648157633841, 'clip_ratio/high_max': 0.02835648157633841, 'clip_ratio/region_mean': 0.049545940244570374, 'reward_total_mean': 0.4485192596912384, 'reward_meter_mean': 0.5467915534973145, 'reward_meter_std': 0.3904469907283783, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6797391176223755, 'reward_repeat_soft_std': 0.3173426687717438, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.2834985554218292, 'reward_total_composite_mean': 0.4485192596912384, 'reward_total_composite_std': 0.33092156052589417, 'epoch': 0.13}

 40%|███▉      | 1311/3300 [3:16:01<5:22:05,  9.72s/it]INFO 04-13 10:36:59 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1312/3300 [3:16:08<4:50:46,  8.78s/it]2026-04-13 10:37:06,039 | INFO | train_grpo_train | metrics_logged mode=train step=1312

                                                       
{'loss': 0.03, 'grad_norm': 11.899909973144531, 'learning_rate': 6.027272727272728e-06, 'num_tokens': 2318847.0, 'completions/mean_length': 37.75, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7507766485214233, 'rewards/meter/std': 0.22238843142986298, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9750610589981079, 'rewards/repeat_soft/std': 0.03612116351723671, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5757731199264526, 'rewards/total_composite/std': 0.07389680296182632, 'reward': 0.5757731199264526, 'reward_std': 0.07389680296182632, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1200757697224617, 'sampling/sampling_logp_difference/max': 2.22994327545166, 'sampling/importance_sampling_ratio/min': 0.10753452777862549, 'sampling/importance_sampling_ratio/mean': 0.9822611808776855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5109836533665657, 'clip_ratio/low_mean': 0.03545725205913186, 'clip_ratio/low_min': 0.03545725205913186, 'clip_ratio/high_mean': 0.06825267663225532, 'clip_ratio/high_max': 0.06825267663225532, 'clip_ratio/region_mean': 0.10370992869138718, 'reward_total_mean': 0.5757731199264526, 'reward_meter_mean': 0.7507766485214233, 'reward_meter_std': 0.22238843142986298, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9750610589981079, 'reward_repeat_soft_std': 0.03612116351723671, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5757731199264526, 'reward_total_composite_std': 0.07389680296182632, 'epoch': 0.13}

 40%|███▉      | 1312/3300 [3:16:08<4:50:46,  8.78s/it]INFO 04-13 10:37:06 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1313/3300 [3:16:14<4:26:16,  8.04s/it]2026-04-13 10:37:12,366 | INFO | train_grpo_train | metrics_logged mode=train step=1313

                                                       
{'loss': 0.0006, 'grad_norm': 13.080678939819336, 'learning_rate': 6.024242424242425e-06, 'num_tokens': 2320426.0, 'completions/mean_length': 25.375, 'completions/min_length': 24.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.7285583019256592, 'rewards/meter/std': 0.31013527512550354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9292786121368408, 'rewards/repeat_soft/std': 0.05441930517554283, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5627000331878662, 'rewards/total_composite/std': 0.09881270676851273, 'reward': 0.5627000331878662, 'reward_std': 0.09881272166967392, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1227404922246933, 'sampling/sampling_logp_difference/max': 1.3473396301269531, 'sampling/importance_sampling_ratio/min': 0.25993087887763977, 'sampling/importance_sampling_ratio/mean': 0.9929416179656982, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5725177004933357, 'clip_ratio/low_mean': 0.03541666641831398, 'clip_ratio/low_min': 0.03541666641831398, 'clip_ratio/high_mean': 0.10167353507131338, 'clip_ratio/high_max': 0.10167353507131338, 'clip_ratio/region_mean': 0.13709020148962736, 'reward_total_mean': 0.5627000331878662, 'reward_meter_mean': 0.7285583019256592, 'reward_meter_std': 0.31013527512550354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9292786121368408, 'reward_repeat_soft_std': 0.05441930517554283, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5627000331878662, 'reward_total_composite_std': 0.09881270676851273, 'epoch': 0.13}

 40%|███▉      | 1313/3300 [3:16:14<4:26:16,  8.04s/it]INFO 04-13 10:37:12 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1314/3300 [3:16:20<4:09:19,  7.53s/it]2026-04-13 10:37:18,710 | INFO | train_grpo_train | metrics_logged mode=train step=1314

                                                       
{'loss': 0.0141, 'grad_norm': 18.13827896118164, 'learning_rate': 6.021212121212122e-06, 'num_tokens': 2322019.0, 'completions/mean_length': 42.125, 'completions/min_length': 36.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.14196985960006714, 'rewards/meter/std': 0.1567562371492386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9805315732955933, 'rewards/repeat_soft/std': 0.02711394801735878, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.3856370151042938, 'rewards/total_composite/std': 0.04208798334002495, 'reward': 0.3856370151042938, 'reward_std': 0.04208798334002495, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17558568716049194, 'sampling/sampling_logp_difference/max': 1.8940314054489136, 'sampling/importance_sampling_ratio/min': 0.1504640132188797, 'sampling/importance_sampling_ratio/mean': 1.0315899848937988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7719564884901047, 'clip_ratio/low_mean': 0.1019074134528637, 'clip_ratio/low_min': 0.1019074134528637, 'clip_ratio/high_mean': 0.03083881549537182, 'clip_ratio/high_max': 0.03083881549537182, 'clip_ratio/region_mean': 0.1327462289482355, 'reward_total_mean': 0.3856370151042938, 'reward_meter_mean': 0.14196985960006714, 'reward_meter_std': 0.1567562371492386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9805315732955933, 'reward_repeat_soft_std': 0.02711394801735878, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.3856370151042938, 'reward_total_composite_std': 0.04208798334002495, 'epoch': 0.13}

 40%|███▉      | 1314/3300 [3:16:20<4:09:19,  7.53s/it]INFO 04-13 10:37:18 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1315/3300 [3:16:26<3:52:45,  7.04s/it]2026-04-13 10:37:24,590 | INFO | train_grpo_train | metrics_logged mode=train step=1315

                                                       
{'loss': 0.0066, 'grad_norm': 21.144624710083008, 'learning_rate': 6.018181818181818e-06, 'num_tokens': 2323419.0, 'completions/mean_length': 27.0, 'completions/min_length': 24.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.0, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.3701016902923584, 'rewards/meter/std': 0.3760126233100891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9868196249008179, 'rewards/repeat_soft/std': 0.011254421435296535, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.20760111510753632, 'rewards/total_composite/mean': 0.5289339423179626, 'rewards/total_composite/std': 0.20768073201179504, 'reward': 0.5289339423179626, 'reward_std': 0.20768074691295624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1722986400127411, 'sampling/sampling_logp_difference/max': 1.8542876243591309, 'sampling/importance_sampling_ratio/min': 0.15656444430351257, 'sampling/importance_sampling_ratio/mean': 1.019766926765442, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9304405450820923, 'clip_ratio/low_mean': 0.1068057743832469, 'clip_ratio/low_min': 0.1068057743832469, 'clip_ratio/high_mean': 0.03017241321504116, 'clip_ratio/high_max': 0.03017241321504116, 'clip_ratio/region_mean': 0.13697818759828806, 'reward_total_mean': 0.5289339423179626, 'reward_meter_mean': 0.3701016902923584, 'reward_meter_std': 0.3760126233100891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9868196249008179, 'reward_repeat_soft_std': 0.011254421435296535, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.20760111510753632, 'reward_total_composite_mean': 0.5289339423179626, 'reward_total_composite_std': 0.20768073201179504, 'epoch': 0.13}

 40%|███▉      | 1315/3300 [3:16:26<3:52:45,  7.04s/it]INFO 04-13 10:37:24 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1316/3300 [3:16:37<4:33:30,  8.27s/it]2026-04-13 10:37:35,743 | INFO | train_grpo_train | metrics_logged mode=train step=1316

                                                       
{'loss': -0.1241, 'grad_norm': 4.055689811706543, 'learning_rate': 6.015151515151516e-06, 'num_tokens': 2325197.0, 'completions/mean_length': 119.25, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 63.142860412597656, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.563539981842041, 'rewards/meter/std': 0.3978107273578644, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.930457353591919, 'rewards/repeat_soft/std': 0.04242706298828125, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.44667279720306396, 'rewards/total_composite/std': 0.20369835197925568, 'reward': 0.44667279720306396, 'reward_std': 0.20369835197925568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16629445552825928, 'sampling/sampling_logp_difference/max': 1.4497017860412598, 'sampling/importance_sampling_ratio/min': 0.23464025557041168, 'sampling/importance_sampling_ratio/mean': 1.0269790887832642, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.208445429801941, 'clip_ratio/low_mean': 0.03451771289110184, 'clip_ratio/low_min': 0.03451771289110184, 'clip_ratio/high_mean': 0.08997233491390944, 'clip_ratio/high_max': 0.08997233491390944, 'clip_ratio/region_mean': 0.12449004780501127, 'reward_total_mean': 0.44667279720306396, 'reward_meter_mean': 0.563539981842041, 'reward_meter_std': 0.3978107273578644, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.930457353591919, 'reward_repeat_soft_std': 0.04242706298828125, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.44667279720306396, 'reward_total_composite_std': 0.20369835197925568, 'epoch': 0.13}

 40%|███▉      | 1316/3300 [3:16:37<4:33:30,  8.27s/it]INFO 04-13 10:37:35 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1317/3300 [3:16:51<5:27:55,  9.92s/it]2026-04-13 10:37:49,520 | INFO | train_grpo_train | metrics_logged mode=train step=1317

                                                       
{'loss': -0.1629, 'grad_norm': 2.7149338722229004, 'learning_rate': 6.012121212121213e-06, 'num_tokens': 2327099.0, 'completions/mean_length': 128.75, 'completions/min_length': 64.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.32614681124687195, 'rewards/meter/std': 0.27816957235336304, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9868057370185852, 'rewards/repeat_soft/std': 0.01320960745215416, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.29731839895248413, 'rewards/total_composite/mean': 0.4235251843929291, 'rewards/total_composite/std': 0.1879330724477768, 'reward': 0.4235251843929291, 'reward_std': 0.1879330724477768, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12032223492860794, 'sampling/sampling_logp_difference/max': 2.3049678802490234, 'sampling/importance_sampling_ratio/min': 0.09976200759410858, 'sampling/importance_sampling_ratio/mean': 1.0027388334274292, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5216909050941467, 'clip_ratio/low_mean': 0.020090090576559305, 'clip_ratio/low_min': 0.020090090576559305, 'clip_ratio/high_mean': 0.06276623345911503, 'clip_ratio/high_max': 0.06276623345911503, 'clip_ratio/region_mean': 0.08285632403567433, 'reward_total_mean': 0.4235251843929291, 'reward_meter_mean': 0.32614681124687195, 'reward_meter_std': 0.27816957235336304, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9868057370185852, 'reward_repeat_soft_std': 0.01320960745215416, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.29731839895248413, 'reward_total_composite_mean': 0.4235251843929291, 'reward_total_composite_std': 0.1879330724477768, 'epoch': 0.13}

 40%|███▉      | 1317/3300 [3:16:51<5:27:55,  9.92s/it]INFO 04-13 10:37:49 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1318/3300 [3:16:57<4:50:56,  8.81s/it]2026-04-13 10:37:55,721 | INFO | train_grpo_train | metrics_logged mode=train step=1318

                                                       
{'loss': -0.0521, 'grad_norm': 12.158468246459961, 'learning_rate': 6.00909090909091e-06, 'num_tokens': 2328686.0, 'completions/mean_length': 54.375, 'completions/min_length': 36.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.4651309549808502, 'rewards/meter/std': 0.2966498136520386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9423577785491943, 'rewards/repeat_soft/std': 0.05423344671726227, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.49903154373168945, 'rewards/total_composite/std': 0.13276229798793793, 'reward': 0.49903154373168945, 'reward_std': 0.13276229798793793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12698955833911896, 'sampling/sampling_logp_difference/max': 1.449594497680664, 'sampling/importance_sampling_ratio/min': 0.23466543853282928, 'sampling/importance_sampling_ratio/mean': 0.9886088967323303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6243526786565781, 'clip_ratio/low_mean': 0.043492653872817755, 'clip_ratio/low_min': 0.043492653872817755, 'clip_ratio/high_mean': 0.08115777466446161, 'clip_ratio/high_max': 0.08115777466446161, 'clip_ratio/region_mean': 0.12465042853727937, 'reward_total_mean': 0.49903154373168945, 'reward_meter_mean': 0.4651309549808502, 'reward_meter_std': 0.2966498136520386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9423577785491943, 'reward_repeat_soft_std': 0.05423344671726227, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.49903154373168945, 'reward_total_composite_std': 0.13276229798793793, 'epoch': 0.13}

 40%|███▉      | 1318/3300 [3:16:57<4:50:56,  8.81s/it]INFO 04-13 10:37:55 [block_pool.py:378] Successfully reset prefix cache

 40%|███▉      | 1319/3300 [3:17:04<4:29:04,  8.15s/it]2026-04-13 10:38:02,352 | INFO | train_grpo_train | metrics_logged mode=train step=1319

                                                       
{'loss': -0.0504, 'grad_norm': 11.741623878479004, 'learning_rate': 6.0060606060606065e-06, 'num_tokens': 2330593.0, 'completions/mean_length': 64.375, 'completions/min_length': 57.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.11510743200778961, 'rewards/meter/std': 0.09939873963594437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9502577185630798, 'rewards/repeat_soft/std': 0.05569569766521454, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.37396299839019775, 'rewards/total_composite/std': 0.02319924347102642, 'reward': 0.37396299839019775, 'reward_std': 0.02319924719631672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17710459232330322, 'sampling/sampling_logp_difference/max': 4.657275676727295, 'sampling/importance_sampling_ratio/min': 0.009492287412285805, 'sampling/importance_sampling_ratio/mean': 0.989753246307373, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.77280592918396, 'clip_ratio/low_mean': 0.06375110056251287, 'clip_ratio/low_min': 0.06375110056251287, 'clip_ratio/high_mean': 0.05356408003717661, 'clip_ratio/high_max': 0.05356408003717661, 'clip_ratio/region_mean': 0.11731518059968948, 'reward_total_mean': 0.37396299839019775, 'reward_meter_mean': 0.11510743200778961, 'reward_meter_std': 0.09939873963594437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9502577185630798, 'reward_repeat_soft_std': 0.05569569766521454, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.37396299839019775, 'reward_total_composite_std': 0.02319924347102642, 'epoch': 0.13}

 40%|███▉      | 1319/3300 [3:17:04<4:29:04,  8.15s/it]INFO 04-13 10:38:02 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1320/3300 [3:17:12<4:24:04,  8.00s/it]2026-04-13 10:38:10,003 | INFO | train_grpo_train | metrics_logged mode=train step=1320

                                                       
{'loss': 0.5899, 'grad_norm': 10.33740234375, 'learning_rate': 6.003030303030304e-06, 'num_tokens': 2332245.0, 'completions/mean_length': 59.5, 'completions/min_length': 24.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.5, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.29521444439888, 'rewards/meter/std': 0.3094203472137451, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.45806270837783813, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9167537093162537, 'rewards/repeat_soft/std': 0.11579932272434235, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.1940544992685318, 'rewards/total_composite/mean': 0.3447820842266083, 'rewards/total_composite/std': 0.11164502054452896, 'reward': 0.3447820842266083, 'reward_std': 0.11164502054452896, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07848409563302994, 'sampling/sampling_logp_difference/max': 1.6388883590698242, 'sampling/importance_sampling_ratio/min': 0.19419580698013306, 'sampling/importance_sampling_ratio/mean': 1.0154354572296143, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.468915194272995, 'clip_ratio/low_mean': 0.0402972879819572, 'clip_ratio/low_min': 0.0402972879819572, 'clip_ratio/high_mean': 0.050527742598205805, 'clip_ratio/high_max': 0.050527742598205805, 'clip_ratio/region_mean': 0.090825030580163, 'reward_total_mean': 0.3447820842266083, 'reward_meter_mean': 0.29521444439888, 'reward_meter_std': 0.3094203472137451, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.45806270837783813, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9167537093162537, 'reward_repeat_soft_std': 0.11579932272434235, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.1940544992685318, 'reward_total_composite_mean': 0.3447820842266083, 'reward_total_composite_std': 0.11164502054452896, 'epoch': 0.13}

 40%|████      | 1320/3300 [3:17:12<4:24:04,  8.00s/it]INFO 04-13 10:38:10 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1321/3300 [3:17:18<4:07:57,  7.52s/it]2026-04-13 10:38:16,388 | INFO | train_grpo_train | metrics_logged mode=train step=1321

                                                       
{'loss': 0.0895, 'grad_norm': 21.839662551879883, 'learning_rate': 6e-06, 'num_tokens': 2333782.0, 'completions/mean_length': 16.125, 'completions/min_length': 12.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.125, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.6248440742492676, 'rewards/meter/std': 0.3955559730529785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.541622519493103, 'rewards/total_composite/std': 0.1336933970451355, 'reward': 0.541622519493103, 'reward_std': 0.1336933970451355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1574832648038864, 'sampling/sampling_logp_difference/max': 1.6570532321929932, 'sampling/importance_sampling_ratio/min': 0.19070009887218475, 'sampling/importance_sampling_ratio/mean': 0.9786161780357361, 'sampling/importance_sampling_ratio/max': 1.7580597400665283, 'entropy': 0.9691708460450172, 'clip_ratio/low_mean': 0.05115089565515518, 'clip_ratio/low_min': 0.05115089565515518, 'clip_ratio/high_mean': 0.13058911636471748, 'clip_ratio/high_max': 0.13058911636471748, 'clip_ratio/region_mean': 0.18174001201987267, 'reward_total_mean': 0.541622519493103, 'reward_meter_mean': 0.6248440742492676, 'reward_meter_std': 0.3955559730529785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.541622519493103, 'reward_total_composite_std': 0.1336933970451355, 'epoch': 0.13}

 40%|████      | 1321/3300 [3:17:18<4:07:57,  7.52s/it]INFO 04-13 10:38:16 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1322/3300 [3:17:25<4:01:44,  7.33s/it]2026-04-13 10:38:23,291 | INFO | train_grpo_train | metrics_logged mode=train step=1322

                                                       
{'loss': -0.0002, 'grad_norm': 9.119623184204102, 'learning_rate': 5.996969696969697e-06, 'num_tokens': 2336046.0, 'completions/mean_length': 87.0, 'completions/min_length': 78.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.2986004650592804, 'rewards/meter/std': 0.33540481328964233, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9160292148590088, 'rewards/repeat_soft/std': 0.022331031039357185, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720350325107574, 'rewards/total_composite/mean': 0.4448355734348297, 'rewards/total_composite/std': 0.13325370848178864, 'reward': 0.4448355734348297, 'reward_std': 0.13325372338294983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16736315190792084, 'sampling/sampling_logp_difference/max': 1.9393364191055298, 'sampling/importance_sampling_ratio/min': 0.14379934966564178, 'sampling/importance_sampling_ratio/mean': 0.9967644214630127, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7418034449219704, 'clip_ratio/low_mean': 0.08746557682752609, 'clip_ratio/low_min': 0.08746557682752609, 'clip_ratio/high_mean': 0.04391687922179699, 'clip_ratio/high_max': 0.04391687922179699, 'clip_ratio/region_mean': 0.13138245604932308, 'reward_total_mean': 0.4448355734348297, 'reward_meter_mean': 0.2986004650592804, 'reward_meter_std': 0.33540481328964233, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9160292148590088, 'reward_repeat_soft_std': 0.022331031039357185, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720350325107574, 'reward_total_composite_mean': 0.4448355734348297, 'reward_total_composite_std': 0.13325370848178864, 'epoch': 0.13}

 40%|████      | 1322/3300 [3:17:25<4:01:44,  7.33s/it]INFO 04-13 10:38:23 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1323/3300 [3:17:31<3:52:59,  7.07s/it]2026-04-13 10:38:29,755 | INFO | train_grpo_train | metrics_logged mode=train step=1323

                                                       
{'loss': -0.0499, 'grad_norm': 14.36684799194336, 'learning_rate': 5.993939393939394e-06, 'num_tokens': 2337489.0, 'completions/mean_length': 37.375, 'completions/min_length': 33.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.7412725687026978, 'rewards/meter/std': 0.3410572409629822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9899576902389526, 'rewards/repeat_soft/std': 0.009993739426136017, 'rewards/judge_quality/mean': 0.5062500238418579, 'rewards/judge_quality/std': 0.15080145001411438, 'rewards/total_composite/mean': 0.5912611484527588, 'rewards/total_composite/std': 0.12845481932163239, 'reward': 0.5912611484527588, 'reward_std': 0.12845481932163239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14535550773143768, 'sampling/sampling_logp_difference/max': 1.510312557220459, 'sampling/importance_sampling_ratio/min': 0.2208409458398819, 'sampling/importance_sampling_ratio/mean': 1.018187403678894, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8165575861930847, 'clip_ratio/low_mean': 0.035386763513088226, 'clip_ratio/low_min': 0.035386763513088226, 'clip_ratio/high_mean': 0.08853822201490402, 'clip_ratio/high_max': 0.08853822201490402, 'clip_ratio/region_mean': 0.12392498552799225, 'reward_total_mean': 0.5912611484527588, 'reward_meter_mean': 0.7412725687026978, 'reward_meter_std': 0.3410572409629822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9899576902389526, 'reward_repeat_soft_std': 0.009993739426136017, 'reward_judge_quality_mean': 0.5062500238418579, 'reward_judge_quality_std': 0.15080145001411438, 'reward_total_composite_mean': 0.5912611484527588, 'reward_total_composite_std': 0.12845481932163239, 'epoch': 0.13}

 40%|████      | 1323/3300 [3:17:31<3:52:59,  7.07s/it]INFO 04-13 10:38:30 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1324/3300 [3:17:38<3:47:03,  6.89s/it]2026-04-13 10:38:36,236 | INFO | train_grpo_train | metrics_logged mode=train step=1324

                                                       
{'loss': 0.0343, 'grad_norm': 11.929206848144531, 'learning_rate': 5.990909090909092e-06, 'num_tokens': 2339127.0, 'completions/mean_length': 37.75, 'completions/min_length': 34.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7784693241119385, 'rewards/meter/std': 0.3361956775188446, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9474538564682007, 'rewards/repeat_soft/std': 0.060473356395959854, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6314690709114075, 'rewards/total_composite/std': 0.19294258952140808, 'reward': 0.6314690709114075, 'reward_std': 0.1929425746202469, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11215405911207199, 'sampling/sampling_logp_difference/max': 1.9784116744995117, 'sampling/importance_sampling_ratio/min': 0.1382887065410614, 'sampling/importance_sampling_ratio/mean': 1.0129324197769165, 'sampling/importance_sampling_ratio/max': 1.9263604879379272, 'entropy': 0.7627578303217888, 'clip_ratio/low_mean': 0.07707995735108852, 'clip_ratio/low_min': 0.07707995735108852, 'clip_ratio/high_mean': 0.040570175275206566, 'clip_ratio/high_max': 0.040570175275206566, 'clip_ratio/region_mean': 0.11765013262629509, 'reward_total_mean': 0.6314690709114075, 'reward_meter_mean': 0.7784693241119385, 'reward_meter_std': 0.3361956775188446, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9474538564682007, 'reward_repeat_soft_std': 0.060473356395959854, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6314690709114075, 'reward_total_composite_std': 0.19294258952140808, 'epoch': 0.13}

 40%|████      | 1324/3300 [3:17:38<3:47:03,  6.89s/it]INFO 04-13 10:38:36 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1325/3300 [3:17:45<3:48:59,  6.96s/it]2026-04-13 10:38:43,355 | INFO | train_grpo_train | metrics_logged mode=train step=1325

                                                       
{'loss': 0.0459, 'grad_norm': 11.517350196838379, 'learning_rate': 5.987878787878788e-06, 'num_tokens': 2341170.0, 'completions/mean_length': 75.375, 'completions/min_length': 64.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.375, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.3564927875995636, 'rewards/meter/std': 0.31763216853141785, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.895226240158081, 'rewards/repeat_soft/std': 0.06435451656579971, 'rewards/judge_quality/mean': 0.5437500476837158, 'rewards/judge_quality/std': 0.1765492856502533, 'rewards/total_composite/mean': 0.41336947679519653, 'rewards/total_composite/std': 0.22498710453510284, 'reward': 0.41336947679519653, 'reward_std': 0.22498708963394165, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13874313235282898, 'sampling/sampling_logp_difference/max': 2.033254861831665, 'sampling/importance_sampling_ratio/min': 0.13090874254703522, 'sampling/importance_sampling_ratio/mean': 1.014022707939148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7146521620452404, 'clip_ratio/low_mean': 0.08022316731512547, 'clip_ratio/low_min': 0.08022316731512547, 'clip_ratio/high_mean': 0.05571772065013647, 'clip_ratio/high_max': 0.05571772065013647, 'clip_ratio/region_mean': 0.13594088796526194, 'reward_total_mean': 0.41336947679519653, 'reward_meter_mean': 0.3564927875995636, 'reward_meter_std': 0.31763216853141785, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.895226240158081, 'reward_repeat_soft_std': 0.06435451656579971, 'reward_judge_quality_mean': 0.5437500476837158, 'reward_judge_quality_std': 0.1765492856502533, 'reward_total_composite_mean': 0.41336947679519653, 'reward_total_composite_std': 0.22498710453510284, 'epoch': 0.13}

 40%|████      | 1325/3300 [3:17:45<3:48:59,  6.96s/it]INFO 04-13 10:38:43 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1326/3300 [3:17:51<3:42:39,  6.77s/it]2026-04-13 10:38:49,666 | INFO | train_grpo_train | metrics_logged mode=train step=1326

                                                       
{'loss': 0.0946, 'grad_norm': 14.53357219696045, 'learning_rate': 5.984848484848486e-06, 'num_tokens': 2343063.0, 'completions/mean_length': 43.625, 'completions/min_length': 38.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.567968487739563, 'rewards/meter/std': 0.41750046610832214, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9909164905548096, 'rewards/repeat_soft/std': 0.011348268948495388, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5037947297096252, 'rewards/total_composite/std': 0.11363547295331955, 'reward': 0.5037947297096252, 'reward_std': 0.11363547295331955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14599864184856415, 'sampling/sampling_logp_difference/max': 2.054823637008667, 'sampling/importance_sampling_ratio/min': 0.1281154304742813, 'sampling/importance_sampling_ratio/mean': 0.9894979000091553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8291291519999504, 'clip_ratio/low_mean': 0.046416791155934334, 'clip_ratio/low_min': 0.046416791155934334, 'clip_ratio/high_mean': 0.09215929359197617, 'clip_ratio/high_max': 0.09215929359197617, 'clip_ratio/region_mean': 0.1385760847479105, 'reward_total_mean': 0.5037947297096252, 'reward_meter_mean': 0.567968487739563, 'reward_meter_std': 0.41750046610832214, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9909164905548096, 'reward_repeat_soft_std': 0.011348268948495388, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5037947297096252, 'reward_total_composite_std': 0.11363547295331955, 'epoch': 0.13}

 40%|████      | 1326/3300 [3:17:51<3:42:39,  6.77s/it]INFO 04-13 10:38:49 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1327/3300 [3:17:57<3:36:31,  6.58s/it]2026-04-13 10:38:55,821 | INFO | train_grpo_train | metrics_logged mode=train step=1327

                                                       
{'loss': -0.0002, 'grad_norm': 16.082365036010742, 'learning_rate': 5.981818181818182e-06, 'num_tokens': 2344393.0, 'completions/mean_length': 18.25, 'completions/min_length': 17.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9682329893112183, 'rewards/meter/std': 0.044429961591959, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.7628706693649292, 'rewards/total_composite/std': 0.16083958745002747, 'reward': 0.7628706693649292, 'reward_std': 0.16083957254886627, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10630126297473907, 'sampling/sampling_logp_difference/max': 1.2814911603927612, 'sampling/importance_sampling_ratio/min': 0.2776229977607727, 'sampling/importance_sampling_ratio/mean': 0.9845349192619324, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4160885140299797, 'clip_ratio/low_mean': 0.04175266623497009, 'clip_ratio/low_min': 0.04175266623497009, 'clip_ratio/high_mean': 0.045573493000119925, 'clip_ratio/high_max': 0.045573493000119925, 'clip_ratio/region_mean': 0.08732615923509002, 'reward_total_mean': 0.7628706693649292, 'reward_meter_mean': 0.9682329893112183, 'reward_meter_std': 0.044429961591959, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.7628706693649292, 'reward_total_composite_std': 0.16083958745002747, 'epoch': 0.13}

 40%|████      | 1327/3300 [3:17:58<3:36:31,  6.58s/it]INFO 04-13 10:38:56 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1328/3300 [3:18:04<3:36:44,  6.59s/it]2026-04-13 10:39:02,450 | INFO | train_grpo_train | metrics_logged mode=train step=1328

                                                       
{'loss': 0.0434, 'grad_norm': 11.687251091003418, 'learning_rate': 5.978787878787879e-06, 'num_tokens': 2346047.0, 'completions/mean_length': 46.75, 'completions/min_length': 43.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6765961647033691, 'rewards/meter/std': 0.35124412178993225, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.989249587059021, 'rewards/repeat_soft/std': 0.014732363633811474, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.5615547299385071, 'rewards/total_composite/std': 0.13680243492126465, 'reward': 0.5615547299385071, 'reward_std': 0.13680243492126465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1515999138355255, 'sampling/sampling_logp_difference/max': 1.685828685760498, 'sampling/importance_sampling_ratio/min': 0.18529081344604492, 'sampling/importance_sampling_ratio/mean': 1.0011440515518188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.849358819425106, 'clip_ratio/low_mean': 0.04327042028307915, 'clip_ratio/low_min': 0.04327042028307915, 'clip_ratio/high_mean': 0.06864966731518507, 'clip_ratio/high_max': 0.06864966731518507, 'clip_ratio/region_mean': 0.11192008759826422, 'reward_total_mean': 0.5615547299385071, 'reward_meter_mean': 0.6765961647033691, 'reward_meter_std': 0.35124412178993225, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.989249587059021, 'reward_repeat_soft_std': 0.014732363633811474, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.5615547299385071, 'reward_total_composite_std': 0.13680243492126465, 'epoch': 0.13}

 40%|████      | 1328/3300 [3:18:04<3:36:44,  6.59s/it]INFO 04-13 10:39:02 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1329/3300 [3:18:11<3:35:25,  6.56s/it]2026-04-13 10:39:08,913 | INFO | train_grpo_train | metrics_logged mode=train step=1329

                                                       
{'loss': -0.0067, 'grad_norm': 15.916736602783203, 'learning_rate': 5.975757575757576e-06, 'num_tokens': 2347335.0, 'completions/mean_length': 23.0, 'completions/min_length': 19.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.6458486318588257, 'rewards/meter/std': 0.4809549152851105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9220465421676636, 'rewards/repeat_soft/std': 0.06529001891613007, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5218911170959473, 'rewards/total_composite/std': 0.14064864814281464, 'reward': 0.5218911170959473, 'reward_std': 0.14064864814281464, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1282486766576767, 'sampling/sampling_logp_difference/max': 1.4058802127838135, 'sampling/importance_sampling_ratio/min': 0.24515117704868317, 'sampling/importance_sampling_ratio/mean': 0.9940204620361328, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8468829244375229, 'clip_ratio/low_mean': 0.02794258389621973, 'clip_ratio/low_min': 0.02794258389621973, 'clip_ratio/high_mean': 0.1299373060464859, 'clip_ratio/high_max': 0.1299373060464859, 'clip_ratio/region_mean': 0.15787988994270563, 'reward_total_mean': 0.5218911170959473, 'reward_meter_mean': 0.6458486318588257, 'reward_meter_std': 0.4809549152851105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9220465421676636, 'reward_repeat_soft_std': 0.06529001891613007, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5218911170959473, 'reward_total_composite_std': 0.14064864814281464, 'epoch': 0.13}

 40%|████      | 1329/3300 [3:18:11<3:35:25,  6.56s/it]INFO 04-13 10:39:09 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1330/3300 [3:18:22<4:24:39,  8.06s/it]2026-04-13 10:39:20,481 | INFO | train_grpo_train | metrics_logged mode=train step=1330

                                                       
{'loss': -0.0218, 'grad_norm': 4.435325622558594, 'learning_rate': 5.972727272727274e-06, 'num_tokens': 2348683.0, 'completions/mean_length': 86.5, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.71428680419922, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.47142189741134644, 'rewards/meter/std': 0.3101181983947754, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9667688608169556, 'rewards/repeat_soft/std': 0.011865193955600262, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.14029940962791443, 'rewards/total_composite/mean': 0.40942153334617615, 'rewards/total_composite/std': 0.20781448483467102, 'reward': 0.40942153334617615, 'reward_std': 0.20781448483467102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13891050219535828, 'sampling/sampling_logp_difference/max': 0.9984486103057861, 'sampling/importance_sampling_ratio/min': 0.3684506118297577, 'sampling/importance_sampling_ratio/mean': 1.0111424922943115, 'sampling/importance_sampling_ratio/max': 1.5134029388427734, 'entropy': 0.8866551592946053, 'clip_ratio/low_mean': 0.04548611119389534, 'clip_ratio/low_min': 0.04548611119389534, 'clip_ratio/high_mean': 0.1330641247332096, 'clip_ratio/high_max': 0.1330641247332096, 'clip_ratio/region_mean': 0.17855023592710495, 'reward_total_mean': 0.40942153334617615, 'reward_meter_mean': 0.47142189741134644, 'reward_meter_std': 0.3101181983947754, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9667688608169556, 'reward_repeat_soft_std': 0.011865193955600262, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.14029940962791443, 'reward_total_composite_mean': 0.40942153334617615, 'reward_total_composite_std': 0.20781448483467102, 'epoch': 0.13}

 40%|████      | 1330/3300 [3:18:22<4:24:39,  8.06s/it]INFO 04-13 10:39:20 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1331/3300 [3:18:29<4:10:55,  7.65s/it]2026-04-13 10:39:27,159 | INFO | train_grpo_train | metrics_logged mode=train step=1331

                                                       
{'loss': 0.0097, 'grad_norm': 16.48143768310547, 'learning_rate': 5.96969696969697e-06, 'num_tokens': 2350438.0, 'completions/mean_length': 63.375, 'completions/min_length': 50.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.375, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.40556466579437256, 'rewards/meter/std': 0.26941341161727905, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9321507215499878, 'rewards/repeat_soft/std': 0.04556303098797798, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4255417585372925, 'rewards/total_composite/std': 0.07310432940721512, 'reward': 0.4255417585372925, 'reward_std': 0.07310432195663452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.150555819272995, 'sampling/sampling_logp_difference/max': 1.848578929901123, 'sampling/importance_sampling_ratio/min': 0.1574607789516449, 'sampling/importance_sampling_ratio/mean': 1.005539894104004, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6611043587327003, 'clip_ratio/low_mean': 0.052912051789462566, 'clip_ratio/low_min': 0.052912051789462566, 'clip_ratio/high_mean': 0.06353189051151276, 'clip_ratio/high_max': 0.06353189051151276, 'clip_ratio/region_mean': 0.11644394230097532, 'reward_total_mean': 0.4255417585372925, 'reward_meter_mean': 0.40556466579437256, 'reward_meter_std': 0.26941341161727905, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9321507215499878, 'reward_repeat_soft_std': 0.04556303098797798, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4255417585372925, 'reward_total_composite_std': 0.07310432940721512, 'epoch': 0.13}

 40%|████      | 1331/3300 [3:18:29<4:10:55,  7.65s/it]INFO 04-13 10:39:27 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1332/3300 [3:18:35<3:59:05,  7.29s/it]2026-04-13 10:39:33,611 | INFO | train_grpo_train | metrics_logged mode=train step=1332

                                                       
{'loss': -0.0754, 'grad_norm': 17.2110652923584, 'learning_rate': 5.966666666666667e-06, 'num_tokens': 2352129.0, 'completions/mean_length': 35.375, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.314901202917099, 'rewards/meter/std': 0.3396137058734894, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8801974058151245, 'rewards/repeat_soft/std': 0.0764533057808876, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.4472670555114746, 'rewards/total_composite/std': 0.14360350370407104, 'reward': 0.4472670555114746, 'reward_std': 0.14360350370407104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1391085535287857, 'sampling/sampling_logp_difference/max': 2.2771387100219727, 'sampling/importance_sampling_ratio/min': 0.10257729142904282, 'sampling/importance_sampling_ratio/mean': 1.01146399974823, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7420665435492992, 'clip_ratio/low_mean': 0.0835973508656025, 'clip_ratio/low_min': 0.0835973508656025, 'clip_ratio/high_mean': 0.0518292672932148, 'clip_ratio/high_max': 0.0518292672932148, 'clip_ratio/region_mean': 0.1354266181588173, 'reward_total_mean': 0.4472670555114746, 'reward_meter_mean': 0.314901202917099, 'reward_meter_std': 0.3396137058734894, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8801974058151245, 'reward_repeat_soft_std': 0.0764533057808876, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.4472670555114746, 'reward_total_composite_std': 0.14360350370407104, 'epoch': 0.13}

 40%|████      | 1332/3300 [3:18:35<3:59:05,  7.29s/it]INFO 04-13 10:39:33 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1333/3300 [3:18:47<4:39:18,  8.52s/it]2026-04-13 10:39:45,005 | INFO | train_grpo_train | metrics_logged mode=train step=1333

                                                       
{'loss': -0.1558, 'grad_norm': 3.7890119552612305, 'learning_rate': 5.963636363636364e-06, 'num_tokens': 2354379.0, 'completions/mean_length': 152.25, 'completions/min_length': 66.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 100.85714721679688, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.2900828421115875, 'rewards/meter/std': 0.2821356952190399, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9511439800262451, 'rewards/repeat_soft/std': 0.051468927413225174, 'rewards/judge_quality/mean': 0.4049999713897705, 'rewards/judge_quality/std': 0.19442224502563477, 'rewards/total_composite/mean': 0.37064436078071594, 'rewards/total_composite/std': 0.18591031432151794, 'reward': 0.37064436078071594, 'reward_std': 0.18591031432151794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1486016809940338, 'sampling/sampling_logp_difference/max': 1.7037146091461182, 'sampling/importance_sampling_ratio/min': 0.18200618028640747, 'sampling/importance_sampling_ratio/mean': 1.001570701599121, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7631268575787544, 'clip_ratio/low_mean': 0.07168418262153864, 'clip_ratio/low_min': 0.07168418262153864, 'clip_ratio/high_mean': 0.061170581728219986, 'clip_ratio/high_max': 0.061170581728219986, 'clip_ratio/region_mean': 0.13285476434975863, 'reward_total_mean': 0.37064436078071594, 'reward_meter_mean': 0.2900828421115875, 'reward_meter_std': 0.2821356952190399, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9511439800262451, 'reward_repeat_soft_std': 0.051468927413225174, 'reward_judge_quality_mean': 0.4049999713897705, 'reward_judge_quality_std': 0.19442224502563477, 'reward_total_composite_mean': 0.37064436078071594, 'reward_total_composite_std': 0.18591031432151794, 'epoch': 0.13}

 40%|████      | 1333/3300 [3:18:47<4:39:18,  8.52s/it]INFO 04-13 10:39:45 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1334/3300 [3:18:54<4:24:27,  8.07s/it]2026-04-13 10:39:52,033 | INFO | train_grpo_train | metrics_logged mode=train step=1334

                                                       
{'loss': 0.0384, 'grad_norm': 10.282875061035156, 'learning_rate': 5.960606060606061e-06, 'num_tokens': 2356568.0, 'completions/mean_length': 80.625, 'completions/min_length': 72.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.625, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6107769012451172, 'rewards/meter/std': 0.34543493390083313, 'rewards/count_adherence/mean': 0.8958333134651184, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8274047374725342, 'rewards/repeat_soft/std': 0.05298303812742233, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4572588801383972, 'rewards/total_composite/std': 0.09525513648986816, 'reward': 0.4572588801383972, 'reward_std': 0.09525512158870697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11038250476121902, 'sampling/sampling_logp_difference/max': 1.768733024597168, 'sampling/importance_sampling_ratio/min': 0.1705489307641983, 'sampling/importance_sampling_ratio/mean': 1.0168070793151855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5664704628288746, 'clip_ratio/low_mean': 0.06878686603158712, 'clip_ratio/low_min': 0.06878686603158712, 'clip_ratio/high_mean': 0.02347811870276928, 'clip_ratio/high_max': 0.02347811870276928, 'clip_ratio/region_mean': 0.0922649847343564, 'reward_total_mean': 0.4572588801383972, 'reward_meter_mean': 0.6107769012451172, 'reward_meter_std': 0.34543493390083313, 'reward_count_adherence_mean': 0.8958333134651184, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8274047374725342, 'reward_repeat_soft_std': 0.05298303812742233, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4572588801383972, 'reward_total_composite_std': 0.09525513648986816, 'epoch': 0.13}

 40%|████      | 1334/3300 [3:18:54<4:24:27,  8.07s/it]INFO 04-13 10:39:52 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1335/3300 [3:19:05<4:57:39,  9.09s/it]2026-04-13 10:40:03,492 | INFO | train_grpo_train | metrics_logged mode=train step=1335

                                                       
{'loss': -0.0998, 'grad_norm': 5.83372163772583, 'learning_rate': 5.9575757575757575e-06, 'num_tokens': 2358608.0, 'completions/mean_length': 126.0, 'completions/min_length': 59.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 70.85714721679688, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.3127872347831726, 'rewards/meter/std': 0.3247506320476532, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.967047393321991, 'rewards/repeat_soft/std': 0.022363485768437386, 'rewards/judge_quality/mean': 0.5074999928474426, 'rewards/judge_quality/std': 0.2748376131057739, 'rewards/total_composite/mean': 0.42580604553222656, 'rewards/total_composite/std': 0.24114061892032623, 'reward': 0.42580604553222656, 'reward_std': 0.24114063382148743, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15268826484680176, 'sampling/sampling_logp_difference/max': 2.7636923789978027, 'sampling/importance_sampling_ratio/min': 0.06305850297212601, 'sampling/importance_sampling_ratio/mean': 1.010711669921875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5862419381737709, 'clip_ratio/low_mean': 0.07036514859646559, 'clip_ratio/low_min': 0.07036514859646559, 'clip_ratio/high_mean': 0.049692231230437756, 'clip_ratio/high_max': 0.049692231230437756, 'clip_ratio/region_mean': 0.12005737982690334, 'reward_total_mean': 0.42580604553222656, 'reward_meter_mean': 0.3127872347831726, 'reward_meter_std': 0.3247506320476532, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.967047393321991, 'reward_repeat_soft_std': 0.022363485768437386, 'reward_judge_quality_mean': 0.5074999928474426, 'reward_judge_quality_std': 0.2748376131057739, 'reward_total_composite_mean': 0.42580604553222656, 'reward_total_composite_std': 0.24114061892032623, 'epoch': 0.13}

 40%|████      | 1335/3300 [3:19:05<4:57:39,  9.09s/it]INFO 04-13 10:40:03 [block_pool.py:378] Successfully reset prefix cache

 40%|████      | 1336/3300 [3:19:12<4:35:50,  8.43s/it]2026-04-13 10:40:10,376 | INFO | train_grpo_train | metrics_logged mode=train step=1336

                                                       
{'loss': 0.0854, 'grad_norm': 9.80910873413086, 'learning_rate': 5.954545454545455e-06, 'num_tokens': 2360746.0, 'completions/mean_length': 83.25, 'completions/min_length': 67.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.8006072640419006, 'rewards/meter/std': 0.3069312870502472, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9120126962661743, 'rewards/repeat_soft/std': 0.04998166114091873, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5647725462913513, 'rewards/total_composite/std': 0.11622896045446396, 'reward': 0.5647725462913513, 'reward_std': 0.11622896045446396, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13410262763500214, 'sampling/sampling_logp_difference/max': 2.931978702545166, 'sampling/importance_sampling_ratio/min': 0.05329148471355438, 'sampling/importance_sampling_ratio/mean': 1.0088557004928589, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7084437683224678, 'clip_ratio/low_mean': 0.046615186147391796, 'clip_ratio/low_min': 0.046615186147391796, 'clip_ratio/high_mean': 0.08461390342563391, 'clip_ratio/high_max': 0.08461390342563391, 'clip_ratio/region_mean': 0.1312290895730257, 'reward_total_mean': 0.5647725462913513, 'reward_meter_mean': 0.8006072640419006, 'reward_meter_std': 0.3069312870502472, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9120126962661743, 'reward_repeat_soft_std': 0.04998166114091873, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5647725462913513, 'reward_total_composite_std': 0.11622896045446396, 'epoch': 0.13}

 40%|████      | 1336/3300 [3:19:12<4:35:50,  8.43s/it]INFO 04-13 10:40:10 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1337/3300 [3:19:18<4:14:04,  7.77s/it]2026-04-13 10:40:16,595 | INFO | train_grpo_train | metrics_logged mode=train step=1337

                                                       
{'loss': 0.0476, 'grad_norm': 15.51915454864502, 'learning_rate': 5.951515151515151e-06, 'num_tokens': 2362513.0, 'completions/mean_length': 44.875, 'completions/min_length': 33.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.710849404335022, 'rewards/meter/std': 0.33453720808029175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9896112680435181, 'rewards/repeat_soft/std': 0.012233623303472996, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5692287683486938, 'rewards/total_composite/std': 0.0953088030219078, 'reward': 0.5692287683486938, 'reward_std': 0.09530878812074661, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1775110512971878, 'sampling/sampling_logp_difference/max': 1.4458134174346924, 'sampling/importance_sampling_ratio/min': 0.23555439710617065, 'sampling/importance_sampling_ratio/mean': 0.9923685193061829, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1398600190877914, 'clip_ratio/low_mean': 0.025348837487399578, 'clip_ratio/low_min': 0.025348837487399578, 'clip_ratio/high_mean': 0.1504751518368721, 'clip_ratio/high_max': 0.1504751518368721, 'clip_ratio/region_mean': 0.17582398932427168, 'reward_total_mean': 0.5692287683486938, 'reward_meter_mean': 0.710849404335022, 'reward_meter_std': 0.33453720808029175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9896112680435181, 'reward_repeat_soft_std': 0.012233623303472996, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5692287683486938, 'reward_total_composite_std': 0.0953088030219078, 'epoch': 0.13}

 41%|████      | 1337/3300 [3:19:18<4:14:04,  7.77s/it]INFO 04-13 10:40:16 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1338/3300 [3:19:24<3:57:30,  7.26s/it]2026-04-13 10:40:22,689 | INFO | train_grpo_train | metrics_logged mode=train step=1338

                                                       
{'loss': 0.0401, 'grad_norm': 15.774927139282227, 'learning_rate': 5.948484848484849e-06, 'num_tokens': 2364045.0, 'completions/mean_length': 37.5, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8714721202850342, 'rewards/meter/std': 0.30579236149787903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.878113865852356, 'rewards/repeat_soft/std': 0.045858755707740784, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6930241584777832, 'rewards/total_composite/std': 0.19361813366413116, 'reward': 0.6930241584777832, 'reward_std': 0.19361810386180878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12548218667507172, 'sampling/sampling_logp_difference/max': 1.1062960624694824, 'sampling/importance_sampling_ratio/min': 0.40296289324760437, 'sampling/importance_sampling_ratio/mean': 1.041352391242981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8229580894112587, 'clip_ratio/low_mean': 0.0958860320970416, 'clip_ratio/low_min': 0.0958860320970416, 'clip_ratio/high_mean': 0.054523564875125885, 'clip_ratio/high_max': 0.054523564875125885, 'clip_ratio/region_mean': 0.1504095969721675, 'reward_total_mean': 0.6930241584777832, 'reward_meter_mean': 0.8714721202850342, 'reward_meter_std': 0.30579236149787903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.878113865852356, 'reward_repeat_soft_std': 0.045858755707740784, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6930241584777832, 'reward_total_composite_std': 0.19361813366413116, 'epoch': 0.13}

 41%|████      | 1338/3300 [3:19:24<3:57:30,  7.26s/it]INFO 04-13 10:40:22 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1339/3300 [3:19:31<3:48:22,  6.99s/it]2026-04-13 10:40:29,035 | INFO | train_grpo_train | metrics_logged mode=train step=1339

                                                       
{'loss': 0.0041, 'grad_norm': 14.936269760131836, 'learning_rate': 5.9454545454545465e-06, 'num_tokens': 2365582.0, 'completions/mean_length': 39.125, 'completions/min_length': 35.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7665152549743652, 'rewards/meter/std': 0.3787933588027954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9921016097068787, 'rewards/repeat_soft/std': 0.007253531366586685, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/total_composite/mean': 0.6462671756744385, 'rewards/total_composite/std': 0.19861456751823425, 'reward': 0.6462671756744385, 'reward_std': 0.19861455261707306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14262601733207703, 'sampling/sampling_logp_difference/max': 1.219222068786621, 'sampling/importance_sampling_ratio/min': 0.31959930062294006, 'sampling/importance_sampling_ratio/mean': 1.0433692932128906, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9027925133705139, 'clip_ratio/low_mean': 0.08850973658263683, 'clip_ratio/low_min': 0.08850973658263683, 'clip_ratio/high_mean': 0.059057971462607384, 'clip_ratio/high_max': 0.059057971462607384, 'clip_ratio/region_mean': 0.14756770804524422, 'reward_total_mean': 0.6462671756744385, 'reward_meter_mean': 0.7665152549743652, 'reward_meter_std': 0.3787933588027954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9921016097068787, 'reward_repeat_soft_std': 0.007253531366586685, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.6462671756744385, 'reward_total_composite_std': 0.19861456751823425, 'epoch': 0.13}

 41%|████      | 1339/3300 [3:19:31<3:48:22,  6.99s/it]INFO 04-13 10:40:29 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1340/3300 [3:19:39<4:01:05,  7.38s/it]2026-04-13 10:40:37,339 | INFO | train_grpo_train | metrics_logged mode=train step=1340

                                                       
{'loss': 0.0193, 'grad_norm': 7.9399895668029785, 'learning_rate': 5.942424242424243e-06, 'num_tokens': 2368188.0, 'completions/mean_length': 117.75, 'completions/min_length': 104.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.75, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.5848919153213501, 'rewards/meter/std': 0.31664523482322693, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8285818099975586, 'rewards/repeat_soft/std': 0.058063607662916183, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.45062944293022156, 'rewards/total_composite/std': 0.08520717173814774, 'reward': 0.45062944293022156, 'reward_std': 0.08520715683698654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13698212802410126, 'sampling/sampling_logp_difference/max': 2.5663671493530273, 'sampling/importance_sampling_ratio/min': 0.07681409269571304, 'sampling/importance_sampling_ratio/mean': 0.9988870620727539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7696087509393692, 'clip_ratio/low_mean': 0.05009881127625704, 'clip_ratio/low_min': 0.05009881127625704, 'clip_ratio/high_mean': 0.0640478739514947, 'clip_ratio/high_max': 0.0640478739514947, 'clip_ratio/region_mean': 0.11414668522775173, 'reward_total_mean': 0.45062944293022156, 'reward_meter_mean': 0.5848919153213501, 'reward_meter_std': 0.31664523482322693, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8285818099975586, 'reward_repeat_soft_std': 0.058063607662916183, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.45062944293022156, 'reward_total_composite_std': 0.08520717173814774, 'epoch': 0.13}

 41%|████      | 1340/3300 [3:19:39<4:01:05,  7.38s/it]INFO 04-13 10:40:37 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1341/3300 [3:19:45<3:47:10,  6.96s/it]2026-04-13 10:40:43,299 | INFO | train_grpo_train | metrics_logged mode=train step=1341

                                                       
{'loss': 0.0377, 'grad_norm': 18.52684211730957, 'learning_rate': 5.93939393939394e-06, 'num_tokens': 2369644.0, 'completions/mean_length': 25.0, 'completions/min_length': 22.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6021800637245178, 'rewards/meter/std': 0.4963541626930237, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9615821838378906, 'rewards/repeat_soft/std': 0.002596014179289341, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5509586930274963, 'rewards/total_composite/std': 0.20226746797561646, 'reward': 0.5509586930274963, 'reward_std': 0.20226743817329407, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1375489979982376, 'sampling/sampling_logp_difference/max': 1.6032088994979858, 'sampling/importance_sampling_ratio/min': 0.20124968886375427, 'sampling/importance_sampling_ratio/mean': 1.0113073587417603, 'sampling/importance_sampling_ratio/max': 1.8520575761795044, 'entropy': 0.9718766510486603, 'clip_ratio/low_mean': 0.04949534125626087, 'clip_ratio/low_min': 0.04949534125626087, 'clip_ratio/high_mean': 0.06608553603291512, 'clip_ratio/high_max': 0.06608553603291512, 'clip_ratio/region_mean': 0.11558087728917599, 'reward_total_mean': 0.5509586930274963, 'reward_meter_mean': 0.6021800637245178, 'reward_meter_std': 0.4963541626930237, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9615821838378906, 'reward_repeat_soft_std': 0.002596014179289341, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5509586930274963, 'reward_total_composite_std': 0.20226746797561646, 'epoch': 0.13}

 41%|████      | 1341/3300 [3:19:45<3:47:10,  6.96s/it]INFO 04-13 10:40:43 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1342/3300 [3:19:51<3:40:06,  6.74s/it]2026-04-13 10:40:49,551 | INFO | train_grpo_train | metrics_logged mode=train step=1342

                                                       
{'loss': -0.0059, 'grad_norm': 15.406861305236816, 'learning_rate': 5.936363636363637e-06, 'num_tokens': 2371339.0, 'completions/mean_length': 44.875, 'completions/min_length': 39.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8774193525314331, 'rewards/meter/std': 0.27800363302230835, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8659125566482544, 'rewards/repeat_soft/std': 0.05217598378658295, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.11310552060604095, 'rewards/total_composite/mean': 0.5258831977844238, 'rewards/total_composite/std': 0.24142670631408691, 'reward': 0.5258831977844238, 'reward_std': 0.24142670631408691, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12861864268779755, 'sampling/sampling_logp_difference/max': 1.7518625259399414, 'sampling/importance_sampling_ratio/min': 0.17345057427883148, 'sampling/importance_sampling_ratio/mean': 1.0132513046264648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8706328645348549, 'clip_ratio/low_mean': 0.019702842459082603, 'clip_ratio/low_min': 0.019702842459082603, 'clip_ratio/high_mean': 0.10141318663954735, 'clip_ratio/high_max': 0.10141318663954735, 'clip_ratio/region_mean': 0.12111602909862995, 'reward_total_mean': 0.5258831977844238, 'reward_meter_mean': 0.8774193525314331, 'reward_meter_std': 0.27800363302230835, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8659125566482544, 'reward_repeat_soft_std': 0.05217598378658295, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.11310552060604095, 'reward_total_composite_mean': 0.5258831977844238, 'reward_total_composite_std': 0.24142670631408691, 'epoch': 0.13}

 41%|████      | 1342/3300 [3:19:51<3:40:06,  6.74s/it]INFO 04-13 10:40:49 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1343/3300 [3:19:57<3:34:23,  6.57s/it]2026-04-13 10:40:55,715 | INFO | train_grpo_train | metrics_logged mode=train step=1343

                                                       
{'loss': 0.0864, 'grad_norm': 14.79797649383545, 'learning_rate': 5.933333333333335e-06, 'num_tokens': 2373021.0, 'completions/mean_length': 40.25, 'completions/min_length': 34.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8833025693893433, 'rewards/meter/std': 0.2559974193572998, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8911500573158264, 'rewards/repeat_soft/std': 0.05381282418966293, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5612891912460327, 'rewards/total_composite/std': 0.06968934088945389, 'reward': 0.5612891912460327, 'reward_std': 0.06968934834003448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14629581570625305, 'sampling/sampling_logp_difference/max': 1.823883056640625, 'sampling/importance_sampling_ratio/min': 0.1613978147506714, 'sampling/importance_sampling_ratio/mean': 1.02433443069458, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7586573287844658, 'clip_ratio/low_mean': 0.038345410488545895, 'clip_ratio/low_min': 0.038345410488545895, 'clip_ratio/high_mean': 0.09234451642259955, 'clip_ratio/high_max': 0.09234451642259955, 'clip_ratio/region_mean': 0.13068992691114545, 'reward_total_mean': 0.5612891912460327, 'reward_meter_mean': 0.8833025693893433, 'reward_meter_std': 0.2559974193572998, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8911500573158264, 'reward_repeat_soft_std': 0.05381282418966293, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5612891912460327, 'reward_total_composite_std': 0.06968934088945389, 'epoch': 0.13}

 41%|████      | 1343/3300 [3:19:57<3:34:23,  6.57s/it]INFO 04-13 10:40:55 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1344/3300 [3:20:03<3:28:59,  6.41s/it]2026-04-13 10:41:01,760 | INFO | train_grpo_train | metrics_logged mode=train step=1344

                                                       
{'loss': 0.0742, 'grad_norm': 11.261421203613281, 'learning_rate': 5.93030303030303e-06, 'num_tokens': 2374570.0, 'completions/mean_length': 41.625, 'completions/min_length': 38.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8662797808647156, 'rewards/meter/std': 0.17917276918888092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9751627445220947, 'rewards/repeat_soft/std': 0.04239398241043091, 'rewards/judge_quality/mean': 0.5849999785423279, 'rewards/judge_quality/std': 0.21967509388923645, 'rewards/total_composite/mean': 0.6761919260025024, 'rewards/total_composite/std': 0.1401374489068985, 'reward': 0.6761919260025024, 'reward_std': 0.1401374489068985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13689248263835907, 'sampling/sampling_logp_difference/max': 2.1067726612091064, 'sampling/importance_sampling_ratio/min': 0.12162987142801285, 'sampling/importance_sampling_ratio/mean': 1.0012785196304321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7067522257566452, 'clip_ratio/low_mean': 0.07602102914825082, 'clip_ratio/low_min': 0.07602102914825082, 'clip_ratio/high_mean': 0.05430820258334279, 'clip_ratio/high_max': 0.05430820258334279, 'clip_ratio/region_mean': 0.1303292317315936, 'reward_total_mean': 0.6761919260025024, 'reward_meter_mean': 0.8662797808647156, 'reward_meter_std': 0.17917276918888092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9751627445220947, 'reward_repeat_soft_std': 0.04239398241043091, 'reward_judge_quality_mean': 0.5849999785423279, 'reward_judge_quality_std': 0.21967509388923645, 'reward_total_composite_mean': 0.6761919260025024, 'reward_total_composite_std': 0.1401374489068985, 'epoch': 0.14}

 41%|████      | 1344/3300 [3:20:03<3:28:59,  6.41s/it]INFO 04-13 10:41:02 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1345/3300 [3:20:10<3:29:12,  6.42s/it]2026-04-13 10:41:08,194 | INFO | train_grpo_train | metrics_logged mode=train step=1345

                                                       
{'loss': 0.0363, 'grad_norm': 9.135940551757812, 'learning_rate': 5.927272727272728e-06, 'num_tokens': 2376381.0, 'completions/mean_length': 55.375, 'completions/min_length': 49.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.375, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9781638383865356, 'rewards/meter/std': 0.006795698311179876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7677721977233887, 'rewards/repeat_soft/std': 0.07447344064712524, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5822045803070068, 'rewards/total_composite/std': 0.010020908899605274, 'reward': 0.5822045803070068, 'reward_std': 0.010020908899605274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08400870859622955, 'sampling/sampling_logp_difference/max': 1.305046558380127, 'sampling/importance_sampling_ratio/min': 0.27115991711616516, 'sampling/importance_sampling_ratio/mean': 1.0251963138580322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49729280546307564, 'clip_ratio/low_mean': 0.04706370458006859, 'clip_ratio/low_min': 0.04706370458006859, 'clip_ratio/high_mean': 0.035152471624314785, 'clip_ratio/high_max': 0.035152471624314785, 'clip_ratio/region_mean': 0.08221617620438337, 'reward_total_mean': 0.5822045803070068, 'reward_meter_mean': 0.9781638383865356, 'reward_meter_std': 0.006795698311179876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7677721977233887, 'reward_repeat_soft_std': 0.07447344064712524, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5822045803070068, 'reward_total_composite_std': 0.010020908899605274, 'epoch': 0.14}

 41%|████      | 1345/3300 [3:20:10<3:29:12,  6.42s/it]INFO 04-13 10:41:08 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1346/3300 [3:20:16<3:30:58,  6.48s/it]2026-04-13 10:41:14,802 | INFO | train_grpo_train | metrics_logged mode=train step=1346

                                                       
{'loss': 0.0451, 'grad_norm': 15.326791763305664, 'learning_rate': 5.924242424242425e-06, 'num_tokens': 2378465.0, 'completions/mean_length': 80.5, 'completions/min_length': 75.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.5, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.8029333353042603, 'rewards/meter/std': 0.30681586265563965, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8098430633544922, 'rewards/repeat_soft/std': 0.06636009365320206, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5424083471298218, 'rewards/total_composite/std': 0.07780204713344574, 'reward': 0.5424083471298218, 'reward_std': 0.07780203223228455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10905911773443222, 'sampling/sampling_logp_difference/max': 1.425297737121582, 'sampling/importance_sampling_ratio/min': 0.2404368668794632, 'sampling/importance_sampling_ratio/mean': 1.0072455406188965, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6222714483737946, 'clip_ratio/low_mean': 0.03178161010146141, 'clip_ratio/low_min': 0.03178161010146141, 'clip_ratio/high_mean': 0.06335204932838678, 'clip_ratio/high_max': 0.06335204932838678, 'clip_ratio/region_mean': 0.0951336594298482, 'reward_total_mean': 0.5424083471298218, 'reward_meter_mean': 0.8029333353042603, 'reward_meter_std': 0.30681586265563965, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8098430633544922, 'reward_repeat_soft_std': 0.06636009365320206, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5424083471298218, 'reward_total_composite_std': 0.07780204713344574, 'epoch': 0.14}

 41%|████      | 1346/3300 [3:20:17<3:30:58,  6.48s/it]INFO 04-13 10:41:15 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1347/3300 [3:20:23<3:27:50,  6.39s/it]2026-04-13 10:41:20,972 | INFO | train_grpo_train | metrics_logged mode=train step=1347

                                                       
{'loss': 0.0868, 'grad_norm': 13.039422035217285, 'learning_rate': 5.921212121212122e-06, 'num_tokens': 2380197.0, 'completions/mean_length': 44.5, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8421345353126526, 'rewards/meter/std': 0.22694624960422516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8642581701278687, 'rewards/repeat_soft/std': 0.07040487229824066, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.6084120869636536, 'rewards/total_composite/std': 0.14150801301002502, 'reward': 0.6084120869636536, 'reward_std': 0.14150801301002502, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13987328112125397, 'sampling/sampling_logp_difference/max': 1.5131187438964844, 'sampling/importance_sampling_ratio/min': 0.2202221006155014, 'sampling/importance_sampling_ratio/mean': 0.9914326071739197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7605642080307007, 'clip_ratio/low_mean': 0.05355572560802102, 'clip_ratio/low_min': 0.05355572560802102, 'clip_ratio/high_mean': 0.06820268929004669, 'clip_ratio/high_max': 0.06820268929004669, 'clip_ratio/region_mean': 0.12175841489806771, 'reward_total_mean': 0.6084120869636536, 'reward_meter_mean': 0.8421345353126526, 'reward_meter_std': 0.22694624960422516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8642581701278687, 'reward_repeat_soft_std': 0.07040487229824066, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.6084120869636536, 'reward_total_composite_std': 0.14150801301002502, 'epoch': 0.14}

 41%|████      | 1347/3300 [3:20:23<3:27:50,  6.39s/it]INFO 04-13 10:41:21 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1348/3300 [3:20:28<3:21:02,  6.18s/it]2026-04-13 10:41:26,677 | INFO | train_grpo_train | metrics_logged mode=train step=1348

                                                       
{'loss': -0.001, 'grad_norm': 20.71615219116211, 'learning_rate': 5.9181818181818184e-06, 'num_tokens': 2381548.0, 'completions/mean_length': 18.875, 'completions/min_length': 14.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9922042489051819, 'rewards/meter/std': 0.0061484333127737045, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.915888786315918, 'rewards/repeat_soft/std': 0.06378740072250366, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.6042660474777222, 'rewards/total_composite/std': 0.04260183125734329, 'reward': 0.6042660474777222, 'reward_std': 0.04260184243321419, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13716576993465424, 'sampling/sampling_logp_difference/max': 0.8954238891601562, 'sampling/importance_sampling_ratio/min': 0.4084344208240509, 'sampling/importance_sampling_ratio/mean': 1.0150609016418457, 'sampling/importance_sampling_ratio/max': 1.8645116090774536, 'entropy': 0.9576961472630501, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0702574853785336, 'clip_ratio/high_max': 0.0702574853785336, 'clip_ratio/region_mean': 0.0702574853785336, 'reward_total_mean': 0.6042660474777222, 'reward_meter_mean': 0.9922042489051819, 'reward_meter_std': 0.0061484333127737045, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.915888786315918, 'reward_repeat_soft_std': 0.06378740072250366, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.6042660474777222, 'reward_total_composite_std': 0.04260183125734329, 'epoch': 0.14}

 41%|████      | 1348/3300 [3:20:28<3:21:02,  6.18s/it]INFO 04-13 10:41:26 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1349/3300 [3:20:35<3:23:01,  6.24s/it]2026-04-13 10:41:33,069 | INFO | train_grpo_train | metrics_logged mode=train step=1349

                                                       
{'loss': 0.0282, 'grad_norm': 11.147529602050781, 'learning_rate': 5.915151515151516e-06, 'num_tokens': 2383138.0, 'completions/mean_length': 47.75, 'completions/min_length': 43.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8343709707260132, 'rewards/meter/std': 0.3385268747806549, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9651442766189575, 'rewards/repeat_soft/std': 0.024144303053617477, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.596796989440918, 'rewards/total_composite/std': 0.15304796397686005, 'reward': 0.596796989440918, 'reward_std': 0.15304794907569885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1502106934785843, 'sampling/sampling_logp_difference/max': 1.7638282775878906, 'sampling/importance_sampling_ratio/min': 0.17138749361038208, 'sampling/importance_sampling_ratio/mean': 1.0072975158691406, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9210467338562012, 'clip_ratio/low_mean': 0.05865122191607952, 'clip_ratio/low_min': 0.05865122191607952, 'clip_ratio/high_mean': 0.0917295441031456, 'clip_ratio/high_max': 0.0917295441031456, 'clip_ratio/region_mean': 0.15038076601922512, 'reward_total_mean': 0.596796989440918, 'reward_meter_mean': 0.8343709707260132, 'reward_meter_std': 0.3385268747806549, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9651442766189575, 'reward_repeat_soft_std': 0.024144303053617477, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.596796989440918, 'reward_total_composite_std': 0.15304796397686005, 'epoch': 0.14}

 41%|████      | 1349/3300 [3:20:35<3:23:01,  6.24s/it]INFO 04-13 10:41:33 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1350/3300 [3:20:41<3:21:25,  6.20s/it]2026-04-13 10:41:39,159 | INFO | train_grpo_train | metrics_logged mode=train step=1350

                                                       
{'loss': 0.0367, 'grad_norm': 11.28422737121582, 'learning_rate': 5.912121212121212e-06, 'num_tokens': 2384844.0, 'completions/mean_length': 46.25, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4069119691848755, 'rewards/meter/std': 0.37338826060295105, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.986085832118988, 'rewards/repeat_soft/std': 0.01686898060142994, 'rewards/judge_quality/mean': 0.7450000047683716, 'rewards/judge_quality/std': 0.16690459847450256, 'rewards/total_composite/mean': 0.5481597185134888, 'rewards/total_composite/std': 0.2011103332042694, 'reward': 0.5481597185134888, 'reward_std': 0.20111030340194702, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13221058249473572, 'sampling/sampling_logp_difference/max': 1.7689342498779297, 'sampling/importance_sampling_ratio/min': 0.17051461338996887, 'sampling/importance_sampling_ratio/mean': 0.9992238879203796, 'sampling/importance_sampling_ratio/max': 1.8740284442901611, 'entropy': 0.9551234468817711, 'clip_ratio/low_mean': 0.08280645590275526, 'clip_ratio/low_min': 0.08280645590275526, 'clip_ratio/high_mean': 0.06803585588932037, 'clip_ratio/high_max': 0.06803585588932037, 'clip_ratio/region_mean': 0.15084231179207563, 'reward_total_mean': 0.5481597185134888, 'reward_meter_mean': 0.4069119691848755, 'reward_meter_std': 0.37338826060295105, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.986085832118988, 'reward_repeat_soft_std': 0.01686898060142994, 'reward_judge_quality_mean': 0.7450000047683716, 'reward_judge_quality_std': 0.16690459847450256, 'reward_total_composite_mean': 0.5481597185134888, 'reward_total_composite_std': 0.2011103332042694, 'epoch': 0.14}

 41%|████      | 1350/3300 [3:20:41<3:21:25,  6.20s/it]INFO 04-13 10:41:39 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.86s/it]

 30%|███       | 3/10 [00:07<00:18,  2.67s/it]

 40%|████      | 4/10 [00:10<00:17,  2.91s/it]

 50%|█████     | 5/10 [00:14<00:16,  3.22s/it]

 60%|██████    | 6/10 [00:18<00:13,  3.49s/it]

 70%|███████   | 7/10 [00:22<00:10,  3.62s/it]

 80%|████████  | 8/10 [00:26<00:07,  3.58s/it]

 90%|█████████ | 9/10 [00:29<00:03,  3.49s/it]

100%|██████████| 10/10 [00:32<00:00,  3.42s/it]2026-04-13 10:42:16,125 | INFO | train_grpo_train | metrics_logged mode=eval step=1350

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 36.8843, 'eval_samples_per_second': 2.169, 'eval_steps_per_second': 0.271, 'eval_num_tokens': 2384844.0, 'eval_completions/mean_length': 64.8625, 'eval_completions/min_length': 29.9, 'eval_completions/max_length': 107.8, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 64.8625, 'eval_completions/min_terminated_length': 29.9, 'eval_completions/max_terminated_length': 107.8, 'eval_rewards/meter/mean': 0.6510038495063781, 'eval_rewards/meter/std': 0.369542495906353, 'eval_rewards/count_adherence/mean': 0.9585416555404663, 'eval_rewards/count_adherence/std': 0.07452610954642296, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/repeat_soft/mean': 0.881782990694046, 'eval_rewards/repeat_soft/std': 0.09381132461130619, 'eval_rewards/judge_quality/mean': 0.467249995470047, 'eval_rewards/judge_quality/std': 0.16026114895939828, 'eval_rewards/total_composite/mean': 0.5154408305883408, 'eval_rewards/total_composite/std': 0.1306465409696102, 'eval_reward': 0.5154408305883408, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06699008122086525, 'eval_sampling/sampling_logp_difference/max': 1.0151650428771972, 'eval_sampling/importance_sampling_ratio/min': 0.3660672038793564, 'eval_sampling/importance_sampling_ratio/mean': 1.0170485734939576, 'eval_sampling/importance_sampling_ratio/max': 1.408859872817993, 'eval_entropy': 0.7701298058032989, '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': 0.5154408305883408, 'eval_reward_meter_mean': 0.6510038495063781, 'eval_reward_meter_std': 0.369542495906353, 'eval_reward_count_adherence_mean': 0.9585416555404663, 'eval_reward_count_adherence_std': 0.07452610954642296, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_repeat_soft_mean': 0.881782990694046, 'eval_reward_repeat_soft_std': 0.09381132461130619, 'eval_reward_judge_quality_mean': 0.467249995470047, 'eval_reward_judge_quality_std': 0.16026114895939828, 'eval_reward_total_composite_mean': 0.5154408305883408, 'eval_reward_total_composite_std': 0.1306465409696102, 'epoch': 0.14}

 41%|████      | 1350/3300 [3:21:18<3:21:25,  6.20s/it]

100%|██████████| 10/10 [00:32<00:00,  3.42s/it]

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

 41%|████      | 1351/3300 [3:21:27<9:49:45, 18.16s/it]2026-04-13 10:42:25,215 | INFO | train_grpo_train | metrics_logged mode=train step=1351

                                                       
{'loss': 0.1548, 'grad_norm': 17.512495040893555, 'learning_rate': 5.90909090909091e-06, 'num_tokens': 2386208.0, 'completions/mean_length': 21.5, 'completions/min_length': 18.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.5794750452041626, 'rewards/meter/std': 0.4747667908668518, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9476755857467651, 'rewards/repeat_soft/std': 0.041929665952920914, 'rewards/judge_quality/mean': 0.32249999046325684, 'rewards/judge_quality/std': 0.14320313930511475, 'rewards/total_composite/mean': 0.4657644033432007, 'rewards/total_composite/std': 0.12241081148386002, 'reward': 0.4657644033432007, 'reward_std': 0.12241081148386002, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12897294759750366, 'sampling/sampling_logp_difference/max': 0.9636764526367188, 'sampling/importance_sampling_ratio/min': 0.38148778676986694, 'sampling/importance_sampling_ratio/mean': 0.9785038828849792, 'sampling/importance_sampling_ratio/max': 1.5874792337417603, 'entropy': 0.8263326734304428, 'clip_ratio/low_mean': 0.06844179518520832, 'clip_ratio/low_min': 0.06844179518520832, 'clip_ratio/high_mean': 0.048245614394545555, 'clip_ratio/high_max': 0.048245614394545555, 'clip_ratio/region_mean': 0.11668740957975388, 'reward_total_mean': 0.4657644033432007, 'reward_meter_mean': 0.5794750452041626, 'reward_meter_std': 0.4747667908668518, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9476755857467651, 'reward_repeat_soft_std': 0.041929665952920914, 'reward_judge_quality_mean': 0.32249999046325684, 'reward_judge_quality_std': 0.14320313930511475, 'reward_total_composite_mean': 0.4657644033432007, 'reward_total_composite_std': 0.12241081148386002, 'epoch': 0.14}

 41%|████      | 1351/3300 [3:21:27<9:49:45, 18.16s/it]INFO 04-13 10:42:25 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1352/3300 [3:21:34<7:58:25, 14.74s/it]2026-04-13 10:42:31,975 | INFO | train_grpo_train | metrics_logged mode=train step=1352

                                                       
{'loss': -0.0074, 'grad_norm': 11.55992603302002, 'learning_rate': 5.906060606060607e-06, 'num_tokens': 2387877.0, 'completions/mean_length': 44.625, 'completions/min_length': 36.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.32143399119377136, 'rewards/meter/std': 0.4152473509311676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9601118564605713, 'rewards/repeat_soft/std': 0.023664511740207672, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.4557591676712036, 'rewards/total_composite/std': 0.14588448405265808, 'reward': 0.4557591676712036, 'reward_std': 0.14588449895381927, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14616696536540985, 'sampling/sampling_logp_difference/max': 1.1126775741577148, 'sampling/importance_sampling_ratio/min': 0.3286777138710022, 'sampling/importance_sampling_ratio/mean': 1.0224692821502686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0526707097887993, 'clip_ratio/low_mean': 0.06728621479123831, 'clip_ratio/low_min': 0.06728621479123831, 'clip_ratio/high_mean': 0.05823863670229912, 'clip_ratio/high_max': 0.05823863670229912, 'clip_ratio/region_mean': 0.12552485149353743, 'reward_total_mean': 0.4557591676712036, 'reward_meter_mean': 0.32143399119377136, 'reward_meter_std': 0.4152473509311676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9601118564605713, 'reward_repeat_soft_std': 0.023664511740207672, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.4557591676712036, 'reward_total_composite_std': 0.14588448405265808, 'epoch': 0.14}

 41%|████      | 1352/3300 [3:21:34<7:58:25, 14.74s/it]INFO 04-13 10:42:32 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1353/3300 [3:21:45<7:26:01, 13.74s/it]2026-04-13 10:42:43,409 | INFO | train_grpo_train | metrics_logged mode=train step=1353

                                                       
{'loss': -0.114, 'grad_norm': 2.1518027782440186, 'learning_rate': 5.903030303030304e-06, 'num_tokens': 2389547.0, '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': 54.0, 'rewards/meter/mean': 0.595986008644104, 'rewards/meter/std': 0.32510465383529663, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8903664946556091, 'rewards/repeat_soft/std': 0.10740083456039429, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.14327171444892883, 'rewards/total_composite/mean': 0.43036508560180664, 'rewards/total_composite/std': 0.19207099080085754, 'reward': 0.43036508560180664, 'reward_std': 0.19207097589969635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11597153544425964, 'sampling/sampling_logp_difference/max': 1.6034255027770996, 'sampling/importance_sampling_ratio/min': 0.20120610296726227, 'sampling/importance_sampling_ratio/mean': 1.011412262916565, 'sampling/importance_sampling_ratio/max': 1.964923620223999, 'entropy': 0.8036029450595379, 'clip_ratio/low_mean': 0.026701120659708977, 'clip_ratio/low_min': 0.026701120659708977, 'clip_ratio/high_mean': 0.07077431888319552, 'clip_ratio/high_max': 0.07077431888319552, 'clip_ratio/region_mean': 0.0974754395429045, 'reward_total_mean': 0.43036508560180664, 'reward_meter_mean': 0.595986008644104, 'reward_meter_std': 0.32510465383529663, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8903664946556091, 'reward_repeat_soft_std': 0.10740083456039429, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.14327171444892883, 'reward_total_composite_mean': 0.43036508560180664, 'reward_total_composite_std': 0.19207099080085754, 'epoch': 0.14}

 41%|████      | 1353/3300 [3:21:45<7:26:01, 13.74s/it]INFO 04-13 10:42:43 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1354/3300 [3:21:55<6:45:20, 12.50s/it]2026-04-13 10:42:52,993 | INFO | train_grpo_train | metrics_logged mode=train step=1354

                                                       
{'loss': 0.01, 'grad_norm': 12.866393089294434, 'learning_rate': 5.9e-06, 'num_tokens': 2391142.0, 'completions/mean_length': 35.375, 'completions/min_length': 33.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8208308219909668, 'rewards/meter/std': 0.25697070360183716, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9737215638160706, 'rewards/repeat_soft/std': 0.025722060352563858, 'rewards/judge_quality/mean': 0.47749996185302734, 'rewards/judge_quality/std': 0.16263456642627716, 'rewards/total_composite/mean': 0.605654776096344, 'rewards/total_composite/std': 0.1339626908302307, 'reward': 0.605654776096344, 'reward_std': 0.13396267592906952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11455649137496948, 'sampling/sampling_logp_difference/max': 1.409034252166748, 'sampling/importance_sampling_ratio/min': 0.2443791776895523, 'sampling/importance_sampling_ratio/mean': 1.0094857215881348, 'sampling/importance_sampling_ratio/max': 1.966202974319458, 'entropy': 0.630085363984108, 'clip_ratio/low_mean': 0.04261363763362169, 'clip_ratio/low_min': 0.04261363763362169, 'clip_ratio/high_mean': 0.07221631123684347, 'clip_ratio/high_max': 0.07221631123684347, 'clip_ratio/region_mean': 0.11482994887046516, 'reward_total_mean': 0.605654776096344, 'reward_meter_mean': 0.8208308219909668, 'reward_meter_std': 0.25697070360183716, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9737215638160706, 'reward_repeat_soft_std': 0.025722060352563858, 'reward_judge_quality_mean': 0.47749996185302734, 'reward_judge_quality_std': 0.16263456642627716, 'reward_total_composite_mean': 0.605654776096344, 'reward_total_composite_std': 0.1339626908302307, 'epoch': 0.14}

 41%|████      | 1354/3300 [3:21:55<6:45:20, 12.50s/it]INFO 04-13 10:42:53 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1355/3300 [3:22:00<5:40:13, 10.50s/it]2026-04-13 10:42:58,818 | INFO | train_grpo_train | metrics_logged mode=train step=1355

                                                       
{'loss': 0.0844, 'grad_norm': 16.96038055419922, 'learning_rate': 5.8969696969696975e-06, 'num_tokens': 2392560.0, 'completions/mean_length': 21.25, 'completions/min_length': 20.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.7864716649055481, 'rewards/meter/std': 0.3567924499511719, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9309078454971313, 'rewards/repeat_soft/std': 0.04106544330716133, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5947766900062561, 'rewards/total_composite/std': 0.1688213050365448, 'reward': 0.5947766900062561, 'reward_std': 0.1688213050365448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13679121434688568, 'sampling/sampling_logp_difference/max': 1.2551536560058594, 'sampling/importance_sampling_ratio/min': 0.2850320339202881, 'sampling/importance_sampling_ratio/mean': 1.0153204202651978, 'sampling/importance_sampling_ratio/max': 1.9045906066894531, 'entropy': 0.9432927444577217, 'clip_ratio/low_mean': 0.03880811156705022, 'clip_ratio/low_min': 0.03880811156705022, 'clip_ratio/high_mean': 0.10338203702121973, 'clip_ratio/high_max': 0.10338203702121973, 'clip_ratio/region_mean': 0.14219014858826995, 'reward_total_mean': 0.5947766900062561, 'reward_meter_mean': 0.7864716649055481, 'reward_meter_std': 0.3567924499511719, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9309078454971313, 'reward_repeat_soft_std': 0.04106544330716133, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5947766900062561, 'reward_total_composite_std': 0.1688213050365448, 'epoch': 0.14}

 41%|████      | 1355/3300 [3:22:01<5:40:13, 10.50s/it]INFO 04-13 10:42:59 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1356/3300 [3:22:06<4:56:18,  9.15s/it]2026-04-13 10:43:04,810 | INFO | train_grpo_train | metrics_logged mode=train step=1356

                                                       
{'loss': 0.0937, 'grad_norm': 17.286178588867188, 'learning_rate': 5.893939393939394e-06, 'num_tokens': 2394154.0, 'completions/mean_length': 39.25, 'completions/min_length': 33.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.33529916405677795, 'rewards/meter/std': 0.3962272107601166, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9275178909301758, 'rewards/repeat_soft/std': 0.05657338723540306, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.4632614552974701, 'rewards/total_composite/std': 0.20912042260169983, 'reward': 0.4632614552974701, 'reward_std': 0.20912039279937744, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1105823665857315, 'sampling/sampling_logp_difference/max': 1.0438995361328125, 'sampling/importance_sampling_ratio/min': 0.38390469551086426, 'sampling/importance_sampling_ratio/mean': 1.0394328832626343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7477941028773785, 'clip_ratio/low_mean': 0.0745665249414742, 'clip_ratio/low_min': 0.0745665249414742, 'clip_ratio/high_mean': 0.022727273404598236, 'clip_ratio/high_max': 0.022727273404598236, 'clip_ratio/region_mean': 0.09729379834607244, 'reward_total_mean': 0.4632614552974701, 'reward_meter_mean': 0.33529916405677795, 'reward_meter_std': 0.3962272107601166, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9275178909301758, 'reward_repeat_soft_std': 0.05657338723540306, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.4632614552974701, 'reward_total_composite_std': 0.20912042260169983, 'epoch': 0.14}

 41%|████      | 1356/3300 [3:22:07<4:56:18,  9.15s/it]INFO 04-13 10:43:05 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1357/3300 [3:22:19<5:25:59, 10.07s/it]2026-04-13 10:43:17,026 | INFO | train_grpo_train | metrics_logged mode=train step=1357

                                                       
{'loss': -0.1563, 'grad_norm': 2.4548747539520264, 'learning_rate': 5.890909090909091e-06, 'num_tokens': 2396176.0, 'completions/mean_length': 248.75, 'completions/min_length': 81.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 90.80000305175781, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.411411315202713, 'rewards/meter/std': 0.2769858241081238, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9301238059997559, 'rewards/repeat_soft/std': 0.07403453439474106, 'rewards/judge_quality/mean': 0.3374999761581421, 'rewards/judge_quality/std': 0.30372685194015503, 'rewards/total_composite/mean': 0.29426518082618713, 'rewards/total_composite/std': 0.2515621781349182, 'reward': 0.29426518082618713, 'reward_std': 0.2515621483325958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12146399170160294, 'sampling/sampling_logp_difference/max': 1.476247787475586, 'sampling/importance_sampling_ratio/min': 0.22849343717098236, 'sampling/importance_sampling_ratio/mean': 1.0232826471328735, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5180720612406731, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0682307556271553, 'clip_ratio/high_max': 0.0682307556271553, 'clip_ratio/region_mean': 0.0682307556271553, 'reward_total_mean': 0.29426518082618713, 'reward_meter_mean': 0.411411315202713, 'reward_meter_std': 0.2769858241081238, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9301238059997559, 'reward_repeat_soft_std': 0.07403453439474106, 'reward_judge_quality_mean': 0.3374999761581421, 'reward_judge_quality_std': 0.30372685194015503, 'reward_total_composite_mean': 0.29426518082618713, 'reward_total_composite_std': 0.2515621781349182, 'epoch': 0.14}

 41%|████      | 1357/3300 [3:22:19<5:25:59, 10.07s/it]INFO 04-13 10:43:17 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1358/3300 [3:22:30<5:37:57, 10.44s/it]2026-04-13 10:43:28,347 | INFO | train_grpo_train | metrics_logged mode=train step=1358

                                                       
{'loss': -0.1726, 'grad_norm': 2.606265068054199, 'learning_rate': 5.887878787878788e-06, 'num_tokens': 2398224.0, 'completions/mean_length': 142.0, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 89.14286041259766, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.5142291784286499, 'rewards/meter/std': 0.27461180090904236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8774094581604004, 'rewards/repeat_soft/std': 0.055666882544755936, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4211280643939972, 'rewards/total_composite/std': 0.1816686987876892, 'reward': 0.4211280643939972, 'reward_std': 0.1816686987876892, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14020542800426483, 'sampling/sampling_logp_difference/max': 1.2822426557540894, 'sampling/importance_sampling_ratio/min': 0.2774144411087036, 'sampling/importance_sampling_ratio/mean': 1.0215743780136108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9193695932626724, 'clip_ratio/low_mean': 0.01923076994717121, 'clip_ratio/low_min': 0.01923076994717121, 'clip_ratio/high_mean': 0.09595720563083887, 'clip_ratio/high_max': 0.09595720563083887, 'clip_ratio/region_mean': 0.11518797557801008, 'reward_total_mean': 0.4211280643939972, 'reward_meter_mean': 0.5142291784286499, 'reward_meter_std': 0.27461180090904236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8774094581604004, 'reward_repeat_soft_std': 0.055666882544755936, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4211280643939972, 'reward_total_composite_std': 0.1816686987876892, 'epoch': 0.14}

 41%|████      | 1358/3300 [3:22:30<5:37:57, 10.44s/it]INFO 04-13 10:43:28 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1359/3300 [3:22:36<4:50:55,  8.99s/it]2026-04-13 10:43:33,955 | INFO | train_grpo_train | metrics_logged mode=train step=1359

                                                       
{'loss': 0.0196, 'grad_norm': 12.106054306030273, 'learning_rate': 5.884848484848486e-06, 'num_tokens': 2399517.0, 'completions/mean_length': 23.625, 'completions/min_length': 20.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8665356040000916, 'rewards/meter/std': 0.3499932289123535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9548467397689819, 'rewards/repeat_soft/std': 0.020502854138612747, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5870592594146729, 'rewards/total_composite/std': 0.09871426224708557, 'reward': 0.5870592594146729, 'reward_std': 0.09871426224708557, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09746392071247101, 'sampling/sampling_logp_difference/max': 1.053166389465332, 'sampling/importance_sampling_ratio/min': 0.348831444978714, 'sampling/importance_sampling_ratio/mean': 1.0364505052566528, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6199148334562778, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.08211747603490949, 'clip_ratio/high_max': 0.08211747603490949, 'clip_ratio/region_mean': 0.09253414301201701, 'reward_total_mean': 0.5870592594146729, 'reward_meter_mean': 0.8665356040000916, 'reward_meter_std': 0.3499932289123535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9548467397689819, 'reward_repeat_soft_std': 0.020502854138612747, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5870592594146729, 'reward_total_composite_std': 0.09871426224708557, 'epoch': 0.14}

 41%|████      | 1359/3300 [3:22:36<4:50:55,  8.99s/it]INFO 04-13 10:43:34 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1360/3300 [3:22:41<4:19:44,  8.03s/it]2026-04-13 10:43:39,750 | INFO | train_grpo_train | metrics_logged mode=train step=1360

                                                       
{'loss': 0.1873, 'grad_norm': 11.9026460647583, 'learning_rate': 5.881818181818182e-06, 'num_tokens': 2400959.0, 'completions/mean_length': 30.25, 'completions/min_length': 21.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.25, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.5686236023902893, 'rewards/meter/std': 0.2765858471393585, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9673341512680054, 'rewards/repeat_soft/std': 0.059099845588207245, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4919586479663849, 'rewards/total_composite/std': 0.09874510020017624, 'reward': 0.4919586479663849, 'reward_std': 0.09874510020017624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11446832865476608, 'sampling/sampling_logp_difference/max': 1.4261484146118164, 'sampling/importance_sampling_ratio/min': 0.2402324229478836, 'sampling/importance_sampling_ratio/mean': 1.0121992826461792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4771619662642479, 'clip_ratio/low_mean': 0.03084415663033724, 'clip_ratio/low_min': 0.03084415663033724, 'clip_ratio/high_mean': 0.06924877362325788, 'clip_ratio/high_max': 0.06924877362325788, 'clip_ratio/region_mean': 0.10009293025359511, 'reward_total_mean': 0.4919586479663849, 'reward_meter_mean': 0.5686236023902893, 'reward_meter_std': 0.2765858471393585, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9673341512680054, 'reward_repeat_soft_std': 0.059099845588207245, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4919586479663849, 'reward_total_composite_std': 0.09874510020017624, 'epoch': 0.14}

 41%|████      | 1360/3300 [3:22:41<4:19:44,  8.03s/it]INFO 04-13 10:43:39 [block_pool.py:378] Successfully reset prefix cache

 41%|████      | 1361/3300 [3:22:48<4:02:50,  7.51s/it]2026-04-13 10:43:46,051 | INFO | train_grpo_train | metrics_logged mode=train step=1361

                                                       
{'loss': 0.0322, 'grad_norm': 10.216547966003418, 'learning_rate': 5.878787878787879e-06, 'num_tokens': 2402972.0, 'completions/mean_length': 64.625, 'completions/min_length': 51.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.29268279671669006, 'rewards/meter/std': 0.2749910354614258, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.92520672082901, 'rewards/repeat_soft/std': 0.04358210787177086, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.4158943295478821, 'rewards/total_composite/std': 0.07025546580553055, 'reward': 0.4158943295478821, 'reward_std': 0.07025546580553055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1270250529050827, 'sampling/sampling_logp_difference/max': 1.696502685546875, 'sampling/importance_sampling_ratio/min': 0.18332353234291077, 'sampling/importance_sampling_ratio/mean': 1.0163600444793701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6991370767354965, 'clip_ratio/low_mean': 0.06225152686238289, 'clip_ratio/low_min': 0.06225152686238289, 'clip_ratio/high_mean': 0.04847604315727949, 'clip_ratio/high_max': 0.04847604315727949, 'clip_ratio/region_mean': 0.11072757001966238, 'reward_total_mean': 0.4158943295478821, 'reward_meter_mean': 0.29268279671669006, 'reward_meter_std': 0.2749910354614258, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.92520672082901, 'reward_repeat_soft_std': 0.04358210787177086, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.4158943295478821, 'reward_total_composite_std': 0.07025546580553055, 'epoch': 0.14}

 41%|████      | 1361/3300 [3:22:48<4:02:50,  7.51s/it]INFO 04-13 10:43:46 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1362/3300 [3:22:54<3:53:37,  7.23s/it]2026-04-13 10:43:52,630 | INFO | train_grpo_train | metrics_logged mode=train step=1362

                                                       
{'loss': 0.2264, 'grad_norm': 26.244260787963867, 'learning_rate': 5.875757575757576e-06, 'num_tokens': 2404385.0, 'completions/mean_length': 25.625, 'completions/min_length': 16.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7774785161018372, 'rewards/meter/std': 0.36153045296669006, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.5175492167472839, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9274853467941284, 'rewards/repeat_soft/std': 0.06705766171216965, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.4675617814064026, 'rewards/total_composite/std': 0.13928283751010895, 'reward': 0.4675617814064026, 'reward_std': 0.13928283751010895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16288518905639648, 'sampling/sampling_logp_difference/max': 1.2759896516799927, 'sampling/importance_sampling_ratio/min': 0.27915456891059875, 'sampling/importance_sampling_ratio/mean': 1.0186101198196411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9245611876249313, 'clip_ratio/low_mean': 0.059489935636520386, 'clip_ratio/low_min': 0.059489935636520386, 'clip_ratio/high_mean': 0.055330883245915174, 'clip_ratio/high_max': 0.055330883245915174, 'clip_ratio/region_mean': 0.11482081888243556, 'reward_total_mean': 0.4675617814064026, 'reward_meter_mean': 0.7774785161018372, 'reward_meter_std': 0.36153045296669006, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.5175492167472839, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9274853467941284, 'reward_repeat_soft_std': 0.06705766171216965, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.4675617814064026, 'reward_total_composite_std': 0.13928283751010895, 'epoch': 0.14}

 41%|████▏     | 1362/3300 [3:22:54<3:53:37,  7.23s/it]INFO 04-13 10:43:52 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1363/3300 [3:23:01<3:47:05,  7.03s/it]2026-04-13 10:43:59,205 | INFO | train_grpo_train | metrics_logged mode=train step=1363

                                                       
{'loss': 0.0196, 'grad_norm': 12.606419563293457, 'learning_rate': 5.872727272727273e-06, 'num_tokens': 2406050.0, 'completions/mean_length': 43.125, 'completions/min_length': 40.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8502081632614136, 'rewards/meter/std': 0.2545948028564453, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8542707562446594, 'rewards/repeat_soft/std': 0.09567058831453323, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5650582313537598, 'rewards/total_composite/std': 0.07314696907997131, 'reward': 0.5650582313537598, 'reward_std': 0.07314696162939072, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14268463850021362, 'sampling/sampling_logp_difference/max': 1.8414602279663086, 'sampling/importance_sampling_ratio/min': 0.15858569741249084, 'sampling/importance_sampling_ratio/mean': 1.026253342628479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9797668755054474, 'clip_ratio/low_mean': 0.05472383834421635, 'clip_ratio/low_min': 0.05472383834421635, 'clip_ratio/high_mean': 0.09119397960603237, 'clip_ratio/high_max': 0.09119397960603237, 'clip_ratio/region_mean': 0.14591781795024872, 'reward_total_mean': 0.5650582313537598, 'reward_meter_mean': 0.8502081632614136, 'reward_meter_std': 0.2545948028564453, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8542707562446594, 'reward_repeat_soft_std': 0.09567058831453323, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5650582313537598, 'reward_total_composite_std': 0.07314696907997131, 'epoch': 0.14}

 41%|████▏     | 1363/3300 [3:23:01<3:47:05,  7.03s/it]INFO 04-13 10:43:59 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1364/3300 [3:23:07<3:37:09,  6.73s/it]2026-04-13 10:44:05,219 | INFO | train_grpo_train | metrics_logged mode=train step=1364

                                                       
{'loss': 0.029, 'grad_norm': 10.385464668273926, 'learning_rate': 5.8696969696969694e-06, 'num_tokens': 2407574.0, 'completions/mean_length': 38.5, 'completions/min_length': 31.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.882033109664917, 'rewards/meter/std': 0.30539917945861816, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7661975026130676, 'rewards/repeat_soft/std': 0.1480618566274643, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5629356503486633, 'rewards/total_composite/std': 0.09463868290185928, 'reward': 0.5629356503486633, 'reward_std': 0.09463869035243988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11570446193218231, 'sampling/sampling_logp_difference/max': 1.1335511207580566, 'sampling/importance_sampling_ratio/min': 0.32188817858695984, 'sampling/importance_sampling_ratio/mean': 1.0171159505844116, 'sampling/importance_sampling_ratio/max': 1.6529345512390137, 'entropy': 0.9708805829286575, 'clip_ratio/low_mean': 0.00872093066573143, 'clip_ratio/low_min': 0.00872093066573143, 'clip_ratio/high_mean': 0.11443972727283835, 'clip_ratio/high_max': 0.11443972727283835, 'clip_ratio/region_mean': 0.12316065793856978, 'reward_total_mean': 0.5629356503486633, 'reward_meter_mean': 0.882033109664917, 'reward_meter_std': 0.30539917945861816, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7661975026130676, 'reward_repeat_soft_std': 0.1480618566274643, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5629356503486633, 'reward_total_composite_std': 0.09463868290185928, 'epoch': 0.14}

 41%|████▏     | 1364/3300 [3:23:07<3:37:09,  6.73s/it]INFO 04-13 10:44:05 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1365/3300 [3:23:13<3:29:44,  6.50s/it]2026-04-13 10:44:11,200 | INFO | train_grpo_train | metrics_logged mode=train step=1365

                                                       
{'loss': 0.0816, 'grad_norm': 14.100876808166504, 'learning_rate': 5.8666666666666675e-06, 'num_tokens': 2409165.0, 'completions/mean_length': 37.875, 'completions/min_length': 34.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.47339460253715515, 'rewards/meter/std': 0.37978094816207886, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9774109125137329, 'rewards/repeat_soft/std': 0.028187068179249763, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.47740429639816284, 'rewards/total_composite/std': 0.10387453436851501, 'reward': 0.47740429639816284, 'reward_std': 0.10387454181909561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13374821841716766, 'sampling/sampling_logp_difference/max': 1.9290554523468018, 'sampling/importance_sampling_ratio/min': 0.14528536796569824, 'sampling/importance_sampling_ratio/mean': 1.0020943880081177, 'sampling/importance_sampling_ratio/max': 1.8436821699142456, 'entropy': 0.8350476697087288, 'clip_ratio/low_mean': 0.029022637056186795, 'clip_ratio/low_min': 0.029022637056186795, 'clip_ratio/high_mean': 0.08177459705621004, 'clip_ratio/high_max': 0.08177459705621004, 'clip_ratio/region_mean': 0.11079723411239684, 'reward_total_mean': 0.47740429639816284, 'reward_meter_mean': 0.47339460253715515, 'reward_meter_std': 0.37978094816207886, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9774109125137329, 'reward_repeat_soft_std': 0.028187068179249763, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.47740429639816284, 'reward_total_composite_std': 0.10387453436851501, 'epoch': 0.14}

 41%|████▏     | 1365/3300 [3:23:13<3:29:44,  6.50s/it]INFO 04-13 10:44:11 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1366/3300 [3:23:19<3:26:08,  6.40s/it]2026-04-13 10:44:17,334 | INFO | train_grpo_train | metrics_logged mode=train step=1366

                                                       
{'loss': 0.0075, 'grad_norm': 15.947189331054688, 'learning_rate': 5.863636363636364e-06, 'num_tokens': 2410658.0, 'completions/mean_length': 38.625, 'completions/min_length': 36.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9263715744018555, 'rewards/meter/std': 0.0999070331454277, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8989310264587402, 'rewards/repeat_soft/std': 0.025368811562657356, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.18192915618419647, 'rewards/total_composite/mean': 0.657719075679779, 'rewards/total_composite/std': 0.10532543063163757, 'reward': 0.657719075679779, 'reward_std': 0.10532545298337936, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12557724118232727, 'sampling/sampling_logp_difference/max': 1.2217206954956055, 'sampling/importance_sampling_ratio/min': 0.2947225868701935, 'sampling/importance_sampling_ratio/mean': 1.012650966644287, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8277973011136055, 'clip_ratio/low_mean': 0.0658501754514873, 'clip_ratio/low_min': 0.0658501754514873, 'clip_ratio/high_mean': 0.022368420846760273, 'clip_ratio/high_max': 0.022368420846760273, 'clip_ratio/region_mean': 0.08821859629824758, 'reward_total_mean': 0.657719075679779, 'reward_meter_mean': 0.9263715744018555, 'reward_meter_std': 0.0999070331454277, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8989310264587402, 'reward_repeat_soft_std': 0.025368811562657356, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.18192915618419647, 'reward_total_composite_mean': 0.657719075679779, 'reward_total_composite_std': 0.10532543063163757, 'epoch': 0.14}

 41%|████▏     | 1366/3300 [3:23:19<3:26:08,  6.40s/it]INFO 04-13 10:44:17 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1367/3300 [3:23:26<3:28:22,  6.47s/it]2026-04-13 10:44:23,975 | INFO | train_grpo_train | metrics_logged mode=train step=1367

                                                       
{'loss': 0.0601, 'grad_norm': 8.43343734741211, 'learning_rate': 5.860606060606061e-06, 'num_tokens': 2412634.0, 'completions/mean_length': 76.0, 'completions/min_length': 67.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9826860427856445, 'rewards/meter/std': 0.006795603781938553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7383664846420288, 'rewards/repeat_soft/std': 0.07205822318792343, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.6053905487060547, 'rewards/total_composite/std': 0.12587036192417145, 'reward': 0.6053905487060547, 'reward_std': 0.12587036192417145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09169119596481323, 'sampling/sampling_logp_difference/max': 1.7090654373168945, 'sampling/importance_sampling_ratio/min': 0.1810348927974701, 'sampling/importance_sampling_ratio/mean': 1.0142838954925537, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6261106580495834, 'clip_ratio/low_mean': 0.07950532995164394, 'clip_ratio/low_min': 0.07950532995164394, 'clip_ratio/high_mean': 0.01689189113676548, 'clip_ratio/high_max': 0.01689189113676548, 'clip_ratio/region_mean': 0.09639722108840942, 'reward_total_mean': 0.6053905487060547, 'reward_meter_mean': 0.9826860427856445, 'reward_meter_std': 0.006795603781938553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7383664846420288, 'reward_repeat_soft_std': 0.07205822318792343, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.6053905487060547, 'reward_total_composite_std': 0.12587036192417145, 'epoch': 0.14}

 41%|████▏     | 1367/3300 [3:23:26<3:28:22,  6.47s/it]INFO 04-13 10:44:24 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1368/3300 [3:23:32<3:28:53,  6.49s/it]2026-04-13 10:44:30,505 | INFO | train_grpo_train | metrics_logged mode=train step=1368

                                                       
{'loss': -0.0401, 'grad_norm': 16.965978622436523, 'learning_rate': 5.8575757575757584e-06, 'num_tokens': 2414299.0, 'completions/mean_length': 36.125, 'completions/min_length': 30.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8173834085464478, 'rewards/meter/std': 0.2950228452682495, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9791263341903687, 'rewards/repeat_soft/std': 0.032146155834198, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.1954299360513687, 'rewards/total_composite/mean': 0.5984628796577454, 'rewards/total_composite/std': 0.09653806686401367, 'reward': 0.5984628796577454, 'reward_std': 0.09653805196285248, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14996446669101715, 'sampling/sampling_logp_difference/max': 1.7358055114746094, 'sampling/importance_sampling_ratio/min': 0.1762581765651703, 'sampling/importance_sampling_ratio/mean': 1.0238789319992065, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8199257031083107, 'clip_ratio/low_mean': 0.052777779288589954, 'clip_ratio/low_min': 0.052777779288589954, 'clip_ratio/high_mean': 0.09061235003173351, 'clip_ratio/high_max': 0.09061235003173351, 'clip_ratio/region_mean': 0.14339012932032347, 'reward_total_mean': 0.5984628796577454, 'reward_meter_mean': 0.8173834085464478, 'reward_meter_std': 0.2950228452682495, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9791263341903687, 'reward_repeat_soft_std': 0.032146155834198, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.1954299360513687, 'reward_total_composite_mean': 0.5984628796577454, 'reward_total_composite_std': 0.09653806686401367, 'epoch': 0.14}

 41%|████▏     | 1368/3300 [3:23:32<3:28:53,  6.49s/it]INFO 04-13 10:44:30 [block_pool.py:378] Successfully reset prefix cache

 41%|████▏     | 1369/3300 [3:23:44<4:19:32,  8.06s/it]2026-04-13 10:44:42,254 | INFO | train_grpo_train | metrics_logged mode=train step=1369

                                                       
{'loss': -0.138, 'grad_norm': 2.002079963684082, 'learning_rate': 5.854545454545455e-06, 'num_tokens': 2416019.0, 'completions/mean_length': 112.0, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 54.857147216796875, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7430543899536133, 'rewards/meter/std': 0.28496313095092773, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7456576824188232, 'rewards/repeat_soft/std': 0.07440777122974396, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.46270573139190674, 'rewards/total_composite/std': 0.19669091701507568, 'reward': 0.46270573139190674, 'reward_std': 0.19669093191623688, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09795010089874268, 'sampling/sampling_logp_difference/max': 1.4368808269500732, 'sampling/importance_sampling_ratio/min': 0.23766793310642242, 'sampling/importance_sampling_ratio/mean': 1.0065462589263916, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4737773537635803, 'clip_ratio/low_mean': 0.010775862261652946, 'clip_ratio/low_min': 0.010775862261652946, 'clip_ratio/high_mean': 0.06216462189331651, 'clip_ratio/high_max': 0.06216462189331651, 'clip_ratio/region_mean': 0.07294048415496945, 'reward_total_mean': 0.46270573139190674, 'reward_meter_mean': 0.7430543899536133, 'reward_meter_std': 0.28496313095092773, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7456576824188232, 'reward_repeat_soft_std': 0.07440777122974396, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.46270573139190674, 'reward_total_composite_std': 0.19669091701507568, 'epoch': 0.14}

 41%|████▏     | 1369/3300 [3:23:44<4:19:32,  8.06s/it]INFO 04-13 10:44:42 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1370/3300 [3:23:51<4:09:18,  7.75s/it]2026-04-13 10:44:49,280 | INFO | train_grpo_train | metrics_logged mode=train step=1370

                                                       
{'loss': 0.0231, 'grad_norm': 10.684398651123047, 'learning_rate': 5.851515151515152e-06, 'num_tokens': 2417878.0, 'completions/mean_length': 67.375, 'completions/min_length': 52.0, 'completions/max_length': 86.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.8285349607467651, 'rewards/meter/std': 0.17354312539100647, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8799509406089783, 'rewards/repeat_soft/std': 0.07719051092863083, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5655603408813477, 'rewards/total_composite/std': 0.10099063813686371, 'reward': 0.5655603408813477, 'reward_std': 0.10099063068628311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12360924482345581, 'sampling/sampling_logp_difference/max': 2.207282543182373, 'sampling/importance_sampling_ratio/min': 0.1099991574883461, 'sampling/importance_sampling_ratio/mean': 1.0059648752212524, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.764433428645134, 'clip_ratio/low_mean': 0.04938335530459881, 'clip_ratio/low_min': 0.04938335530459881, 'clip_ratio/high_mean': 0.05147484363988042, 'clip_ratio/high_max': 0.05147484363988042, 'clip_ratio/region_mean': 0.10085819894447923, 'reward_total_mean': 0.5655603408813477, 'reward_meter_mean': 0.8285349607467651, 'reward_meter_std': 0.17354312539100647, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8799509406089783, 'reward_repeat_soft_std': 0.07719051092863083, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5655603408813477, 'reward_total_composite_std': 0.10099063813686371, 'epoch': 0.14}

 42%|████▏     | 1370/3300 [3:23:51<4:09:18,  7.75s/it]INFO 04-13 10:44:49 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1371/3300 [3:23:57<3:49:24,  7.14s/it]2026-04-13 10:44:54,975 | INFO | train_grpo_train | metrics_logged mode=train step=1371

                                                       
{'loss': -0.0581, 'grad_norm': 15.889586448669434, 'learning_rate': 5.8484848484848485e-06, 'num_tokens': 2419231.0, 'completions/mean_length': 20.125, 'completions/min_length': 16.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.125, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9839861989021301, 'rewards/meter/std': 0.017121868208050728, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9169131517410278, 'rewards/repeat_soft/std': 0.048553213477134705, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.5842519998550415, 'rewards/total_composite/std': 0.05528772622346878, 'reward': 0.5842519998550415, 'reward_std': 0.05528773367404938, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14024120569229126, 'sampling/sampling_logp_difference/max': 1.1625232696533203, 'sampling/importance_sampling_ratio/min': 0.3126961886882782, 'sampling/importance_sampling_ratio/mean': 1.0222989320755005, 'sampling/importance_sampling_ratio/max': 1.7781554460525513, 'entropy': 1.002502478659153, 'clip_ratio/low_mean': 0.022518382407724857, 'clip_ratio/low_min': 0.022518382407724857, 'clip_ratio/high_mean': 0.12860644469037652, 'clip_ratio/high_max': 0.12860644469037652, 'clip_ratio/region_mean': 0.15112482709810138, 'reward_total_mean': 0.5842519998550415, 'reward_meter_mean': 0.9839861989021301, 'reward_meter_std': 0.017121868208050728, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9169131517410278, 'reward_repeat_soft_std': 0.048553213477134705, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.5842519998550415, 'reward_total_composite_std': 0.05528772622346878, 'epoch': 0.14}

 42%|████▏     | 1371/3300 [3:23:57<3:49:24,  7.14s/it]INFO 04-13 10:44:55 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1372/3300 [3:24:03<3:38:08,  6.79s/it]2026-04-13 10:45:00,955 | INFO | train_grpo_train | metrics_logged mode=train step=1372

                                                       
{'loss': 0.0497, 'grad_norm': 11.172639846801758, 'learning_rate': 5.845454545454547e-06, 'num_tokens': 2420871.0, 'completions/mean_length': 32.0, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.0, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9418027400970459, 'rewards/meter/std': 0.08665227890014648, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7936781644821167, 'rewards/repeat_soft/std': 0.04992871358990669, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.8017357587814331, 'rewards/total_composite/std': 0.14127589762210846, 'reward': 0.8017357587814331, 'reward_std': 0.14127591252326965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08830776065587997, 'sampling/sampling_logp_difference/max': 1.2683124542236328, 'sampling/importance_sampling_ratio/min': 0.2813059389591217, 'sampling/importance_sampling_ratio/mean': 1.003429889678955, 'sampling/importance_sampling_ratio/max': 1.9754638671875, 'entropy': 0.5145772732794285, 'clip_ratio/low_mean': 0.042360677383840084, 'clip_ratio/low_min': 0.042360677383840084, 'clip_ratio/high_mean': 0.023839198518544436, 'clip_ratio/high_max': 0.023839198518544436, 'clip_ratio/region_mean': 0.06619987590238452, 'reward_total_mean': 0.8017357587814331, 'reward_meter_mean': 0.9418027400970459, 'reward_meter_std': 0.08665227890014648, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7936781644821167, 'reward_repeat_soft_std': 0.04992871358990669, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.8017357587814331, 'reward_total_composite_std': 0.14127589762210846, 'epoch': 0.14}

 42%|████▏     | 1372/3300 [3:24:03<3:38:08,  6.79s/it]INFO 04-13 10:45:01 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1373/3300 [3:24:09<3:36:52,  6.75s/it]2026-04-13 10:45:07,625 | INFO | train_grpo_train | metrics_logged mode=train step=1373

                                                       
{'loss': 0.0186, 'grad_norm': 9.985396385192871, 'learning_rate': 5.842424242424243e-06, 'num_tokens': 2423047.0, 'completions/mean_length': 95.0, 'completions/min_length': 86.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.0, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.5731029510498047, 'rewards/meter/std': 0.37828680872917175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.838218092918396, 'rewards/repeat_soft/std': 0.07607004791498184, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.482189804315567, 'rewards/total_composite/std': 0.09745136648416519, 'reward': 0.482189804315567, 'reward_std': 0.0974513590335846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0944228395819664, 'sampling/sampling_logp_difference/max': 1.4184298515319824, 'sampling/importance_sampling_ratio/min': 0.24209386110305786, 'sampling/importance_sampling_ratio/mean': 1.0090051889419556, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.587487243115902, 'clip_ratio/low_mean': 0.036595333367586136, 'clip_ratio/low_min': 0.036595333367586136, 'clip_ratio/high_mean': 0.050420207902789116, 'clip_ratio/high_max': 0.050420207902789116, 'clip_ratio/region_mean': 0.08701554127037525, 'reward_total_mean': 0.482189804315567, 'reward_meter_mean': 0.5731029510498047, 'reward_meter_std': 0.37828680872917175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.838218092918396, 'reward_repeat_soft_std': 0.07607004791498184, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.482189804315567, 'reward_total_composite_std': 0.09745136648416519, 'epoch': 0.14}

 42%|████▏     | 1373/3300 [3:24:09<3:36:52,  6.75s/it]INFO 04-13 10:45:07 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1374/3300 [3:24:16<3:38:46,  6.82s/it]2026-04-13 10:45:14,586 | INFO | train_grpo_train | metrics_logged mode=train step=1374

                                                       
{'loss': 0.0507, 'grad_norm': 12.163618087768555, 'learning_rate': 5.83939393939394e-06, 'num_tokens': 2425055.0, 'completions/mean_length': 79.0, 'completions/min_length': 70.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.0, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.5774952173233032, 'rewards/meter/std': 0.26382485032081604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8827155232429504, 'rewards/repeat_soft/std': 0.06778127700090408, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.49006354808807373, 'rewards/total_composite/std': 0.06766549497842789, 'reward': 0.49006354808807373, 'reward_std': 0.06766548752784729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11644869297742844, 'sampling/sampling_logp_difference/max': 2.140930652618408, 'sampling/importance_sampling_ratio/min': 0.11754540354013443, 'sampling/importance_sampling_ratio/mean': 0.9974746704101562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5163152329623699, 'clip_ratio/low_mean': 0.0407986119389534, 'clip_ratio/low_min': 0.0407986119389534, 'clip_ratio/high_mean': 0.07150247041136026, 'clip_ratio/high_max': 0.07150247041136026, 'clip_ratio/region_mean': 0.11230108235031366, 'reward_total_mean': 0.49006354808807373, 'reward_meter_mean': 0.5774952173233032, 'reward_meter_std': 0.26382485032081604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8827155232429504, 'reward_repeat_soft_std': 0.06778127700090408, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.49006354808807373, 'reward_total_composite_std': 0.06766549497842789, 'epoch': 0.14}

 42%|████▏     | 1374/3300 [3:24:16<3:38:46,  6.82s/it]INFO 04-13 10:45:14 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1375/3300 [3:24:23<3:40:10,  6.86s/it]2026-04-13 10:45:21,557 | INFO | train_grpo_train | metrics_logged mode=train step=1375

                                                       
{'loss': 0.1259, 'grad_norm': 9.925308227539062, 'learning_rate': 5.836363636363637e-06, 'num_tokens': 2427042.0, 'completions/mean_length': 85.375, 'completions/min_length': 68.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.5373973846435547, 'rewards/meter/std': 0.3738399147987366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8409735560417175, 'rewards/repeat_soft/std': 0.06459219008684158, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4707145690917969, 'rewards/total_composite/std': 0.10587940365076065, 'reward': 0.4707145690917969, 'reward_std': 0.10587940365076065, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10384798049926758, 'sampling/sampling_logp_difference/max': 2.5874240398406982, 'sampling/importance_sampling_ratio/min': 0.07521353662014008, 'sampling/importance_sampling_ratio/mean': 1.0010960102081299, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5792411118745804, 'clip_ratio/low_mean': 0.04503625771030784, 'clip_ratio/low_min': 0.04503625771030784, 'clip_ratio/high_mean': 0.05577491782605648, 'clip_ratio/high_max': 0.05577491782605648, 'clip_ratio/region_mean': 0.10081117553636432, 'reward_total_mean': 0.4707145690917969, 'reward_meter_mean': 0.5373973846435547, 'reward_meter_std': 0.3738399147987366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8409735560417175, 'reward_repeat_soft_std': 0.06459219008684158, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4707145690917969, 'reward_total_composite_std': 0.10587940365076065, 'epoch': 0.14}

 42%|████▏     | 1375/3300 [3:24:23<3:40:10,  6.86s/it]INFO 04-13 10:45:21 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1376/3300 [3:24:34<4:20:08,  8.11s/it]2026-04-13 10:45:32,580 | INFO | train_grpo_train | metrics_logged mode=train step=1376

                                                       
{'loss': -0.1087, 'grad_norm': 1.4767781496047974, 'learning_rate': 5.833333333333334e-06, 'num_tokens': 2428574.0, 'completions/mean_length': 93.5, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.71428680419922, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8505613803863525, 'rewards/meter/std': 0.326339989900589, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8759030103683472, 'rewards/repeat_soft/std': 0.07500092685222626, 'rewards/judge_quality/mean': 0.39249998331069946, 'rewards/judge_quality/std': 0.13905291259288788, 'rewards/total_composite/mean': 0.530614972114563, 'rewards/total_composite/std': 0.215074360370636, 'reward': 0.530614972114563, 'reward_std': 0.21507437527179718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12482795119285583, 'sampling/sampling_logp_difference/max': 1.912602186203003, 'sampling/importance_sampling_ratio/min': 0.14769555628299713, 'sampling/importance_sampling_ratio/mean': 0.9998764991760254, 'sampling/importance_sampling_ratio/max': 1.6154453754425049, 'entropy': 0.688353419303894, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10693441238254309, 'clip_ratio/high_max': 0.10693441238254309, 'clip_ratio/region_mean': 0.10693441238254309, 'reward_total_mean': 0.530614972114563, 'reward_meter_mean': 0.8505613803863525, 'reward_meter_std': 0.326339989900589, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8759030103683472, 'reward_repeat_soft_std': 0.07500092685222626, 'reward_judge_quality_mean': 0.39249998331069946, 'reward_judge_quality_std': 0.13905291259288788, 'reward_total_composite_mean': 0.530614972114563, 'reward_total_composite_std': 0.215074360370636, 'epoch': 0.14}

 42%|████▏     | 1376/3300 [3:24:34<4:20:08,  8.11s/it]INFO 04-13 10:45:32 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1377/3300 [3:24:46<4:52:13,  9.12s/it]2026-04-13 10:45:44,043 | INFO | train_grpo_train | metrics_logged mode=train step=1377

                                                       
{'loss': -0.1533, 'grad_norm': 1.7564753293991089, 'learning_rate': 5.83030303030303e-06, 'num_tokens': 2430299.0, 'completions/mean_length': 115.625, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.000003814697266, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8179754018783569, 'rewards/meter/std': 0.17808018624782562, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.6838929653167725, 'rewards/repeat_soft/std': 0.06658895313739777, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4737984240055084, 'rewards/total_composite/std': 0.19232341647148132, 'reward': 0.4737984240055084, 'reward_std': 0.19232341647148132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09157626330852509, 'sampling/sampling_logp_difference/max': 2.145017623901367, 'sampling/importance_sampling_ratio/min': 0.11706597357988358, 'sampling/importance_sampling_ratio/mean': 1.0068401098251343, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4960397928953171, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08159125922247767, 'clip_ratio/high_max': 0.08159125922247767, 'clip_ratio/region_mean': 0.08159125922247767, 'reward_total_mean': 0.4737984240055084, 'reward_meter_mean': 0.8179754018783569, 'reward_meter_std': 0.17808018624782562, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.6838929653167725, 'reward_repeat_soft_std': 0.06658895313739777, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4737984240055084, 'reward_total_composite_std': 0.19232341647148132, 'epoch': 0.14}

 42%|████▏     | 1377/3300 [3:24:46<4:52:13,  9.12s/it]INFO 04-13 10:45:44 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1378/3300 [3:24:57<5:13:44,  9.79s/it]2026-04-13 10:45:55,417 | INFO | train_grpo_train | metrics_logged mode=train step=1378

                                                       
{'loss': -0.1543, 'grad_norm': 2.857304334640503, 'learning_rate': 5.8272727272727285e-06, 'num_tokens': 2432266.0, 'completions/mean_length': 127.875, 'completions/min_length': 69.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.0, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.6391924619674683, 'rewards/meter/std': 0.36736181378364563, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.80765700340271, 'rewards/repeat_soft/std': 0.09302859008312225, 'rewards/judge_quality/mean': 0.38999998569488525, 'rewards/judge_quality/std': 0.18845234811306, 'rewards/total_composite/mean': 0.4120762348175049, 'rewards/total_composite/std': 0.18186332285404205, 'reward': 0.4120762348175049, 'reward_std': 0.18186332285404205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13007330894470215, 'sampling/sampling_logp_difference/max': 1.6582809686660767, 'sampling/importance_sampling_ratio/min': 0.19046610593795776, 'sampling/importance_sampling_ratio/mean': 1.0254606008529663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6915091574192047, 'clip_ratio/low_mean': 0.025597664527595043, 'clip_ratio/low_min': 0.025597664527595043, 'clip_ratio/high_mean': 0.0639688279479742, 'clip_ratio/high_max': 0.0639688279479742, 'clip_ratio/region_mean': 0.08956649247556925, 'reward_total_mean': 0.4120762348175049, 'reward_meter_mean': 0.6391924619674683, 'reward_meter_std': 0.36736181378364563, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.80765700340271, 'reward_repeat_soft_std': 0.09302859008312225, 'reward_judge_quality_mean': 0.38999998569488525, 'reward_judge_quality_std': 0.18845234811306, 'reward_total_composite_mean': 0.4120762348175049, 'reward_total_composite_std': 0.18186332285404205, 'epoch': 0.14}

 42%|████▏     | 1378/3300 [3:24:57<5:13:44,  9.79s/it]INFO 04-13 10:45:55 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1379/3300 [3:25:03<4:35:17,  8.60s/it]2026-04-13 10:46:01,228 | INFO | train_grpo_train | metrics_logged mode=train step=1379

                                                       
{'loss': 0.0187, 'grad_norm': 14.179174423217773, 'learning_rate': 5.824242424242425e-06, 'num_tokens': 2433671.0, 'completions/mean_length': 19.625, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.980178952217102, 'rewards/meter/std': 0.017566950991749763, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8862823247909546, 'rewards/repeat_soft/std': 0.06156831234693527, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.6125080585479736, 'rewards/total_composite/std': 0.016003286466002464, 'reward': 0.6125080585479736, 'reward_std': 0.016003292053937912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08959906548261642, 'sampling/sampling_logp_difference/max': 0.9062271118164062, 'sampling/importance_sampling_ratio/min': 0.4040457606315613, 'sampling/importance_sampling_ratio/mean': 1.0124434232711792, 'sampling/importance_sampling_ratio/max': 1.6209818124771118, 'entropy': 0.6474045962095261, 'clip_ratio/low_mean': 0.03131265705451369, 'clip_ratio/low_min': 0.03131265705451369, 'clip_ratio/high_mean': 0.052736007142812014, 'clip_ratio/high_max': 0.052736007142812014, 'clip_ratio/region_mean': 0.0840486641973257, 'reward_total_mean': 0.6125080585479736, 'reward_meter_mean': 0.980178952217102, 'reward_meter_std': 0.017566950991749763, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8862823247909546, 'reward_repeat_soft_std': 0.06156831234693527, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.6125080585479736, 'reward_total_composite_std': 0.016003286466002464, 'epoch': 0.14}

 42%|████▏     | 1379/3300 [3:25:03<4:35:17,  8.60s/it]INFO 04-13 10:46:01 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1380/3300 [3:25:10<4:17:51,  8.06s/it]2026-04-13 10:46:08,029 | INFO | train_grpo_train | metrics_logged mode=train step=1380

                                                       
{'loss': -0.0135, 'grad_norm': 6.226532936096191, 'learning_rate': 5.821212121212122e-06, 'num_tokens': 2435696.0, 'completions/mean_length': 79.125, 'completions/min_length': 71.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.9815012812614441, 'rewards/meter/std': 0.021678613498806953, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7070019841194153, 'rewards/repeat_soft/std': 0.11901603639125824, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.546882152557373, 'rewards/total_composite/std': 0.06000892072916031, 'reward': 0.546882152557373, 'reward_std': 0.06000891700387001, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10609939694404602, 'sampling/sampling_logp_difference/max': 1.5791711807250977, 'sampling/importance_sampling_ratio/min': 0.20614588260650635, 'sampling/importance_sampling_ratio/mean': 1.0092766284942627, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6289454698562622, 'clip_ratio/low_mean': 0.017489806516095996, 'clip_ratio/low_min': 0.017489806516095996, 'clip_ratio/high_mean': 0.07526215631514788, 'clip_ratio/high_max': 0.07526215631514788, 'clip_ratio/region_mean': 0.09275196283124387, 'reward_total_mean': 0.546882152557373, 'reward_meter_mean': 0.9815012812614441, 'reward_meter_std': 0.021678613498806953, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7070019841194153, 'reward_repeat_soft_std': 0.11901603639125824, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.546882152557373, 'reward_total_composite_std': 0.06000892072916031, 'epoch': 0.14}

 42%|████▏     | 1380/3300 [3:25:10<4:17:51,  8.06s/it]INFO 04-13 10:46:08 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1381/3300 [3:25:16<4:01:57,  7.56s/it]2026-04-13 10:46:14,441 | INFO | train_grpo_train | metrics_logged mode=train step=1381

                                                       
{'loss': 0.0016, 'grad_norm': 9.563642501831055, 'learning_rate': 5.8181818181818185e-06, 'num_tokens': 2437704.0, 'completions/mean_length': 72.0, 'completions/min_length': 64.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9264421463012695, 'rewards/meter/std': 0.095173180103302, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8494632244110107, 'rewards/repeat_soft/std': 0.08679043501615524, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5803382396697998, 'rewards/total_composite/std': 0.024983327835798264, 'reward': 0.5803382396697998, 'reward_std': 0.024983324110507965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11786331236362457, 'sampling/sampling_logp_difference/max': 1.7849845886230469, 'sampling/importance_sampling_ratio/min': 0.16779965162277222, 'sampling/importance_sampling_ratio/mean': 1.0110652446746826, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7982779815793037, 'clip_ratio/low_mean': 0.05310388561338186, 'clip_ratio/low_min': 0.05310388561338186, 'clip_ratio/high_mean': 0.06156429462134838, 'clip_ratio/high_max': 0.06156429462134838, 'clip_ratio/region_mean': 0.11466818023473024, 'reward_total_mean': 0.5803382396697998, 'reward_meter_mean': 0.9264421463012695, 'reward_meter_std': 0.095173180103302, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8494632244110107, 'reward_repeat_soft_std': 0.08679043501615524, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5803382396697998, 'reward_total_composite_std': 0.024983327835798264, 'epoch': 0.14}

 42%|████▏     | 1381/3300 [3:25:16<4:01:57,  7.56s/it]INFO 04-13 10:46:14 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1382/3300 [3:25:23<3:52:28,  7.27s/it]2026-04-13 10:46:21,057 | INFO | train_grpo_train | metrics_logged mode=train step=1382

                                                       
{'loss': -0.0187, 'grad_norm': 7.579502582550049, 'learning_rate': 5.815151515151516e-06, 'num_tokens': 2439593.0, 'completions/mean_length': 59.125, 'completions/min_length': 54.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8978123664855957, 'rewards/meter/std': 0.1514429748058319, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6922234296798706, 'rewards/repeat_soft/std': 0.048303522169589996, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.21931305527687073, 'rewards/total_composite/mean': 0.5479089021682739, 'rewards/total_composite/std': 0.1442357897758484, 'reward': 0.5479089021682739, 'reward_std': 0.1442357748746872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08081500232219696, 'sampling/sampling_logp_difference/max': 1.7496185302734375, 'sampling/importance_sampling_ratio/min': 0.1738402396440506, 'sampling/importance_sampling_ratio/mean': 0.9880836606025696, 'sampling/importance_sampling_ratio/max': 1.9317129850387573, 'entropy': 0.3957347720861435, 'clip_ratio/low_mean': 0.04596590343862772, 'clip_ratio/low_min': 0.04596590343862772, 'clip_ratio/high_mean': 0.037464292254298925, 'clip_ratio/high_max': 0.037464292254298925, 'clip_ratio/region_mean': 0.08343019569292665, 'reward_total_mean': 0.5479089021682739, 'reward_meter_mean': 0.8978123664855957, 'reward_meter_std': 0.1514429748058319, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6922234296798706, 'reward_repeat_soft_std': 0.048303522169589996, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.21931305527687073, 'reward_total_composite_mean': 0.5479089021682739, 'reward_total_composite_std': 0.1442357897758484, 'epoch': 0.14}

 42%|████▏     | 1382/3300 [3:25:23<3:52:28,  7.27s/it]INFO 04-13 10:46:21 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1383/3300 [3:25:30<3:49:38,  7.19s/it]2026-04-13 10:46:28,267 | INFO | train_grpo_train | metrics_logged mode=train step=1383

                                                       
{'loss': 0.01, 'grad_norm': 7.723628520965576, 'learning_rate': 5.812121212121212e-06, 'num_tokens': 2441537.0, 'completions/mean_length': 70.0, 'completions/min_length': 66.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.0, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8299963474273682, 'rewards/meter/std': 0.13433019816875458, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7724564075469971, 'rewards/repeat_soft/std': 0.028093140572309494, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.531490683555603, 'rewards/total_composite/std': 0.04767953231930733, 'reward': 0.531490683555603, 'reward_std': 0.04767955094575882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0866677388548851, 'sampling/sampling_logp_difference/max': 2.966364860534668, 'sampling/importance_sampling_ratio/min': 0.05149014666676521, 'sampling/importance_sampling_ratio/mean': 1.0031169652938843, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42300498485565186, 'clip_ratio/low_mean': 0.02848917990922928, 'clip_ratio/low_min': 0.02848917990922928, 'clip_ratio/high_mean': 0.02885281457565725, 'clip_ratio/high_max': 0.02885281457565725, 'clip_ratio/region_mean': 0.05734199448488653, 'reward_total_mean': 0.531490683555603, 'reward_meter_mean': 0.8299963474273682, 'reward_meter_std': 0.13433019816875458, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7724564075469971, 'reward_repeat_soft_std': 0.028093140572309494, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.531490683555603, 'reward_total_composite_std': 0.04767953231930733, 'epoch': 0.14}

 42%|████▏     | 1383/3300 [3:25:30<3:49:38,  7.19s/it]INFO 04-13 10:46:28 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1384/3300 [3:25:42<4:39:17,  8.75s/it]2026-04-13 10:46:40,438 | INFO | train_grpo_train | metrics_logged mode=train step=1384

                                                       
{'loss': -0.1201, 'grad_norm': 2.4333841800689697, 'learning_rate': 5.8090909090909095e-06, 'num_tokens': 2443120.0, 'completions/mean_length': 106.875, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.26074719429016113, 'rewards/meter/std': 0.245651975274086, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.793143630027771, 'rewards/repeat_soft/std': 0.07925429940223694, 'rewards/judge_quality/mean': 0.33124998211860657, 'rewards/judge_quality/std': 0.13715866208076477, 'rewards/total_composite/mean': 0.3378831744194031, 'rewards/total_composite/std': 0.1523023545742035, 'reward': 0.3378831744194031, 'reward_std': 0.1523023545742035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10070676356554031, 'sampling/sampling_logp_difference/max': 2.923140525817871, 'sampling/importance_sampling_ratio/min': 0.053764570504426956, 'sampling/importance_sampling_ratio/mean': 1.0106135606765747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36102766171097755, 'clip_ratio/low_mean': 0.014150943607091904, 'clip_ratio/low_min': 0.014150943607091904, 'clip_ratio/high_mean': 0.06851313123479486, 'clip_ratio/high_max': 0.06851313123479486, 'clip_ratio/region_mean': 0.08266407484188676, 'reward_total_mean': 0.3378831744194031, 'reward_meter_mean': 0.26074719429016113, 'reward_meter_std': 0.245651975274086, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.793143630027771, 'reward_repeat_soft_std': 0.07925429940223694, 'reward_judge_quality_mean': 0.33124998211860657, 'reward_judge_quality_std': 0.13715866208076477, 'reward_total_composite_mean': 0.3378831744194031, 'reward_total_composite_std': 0.1523023545742035, 'epoch': 0.14}

 42%|████▏     | 1384/3300 [3:25:42<4:39:17,  8.75s/it]INFO 04-13 10:46:40 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1385/3300 [3:25:48<4:14:24,  7.97s/it]2026-04-13 10:46:46,631 | INFO | train_grpo_train | metrics_logged mode=train step=1385

                                                       
{'loss': -0.0587, 'grad_norm': 15.846782684326172, 'learning_rate': 5.806060606060606e-06, 'num_tokens': 2444513.0, 'completions/mean_length': 27.125, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9157686233520508, 'rewards/meter/std': 0.10688792914152145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.713201642036438, 'rewards/repeat_soft/std': 0.07831180840730667, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5479727983474731, 'rewards/total_composite/std': 0.06456225365400314, 'reward': 0.5479727983474731, 'reward_std': 0.06456225365400314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09401027858257294, 'sampling/sampling_logp_difference/max': 1.2109918594360352, 'sampling/importance_sampling_ratio/min': 0.2979016602039337, 'sampling/importance_sampling_ratio/mean': 0.9902519583702087, 'sampling/importance_sampling_ratio/max': 1.700728416442871, 'entropy': 0.46405426785349846, 'clip_ratio/low_mean': 0.005434782709926367, 'clip_ratio/low_min': 0.005434782709926367, 'clip_ratio/high_mean': 0.049688469618558884, 'clip_ratio/high_max': 0.049688469618558884, 'clip_ratio/region_mean': 0.05512325232848525, 'reward_total_mean': 0.5479727983474731, 'reward_meter_mean': 0.9157686233520508, 'reward_meter_std': 0.10688792914152145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.713201642036438, 'reward_repeat_soft_std': 0.07831180840730667, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5479727983474731, 'reward_total_composite_std': 0.06456225365400314, 'epoch': 0.14}

 42%|████▏     | 1385/3300 [3:25:48<4:14:24,  7.97s/it]INFO 04-13 10:46:46 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1386/3300 [3:25:55<4:04:32,  7.67s/it]2026-04-13 10:46:53,722 | INFO | train_grpo_train | metrics_logged mode=train step=1386

                                                       
{'loss': 0.0188, 'grad_norm': 10.31757640838623, 'learning_rate': 5.803030303030304e-06, 'num_tokens': 2446218.0, 'completions/mean_length': 53.125, 'completions/min_length': 45.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8301163911819458, 'rewards/meter/std': 0.29177701473236084, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7438743710517883, 'rewards/repeat_soft/std': 0.0749485120177269, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.5660555958747864, 'rewards/total_composite/std': 0.14403384923934937, 'reward': 0.5660555958747864, 'reward_std': 0.14403383433818817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0759248211979866, 'sampling/sampling_logp_difference/max': 1.408792495727539, 'sampling/importance_sampling_ratio/min': 0.2444382756948471, 'sampling/importance_sampling_ratio/mean': 1.0056344270706177, 'sampling/importance_sampling_ratio/max': 1.8468225002288818, 'entropy': 0.4119080901145935, 'clip_ratio/low_mean': 0.009722222341224551, 'clip_ratio/low_min': 0.009722222341224551, 'clip_ratio/high_mean': 0.04207471385598183, 'clip_ratio/high_max': 0.04207471385598183, 'clip_ratio/region_mean': 0.05179693619720638, 'reward_total_mean': 0.5660555958747864, 'reward_meter_mean': 0.8301163911819458, 'reward_meter_std': 0.29177701473236084, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7438743710517883, 'reward_repeat_soft_std': 0.0749485120177269, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.5660555958747864, 'reward_total_composite_std': 0.14403384923934937, 'epoch': 0.14}

 42%|████▏     | 1386/3300 [3:25:55<4:04:32,  7.67s/it]INFO 04-13 10:46:53 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1387/3300 [3:26:07<4:41:27,  8.83s/it]2026-04-13 10:47:05,054 | INFO | train_grpo_train | metrics_logged mode=train step=1387

                                                       
{'loss': -0.1763, 'grad_norm': 1.929757833480835, 'learning_rate': 5.8e-06, 'num_tokens': 2448452.0, 'completions/mean_length': 140.25, 'completions/min_length': 79.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 87.14286041259766, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9144402146339417, 'rewards/meter/std': 0.1888921856880188, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8584533333778381, 'rewards/repeat_soft/std': 0.05335886776447296, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5105130672454834, 'rewards/total_composite/std': 0.21891987323760986, 'reward': 0.5105130672454834, 'reward_std': 0.21891987323760986, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10738111287355423, 'sampling/sampling_logp_difference/max': 1.695192813873291, 'sampling/importance_sampling_ratio/min': 0.18356384336948395, 'sampling/importance_sampling_ratio/mean': 1.0207411050796509, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5428299009799957, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0956077417358756, 'clip_ratio/high_max': 0.0956077417358756, 'clip_ratio/region_mean': 0.0956077417358756, 'reward_total_mean': 0.5105130672454834, 'reward_meter_mean': 0.9144402146339417, 'reward_meter_std': 0.1888921856880188, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8584533333778381, 'reward_repeat_soft_std': 0.05335886776447296, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5105130672454834, 'reward_total_composite_std': 0.21891987323760986, 'epoch': 0.14}

 42%|████▏     | 1387/3300 [3:26:07<4:41:27,  8.83s/it]INFO 04-13 10:47:05 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1388/3300 [3:26:13<4:19:39,  8.15s/it]2026-04-13 10:47:11,620 | INFO | train_grpo_train | metrics_logged mode=train step=1388

                                                       
{'loss': -0.0345, 'grad_norm': 9.070403099060059, 'learning_rate': 5.796969696969698e-06, 'num_tokens': 2450718.0, 'completions/mean_length': 93.25, 'completions/min_length': 82.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.25, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.5084026455879211, 'rewards/meter/std': 0.409898042678833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7606096863746643, 'rewards/repeat_soft/std': 0.05007196590304375, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.4384540021419525, 'rewards/total_composite/std': 0.1053735613822937, 'reward': 0.4384540021419525, 'reward_std': 0.1053735539317131, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10059534758329391, 'sampling/sampling_logp_difference/max': 1.3359971046447754, 'sampling/importance_sampling_ratio/min': 0.2628959119319916, 'sampling/importance_sampling_ratio/mean': 1.0168665647506714, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5884588398039341, 'clip_ratio/low_mean': 0.038291629403829575, 'clip_ratio/low_min': 0.038291629403829575, 'clip_ratio/high_mean': 0.046265526209026575, 'clip_ratio/high_max': 0.046265526209026575, 'clip_ratio/region_mean': 0.08455715561285615, 'reward_total_mean': 0.4384540021419525, 'reward_meter_mean': 0.5084026455879211, 'reward_meter_std': 0.409898042678833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7606096863746643, 'reward_repeat_soft_std': 0.05007196590304375, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.4384540021419525, 'reward_total_composite_std': 0.1053735613822937, 'epoch': 0.14}

 42%|████▏     | 1388/3300 [3:26:13<4:19:39,  8.15s/it]INFO 04-13 10:47:11 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1389/3300 [3:26:24<4:46:07,  8.98s/it]2026-04-13 10:47:22,553 | INFO | train_grpo_train | metrics_logged mode=train step=1389

                                                       
{'loss': -0.1142, 'grad_norm': 2.0049874782562256, 'learning_rate': 5.793939393939394e-06, 'num_tokens': 2452175.0, 'completions/mean_length': 96.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 36.71428680419922, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6961934566497803, 'rewards/meter/std': 0.27985823154449463, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.817654013633728, 'rewards/repeat_soft/std': 0.04878468066453934, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.4702560305595398, 'rewards/total_composite/std': 0.19222553074359894, 'reward': 0.4702560305595398, 'reward_std': 0.19222553074359894, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13873133063316345, 'sampling/sampling_logp_difference/max': 1.80045747756958, 'sampling/importance_sampling_ratio/min': 0.16522328555583954, 'sampling/importance_sampling_ratio/mean': 1.0392733812332153, 'sampling/importance_sampling_ratio/max': 1.945328712463379, 'entropy': 0.7405487895011902, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08850523363798857, 'clip_ratio/high_max': 0.08850523363798857, 'clip_ratio/region_mean': 0.08850523363798857, 'reward_total_mean': 0.4702560305595398, 'reward_meter_mean': 0.6961934566497803, 'reward_meter_std': 0.27985823154449463, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.817654013633728, 'reward_repeat_soft_std': 0.04878468066453934, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.4702560305595398, 'reward_total_composite_std': 0.19222553074359894, 'epoch': 0.14}

 42%|████▏     | 1389/3300 [3:26:24<4:46:07,  8.98s/it]INFO 04-13 10:47:22 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1390/3300 [3:26:30<4:15:59,  8.04s/it]2026-04-13 10:47:28,406 | INFO | train_grpo_train | metrics_logged mode=train step=1390

                                                       
{'loss': -0.0411, 'grad_norm': 11.59184455871582, 'learning_rate': 5.790909090909091e-06, 'num_tokens': 2453881.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.0587540939450264, 'rewards/meter/std': 0.036901943385601044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8439459800720215, 'rewards/repeat_soft/std': 0.0702349916100502, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.3458341658115387, 'rewards/total_composite/std': 0.016634918749332428, 'reward': 0.3458341658115387, 'reward_std': 0.01663491502404213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09202395379543304, 'sampling/sampling_logp_difference/max': 1.4838995933532715, 'sampling/importance_sampling_ratio/min': 0.22675172984600067, 'sampling/importance_sampling_ratio/mean': 1.01513671875, 'sampling/importance_sampling_ratio/max': 1.8051742315292358, 'entropy': 0.5906520001590252, 'clip_ratio/low_mean': 0.027295285370200872, 'clip_ratio/low_min': 0.027295285370200872, 'clip_ratio/high_mean': 0.06077552307397127, 'clip_ratio/high_max': 0.06077552307397127, 'clip_ratio/region_mean': 0.08807080844417214, 'reward_total_mean': 0.3458341658115387, 'reward_meter_mean': 0.0587540939450264, 'reward_meter_std': 0.036901943385601044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8439459800720215, 'reward_repeat_soft_std': 0.0702349916100502, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.3458341658115387, 'reward_total_composite_std': 0.016634918749332428, 'epoch': 0.14}

 42%|████▏     | 1390/3300 [3:26:30<4:15:59,  8.04s/it]INFO 04-13 10:47:28 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1391/3300 [3:26:36<3:59:10,  7.52s/it]2026-04-13 10:47:34,697 | INFO | train_grpo_train | metrics_logged mode=train step=1391

                                                       
{'loss': 0.0251, 'grad_norm': 10.879137992858887, 'learning_rate': 5.787878787878788e-06, 'num_tokens': 2455597.0, 'completions/mean_length': 43.5, 'completions/min_length': 37.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9572920203208923, 'rewards/meter/std': 0.021061226725578308, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7102234363555908, 'rewards/repeat_soft/std': 0.09355749189853668, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.1249857097864151, 'rewards/total_composite/mean': 0.5172368288040161, 'rewards/total_composite/std': 0.07413040846586227, 'reward': 0.5172368288040161, 'reward_std': 0.07413038611412048, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08461681008338928, 'sampling/sampling_logp_difference/max': 1.3330113887786865, 'sampling/importance_sampling_ratio/min': 0.2636820375919342, 'sampling/importance_sampling_ratio/mean': 1.00737464427948, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4708849526941776, 'clip_ratio/low_mean': 0.025728388223797083, 'clip_ratio/low_min': 0.025728388223797083, 'clip_ratio/high_mean': 0.059031757060438395, 'clip_ratio/high_max': 0.059031757060438395, 'clip_ratio/region_mean': 0.08476014528423548, 'reward_total_mean': 0.5172368288040161, 'reward_meter_mean': 0.9572920203208923, 'reward_meter_std': 0.021061226725578308, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7102234363555908, 'reward_repeat_soft_std': 0.09355749189853668, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.1249857097864151, 'reward_total_composite_mean': 0.5172368288040161, 'reward_total_composite_std': 0.07413040846586227, 'epoch': 0.14}

 42%|████▏     | 1391/3300 [3:26:36<3:59:10,  7.52s/it]INFO 04-13 10:47:34 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1392/3300 [3:26:42<3:41:13,  6.96s/it]2026-04-13 10:47:40,349 | INFO | train_grpo_train | metrics_logged mode=train step=1392

                                                       
{'loss': -0.0233, 'grad_norm': 17.71356773376465, 'learning_rate': 5.784848484848486e-06, 'num_tokens': 2456934.0, 'completions/mean_length': 21.125, 'completions/min_length': 17.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.8005595207214355, 'rewards/meter/std': 0.3589307367801666, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9440689086914062, 'rewards/repeat_soft/std': 0.029856212437152863, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.24976776540279388, 'rewards/total_composite/mean': 0.6497457027435303, 'rewards/total_composite/std': 0.20137529075145721, 'reward': 0.6497457027435303, 'reward_std': 0.20137529075145721, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11411852389574051, 'sampling/sampling_logp_difference/max': 0.8509979248046875, 'sampling/importance_sampling_ratio/min': 0.4269886016845703, 'sampling/importance_sampling_ratio/mean': 1.0057789087295532, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8141906559467316, 'clip_ratio/low_mean': 0.07073577027767897, 'clip_ratio/low_min': 0.07073577027767897, 'clip_ratio/high_mean': 0.017316017765551805, 'clip_ratio/high_max': 0.017316017765551805, 'clip_ratio/region_mean': 0.08805178804323077, 'reward_total_mean': 0.6497457027435303, 'reward_meter_mean': 0.8005595207214355, 'reward_meter_std': 0.3589307367801666, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9440689086914062, 'reward_repeat_soft_std': 0.029856212437152863, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.24976776540279388, 'reward_total_composite_mean': 0.6497457027435303, 'reward_total_composite_std': 0.20137529075145721, 'epoch': 0.14}

 42%|████▏     | 1392/3300 [3:26:42<3:41:13,  6.96s/it]INFO 04-13 10:47:40 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1393/3300 [3:26:48<3:32:32,  6.69s/it]2026-04-13 10:47:46,406 | INFO | train_grpo_train | metrics_logged mode=train step=1393

                                                       
{'loss': 0.0317, 'grad_norm': 22.41649627685547, 'learning_rate': 5.781818181818181e-06, 'num_tokens': 2458266.0, 'completions/mean_length': 22.5, 'completions/min_length': 17.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8784810304641724, 'rewards/meter/std': 0.2654203772544861, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8913486003875732, 'rewards/repeat_soft/std': 0.07164983451366425, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.565843939781189, 'rewards/total_composite/std': 0.07566385716199875, 'reward': 0.565843939781189, 'reward_std': 0.07566386461257935, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13356241583824158, 'sampling/sampling_logp_difference/max': 1.678497076034546, 'sampling/importance_sampling_ratio/min': 0.18665429949760437, 'sampling/importance_sampling_ratio/mean': 1.012643814086914, 'sampling/importance_sampling_ratio/max': 1.7101576328277588, 'entropy': 0.9071509465575218, 'clip_ratio/low_mean': 0.03941993601620197, 'clip_ratio/low_min': 0.03941993601620197, 'clip_ratio/high_mean': 0.05967909004539251, 'clip_ratio/high_max': 0.05967909004539251, 'clip_ratio/region_mean': 0.09909902606159449, 'reward_total_mean': 0.565843939781189, 'reward_meter_mean': 0.8784810304641724, 'reward_meter_std': 0.2654203772544861, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8913486003875732, 'reward_repeat_soft_std': 0.07164983451366425, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.565843939781189, 'reward_total_composite_std': 0.07566385716199875, 'epoch': 0.14}

 42%|████▏     | 1393/3300 [3:26:48<3:32:32,  6.69s/it]INFO 04-13 10:47:46 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1394/3300 [3:26:54<3:28:45,  6.57s/it]2026-04-13 10:47:52,708 | INFO | train_grpo_train | metrics_logged mode=train step=1394

                                                       
{'loss': -0.0054, 'grad_norm': 11.413646697998047, 'learning_rate': 5.7787878787878795e-06, 'num_tokens': 2459885.0, 'completions/mean_length': 42.375, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9350890517234802, 'rewards/meter/std': 0.12849915027618408, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8567144870758057, 'rewards/repeat_soft/std': 0.08090715855360031, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.6157437562942505, 'rewards/total_composite/std': 0.07533933222293854, 'reward': 0.6157437562942505, 'reward_std': 0.07533932477235794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13664910197257996, 'sampling/sampling_logp_difference/max': 1.7999773025512695, 'sampling/importance_sampling_ratio/min': 0.16530264914035797, 'sampling/importance_sampling_ratio/mean': 0.99691241979599, 'sampling/importance_sampling_ratio/max': 1.8642349243164062, 'entropy': 0.7245934903621674, 'clip_ratio/low_mean': 0.07677233032882214, 'clip_ratio/low_min': 0.07677233032882214, 'clip_ratio/high_mean': 0.020740161649882793, 'clip_ratio/high_max': 0.020740161649882793, 'clip_ratio/region_mean': 0.09751249197870493, 'reward_total_mean': 0.6157437562942505, 'reward_meter_mean': 0.9350890517234802, 'reward_meter_std': 0.12849915027618408, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8567144870758057, 'reward_repeat_soft_std': 0.08090715855360031, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.6157437562942505, 'reward_total_composite_std': 0.07533933222293854, 'epoch': 0.14}

 42%|████▏     | 1394/3300 [3:26:54<3:28:45,  6.57s/it]INFO 04-13 10:47:52 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1395/3300 [3:27:00<3:24:34,  6.44s/it]2026-04-13 10:47:58,855 | INFO | train_grpo_train | metrics_logged mode=train step=1395

                                                       
{'loss': -0.0168, 'grad_norm': 9.653735160827637, 'learning_rate': 5.775757575757577e-06, 'num_tokens': 2461346.0, 'completions/mean_length': 37.625, 'completions/min_length': 34.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9894571900367737, 'rewards/meter/std': 0.009738119319081306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7672044038772583, 'rewards/repeat_soft/std': 0.07762787491083145, 'rewards/judge_quality/mean': 0.24250000715255737, 'rewards/judge_quality/std': 0.11792854964733124, 'rewards/total_composite/mean': 0.47067469358444214, 'rewards/total_composite/std': 0.08534766733646393, 'reward': 0.47067469358444214, 'reward_std': 0.08534765243530273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09565913677215576, 'sampling/sampling_logp_difference/max': 1.7760858535766602, 'sampling/importance_sampling_ratio/min': 0.16929951310157776, 'sampling/importance_sampling_ratio/mean': 1.0008505582809448, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5285905115306377, 'clip_ratio/low_mean': 0.026665922487154603, 'clip_ratio/low_min': 0.026665922487154603, 'clip_ratio/high_mean': 0.0553113566711545, 'clip_ratio/high_max': 0.0553113566711545, 'clip_ratio/region_mean': 0.0819772791583091, 'reward_total_mean': 0.47067469358444214, 'reward_meter_mean': 0.9894571900367737, 'reward_meter_std': 0.009738119319081306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7672044038772583, 'reward_repeat_soft_std': 0.07762787491083145, 'reward_judge_quality_mean': 0.24250000715255737, 'reward_judge_quality_std': 0.11792854964733124, 'reward_total_composite_mean': 0.47067469358444214, 'reward_total_composite_std': 0.08534766733646393, 'epoch': 0.14}

 42%|████▏     | 1395/3300 [3:27:01<3:24:34,  6.44s/it]INFO 04-13 10:47:59 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1396/3300 [3:27:07<3:28:50,  6.58s/it]2026-04-13 10:48:05,758 | INFO | train_grpo_train | metrics_logged mode=train step=1396

                                                       
{'loss': 0.0376, 'grad_norm': 11.208836555480957, 'learning_rate': 5.772727272727273e-06, 'num_tokens': 2462702.0, 'completions/mean_length': 26.5, 'completions/min_length': 21.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.25748664140701294, 'rewards/meter/std': 0.43729761242866516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.8025000095367432, 'rewards/judge_quality/std': 0.21756774187088013, 'rewards/total_composite/mean': 0.4982941150665283, 'rewards/total_composite/std': 0.2615428566932678, 'reward': 0.4982941150665283, 'reward_std': 0.2615428566932678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1218995675444603, 'sampling/sampling_logp_difference/max': 1.2938337326049805, 'sampling/importance_sampling_ratio/min': 0.27421748638153076, 'sampling/importance_sampling_ratio/mean': 1.0340930223464966, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8053631111979485, 'clip_ratio/low_mean': 0.11612405255436897, 'clip_ratio/low_min': 0.11612405255436897, 'clip_ratio/high_mean': 0.02425925899296999, 'clip_ratio/high_max': 0.02425925899296999, 'clip_ratio/region_mean': 0.14038331154733896, 'reward_total_mean': 0.4982941150665283, 'reward_meter_mean': 0.25748664140701294, 'reward_meter_std': 0.43729761242866516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.8025000095367432, 'reward_judge_quality_std': 0.21756774187088013, 'reward_total_composite_mean': 0.4982941150665283, 'reward_total_composite_std': 0.2615428566932678, 'epoch': 0.14}

 42%|████▏     | 1396/3300 [3:27:07<3:28:50,  6.58s/it]INFO 04-13 10:48:06 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1397/3300 [3:27:14<3:29:07,  6.59s/it]2026-04-13 10:48:12,380 | INFO | train_grpo_train | metrics_logged mode=train step=1397

                                                       
{'loss': -0.0281, 'grad_norm': 7.8243865966796875, 'learning_rate': 5.76969696969697e-06, 'num_tokens': 2464378.0, 'completions/mean_length': 44.5, 'completions/min_length': 35.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9040476083755493, 'rewards/meter/std': 0.21422027051448822, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8503400087356567, 'rewards/repeat_soft/std': 0.05986413732171059, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5766725540161133, 'rewards/total_composite/std': 0.06210015341639519, 'reward': 0.5766725540161133, 'reward_std': 0.06210014224052429, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1364356279373169, 'sampling/sampling_logp_difference/max': 2.3001978397369385, 'sampling/importance_sampling_ratio/min': 0.10023900866508484, 'sampling/importance_sampling_ratio/mean': 0.9939990043640137, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6070137321949005, 'clip_ratio/low_mean': 0.016447369009256363, 'clip_ratio/low_min': 0.016447369009256363, 'clip_ratio/high_mean': 0.10732440184801817, 'clip_ratio/high_max': 0.10732440184801817, 'clip_ratio/region_mean': 0.12377177085727453, 'reward_total_mean': 0.5766725540161133, 'reward_meter_mean': 0.9040476083755493, 'reward_meter_std': 0.21422027051448822, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8503400087356567, 'reward_repeat_soft_std': 0.05986413732171059, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5766725540161133, 'reward_total_composite_std': 0.06210015341639519, 'epoch': 0.14}

 42%|████▏     | 1397/3300 [3:27:14<3:29:07,  6.59s/it]INFO 04-13 10:48:12 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1398/3300 [3:27:20<3:25:21,  6.48s/it]2026-04-13 10:48:18,583 | INFO | train_grpo_train | metrics_logged mode=train step=1398

                                                       
{'loss': 0.0651, 'grad_norm': 13.051623344421387, 'learning_rate': 5.766666666666667e-06, 'num_tokens': 2466024.0, 'completions/mean_length': 37.75, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.6293722987174988, 'rewards/meter/std': 0.3659643232822418, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9026373624801636, 'rewards/repeat_soft/std': 0.023215971887111664, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5114091634750366, 'rewards/total_composite/std': 0.09964381158351898, 'reward': 0.5114091634750366, 'reward_std': 0.09964380413293839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15148068964481354, 'sampling/sampling_logp_difference/max': 1.5002524852752686, 'sampling/importance_sampling_ratio/min': 0.22307385504245758, 'sampling/importance_sampling_ratio/mean': 0.9983512163162231, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.840466633439064, 'clip_ratio/low_mean': 0.0661195544525981, 'clip_ratio/low_min': 0.0661195544525981, 'clip_ratio/high_mean': 0.06789673771709204, 'clip_ratio/high_max': 0.06789673771709204, 'clip_ratio/region_mean': 0.13401629216969013, 'reward_total_mean': 0.5114091634750366, 'reward_meter_mean': 0.6293722987174988, 'reward_meter_std': 0.3659643232822418, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9026373624801636, 'reward_repeat_soft_std': 0.023215971887111664, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5114091634750366, 'reward_total_composite_std': 0.09964381158351898, 'epoch': 0.14}

 42%|████▏     | 1398/3300 [3:27:20<3:25:21,  6.48s/it]INFO 04-13 10:48:18 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1399/3300 [3:27:27<3:24:20,  6.45s/it]2026-04-13 10:48:24,971 | INFO | train_grpo_train | metrics_logged mode=train step=1399

                                                       
{'loss': -0.0003, 'grad_norm': 9.11116886138916, 'learning_rate': 5.763636363636365e-06, 'num_tokens': 2467635.0, 'completions/mean_length': 43.375, 'completions/min_length': 39.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9829332828521729, 'rewards/meter/std': 0.010218682698905468, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8421238660812378, 'rewards/repeat_soft/std': 0.06742725521326065, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.5905901193618774, 'rewards/total_composite/std': 0.04515933245420456, 'reward': 0.5905901193618774, 'reward_std': 0.045159339904785156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11853578686714172, 'sampling/sampling_logp_difference/max': 1.983877182006836, 'sampling/importance_sampling_ratio/min': 0.137534961104393, 'sampling/importance_sampling_ratio/mean': 1.008230447769165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6134412474930286, 'clip_ratio/low_mean': 0.026244589127600193, 'clip_ratio/low_min': 0.026244589127600193, 'clip_ratio/high_mean': 0.0931791178882122, 'clip_ratio/high_max': 0.0931791178882122, 'clip_ratio/region_mean': 0.1194237070158124, 'reward_total_mean': 0.5905901193618774, 'reward_meter_mean': 0.9829332828521729, 'reward_meter_std': 0.010218682698905468, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8421238660812378, 'reward_repeat_soft_std': 0.06742725521326065, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.5905901193618774, 'reward_total_composite_std': 0.04515933245420456, 'epoch': 0.14}

 42%|████▏     | 1399/3300 [3:27:27<3:24:20,  6.45s/it]INFO 04-13 10:48:25 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1400/3300 [3:27:37<4:03:21,  7.68s/it]2026-04-13 10:48:35,541 | INFO | train_grpo_train | metrics_logged mode=train step=1400

                                                       
{'loss': 0.0561, 'grad_norm': 11.920273780822754, 'learning_rate': 5.760606060606061e-06, 'num_tokens': 2469500.0, 'completions/mean_length': 61.125, 'completions/min_length': 53.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9751566648483276, 'rewards/meter/std': 0.01855885051190853, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8131760954856873, 'rewards/repeat_soft/std': 0.07321101427078247, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5881941914558411, 'rewards/total_composite/std': 0.012843296863138676, 'reward': 0.5881941914558411, 'reward_std': 0.01284329779446125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14106321334838867, 'sampling/sampling_logp_difference/max': 2.0515499114990234, 'sampling/importance_sampling_ratio/min': 0.12853553891181946, 'sampling/importance_sampling_ratio/mean': 0.9969595670700073, 'sampling/importance_sampling_ratio/max': 1.8442491292953491, 'entropy': 0.721710816025734, 'clip_ratio/low_mean': 0.06279363110661507, 'clip_ratio/low_min': 0.06279363110661507, 'clip_ratio/high_mean': 0.07298974320292473, 'clip_ratio/high_max': 0.07298974320292473, 'clip_ratio/region_mean': 0.1357833743095398, 'reward_total_mean': 0.5881941914558411, 'reward_meter_mean': 0.9751566648483276, 'reward_meter_std': 0.01855885051190853, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8131760954856873, 'reward_repeat_soft_std': 0.07321101427078247, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5881941914558411, 'reward_total_composite_std': 0.012843296863138676, 'epoch': 0.14}

 42%|████▏     | 1400/3300 [3:27:37<4:03:21,  7.68s/it]INFO 04-13 10:48:35 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:13,  1.66s/it]

 30%|███       | 3/10 [00:07<00:19,  2.72s/it]

 40%|████      | 4/10 [00:10<00:17,  2.92s/it]

 50%|█████     | 5/10 [00:14<00:16,  3.30s/it]

 60%|██████    | 6/10 [00:18<00:14,  3.51s/it]

 70%|███████   | 7/10 [00:23<00:11,  3.76s/it]

 80%|████████  | 8/10 [00:26<00:07,  3.75s/it]

 90%|█████████ | 9/10 [00:30<00:03,  3.65s/it]

100%|██████████| 10/10 [00:33<00:00,  3.55s/it]2026-04-13 10:49:13,707 | INFO | train_grpo_train | metrics_logged mode=eval step=1400

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 38.0793, 'eval_samples_per_second': 2.101, 'eval_steps_per_second': 0.263, 'eval_num_tokens': 2469500.0, 'eval_completions/mean_length': 66.6375, 'eval_completions/min_length': 32.7, 'eval_completions/max_length': 109.7, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 66.6375, 'eval_completions/min_terminated_length': 32.7, 'eval_completions/max_terminated_length': 109.7, 'eval_rewards/meter/mean': 0.8095608830451966, 'eval_rewards/meter/std': 0.2641849434003234, 'eval_rewards/count_adherence/mean': 0.96583331823349, 'eval_rewards/count_adherence/std': 0.06645093113183975, 'eval_rewards/hard_gate/mean': 0.9875, 'eval_rewards/hard_gate/std': 0.03535533845424652, 'eval_rewards/repeat_soft/mean': 0.8524089634418488, 'eval_rewards/repeat_soft/std': 0.10649546608328819, 'eval_rewards/judge_quality/mean': 0.45062499940395356, 'eval_rewards/judge_quality/std': 0.15118036419153214, 'eval_rewards/total_composite/mean': 0.5535447299480438, 'eval_rewards/total_composite/std': 0.1373872399330139, 'eval_reward': 0.5535447299480438, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.056691231578588484, 'eval_sampling/sampling_logp_difference/max': 0.9773582935333252, 'eval_sampling/importance_sampling_ratio/min': 0.3831986427307129, 'eval_sampling/importance_sampling_ratio/mean': 1.0140406012535095, 'eval_sampling/importance_sampling_ratio/max': 1.4029753446578979, 'eval_entropy': 0.6156320989131927, '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': 0.5535447299480438, 'eval_reward_meter_mean': 0.8095608830451966, 'eval_reward_meter_std': 0.2641849434003234, 'eval_reward_count_adherence_mean': 0.96583331823349, 'eval_reward_count_adherence_std': 0.06645093113183975, 'eval_reward_hard_gate_mean': 0.9875, 'eval_reward_hard_gate_std': 0.03535533845424652, 'eval_reward_repeat_soft_mean': 0.8524089634418488, 'eval_reward_repeat_soft_std': 0.10649546608328819, 'eval_reward_judge_quality_mean': 0.45062499940395356, 'eval_reward_judge_quality_std': 0.15118036419153214, 'eval_reward_total_composite_mean': 0.5535447299480438, 'eval_reward_total_composite_std': 0.1373872399330139, 'epoch': 0.14}

 42%|████▏     | 1400/3300 [3:28:15<4:03:21,  7.68s/it]

100%|██████████| 10/10 [00:33<00:00,  3.55s/it]

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

 42%|████▏     | 1401/3300 [3:28:25<10:20:39, 19.61s/it]2026-04-13 10:49:22,970 | INFO | train_grpo_train | metrics_logged mode=train step=1401

                                                        
{'loss': 0.0659, 'grad_norm': 17.8997802734375, 'learning_rate': 5.7575757575757586e-06, 'num_tokens': 2470883.0, 'completions/mean_length': 37.875, 'completions/min_length': 33.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.5351673364639282, 'rewards/meter/std': 0.3454230725765228, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8383328914642334, 'rewards/repeat_soft/std': 0.1236555278301239, 'rewards/judge_quality/mean': 0.5112500190734863, 'rewards/judge_quality/std': 0.268936425447464, 'rewards/total_composite/mean': 0.5070319771766663, 'rewards/total_composite/std': 0.17160125076770782, 'reward': 0.5070319771766663, 'reward_std': 0.17160125076770782, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11590097099542618, 'sampling/sampling_logp_difference/max': 1.930302619934082, 'sampling/importance_sampling_ratio/min': 0.1451042741537094, 'sampling/importance_sampling_ratio/mean': 0.9949010610580444, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6245596781373024, 'clip_ratio/low_mean': 0.04386756452731788, 'clip_ratio/low_min': 0.04386756452731788, 'clip_ratio/high_mean': 0.05764055158942938, 'clip_ratio/high_max': 0.05764055158942938, 'clip_ratio/region_mean': 0.10150811611674726, 'reward_total_mean': 0.5070319771766663, 'reward_meter_mean': 0.5351673364639282, 'reward_meter_std': 0.3454230725765228, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8383328914642334, 'reward_repeat_soft_std': 0.1236555278301239, 'reward_judge_quality_mean': 0.5112500190734863, 'reward_judge_quality_std': 0.268936425447464, 'reward_total_composite_mean': 0.5070319771766663, 'reward_total_composite_std': 0.17160125076770782, 'epoch': 0.14}

 42%|████▏     | 1401/3300 [3:28:25<10:20:39, 19.61s/it]INFO 04-13 10:49:23 [block_pool.py:378] Successfully reset prefix cache

 42%|████▏     | 1402/3300 [3:28:32<8:21:09, 15.84s/it] 2026-04-13 10:49:30,020 | INFO | train_grpo_train | metrics_logged mode=train step=1402

                                                       
{'loss': 0.0051, 'grad_norm': 10.03658390045166, 'learning_rate': 5.754545454545455e-06, 'num_tokens': 2472649.0, 'completions/mean_length': 63.75, 'completions/min_length': 58.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9593009948730469, 'rewards/meter/std': 0.08032622188329697, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9324356317520142, 'rewards/repeat_soft/std': 0.042951393872499466, 'rewards/judge_quality/mean': 0.5149999856948853, 'rewards/judge_quality/std': 0.1810288280248642, 'rewards/total_composite/mean': 0.6627146005630493, 'rewards/total_composite/std': 0.12697963416576385, 'reward': 0.6627146005630493, 'reward_std': 0.12697963416576385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11571583896875381, 'sampling/sampling_logp_difference/max': 3.0238711833953857, 'sampling/importance_sampling_ratio/min': 0.04861266538500786, 'sampling/importance_sampling_ratio/mean': 1.010787010192871, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.593965407460928, 'clip_ratio/low_mean': 0.08435330679640174, 'clip_ratio/low_min': 0.08435330679640174, 'clip_ratio/high_mean': 0.03256302606314421, 'clip_ratio/high_max': 0.03256302606314421, 'clip_ratio/region_mean': 0.11691633285954595, 'reward_total_mean': 0.6627146005630493, 'reward_meter_mean': 0.9593009948730469, 'reward_meter_std': 0.08032622188329697, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9324356317520142, 'reward_repeat_soft_std': 0.042951393872499466, 'reward_judge_quality_mean': 0.5149999856948853, 'reward_judge_quality_std': 0.1810288280248642, 'reward_total_composite_mean': 0.6627146005630493, 'reward_total_composite_std': 0.12697963416576385, 'epoch': 0.14}

 42%|████▏     | 1402/3300 [3:28:32<8:21:09, 15.84s/it]INFO 04-13 10:49:30 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1403/3300 [3:28:38<6:50:10, 12.97s/it]2026-04-13 10:49:36,305 | INFO | train_grpo_train | metrics_logged mode=train step=1403

                                                       
{'loss': -0.011, 'grad_norm': 15.948739051818848, 'learning_rate': 5.751515151515152e-06, 'num_tokens': 2474383.0, 'completions/mean_length': 36.75, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9653463363647461, 'rewards/meter/std': 0.028619877994060516, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8667511343955994, 'rewards/repeat_soft/std': 0.09769817441701889, 'rewards/judge_quality/mean': 0.4050000011920929, 'rewards/judge_quality/std': 0.10392304509878159, 'rewards/total_composite/mean': 0.5837519764900208, 'rewards/total_composite/std': 0.07472193986177444, 'reward': 0.5837519764900208, 'reward_std': 0.07472193986177444, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12553322315216064, 'sampling/sampling_logp_difference/max': 1.3544254302978516, 'sampling/importance_sampling_ratio/min': 0.25809556245803833, 'sampling/importance_sampling_ratio/mean': 1.013904094696045, 'sampling/importance_sampling_ratio/max': 1.9975882768630981, 'entropy': 0.6746665462851524, 'clip_ratio/low_mean': 0.0234375, 'clip_ratio/low_min': 0.0234375, 'clip_ratio/high_mean': 0.09110360452905297, 'clip_ratio/high_max': 0.09110360452905297, 'clip_ratio/region_mean': 0.11454110452905297, 'reward_total_mean': 0.5837519764900208, 'reward_meter_mean': 0.9653463363647461, 'reward_meter_std': 0.028619877994060516, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8667511343955994, 'reward_repeat_soft_std': 0.09769817441701889, 'reward_judge_quality_mean': 0.4050000011920929, 'reward_judge_quality_std': 0.10392304509878159, 'reward_total_composite_mean': 0.5837519764900208, 'reward_total_composite_std': 0.07472193986177444, 'epoch': 0.14}

 43%|████▎     | 1403/3300 [3:28:38<6:50:10, 12.97s/it]INFO 04-13 10:49:36 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1404/3300 [3:28:46<6:01:13, 11.43s/it]2026-04-13 10:49:44,130 | INFO | train_grpo_train | metrics_logged mode=train step=1404

                                                       
{'loss': 0.0233, 'grad_norm': 10.28912353515625, 'learning_rate': 5.748484848484849e-06, 'num_tokens': 2476611.0, 'completions/mean_length': 78.5, 'completions/min_length': 70.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.5, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.5133029222488403, 'rewards/meter/std': 0.36811619997024536, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8736454844474792, 'rewards/repeat_soft/std': 0.06013016775250435, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5057510137557983, 'rewards/total_composite/std': 0.1692267805337906, 'reward': 0.5057510137557983, 'reward_std': 0.1692267656326294, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11171042174100876, 'sampling/sampling_logp_difference/max': 2.0017662048339844, 'sampling/importance_sampling_ratio/min': 0.1350964605808258, 'sampling/importance_sampling_ratio/mean': 1.0089904069900513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4880719818174839, 'clip_ratio/low_mean': 0.05477775679901242, 'clip_ratio/low_min': 0.05477775679901242, 'clip_ratio/high_mean': 0.033478156197816133, 'clip_ratio/high_max': 0.033478156197816133, 'clip_ratio/region_mean': 0.08825591299682856, 'reward_total_mean': 0.5057510137557983, 'reward_meter_mean': 0.5133029222488403, 'reward_meter_std': 0.36811619997024536, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8736454844474792, 'reward_repeat_soft_std': 0.06013016775250435, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5057510137557983, 'reward_total_composite_std': 0.1692267805337906, 'epoch': 0.14}

 43%|████▎     | 1404/3300 [3:28:46<6:01:13, 11.43s/it]INFO 04-13 10:49:44 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1405/3300 [3:28:52<5:14:09,  9.95s/it]2026-04-13 10:49:50,619 | INFO | train_grpo_train | metrics_logged mode=train step=1405

                                                       
{'loss': 0.0774, 'grad_norm': 14.689164161682129, 'learning_rate': 5.745454545454546e-06, 'num_tokens': 2478191.0, 'completions/mean_length': 40.5, 'completions/min_length': 34.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7688623070716858, 'rewards/meter/std': 0.3516474962234497, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8841646909713745, 'rewards/repeat_soft/std': 0.08102813363075256, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5471103191375732, 'rewards/total_composite/std': 0.09307053685188293, 'reward': 0.5471103191375732, 'reward_std': 0.09307053685188293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1309376060962677, 'sampling/sampling_logp_difference/max': 1.5793876647949219, 'sampling/importance_sampling_ratio/min': 0.20610125362873077, 'sampling/importance_sampling_ratio/mean': 1.02289879322052, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8622675091028214, 'clip_ratio/low_mean': 0.03683094959706068, 'clip_ratio/low_min': 0.03683094959706068, 'clip_ratio/high_mean': 0.10103944409638643, 'clip_ratio/high_max': 0.10103944409638643, 'clip_ratio/region_mean': 0.1378703936934471, 'reward_total_mean': 0.5471103191375732, 'reward_meter_mean': 0.7688623070716858, 'reward_meter_std': 0.3516474962234497, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8841646909713745, 'reward_repeat_soft_std': 0.08102813363075256, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5471103191375732, 'reward_total_composite_std': 0.09307053685188293, 'epoch': 0.14}

 43%|████▎     | 1405/3300 [3:28:52<5:14:09,  9.95s/it]INFO 04-13 10:49:50 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1406/3300 [3:28:59<4:41:24,  8.91s/it]2026-04-13 10:49:57,123 | INFO | train_grpo_train | metrics_logged mode=train step=1406

                                                       
{'loss': 0.0302, 'grad_norm': 12.480927467346191, 'learning_rate': 5.742424242424242e-06, 'num_tokens': 2480060.0, 'completions/mean_length': 58.625, 'completions/min_length': 48.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.7705498933792114, 'rewards/meter/std': 0.32462453842163086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7663149237632751, 'rewards/repeat_soft/std': 0.0751347616314888, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5253073573112488, 'rewards/total_composite/std': 0.08485578000545502, 'reward': 0.5253073573112488, 'reward_std': 0.08485576510429382, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11231084167957306, 'sampling/sampling_logp_difference/max': 1.8563451766967773, 'sampling/importance_sampling_ratio/min': 0.15624262392520905, 'sampling/importance_sampling_ratio/mean': 0.9956095218658447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5471113361418247, 'clip_ratio/low_mean': 0.04088320955634117, 'clip_ratio/low_min': 0.04088320955634117, 'clip_ratio/high_mean': 0.07989784982055426, 'clip_ratio/high_max': 0.07989784982055426, 'clip_ratio/region_mean': 0.12078105937689543, 'reward_total_mean': 0.5253073573112488, 'reward_meter_mean': 0.7705498933792114, 'reward_meter_std': 0.32462453842163086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7663149237632751, 'reward_repeat_soft_std': 0.0751347616314888, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5253073573112488, 'reward_total_composite_std': 0.08485578000545502, 'epoch': 0.14}

 43%|████▎     | 1406/3300 [3:28:59<4:41:24,  8.91s/it]INFO 04-13 10:49:57 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1407/3300 [3:29:05<4:11:37,  7.98s/it]2026-04-13 10:50:02,912 | INFO | train_grpo_train | metrics_logged mode=train step=1407

                                                       
{'loss': 0.0341, 'grad_norm': 14.56716537475586, 'learning_rate': 5.73939393939394e-06, 'num_tokens': 2481489.0, 'completions/mean_length': 23.625, 'completions/min_length': 22.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.5254626870155334, 'rewards/meter/std': 0.3642630875110626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.951632022857666, 'rewards/repeat_soft/std': 0.025855133309960365, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.44512939453125, 'rewards/total_composite/std': 0.0740780457854271, 'reward': 0.44512939453125, 'reward_std': 0.0740780457854271, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1347288340330124, 'sampling/sampling_logp_difference/max': 1.4268412590026855, 'sampling/importance_sampling_ratio/min': 0.2400660365819931, 'sampling/importance_sampling_ratio/mean': 1.0078125, 'sampling/importance_sampling_ratio/max': 1.599117636680603, 'entropy': 0.8280720114707947, 'clip_ratio/low_mean': 0.055887049064040184, 'clip_ratio/low_min': 0.055887049064040184, 'clip_ratio/high_mean': 0.043560607358813286, 'clip_ratio/high_max': 0.043560607358813286, 'clip_ratio/region_mean': 0.09944765642285347, 'reward_total_mean': 0.44512939453125, 'reward_meter_mean': 0.5254626870155334, 'reward_meter_std': 0.3642630875110626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.951632022857666, 'reward_repeat_soft_std': 0.025855133309960365, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.44512939453125, 'reward_total_composite_std': 0.0740780457854271, 'epoch': 0.14}

 43%|████▎     | 1407/3300 [3:29:05<4:11:37,  7.98s/it]INFO 04-13 10:50:03 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1408/3300 [3:29:11<3:55:35,  7.47s/it]2026-04-13 10:50:09,205 | INFO | train_grpo_train | metrics_logged mode=train step=1408

                                                       
{'loss': -0.0026, 'grad_norm': 12.827202796936035, 'learning_rate': 5.736363636363637e-06, 'num_tokens': 2483058.0, 'completions/mean_length': 40.125, 'completions/min_length': 29.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.715053141117096, 'rewards/meter/std': 0.29322072863578796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9744253158569336, 'rewards/repeat_soft/std': 0.020732590928673744, 'rewards/judge_quality/mean': 0.565000057220459, 'rewards/judge_quality/std': 0.194054514169693, 'rewards/total_composite/mean': 0.6142382621765137, 'rewards/total_composite/std': 0.14438693225383759, 'reward': 0.6142382621765137, 'reward_std': 0.14438693225383759, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15014910697937012, 'sampling/sampling_logp_difference/max': 1.556006908416748, 'sampling/importance_sampling_ratio/min': 0.21097683906555176, 'sampling/importance_sampling_ratio/mean': 0.9943320751190186, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5537070706486702, 'clip_ratio/low_mean': 0.06867803446948528, 'clip_ratio/low_min': 0.06867803446948528, 'clip_ratio/high_mean': 0.05681167170405388, 'clip_ratio/high_max': 0.05681167170405388, 'clip_ratio/region_mean': 0.12548970617353916, 'reward_total_mean': 0.6142382621765137, 'reward_meter_mean': 0.715053141117096, 'reward_meter_std': 0.29322072863578796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9744253158569336, 'reward_repeat_soft_std': 0.020732590928673744, 'reward_judge_quality_mean': 0.565000057220459, 'reward_judge_quality_std': 0.194054514169693, 'reward_total_composite_mean': 0.6142382621765137, 'reward_total_composite_std': 0.14438693225383759, 'epoch': 0.14}

 43%|████▎     | 1408/3300 [3:29:11<3:55:35,  7.47s/it]INFO 04-13 10:50:09 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1409/3300 [3:29:18<3:50:25,  7.31s/it]2026-04-13 10:50:16,147 | INFO | train_grpo_train | metrics_logged mode=train step=1409

                                                       
{'loss': -0.0022, 'grad_norm': 12.561360359191895, 'learning_rate': 5.733333333333334e-06, 'num_tokens': 2485080.0, 'completions/mean_length': 74.75, 'completions/min_length': 71.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.8560695648193359, 'rewards/meter/std': 0.27132144570350647, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9129428863525391, 'rewards/repeat_soft/std': 0.034970927983522415, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.510246992111206, 'rewards/total_composite/std': 0.21708330512046814, 'reward': 0.510246992111206, 'reward_std': 0.21708329021930695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12822890281677246, 'sampling/sampling_logp_difference/max': 2.5206103324890137, 'sampling/importance_sampling_ratio/min': 0.20428669452667236, 'sampling/importance_sampling_ratio/mean': 1.0028680562973022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6156154870986938, 'clip_ratio/low_mean': 0.010416666977107525, 'clip_ratio/low_min': 0.010416666977107525, 'clip_ratio/high_mean': 0.09973405580967665, 'clip_ratio/high_max': 0.09973405580967665, 'clip_ratio/region_mean': 0.11015072278678417, 'reward_total_mean': 0.510246992111206, 'reward_meter_mean': 0.8560695648193359, 'reward_meter_std': 0.27132144570350647, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9129428863525391, 'reward_repeat_soft_std': 0.034970927983522415, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.510246992111206, 'reward_total_composite_std': 0.21708330512046814, 'epoch': 0.14}

 43%|████▎     | 1409/3300 [3:29:18<3:50:25,  7.31s/it]INFO 04-13 10:50:16 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1410/3300 [3:29:25<3:45:30,  7.16s/it]2026-04-13 10:50:22,946 | INFO | train_grpo_train | metrics_logged mode=train step=1410

                                                       
{'loss': -0.053, 'grad_norm': 14.60198974609375, 'learning_rate': 5.7303030303030305e-06, 'num_tokens': 2486748.0, 'completions/mean_length': 40.5, 'completions/min_length': 30.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8249857425689697, 'rewards/meter/std': 0.33494699001312256, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.950890064239502, 'rewards/repeat_soft/std': 0.051062971353530884, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.6086096167564392, 'rewards/total_composite/std': 0.15268997848033905, 'reward': 0.6086096167564392, 'reward_std': 0.15268996357917786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16327844560146332, 'sampling/sampling_logp_difference/max': 3.4548776149749756, 'sampling/importance_sampling_ratio/min': 0.03159116953611374, 'sampling/importance_sampling_ratio/mean': 0.9968622326850891, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8403591066598892, 'clip_ratio/low_mean': 0.048809911124408245, 'clip_ratio/low_min': 0.048809911124408245, 'clip_ratio/high_mean': 0.06952887680381536, 'clip_ratio/high_max': 0.06952887680381536, 'clip_ratio/region_mean': 0.11833878792822361, 'reward_total_mean': 0.6086096167564392, 'reward_meter_mean': 0.8249857425689697, 'reward_meter_std': 0.33494699001312256, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.950890064239502, 'reward_repeat_soft_std': 0.051062971353530884, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.6086096167564392, 'reward_total_composite_std': 0.15268997848033905, 'epoch': 0.14}

 43%|████▎     | 1410/3300 [3:29:25<3:45:30,  7.16s/it]INFO 04-13 10:50:23 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1411/3300 [3:29:31<3:41:47,  7.04s/it]2026-04-13 10:50:29,723 | INFO | train_grpo_train | metrics_logged mode=train step=1411

                                                       
{'loss': 0.0223, 'grad_norm': 20.06722640991211, 'learning_rate': 5.727272727272728e-06, 'num_tokens': 2488469.0, 'completions/mean_length': 45.125, 'completions/min_length': 42.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.75612473487854, 'rewards/meter/std': 0.3957335948944092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9366524815559387, 'rewards/repeat_soft/std': 0.07409165799617767, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.16070716083049774, 'rewards/total_composite/mean': 0.5985976457595825, 'rewards/total_composite/std': 0.16009141504764557, 'reward': 0.5985976457595825, 'reward_std': 0.16009140014648438, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11883072555065155, 'sampling/sampling_logp_difference/max': 2.6178369522094727, 'sampling/importance_sampling_ratio/min': 0.0729605108499527, 'sampling/importance_sampling_ratio/mean': 1.0040268898010254, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7764797955751419, 'clip_ratio/low_mean': 0.027832756750285625, 'clip_ratio/low_min': 0.027832756750285625, 'clip_ratio/high_mean': 0.09952107258141041, 'clip_ratio/high_max': 0.09952107258141041, 'clip_ratio/region_mean': 0.12735382933169603, 'reward_total_mean': 0.5985976457595825, 'reward_meter_mean': 0.75612473487854, 'reward_meter_std': 0.3957335948944092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9366524815559387, 'reward_repeat_soft_std': 0.07409165799617767, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.16070716083049774, 'reward_total_composite_mean': 0.5985976457595825, 'reward_total_composite_std': 0.16009141504764557, 'epoch': 0.14}

 43%|████▎     | 1411/3300 [3:29:31<3:41:47,  7.04s/it]INFO 04-13 10:50:29 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1412/3300 [3:29:38<3:41:16,  7.03s/it]2026-04-13 10:50:36,727 | INFO | train_grpo_train | metrics_logged mode=train step=1412

                                                       
{'loss': 0.0228, 'grad_norm': 19.718990325927734, 'learning_rate': 5.724242424242424e-06, 'num_tokens': 2490611.0, 'completions/mean_length': 94.75, 'completions/min_length': 80.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.75, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.6265571713447571, 'rewards/meter/std': 0.35264143347740173, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9424885511398315, 'rewards/repeat_soft/std': 0.036095939576625824, 'rewards/judge_quality/mean': 0.5324999690055847, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.5623034834861755, 'rewards/total_composite/std': 0.15559940040111542, 'reward': 0.5623034834861755, 'reward_std': 0.15559938549995422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13852238655090332, 'sampling/sampling_logp_difference/max': 3.2659621238708496, 'sampling/importance_sampling_ratio/min': 0.03816020488739014, 'sampling/importance_sampling_ratio/mean': 1.0092533826828003, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7166869342327118, 'clip_ratio/low_mean': 0.06569533981382847, 'clip_ratio/low_min': 0.06569533981382847, 'clip_ratio/high_mean': 0.057803040370345116, 'clip_ratio/high_max': 0.057803040370345116, 'clip_ratio/region_mean': 0.12349838018417358, 'reward_total_mean': 0.5623034834861755, 'reward_meter_mean': 0.6265571713447571, 'reward_meter_std': 0.35264143347740173, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9424885511398315, 'reward_repeat_soft_std': 0.036095939576625824, 'reward_judge_quality_mean': 0.5324999690055847, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.5623034834861755, 'reward_total_composite_std': 0.15559940040111542, 'epoch': 0.14}

 43%|████▎     | 1412/3300 [3:29:38<3:41:16,  7.03s/it]INFO 04-13 10:50:36 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1413/3300 [3:29:46<3:47:58,  7.25s/it]2026-04-13 10:50:44,482 | INFO | train_grpo_train | metrics_logged mode=train step=1413

                                                       
{'loss': 0.0103, 'grad_norm': 9.716988563537598, 'learning_rate': 5.721212121212122e-06, 'num_tokens': 2492560.0, 'completions/mean_length': 68.625, 'completions/min_length': 61.0, 'completions/max_length': 91.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 68.625, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 91.0, 'rewards/meter/mean': 0.9647419452667236, 'rewards/meter/std': 0.047957222908735275, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7776961326599121, 'rewards/repeat_soft/std': 0.05703750625252724, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.09086881577968597, 'rewards/total_composite/mean': 0.5277512669563293, 'rewards/total_composite/std': 0.06632345169782639, 'reward': 0.5277512669563293, 'reward_std': 0.06632344424724579, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11125600337982178, 'sampling/sampling_logp_difference/max': 1.4070215225219727, 'sampling/importance_sampling_ratio/min': 0.24487155675888062, 'sampling/importance_sampling_ratio/mean': 1.0192159414291382, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7872535884380341, 'clip_ratio/low_mean': 0.043009101413190365, 'clip_ratio/low_min': 0.043009101413190365, 'clip_ratio/high_mean': 0.05227456334978342, 'clip_ratio/high_max': 0.05227456334978342, 'clip_ratio/region_mean': 0.09528366476297379, 'reward_total_mean': 0.5277512669563293, 'reward_meter_mean': 0.9647419452667236, 'reward_meter_std': 0.047957222908735275, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7776961326599121, 'reward_repeat_soft_std': 0.05703750625252724, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.09086881577968597, 'reward_total_composite_mean': 0.5277512669563293, 'reward_total_composite_std': 0.06632345169782639, 'epoch': 0.14}

 43%|████▎     | 1413/3300 [3:29:46<3:47:58,  7.25s/it]INFO 04-13 10:50:44 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1414/3300 [3:29:52<3:39:28,  6.98s/it]2026-04-13 10:50:50,843 | INFO | train_grpo_train | metrics_logged mode=train step=1414

                                                       
{'loss': -0.0177, 'grad_norm': 14.703536033630371, 'learning_rate': 5.718181818181819e-06, 'num_tokens': 2494106.0, 'completions/mean_length': 46.25, 'completions/min_length': 36.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.301479697227478, 'rewards/meter/std': 0.358011394739151, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9754935503005981, 'rewards/repeat_soft/std': 0.029871217906475067, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.47409993410110474, 'rewards/total_composite/std': 0.1980825662612915, 'reward': 0.47409993410110474, 'reward_std': 0.1980825513601303, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16618722677230835, 'sampling/sampling_logp_difference/max': 1.4385216236114502, 'sampling/importance_sampling_ratio/min': 0.2372782826423645, 'sampling/importance_sampling_ratio/mean': 1.0048731565475464, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8378394320607185, 'clip_ratio/low_mean': 0.0703724012710154, 'clip_ratio/low_min': 0.0703724012710154, 'clip_ratio/high_mean': 0.06799311935901642, 'clip_ratio/high_max': 0.06799311935901642, 'clip_ratio/region_mean': 0.13836552063003182, 'reward_total_mean': 0.47409993410110474, 'reward_meter_mean': 0.301479697227478, 'reward_meter_std': 0.358011394739151, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9754935503005981, 'reward_repeat_soft_std': 0.029871217906475067, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.47409993410110474, 'reward_total_composite_std': 0.1980825662612915, 'epoch': 0.14}

 43%|████▎     | 1414/3300 [3:29:53<3:39:28,  6.98s/it]INFO 04-13 10:50:51 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1415/3300 [3:29:59<3:33:30,  6.80s/it]2026-04-13 10:50:57,190 | INFO | train_grpo_train | metrics_logged mode=train step=1415

                                                       
{'loss': 0.013, 'grad_norm': 13.053173065185547, 'learning_rate': 5.715151515151516e-06, 'num_tokens': 2495743.0, 'completions/mean_length': 47.625, 'completions/min_length': 43.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8571093082427979, 'rewards/meter/std': 0.27939456701278687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9783151149749756, 'rewards/repeat_soft/std': 0.0327603854238987, 'rewards/judge_quality/mean': 0.5687500238418579, 'rewards/judge_quality/std': 0.19773270189762115, 'rewards/total_composite/mean': 0.6719194650650024, 'rewards/total_composite/std': 0.1636134833097458, 'reward': 0.6719194650650024, 'reward_std': 0.1636134833097458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11143966019153595, 'sampling/sampling_logp_difference/max': 1.3081310987472534, 'sampling/importance_sampling_ratio/min': 0.27032479643821716, 'sampling/importance_sampling_ratio/mean': 1.005942702293396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7236667647957802, 'clip_ratio/low_mean': 0.08082544151693583, 'clip_ratio/low_min': 0.08082544151693583, 'clip_ratio/high_mean': 0.05500944051891565, 'clip_ratio/high_max': 0.05500944051891565, 'clip_ratio/region_mean': 0.13583488203585148, 'reward_total_mean': 0.6719194650650024, 'reward_meter_mean': 0.8571093082427979, 'reward_meter_std': 0.27939456701278687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9783151149749756, 'reward_repeat_soft_std': 0.0327603854238987, 'reward_judge_quality_mean': 0.5687500238418579, 'reward_judge_quality_std': 0.19773270189762115, 'reward_total_composite_mean': 0.6719194650650024, 'reward_total_composite_std': 0.1636134833097458, 'epoch': 0.14}

 43%|████▎     | 1415/3300 [3:29:59<3:33:30,  6.80s/it]INFO 04-13 10:50:57 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1416/3300 [3:30:05<3:28:09,  6.63s/it]2026-04-13 10:51:03,430 | INFO | train_grpo_train | metrics_logged mode=train step=1416

                                                       
{'loss': 0.0465, 'grad_norm': 27.83479881286621, 'learning_rate': 5.712121212121212e-06, 'num_tokens': 2497121.0, 'completions/mean_length': 20.25, 'completions/min_length': 18.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9716972708702087, 'rewards/meter/std': 0.035730406641960144, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.928876519203186, 'rewards/repeat_soft/std': 0.08303026854991913, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.593076229095459, 'rewards/total_composite/std': 0.03745569661259651, 'reward': 0.593076229095459, 'reward_std': 0.03745569661259651, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13063624501228333, 'sampling/sampling_logp_difference/max': 1.1894512176513672, 'sampling/importance_sampling_ratio/min': 0.30438825488090515, 'sampling/importance_sampling_ratio/mean': 1.0124963521957397, 'sampling/importance_sampling_ratio/max': 1.8049801588058472, 'entropy': 0.6687679588794708, 'clip_ratio/low_mean': 0.022997836116701365, 'clip_ratio/low_min': 0.022997836116701365, 'clip_ratio/high_mean': 0.057428672444075346, 'clip_ratio/high_max': 0.057428672444075346, 'clip_ratio/region_mean': 0.08042650856077671, 'reward_total_mean': 0.593076229095459, 'reward_meter_mean': 0.9716972708702087, 'reward_meter_std': 0.035730406641960144, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.928876519203186, 'reward_repeat_soft_std': 0.08303026854991913, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.593076229095459, 'reward_total_composite_std': 0.03745569661259651, 'epoch': 0.14}

 43%|████▎     | 1416/3300 [3:30:05<3:28:09,  6.63s/it]INFO 04-13 10:51:03 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1417/3300 [3:30:11<3:21:27,  6.42s/it]2026-04-13 10:51:09,358 | INFO | train_grpo_train | metrics_logged mode=train step=1417

                                                       
{'loss': 0.0011, 'grad_norm': 14.798911094665527, 'learning_rate': 5.7090909090909096e-06, 'num_tokens': 2498561.0, 'completions/mean_length': 30.0, 'completions/min_length': 28.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.9242525100708008, 'rewards/meter/std': 0.07899387180805206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8455250859260559, 'rewards/repeat_soft/std': 0.04145512357354164, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.6020537614822388, 'rewards/total_composite/std': 0.11468619108200073, 'reward': 0.6020537614822388, 'reward_std': 0.11468618363142014, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0789555236697197, 'sampling/sampling_logp_difference/max': 1.3896409273147583, 'sampling/importance_sampling_ratio/min': 0.24916476011276245, 'sampling/importance_sampling_ratio/mean': 1.0224395990371704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5047766417264938, 'clip_ratio/low_mean': 0.046301571652293205, 'clip_ratio/low_min': 0.046301571652293205, 'clip_ratio/high_mean': 0.01953125, 'clip_ratio/high_max': 0.01953125, 'clip_ratio/region_mean': 0.0658328216522932, 'reward_total_mean': 0.6020537614822388, 'reward_meter_mean': 0.9242525100708008, 'reward_meter_std': 0.07899387180805206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8455250859260559, 'reward_repeat_soft_std': 0.04145512357354164, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.6020537614822388, 'reward_total_composite_std': 0.11468619108200073, 'epoch': 0.14}

 43%|████▎     | 1417/3300 [3:30:11<3:21:27,  6.42s/it]INFO 04-13 10:51:09 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1418/3300 [3:30:17<3:19:33,  6.36s/it]2026-04-13 10:51:15,589 | INFO | train_grpo_train | metrics_logged mode=train step=1418

                                                       
{'loss': 0.0091, 'grad_norm': 15.589071273803711, 'learning_rate': 5.706060606060606e-06, 'num_tokens': 2500117.0, 'completions/mean_length': 43.5, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.5782667398452759, 'rewards/meter/std': 0.4150027632713318, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9840254783630371, 'rewards/repeat_soft/std': 0.011035393923521042, 'rewards/judge_quality/mean': 0.5475000143051147, 'rewards/judge_quality/std': 0.14320316910743713, 'rewards/total_composite/mean': 0.5291829109191895, 'rewards/total_composite/std': 0.10961486399173737, 'reward': 0.5291829109191895, 'reward_std': 0.10961487144231796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11890725791454315, 'sampling/sampling_logp_difference/max': 1.7466232776641846, 'sampling/importance_sampling_ratio/min': 0.1743617206811905, 'sampling/importance_sampling_ratio/mean': 1.0020751953125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7522838935256004, 'clip_ratio/low_mean': 0.04423077031970024, 'clip_ratio/low_min': 0.04423077031970024, 'clip_ratio/high_mean': 0.06876990105956793, 'clip_ratio/high_max': 0.06876990105956793, 'clip_ratio/region_mean': 0.11300067137926817, 'reward_total_mean': 0.5291829109191895, 'reward_meter_mean': 0.5782667398452759, 'reward_meter_std': 0.4150027632713318, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9840254783630371, 'reward_repeat_soft_std': 0.011035393923521042, 'reward_judge_quality_mean': 0.5475000143051147, 'reward_judge_quality_std': 0.14320316910743713, 'reward_total_composite_mean': 0.5291829109191895, 'reward_total_composite_std': 0.10961486399173737, 'epoch': 0.14}

 43%|████▎     | 1418/3300 [3:30:17<3:19:33,  6.36s/it]INFO 04-13 10:51:15 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1419/3300 [3:30:23<3:15:12,  6.23s/it]2026-04-13 10:51:21,500 | INFO | train_grpo_train | metrics_logged mode=train step=1419

                                                       
{'loss': 0.0123, 'grad_norm': 18.427335739135742, 'learning_rate': 5.703030303030303e-06, 'num_tokens': 2501696.0, 'completions/mean_length': 23.375, 'completions/min_length': 20.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9696396589279175, 'rewards/meter/std': 0.03659556433558464, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.948153018951416, 'rewards/repeat_soft/std': 0.0210404172539711, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.22720351815223694, 'rewards/total_composite/mean': 0.6885734796524048, 'rewards/total_composite/std': 0.13805341720581055, 'reward': 0.6885734796524048, 'reward_std': 0.13805341720581055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11416921764612198, 'sampling/sampling_logp_difference/max': 1.1906063556671143, 'sampling/importance_sampling_ratio/min': 0.31950920820236206, 'sampling/importance_sampling_ratio/mean': 1.0270253419876099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5909772589802742, 'clip_ratio/low_mean': 0.07892339676618576, 'clip_ratio/low_min': 0.07892339676618576, 'clip_ratio/high_mean': 0.02116402145475149, 'clip_ratio/high_max': 0.02116402145475149, 'clip_ratio/region_mean': 0.10008741822093725, 'reward_total_mean': 0.6885734796524048, 'reward_meter_mean': 0.9696396589279175, 'reward_meter_std': 0.03659556433558464, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.948153018951416, 'reward_repeat_soft_std': 0.0210404172539711, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.22720351815223694, 'reward_total_composite_mean': 0.6885734796524048, 'reward_total_composite_std': 0.13805341720581055, 'epoch': 0.14}

 43%|████▎     | 1419/3300 [3:30:23<3:15:12,  6.23s/it]INFO 04-13 10:51:21 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1420/3300 [3:30:29<3:13:32,  6.18s/it]2026-04-13 10:51:27,568 | INFO | train_grpo_train | metrics_logged mode=train step=1420

                                                       
{'loss': 0.0516, 'grad_norm': 12.975893020629883, 'learning_rate': 5.7e-06, 'num_tokens': 2503264.0, 'completions/mean_length': 42.0, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9350648522377014, 'rewards/meter/std': 0.08645668625831604, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9639267921447754, 'rewards/repeat_soft/std': 0.03511298447847366, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.7239274978637695, 'rewards/total_composite/std': 0.1742723286151886, 'reward': 0.7239274978637695, 'reward_std': 0.1742723137140274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12830279767513275, 'sampling/sampling_logp_difference/max': 1.9506807327270508, 'sampling/importance_sampling_ratio/min': 0.1421772539615631, 'sampling/importance_sampling_ratio/mean': 1.0142592191696167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8880967274308205, 'clip_ratio/low_mean': 0.08618996618315578, 'clip_ratio/low_min': 0.08618996618315578, 'clip_ratio/high_mean': 0.043819112004712224, 'clip_ratio/high_max': 0.043819112004712224, 'clip_ratio/region_mean': 0.130009078187868, 'reward_total_mean': 0.7239274978637695, 'reward_meter_mean': 0.9350648522377014, 'reward_meter_std': 0.08645668625831604, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9639267921447754, 'reward_repeat_soft_std': 0.03511298447847366, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.7239274978637695, 'reward_total_composite_std': 0.1742723286151886, 'epoch': 0.14}

 43%|████▎     | 1420/3300 [3:30:29<3:13:32,  6.18s/it]INFO 04-13 10:51:27 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1421/3300 [3:30:36<3:14:30,  6.21s/it]2026-04-13 10:51:33,864 | INFO | train_grpo_train | metrics_logged mode=train step=1421

                                                       
{'loss': 0.001, 'grad_norm': 14.406489372253418, 'learning_rate': 5.696969696969698e-06, 'num_tokens': 2504573.0, 'completions/mean_length': 21.625, 'completions/min_length': 19.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.625, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8017401695251465, 'rewards/meter/std': 0.21783582866191864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9435555934906006, 'rewards/repeat_soft/std': 0.036296285688877106, 'rewards/judge_quality/mean': 0.44749999046325684, 'rewards/judge_quality/std': 0.20824094116687775, 'rewards/total_composite/mean': 0.4621683955192566, 'rewards/total_composite/std': 0.2980007827281952, 'reward': 0.4621683955192566, 'reward_std': 0.2980007827281952, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12893855571746826, 'sampling/sampling_logp_difference/max': 0.9362697601318359, 'sampling/importance_sampling_ratio/min': 0.3920876979827881, 'sampling/importance_sampling_ratio/mean': 0.997369110584259, 'sampling/importance_sampling_ratio/max': 1.6800825595855713, 'entropy': 0.858201414346695, 'clip_ratio/low_mean': 0.023863636888563633, 'clip_ratio/low_min': 0.023863636888563633, 'clip_ratio/high_mean': 0.13010139018297195, 'clip_ratio/high_max': 0.13010139018297195, 'clip_ratio/region_mean': 0.1539650270715356, 'reward_total_mean': 0.4621683955192566, 'reward_meter_mean': 0.8017401695251465, 'reward_meter_std': 0.21783582866191864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9435555934906006, 'reward_repeat_soft_std': 0.036296285688877106, 'reward_judge_quality_mean': 0.44749999046325684, 'reward_judge_quality_std': 0.20824094116687775, 'reward_total_composite_mean': 0.4621683955192566, 'reward_total_composite_std': 0.2980007827281952, 'epoch': 0.14}

 43%|████▎     | 1421/3300 [3:30:36<3:14:30,  6.21s/it]INFO 04-13 10:51:34 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1422/3300 [3:30:42<3:17:32,  6.31s/it]2026-04-13 10:51:40,399 | INFO | train_grpo_train | metrics_logged mode=train step=1422

                                                       
{'loss': 0.0346, 'grad_norm': 9.34439754486084, 'learning_rate': 5.693939393939394e-06, 'num_tokens': 2506629.0, 'completions/mean_length': 88.0, 'completions/min_length': 83.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.0, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.7999811172485352, 'rewards/meter/std': 0.18497784435749054, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8585450053215027, 'rewards/repeat_soft/std': 0.036236103624105453, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5709871053695679, 'rewards/total_composite/std': 0.09682025015354156, 'reward': 0.5709871053695679, 'reward_std': 0.09682025015354156, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1265259087085724, 'sampling/sampling_logp_difference/max': 1.78059720993042, 'sampling/importance_sampling_ratio/min': 0.1685374677181244, 'sampling/importance_sampling_ratio/mean': 1.015747308731079, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6572273150086403, 'clip_ratio/low_mean': 0.07174063427373767, 'clip_ratio/low_min': 0.07174063427373767, 'clip_ratio/high_mean': 0.04543060716241598, 'clip_ratio/high_max': 0.04543060716241598, 'clip_ratio/region_mean': 0.11717124143615365, 'reward_total_mean': 0.5709871053695679, 'reward_meter_mean': 0.7999811172485352, 'reward_meter_std': 0.18497784435749054, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8585450053215027, 'reward_repeat_soft_std': 0.036236103624105453, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5709871053695679, 'reward_total_composite_std': 0.09682025015354156, 'epoch': 0.14}

 43%|████▎     | 1422/3300 [3:30:42<3:17:32,  6.31s/it]INFO 04-13 10:51:40 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1423/3300 [3:30:49<3:25:04,  6.56s/it]2026-04-13 10:51:47,537 | INFO | train_grpo_train | metrics_logged mode=train step=1423

                                                       
{'loss': 0.0202, 'grad_norm': 8.207839965820312, 'learning_rate': 5.690909090909091e-06, 'num_tokens': 2508931.0, 'completions/mean_length': 96.75, 'completions/min_length': 77.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7195549607276917, 'rewards/meter/std': 0.3614080846309662, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9357661008834839, 'rewards/repeat_soft/std': 0.04376447945833206, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5204696655273438, 'rewards/total_composite/std': 0.13906559348106384, 'reward': 0.5204696655273438, 'reward_std': 0.13906559348106384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1431320309638977, 'sampling/sampling_logp_difference/max': 4.391334533691406, 'sampling/importance_sampling_ratio/min': 0.012384191155433655, 'sampling/importance_sampling_ratio/mean': 1.0087347030639648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7626270726323128, 'clip_ratio/low_mean': 0.0331280380487442, 'clip_ratio/low_min': 0.0331280380487442, 'clip_ratio/high_mean': 0.08710999134927988, 'clip_ratio/high_max': 0.08710999134927988, 'clip_ratio/region_mean': 0.12023802939802408, 'reward_total_mean': 0.5204696655273438, 'reward_meter_mean': 0.7195549607276917, 'reward_meter_std': 0.3614080846309662, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9357661008834839, 'reward_repeat_soft_std': 0.04376447945833206, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5204696655273438, 'reward_total_composite_std': 0.13906559348106384, 'epoch': 0.14}

 43%|████▎     | 1423/3300 [3:30:49<3:25:04,  6.56s/it]INFO 04-13 10:51:47 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1424/3300 [3:30:56<3:25:05,  6.56s/it]2026-04-13 10:51:54,092 | INFO | train_grpo_train | metrics_logged mode=train step=1424

                                                       
{'loss': 0.0055, 'grad_norm': 8.282552719116211, 'learning_rate': 5.687878787878789e-06, 'num_tokens': 2511202.0, 'completions/mean_length': 93.875, 'completions/min_length': 87.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.875, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.6325433254241943, 'rewards/meter/std': 0.3539314270019531, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8354396820068359, 'rewards/repeat_soft/std': 0.05221690610051155, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.49850624799728394, 'rewards/total_composite/std': 0.09172514826059341, 'reward': 0.49850624799728394, 'reward_std': 0.09172514081001282, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11522486060857773, 'sampling/sampling_logp_difference/max': 2.3873190879821777, 'sampling/importance_sampling_ratio/min': 0.09187566488981247, 'sampling/importance_sampling_ratio/mean': 1.00412917137146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6340465322136879, 'clip_ratio/low_mean': 0.05321041587740183, 'clip_ratio/low_min': 0.05321041587740183, 'clip_ratio/high_mean': 0.055305938702076674, 'clip_ratio/high_max': 0.055305938702076674, 'clip_ratio/region_mean': 0.1085163545794785, 'reward_total_mean': 0.49850624799728394, 'reward_meter_mean': 0.6325433254241943, 'reward_meter_std': 0.3539314270019531, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8354396820068359, 'reward_repeat_soft_std': 0.05221690610051155, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.49850624799728394, 'reward_total_composite_std': 0.09172514826059341, 'epoch': 0.14}

 43%|████▎     | 1424/3300 [3:30:56<3:25:05,  6.56s/it]INFO 04-13 10:51:54 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1425/3300 [3:31:03<3:28:42,  6.68s/it]2026-04-13 10:52:01,048 | INFO | train_grpo_train | metrics_logged mode=train step=1425

                                                       
{'loss': 0.0471, 'grad_norm': 9.898316383361816, 'learning_rate': 5.684848484848485e-06, 'num_tokens': 2513556.0, 'completions/mean_length': 93.25, 'completions/min_length': 87.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.25, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.9084113240242004, 'rewards/meter/std': 0.0786547064781189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8716374039649963, 'rewards/repeat_soft/std': 0.05684646591544151, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6510645151138306, 'rewards/total_composite/std': 0.13250632584095, 'reward': 0.6510645151138306, 'reward_std': 0.13250631093978882, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13459469377994537, 'sampling/sampling_logp_difference/max': 1.9838438034057617, 'sampling/importance_sampling_ratio/min': 0.1375395506620407, 'sampling/importance_sampling_ratio/mean': 1.0100266933441162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7431213334202766, 'clip_ratio/low_mean': 0.07877854350954294, 'clip_ratio/low_min': 0.07877854350954294, 'clip_ratio/high_mean': 0.019875479396432638, 'clip_ratio/high_max': 0.019875479396432638, 'clip_ratio/region_mean': 0.09865402290597558, 'reward_total_mean': 0.6510645151138306, 'reward_meter_mean': 0.9084113240242004, 'reward_meter_std': 0.0786547064781189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8716374039649963, 'reward_repeat_soft_std': 0.05684646591544151, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6510645151138306, 'reward_total_composite_std': 0.13250632584095, 'epoch': 0.14}

 43%|████▎     | 1425/3300 [3:31:03<3:28:42,  6.68s/it]INFO 04-13 10:52:01 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1426/3300 [3:31:10<3:35:38,  6.90s/it]2026-04-13 10:52:08,479 | INFO | train_grpo_train | metrics_logged mode=train step=1426

                                                       
{'loss': 0.1294, 'grad_norm': 8.25341796875, 'learning_rate': 5.681818181818183e-06, 'num_tokens': 2515217.0, 'completions/mean_length': 58.625, 'completions/min_length': 36.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.6144726276397705, 'rewards/meter/std': 0.373422235250473, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9629454612731934, 'rewards/repeat_soft/std': 0.036584775894880295, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.21963852643966675, 'rewards/total_composite/mean': 0.5260547399520874, 'rewards/total_composite/std': 0.12817879021167755, 'reward': 0.5260547399520874, 'reward_std': 0.12817879021167755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09847000986337662, 'sampling/sampling_logp_difference/max': 1.1577715873718262, 'sampling/importance_sampling_ratio/min': 0.31418555974960327, 'sampling/importance_sampling_ratio/mean': 1.0045069456100464, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7123184055089951, 'clip_ratio/low_mean': 0.07798398123122752, 'clip_ratio/low_min': 0.07798398123122752, 'clip_ratio/high_mean': 0.051209207624197006, 'clip_ratio/high_max': 0.051209207624197006, 'clip_ratio/region_mean': 0.12919318885542452, 'reward_total_mean': 0.5260547399520874, 'reward_meter_mean': 0.6144726276397705, 'reward_meter_std': 0.373422235250473, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9629454612731934, 'reward_repeat_soft_std': 0.036584775894880295, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.21963852643966675, 'reward_total_composite_mean': 0.5260547399520874, 'reward_total_composite_std': 0.12817879021167755, 'epoch': 0.14}

 43%|████▎     | 1426/3300 [3:31:10<3:35:38,  6.90s/it]INFO 04-13 10:52:08 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1427/3300 [3:31:22<4:17:31,  8.25s/it]2026-04-13 10:52:19,873 | INFO | train_grpo_train | metrics_logged mode=train step=1427

                                                       
{'loss': -0.0972, 'grad_norm': 3.392662525177002, 'learning_rate': 5.67878787878788e-06, 'num_tokens': 2516861.0, 'completions/mean_length': 109.5, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.000003814697266, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4390566945075989, 'rewards/meter/std': 0.35161668062210083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9917404651641846, 'rewards/repeat_soft/std': 0.006031770259141922, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.20982563495635986, 'rewards/total_composite/mean': 0.45237046480178833, 'rewards/total_composite/std': 0.22913609445095062, 'reward': 0.45237046480178833, 'reward_std': 0.22913609445095062, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1584887057542801, 'sampling/sampling_logp_difference/max': 2.5816755294799805, 'sampling/importance_sampling_ratio/min': 0.07564714550971985, 'sampling/importance_sampling_ratio/mean': 1.0274124145507812, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.973654605448246, 'clip_ratio/low_mean': 0.03229166753590107, 'clip_ratio/low_min': 0.03229166753590107, 'clip_ratio/high_mean': 0.07325901370495558, 'clip_ratio/high_max': 0.07325901370495558, 'clip_ratio/region_mean': 0.10555068124085665, 'reward_total_mean': 0.45237046480178833, 'reward_meter_mean': 0.4390566945075989, 'reward_meter_std': 0.35161668062210083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9917404651641846, 'reward_repeat_soft_std': 0.006031770259141922, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.20982563495635986, 'reward_total_composite_mean': 0.45237046480178833, 'reward_total_composite_std': 0.22913609445095062, 'epoch': 0.14}

 43%|████▎     | 1427/3300 [3:31:22<4:17:31,  8.25s/it]INFO 04-13 10:52:20 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1428/3300 [3:31:29<4:08:43,  7.97s/it]2026-04-13 10:52:27,195 | INFO | train_grpo_train | metrics_logged mode=train step=1428

                                                       
{'loss': 0.0098, 'grad_norm': 10.655179023742676, 'learning_rate': 5.675757575757577e-06, 'num_tokens': 2518932.0, 'completions/mean_length': 79.875, 'completions/min_length': 55.0, 'completions/max_length': 108.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 108.0, 'rewards/meter/mean': 0.5962883234024048, 'rewards/meter/std': 0.327947735786438, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8625436425209045, 'rewards/repeat_soft/std': 0.0839657112956047, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5221672058105469, 'rewards/total_composite/std': 0.13741447031497955, 'reward': 0.5221672058105469, 'reward_std': 0.13741447031497955, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1376197636127472, 'sampling/sampling_logp_difference/max': 2.1869306564331055, 'sampling/importance_sampling_ratio/min': 0.11226078122854233, 'sampling/importance_sampling_ratio/mean': 1.0101064443588257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7594892457127571, 'clip_ratio/low_mean': 0.03768882201984525, 'clip_ratio/low_min': 0.03768882201984525, 'clip_ratio/high_mean': 0.06279271235689521, 'clip_ratio/high_max': 0.06279271235689521, 'clip_ratio/region_mean': 0.10048153437674046, 'reward_total_mean': 0.5221672058105469, 'reward_meter_mean': 0.5962883234024048, 'reward_meter_std': 0.327947735786438, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8625436425209045, 'reward_repeat_soft_std': 0.0839657112956047, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5221672058105469, 'reward_total_composite_std': 0.13741447031497955, 'epoch': 0.14}

 43%|████▎     | 1428/3300 [3:31:29<4:08:43,  7.97s/it]INFO 04-13 10:52:27 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1429/3300 [3:31:36<3:57:57,  7.63s/it]2026-04-13 10:52:34,033 | INFO | train_grpo_train | metrics_logged mode=train step=1429

                                                       
{'loss': -0.0631, 'grad_norm': 10.576534271240234, 'learning_rate': 5.672727272727273e-06, 'num_tokens': 2521051.0, 'completions/mean_length': 76.875, 'completions/min_length': 61.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.875, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.4533776044845581, 'rewards/meter/std': 0.2749289572238922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9741709232330322, 'rewards/repeat_soft/std': 0.028394293040037155, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4698977470397949, 'rewards/total_composite/std': 0.07737768441438675, 'reward': 0.4698977470397949, 'reward_std': 0.07737768441438675, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13355378806591034, 'sampling/sampling_logp_difference/max': 1.7773722410202026, 'sampling/importance_sampling_ratio/min': 0.1690818816423416, 'sampling/importance_sampling_ratio/mean': 1.0167454481124878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9134811013936996, 'clip_ratio/low_mean': 0.05188595782965422, 'clip_ratio/low_min': 0.05188595782965422, 'clip_ratio/high_mean': 0.0636458182707429, 'clip_ratio/high_max': 0.0636458182707429, 'clip_ratio/region_mean': 0.11553177610039711, 'reward_total_mean': 0.4698977470397949, 'reward_meter_mean': 0.4533776044845581, 'reward_meter_std': 0.2749289572238922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9741709232330322, 'reward_repeat_soft_std': 0.028394293040037155, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4698977470397949, 'reward_total_composite_std': 0.07737768441438675, 'epoch': 0.14}

 43%|████▎     | 1429/3300 [3:31:36<3:57:57,  7.63s/it]INFO 04-13 10:52:34 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1430/3300 [3:31:42<3:45:51,  7.25s/it]2026-04-13 10:52:40,380 | INFO | train_grpo_train | metrics_logged mode=train step=1430

                                                       
{'loss': -0.0189, 'grad_norm': 13.204424858093262, 'learning_rate': 5.6696969696969705e-06, 'num_tokens': 2522840.0, 'completions/mean_length': 41.625, 'completions/min_length': 37.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8196427822113037, 'rewards/meter/std': 0.31651848554611206, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9179916381835938, 'rewards/repeat_soft/std': 0.0478409081697464, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.563774824142456, 'rewards/total_composite/std': 0.08826941251754761, 'reward': 0.563774824142456, 'reward_std': 0.08826940506696701, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13667036592960358, 'sampling/sampling_logp_difference/max': 1.5471887588500977, 'sampling/importance_sampling_ratio/min': 0.21284548938274384, 'sampling/importance_sampling_ratio/mean': 1.0013889074325562, 'sampling/importance_sampling_ratio/max': 1.981207013130188, 'entropy': 0.688811220228672, 'clip_ratio/low_mean': 0.04380293842405081, 'clip_ratio/low_min': 0.04380293842405081, 'clip_ratio/high_mean': 0.061815458349883556, 'clip_ratio/high_max': 0.061815458349883556, 'clip_ratio/region_mean': 0.10561839677393436, 'reward_total_mean': 0.563774824142456, 'reward_meter_mean': 0.8196427822113037, 'reward_meter_std': 0.31651848554611206, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9179916381835938, 'reward_repeat_soft_std': 0.0478409081697464, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.563774824142456, 'reward_total_composite_std': 0.08826941251754761, 'epoch': 0.14}

 43%|████▎     | 1430/3300 [3:31:42<3:45:51,  7.25s/it]INFO 04-13 10:52:40 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1431/3300 [3:31:48<3:33:39,  6.86s/it]2026-04-13 10:52:46,329 | INFO | train_grpo_train | metrics_logged mode=train step=1431

                                                       
{'loss': 0.0257, 'grad_norm': 12.921917915344238, 'learning_rate': 5.666666666666667e-06, 'num_tokens': 2524347.0, 'completions/mean_length': 40.375, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.954035758972168, 'rewards/meter/std': 0.034081488847732544, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9922205209732056, 'rewards/repeat_soft/std': 0.008540960028767586, 'rewards/judge_quality/mean': 0.581250011920929, 'rewards/judge_quality/std': 0.22183892130851746, 'rewards/total_composite/mean': 0.7052081227302551, 'rewards/total_composite/std': 0.12278620153665543, 'reward': 0.7052081227302551, 'reward_std': 0.12278620153665543, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12639567255973816, 'sampling/sampling_logp_difference/max': 1.8599350452423096, 'sampling/importance_sampling_ratio/min': 0.1556827425956726, 'sampling/importance_sampling_ratio/mean': 0.9986779093742371, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7381427623331547, 'clip_ratio/low_mean': 0.07170370407402515, 'clip_ratio/low_min': 0.07170370407402515, 'clip_ratio/high_mean': 0.05122265079990029, 'clip_ratio/high_max': 0.05122265079990029, 'clip_ratio/region_mean': 0.12292635487392545, 'reward_total_mean': 0.7052081227302551, 'reward_meter_mean': 0.954035758972168, 'reward_meter_std': 0.034081488847732544, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9922205209732056, 'reward_repeat_soft_std': 0.008540960028767586, 'reward_judge_quality_mean': 0.581250011920929, 'reward_judge_quality_std': 0.22183892130851746, 'reward_total_composite_mean': 0.7052081227302551, 'reward_total_composite_std': 0.12278620153665543, 'epoch': 0.14}

 43%|████▎     | 1431/3300 [3:31:48<3:33:39,  6.86s/it]INFO 04-13 10:52:46 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1432/3300 [3:31:59<4:15:50,  8.22s/it]2026-04-13 10:52:57,718 | INFO | train_grpo_train | metrics_logged mode=train step=1432

                                                       
{'loss': -0.1825, 'grad_norm': 2.517871141433716, 'learning_rate': 5.663636363636364e-06, 'num_tokens': 2526584.0, 'completions/mean_length': 145.625, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 93.28572082519531, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.66487056016922, 'rewards/meter/std': 0.30841493606567383, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8946514129638672, 'rewards/repeat_soft/std': 0.06716026365756989, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4258626103401184, 'rewards/total_composite/std': 0.1878291219472885, 'reward': 0.4258626103401184, 'reward_std': 0.1878291219472885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13334038853645325, 'sampling/sampling_logp_difference/max': 3.3211793899536133, 'sampling/importance_sampling_ratio/min': 0.03611021861433983, 'sampling/importance_sampling_ratio/mean': 1.012913465499878, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5987378433346748, 'clip_ratio/low_mean': 0.028645833022892475, 'clip_ratio/low_min': 0.028645833022892475, 'clip_ratio/high_mean': 0.06977557484060526, 'clip_ratio/high_max': 0.06977557484060526, 'clip_ratio/region_mean': 0.09842140786349773, 'reward_total_mean': 0.4258626103401184, 'reward_meter_mean': 0.66487056016922, 'reward_meter_std': 0.30841493606567383, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8946514129638672, 'reward_repeat_soft_std': 0.06716026365756989, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4258626103401184, 'reward_total_composite_std': 0.1878291219472885, 'epoch': 0.14}

 43%|████▎     | 1432/3300 [3:31:59<4:15:50,  8.22s/it]INFO 04-13 10:52:57 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1433/3300 [3:32:06<3:57:47,  7.64s/it]2026-04-13 10:53:04,021 | INFO | train_grpo_train | metrics_logged mode=train step=1433

                                                       
{'loss': 0.0887, 'grad_norm': 12.529499053955078, 'learning_rate': 5.6606060606060606e-06, 'num_tokens': 2528272.0, 'completions/mean_length': 50.0, 'completions/min_length': 42.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6653802394866943, 'rewards/meter/std': 0.3758772611618042, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9062150716781616, 'rewards/repeat_soft/std': 0.12127143889665604, 'rewards/judge_quality/mean': 0.6862500309944153, 'rewards/judge_quality/std': 0.27406659722328186, 'rewards/total_composite/mean': 0.6130000352859497, 'rewards/total_composite/std': 0.20596526563167572, 'reward': 0.6130000352859497, 'reward_std': 0.20596528053283691, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1393696516752243, 'sampling/sampling_logp_difference/max': 1.537881851196289, 'sampling/importance_sampling_ratio/min': 0.21483567357063293, 'sampling/importance_sampling_ratio/mean': 0.9889280200004578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.724241353571415, 'clip_ratio/low_mean': 0.05138481268659234, 'clip_ratio/low_min': 0.05138481268659234, 'clip_ratio/high_mean': 0.07794269174337387, 'clip_ratio/high_max': 0.07794269174337387, 'clip_ratio/region_mean': 0.1293275044299662, 'reward_total_mean': 0.6130000352859497, 'reward_meter_mean': 0.6653802394866943, 'reward_meter_std': 0.3758772611618042, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9062150716781616, 'reward_repeat_soft_std': 0.12127143889665604, 'reward_judge_quality_mean': 0.6862500309944153, 'reward_judge_quality_std': 0.27406659722328186, 'reward_total_composite_mean': 0.6130000352859497, 'reward_total_composite_std': 0.20596526563167572, 'epoch': 0.14}

 43%|████▎     | 1433/3300 [3:32:06<3:57:47,  7.64s/it]INFO 04-13 10:53:04 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1434/3300 [3:32:12<3:43:48,  7.20s/it]2026-04-13 10:53:10,176 | INFO | train_grpo_train | metrics_logged mode=train step=1434

                                                       
{'loss': -0.0152, 'grad_norm': 12.750229835510254, 'learning_rate': 5.657575757575759e-06, 'num_tokens': 2529611.0, 'completions/mean_length': 43.375, 'completions/min_length': 36.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.375, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.2933866083621979, 'rewards/meter/std': 0.313712477684021, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9353867769241333, 'rewards/repeat_soft/std': 0.08880577236413956, 'rewards/judge_quality/mean': 0.47749999165534973, 'rewards/judge_quality/std': 0.16184209287166595, 'rewards/total_composite/mean': 0.42346441745758057, 'rewards/total_composite/std': 0.0754440650343895, 'reward': 0.42346441745758057, 'reward_std': 0.0754440575838089, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1647147387266159, 'sampling/sampling_logp_difference/max': 1.2110676765441895, 'sampling/importance_sampling_ratio/min': 0.29787907004356384, 'sampling/importance_sampling_ratio/mean': 1.0240682363510132, 'sampling/importance_sampling_ratio/max': 1.938675880432129, 'entropy': 1.508640632033348, 'clip_ratio/low_mean': 0.07083119731396437, 'clip_ratio/low_min': 0.07083119731396437, 'clip_ratio/high_mean': 0.05808390025049448, 'clip_ratio/high_max': 0.05808390025049448, 'clip_ratio/region_mean': 0.12891509756445885, 'reward_total_mean': 0.42346441745758057, 'reward_meter_mean': 0.2933866083621979, 'reward_meter_std': 0.313712477684021, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9353867769241333, 'reward_repeat_soft_std': 0.08880577236413956, 'reward_judge_quality_mean': 0.47749999165534973, 'reward_judge_quality_std': 0.16184209287166595, 'reward_total_composite_mean': 0.42346441745758057, 'reward_total_composite_std': 0.0754440650343895, 'epoch': 0.14}

 43%|████▎     | 1434/3300 [3:32:12<3:43:48,  7.20s/it]INFO 04-13 10:53:10 [block_pool.py:378] Successfully reset prefix cache

 43%|████▎     | 1435/3300 [3:32:18<3:35:08,  6.92s/it]2026-04-13 10:53:16,458 | INFO | train_grpo_train | metrics_logged mode=train step=1435

                                                       
{'loss': 0.0928, 'grad_norm': 17.791757583618164, 'learning_rate': 5.654545454545455e-06, 'num_tokens': 2531193.0, 'completions/mean_length': 38.75, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.75, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6358752250671387, 'rewards/meter/std': 0.29427236318588257, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9202096462249756, 'rewards/repeat_soft/std': 0.05021527409553528, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.49828529357910156, 'rewards/total_composite/std': 0.07509937137365341, 'reward': 0.49828529357910156, 'reward_std': 0.07509937137365341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1439976990222931, 'sampling/sampling_logp_difference/max': 1.0636416673660278, 'sampling/importance_sampling_ratio/min': 0.3451964259147644, 'sampling/importance_sampling_ratio/mean': 1.0063138008117676, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7885511443018913, 'clip_ratio/low_mean': 0.07318278471939266, 'clip_ratio/low_min': 0.07318278471939266, 'clip_ratio/high_mean': 0.06283324770629406, 'clip_ratio/high_max': 0.06283324770629406, 'clip_ratio/region_mean': 0.13601603242568672, 'reward_total_mean': 0.49828529357910156, 'reward_meter_mean': 0.6358752250671387, 'reward_meter_std': 0.29427236318588257, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9202096462249756, 'reward_repeat_soft_std': 0.05021527409553528, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.49828529357910156, 'reward_total_composite_std': 0.07509937137365341, 'epoch': 0.14}

 43%|████▎     | 1435/3300 [3:32:18<3:35:08,  6.92s/it]INFO 04-13 10:53:16 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1436/3300 [3:32:24<3:22:52,  6.53s/it]2026-04-13 10:53:22,087 | INFO | train_grpo_train | metrics_logged mode=train step=1436

                                                       
{'loss': 0.0895, 'grad_norm': 17.731231689453125, 'learning_rate': 5.651515151515152e-06, 'num_tokens': 2532600.0, 'completions/mean_length': 19.875, 'completions/min_length': 17.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.875, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.8622947335243225, 'rewards/meter/std': 0.32788988947868347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9495408535003662, 'rewards/repeat_soft/std': 0.024242952466011047, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5778375864028931, 'rewards/total_composite/std': 0.0893905982375145, 'reward': 0.5778375864028931, 'reward_std': 0.0893905833363533, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10792140662670135, 'sampling/sampling_logp_difference/max': 1.2819077968597412, 'sampling/importance_sampling_ratio/min': 0.27750736474990845, 'sampling/importance_sampling_ratio/mean': 0.998196542263031, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.515293225646019, 'clip_ratio/low_mean': 0.010869565419852734, 'clip_ratio/low_min': 0.010869565419852734, 'clip_ratio/high_mean': 0.09196286275982857, 'clip_ratio/high_max': 0.09196286275982857, 'clip_ratio/region_mean': 0.1028324281796813, 'reward_total_mean': 0.5778375864028931, 'reward_meter_mean': 0.8622947335243225, 'reward_meter_std': 0.32788988947868347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9495408535003662, 'reward_repeat_soft_std': 0.024242952466011047, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5778375864028931, 'reward_total_composite_std': 0.0893905982375145, 'epoch': 0.14}

 44%|████▎     | 1436/3300 [3:32:24<3:22:52,  6.53s/it]INFO 04-13 10:53:22 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1437/3300 [3:32:30<3:18:00,  6.38s/it]2026-04-13 10:53:28,093 | INFO | train_grpo_train | metrics_logged mode=train step=1437

                                                       
{'loss': 0.0494, 'grad_norm': 12.390629768371582, 'learning_rate': 5.648484848484849e-06, 'num_tokens': 2534172.0, 'completions/mean_length': 38.5, 'completions/min_length': 36.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7535187005996704, 'rewards/meter/std': 0.3593631088733673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9046169519424438, 'rewards/repeat_soft/std': 0.07597479969263077, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.631027340888977, 'rewards/total_composite/std': 0.17327415943145752, 'reward': 0.631027340888977, 'reward_std': 0.17327415943145752, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1142137199640274, 'sampling/sampling_logp_difference/max': 2.5830631256103516, 'sampling/importance_sampling_ratio/min': 0.07554225623607635, 'sampling/importance_sampling_ratio/mean': 1.008995771408081, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6440048366785049, 'clip_ratio/low_mean': 0.07128219399601221, 'clip_ratio/low_min': 0.07128219399601221, 'clip_ratio/high_mean': 0.032894738018512726, 'clip_ratio/high_max': 0.032894738018512726, 'clip_ratio/region_mean': 0.10417693201452494, 'reward_total_mean': 0.631027340888977, 'reward_meter_mean': 0.7535187005996704, 'reward_meter_std': 0.3593631088733673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9046169519424438, 'reward_repeat_soft_std': 0.07597479969263077, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.631027340888977, 'reward_total_composite_std': 0.17327415943145752, 'epoch': 0.14}

 44%|████▎     | 1437/3300 [3:32:30<3:18:00,  6.38s/it]INFO 04-13 10:53:28 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1438/3300 [3:32:37<3:22:17,  6.52s/it]2026-04-13 10:53:34,940 | INFO | train_grpo_train | metrics_logged mode=train step=1438

                                                       
{'loss': 0.0082, 'grad_norm': 8.533262252807617, 'learning_rate': 5.645454545454546e-06, 'num_tokens': 2536644.0, 'completions/mean_length': 93.0, 'completions/min_length': 84.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 93.0, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.6551315784454346, 'rewards/meter/std': 0.4282614588737488, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7912271618843079, 'rewards/repeat_soft/std': 0.07064896076917648, 'rewards/judge_quality/mean': 0.47749996185302734, 'rewards/judge_quality/std': 0.16263456642627716, 'rewards/total_composite/mean': 0.4958537817001343, 'rewards/total_composite/std': 0.1731943041086197, 'reward': 0.4958537817001343, 'reward_std': 0.1731942892074585, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1107155978679657, 'sampling/sampling_logp_difference/max': 1.740966558456421, 'sampling/importance_sampling_ratio/min': 0.1753508448600769, 'sampling/importance_sampling_ratio/mean': 1.0029927492141724, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5855625607073307, 'clip_ratio/low_mean': 0.039472698234021664, 'clip_ratio/low_min': 0.039472698234021664, 'clip_ratio/high_mean': 0.06870181625708938, 'clip_ratio/high_max': 0.06870181625708938, 'clip_ratio/region_mean': 0.10817451449111104, 'reward_total_mean': 0.4958537817001343, 'reward_meter_mean': 0.6551315784454346, 'reward_meter_std': 0.4282614588737488, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7912271618843079, 'reward_repeat_soft_std': 0.07064896076917648, 'reward_judge_quality_mean': 0.47749996185302734, 'reward_judge_quality_std': 0.16263456642627716, 'reward_total_composite_mean': 0.4958537817001343, 'reward_total_composite_std': 0.1731943041086197, 'epoch': 0.14}

 44%|████▎     | 1438/3300 [3:32:37<3:22:17,  6.52s/it]INFO 04-13 10:53:35 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1439/3300 [3:32:43<3:22:22,  6.52s/it]2026-04-13 10:53:41,482 | INFO | train_grpo_train | metrics_logged mode=train step=1439

                                                       
{'loss': 0.0215, 'grad_norm': 10.35162353515625, 'learning_rate': 5.642424242424242e-06, 'num_tokens': 2538462.0, 'completions/mean_length': 65.25, 'completions/min_length': 61.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.3177836239337921, 'rewards/meter/std': 0.27080944180488586, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8657203316688538, 'rewards/repeat_soft/std': 0.06126207858324051, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.4653521776199341, 'rewards/total_composite/std': 0.15516647696495056, 'reward': 0.4653521776199341, 'reward_std': 0.15516646206378937, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11212596297264099, 'sampling/sampling_logp_difference/max': 1.9287055730819702, 'sampling/importance_sampling_ratio/min': 0.14533619582653046, 'sampling/importance_sampling_ratio/mean': 1.0013866424560547, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36079560965299606, 'clip_ratio/low_mean': 0.07574523240327835, 'clip_ratio/low_min': 0.07574523240327835, 'clip_ratio/high_mean': 0.03560520429164171, 'clip_ratio/high_max': 0.03560520429164171, 'clip_ratio/region_mean': 0.11135043669492006, 'reward_total_mean': 0.4653521776199341, 'reward_meter_mean': 0.3177836239337921, 'reward_meter_std': 0.27080944180488586, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8657203316688538, 'reward_repeat_soft_std': 0.06126207858324051, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.4653521776199341, 'reward_total_composite_std': 0.15516647696495056, 'epoch': 0.14}

 44%|████▎     | 1439/3300 [3:32:43<3:22:22,  6.52s/it]INFO 04-13 10:53:41 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1440/3300 [3:32:51<3:33:19,  6.88s/it]2026-04-13 10:53:49,193 | INFO | train_grpo_train | metrics_logged mode=train step=1440

                                                       
{'loss': 0.0367, 'grad_norm': 7.798024654388428, 'learning_rate': 5.6393939393939405e-06, 'num_tokens': 2540855.0, 'completions/mean_length': 94.125, 'completions/min_length': 81.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.125, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.09384974837303162, 'rewards/meter/std': 0.07115980982780457, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8989412784576416, 'rewards/repeat_soft/std': 0.0413278266787529, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.32009556889533997, 'rewards/total_composite/std': 0.019734904170036316, 'reward': 0.32009556889533997, 'reward_std': 0.019734907895326614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12110128998756409, 'sampling/sampling_logp_difference/max': 3.317157745361328, 'sampling/importance_sampling_ratio/min': 0.03625573590397835, 'sampling/importance_sampling_ratio/mean': 1.006064534187317, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6290554516017437, 'clip_ratio/low_mean': 0.07291316520422697, 'clip_ratio/low_min': 0.07291316520422697, 'clip_ratio/high_mean': 0.03113451600074768, 'clip_ratio/high_max': 0.03113451600074768, 'clip_ratio/region_mean': 0.10404768120497465, 'reward_total_mean': 0.32009556889533997, 'reward_meter_mean': 0.09384974837303162, 'reward_meter_std': 0.07115980982780457, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8989412784576416, 'reward_repeat_soft_std': 0.0413278266787529, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.32009556889533997, 'reward_total_composite_std': 0.019734904170036316, 'epoch': 0.14}

 44%|████▎     | 1440/3300 [3:32:51<3:33:19,  6.88s/it]INFO 04-13 10:53:49 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1441/3300 [3:33:02<4:12:55,  8.16s/it]2026-04-13 10:54:00,360 | INFO | train_grpo_train | metrics_logged mode=train step=1441

                                                       
{'loss': -0.0692, 'grad_norm': 3.9723446369171143, 'learning_rate': 5.636363636363636e-06, 'num_tokens': 2542352.0, 'completions/mean_length': 93.125, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.28571701049805, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.3827114999294281, 'rewards/meter/std': 0.3394012451171875, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9972173571586609, 'rewards/repeat_soft/std': 0.005236493889242411, 'rewards/judge_quality/mean': 0.6187499761581421, 'rewards/judge_quality/std': 0.33434104919433594, 'rewards/total_composite/mean': 0.48450589179992676, 'rewards/total_composite/std': 0.2517508268356323, 'reward': 0.48450589179992676, 'reward_std': 0.2517508268356323, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12247990816831589, 'sampling/sampling_logp_difference/max': 1.6840438842773438, 'sampling/importance_sampling_ratio/min': 0.18562181293964386, 'sampling/importance_sampling_ratio/mean': 1.0007109642028809, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3911283649504185, 'clip_ratio/low_mean': 0.047941374592483044, 'clip_ratio/low_min': 0.047941374592483044, 'clip_ratio/high_mean': 0.04840888362377882, 'clip_ratio/high_max': 0.04840888362377882, 'clip_ratio/region_mean': 0.09635025821626186, 'reward_total_mean': 0.48450589179992676, 'reward_meter_mean': 0.3827114999294281, 'reward_meter_std': 0.3394012451171875, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9972173571586609, 'reward_repeat_soft_std': 0.005236493889242411, 'reward_judge_quality_mean': 0.6187499761581421, 'reward_judge_quality_std': 0.33434104919433594, 'reward_total_composite_mean': 0.48450589179992676, 'reward_total_composite_std': 0.2517508268356323, 'epoch': 0.14}

 44%|████▎     | 1441/3300 [3:33:02<4:12:55,  8.16s/it]INFO 04-13 10:54:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1442/3300 [3:33:08<3:53:30,  7.54s/it]2026-04-13 10:54:06,440 | INFO | train_grpo_train | metrics_logged mode=train step=1442

                                                       
{'loss': 0.1372, 'grad_norm': 12.586616516113281, 'learning_rate': 5.633333333333334e-06, 'num_tokens': 2543784.0, 'completions/mean_length': 40.0, 'completions/min_length': 33.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.4086367189884186, 'rewards/meter/std': 0.3910621404647827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9808080196380615, 'rewards/repeat_soft/std': 0.021186163648962975, 'rewards/judge_quality/mean': 0.5262500047683716, 'rewards/judge_quality/std': 0.17079123854637146, 'rewards/total_composite/mean': 0.4698314368724823, 'rewards/total_composite/std': 0.11303868889808655, 'reward': 0.4698314368724823, 'reward_std': 0.11303868889808655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14552311599254608, 'sampling/sampling_logp_difference/max': 1.7716580629348755, 'sampling/importance_sampling_ratio/min': 0.17005079984664917, 'sampling/importance_sampling_ratio/mean': 1.0104730129241943, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7147787809371948, 'clip_ratio/low_mean': 0.08094080351293087, 'clip_ratio/low_min': 0.08094080351293087, 'clip_ratio/high_mean': 0.056038325652480125, 'clip_ratio/high_max': 0.056038325652480125, 'clip_ratio/region_mean': 0.136979129165411, 'reward_total_mean': 0.4698314368724823, 'reward_meter_mean': 0.4086367189884186, 'reward_meter_std': 0.3910621404647827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9808080196380615, 'reward_repeat_soft_std': 0.021186163648962975, 'reward_judge_quality_mean': 0.5262500047683716, 'reward_judge_quality_std': 0.17079123854637146, 'reward_total_composite_mean': 0.4698314368724823, 'reward_total_composite_std': 0.11303868889808655, 'epoch': 0.14}

 44%|████▎     | 1442/3300 [3:33:08<3:53:30,  7.54s/it]INFO 04-13 10:54:06 [block_pool.py:378] Successfully reset prefix cache

 44%|████▎     | 1443/3300 [3:33:19<4:27:02,  8.63s/it]2026-04-13 10:54:17,604 | INFO | train_grpo_train | metrics_logged mode=train step=1443

                                                       
{'loss': -0.1288, 'grad_norm': 4.243305206298828, 'learning_rate': 5.630303030303031e-06, 'num_tokens': 2545698.0, 'completions/mean_length': 120.25, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.28572082519531, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6812485456466675, 'rewards/meter/std': 0.2924301028251648, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9527044296264648, 'rewards/repeat_soft/std': 0.013067924417555332, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.23046152293682098, 'rewards/total_composite/mean': 0.5220375657081604, 'rewards/total_composite/std': 0.2517196238040924, 'reward': 0.5220375657081604, 'reward_std': 0.25171959400177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16307465732097626, 'sampling/sampling_logp_difference/max': 1.9829330444335938, 'sampling/importance_sampling_ratio/min': 0.13766486942768097, 'sampling/importance_sampling_ratio/mean': 0.9993640184402466, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7299496084451675, 'clip_ratio/low_mean': 0.036057693883776665, 'clip_ratio/low_min': 0.036057693883776665, 'clip_ratio/high_mean': 0.11155798844993114, 'clip_ratio/high_max': 0.11155798844993114, 'clip_ratio/region_mean': 0.1476156823337078, 'reward_total_mean': 0.5220375657081604, 'reward_meter_mean': 0.6812485456466675, 'reward_meter_std': 0.2924301028251648, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9527044296264648, 'reward_repeat_soft_std': 0.013067924417555332, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.23046152293682098, 'reward_total_composite_mean': 0.5220375657081604, 'reward_total_composite_std': 0.2517196238040924, 'epoch': 0.14}

 44%|████▎     | 1443/3300 [3:33:19<4:27:02,  8.63s/it]INFO 04-13 10:54:17 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1444/3300 [3:33:27<4:14:49,  8.24s/it]2026-04-13 10:54:24,925 | INFO | train_grpo_train | metrics_logged mode=train step=1444

                                                       
{'loss': 0.0471, 'grad_norm': 6.888050079345703, 'learning_rate': 5.627272727272728e-06, 'num_tokens': 2548026.0, 'completions/mean_length': 92.0, 'completions/min_length': 78.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.0, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.4131786823272705, 'rewards/meter/std': 0.24545283615589142, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8094701766967773, 'rewards/repeat_soft/std': 0.05194660276174545, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.3885853886604309, 'rewards/total_composite/std': 0.05554511770606041, 'reward': 0.3885853886604309, 'reward_std': 0.05554512143135071, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10161849856376648, 'sampling/sampling_logp_difference/max': 2.4903271198272705, 'sampling/importance_sampling_ratio/min': 0.0828828513622284, 'sampling/importance_sampling_ratio/mean': 1.0064361095428467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6244943141937256, 'clip_ratio/low_mean': 0.03710805019363761, 'clip_ratio/low_min': 0.03710805019363761, 'clip_ratio/high_mean': 0.04541767109185457, 'clip_ratio/high_max': 0.04541767109185457, 'clip_ratio/region_mean': 0.08252572128549218, 'reward_total_mean': 0.3885853886604309, 'reward_meter_mean': 0.4131786823272705, 'reward_meter_std': 0.24545283615589142, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8094701766967773, 'reward_repeat_soft_std': 0.05194660276174545, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.3885853886604309, 'reward_total_composite_std': 0.05554511770606041, 'epoch': 0.15}

 44%|████▍     | 1444/3300 [3:33:27<4:14:49,  8.24s/it]INFO 04-13 10:54:25 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1445/3300 [3:33:33<3:55:50,  7.63s/it]2026-04-13 10:54:31,136 | INFO | train_grpo_train | metrics_logged mode=train step=1445

                                                       
{'loss': 0.1322, 'grad_norm': 16.269306182861328, 'learning_rate': 5.624242424242424e-06, 'num_tokens': 2549506.0, 'completions/mean_length': 26.0, 'completions/min_length': 23.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.43530893325805664, 'rewards/meter/std': 0.35694068670272827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9530594348907471, 'rewards/repeat_soft/std': 0.026701925322413445, 'rewards/judge_quality/mean': 0.4437499940395355, 'rewards/judge_quality/std': 0.2084594964981079, 'rewards/total_composite/mean': 0.4565836191177368, 'rewards/total_composite/std': 0.09808404743671417, 'reward': 0.4565836191177368, 'reward_std': 0.09808403998613358, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14200277626514435, 'sampling/sampling_logp_difference/max': 1.2099218368530273, 'sampling/importance_sampling_ratio/min': 0.29822060465812683, 'sampling/importance_sampling_ratio/mean': 1.0090938806533813, 'sampling/importance_sampling_ratio/max': 1.6397645473480225, 'entropy': 0.9584878236055374, 'clip_ratio/low_mean': 0.04872150160372257, 'clip_ratio/low_min': 0.04872150160372257, 'clip_ratio/high_mean': 0.05713768117129803, 'clip_ratio/high_max': 0.05713768117129803, 'clip_ratio/region_mean': 0.1058591827750206, 'reward_total_mean': 0.4565836191177368, 'reward_meter_mean': 0.43530893325805664, 'reward_meter_std': 0.35694068670272827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9530594348907471, 'reward_repeat_soft_std': 0.026701925322413445, 'reward_judge_quality_mean': 0.4437499940395355, 'reward_judge_quality_std': 0.2084594964981079, 'reward_total_composite_mean': 0.4565836191177368, 'reward_total_composite_std': 0.09808404743671417, 'epoch': 0.15}

 44%|████▍     | 1445/3300 [3:33:33<3:55:50,  7.63s/it]INFO 04-13 10:54:31 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1446/3300 [3:33:39<3:38:41,  7.08s/it]2026-04-13 10:54:36,924 | INFO | train_grpo_train | metrics_logged mode=train step=1446

                                                       
{'loss': -0.0025, 'grad_norm': 12.225137710571289, 'learning_rate': 5.6212121212121215e-06, 'num_tokens': 2551060.0, 'completions/mean_length': 32.25, 'completions/min_length': 28.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9132502675056458, 'rewards/meter/std': 0.09246927499771118, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9432229995727539, 'rewards/repeat_soft/std': 0.05181838199496269, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.5862990617752075, 'rewards/total_composite/std': 0.04857126250863075, 'reward': 0.5862990617752075, 'reward_std': 0.04857126623392105, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09656811505556107, 'sampling/sampling_logp_difference/max': 1.240281581878662, 'sampling/importance_sampling_ratio/min': 0.2893027663230896, 'sampling/importance_sampling_ratio/mean': 1.0012781620025635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5067105516791344, 'clip_ratio/low_mean': 0.04288594517856836, 'clip_ratio/low_min': 0.04288594517856836, 'clip_ratio/high_mean': 0.04291552724316716, 'clip_ratio/high_max': 0.04291552724316716, 'clip_ratio/region_mean': 0.08580147242173553, 'reward_total_mean': 0.5862990617752075, 'reward_meter_mean': 0.9132502675056458, 'reward_meter_std': 0.09246927499771118, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9432229995727539, 'reward_repeat_soft_std': 0.05181838199496269, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.5862990617752075, 'reward_total_composite_std': 0.04857126250863075, 'epoch': 0.15}

 44%|████▍     | 1446/3300 [3:33:39<3:38:41,  7.08s/it]INFO 04-13 10:54:37 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1447/3300 [3:33:45<3:28:06,  6.74s/it]2026-04-13 10:54:42,873 | INFO | train_grpo_train | metrics_logged mode=train step=1447

                                                       
{'loss': 0.068, 'grad_norm': 13.407052040100098, 'learning_rate': 5.618181818181818e-06, 'num_tokens': 2552642.0, 'completions/mean_length': 37.75, 'completions/min_length': 32.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.4962160289287567, 'rewards/meter/std': 0.3201512396335602, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9090368747711182, 'rewards/repeat_soft/std': 0.04927627369761467, 'rewards/judge_quality/mean': 0.7112500071525574, 'rewards/judge_quality/std': 0.24793073534965515, 'rewards/total_composite/mean': 0.5681751370429993, 'rewards/total_composite/std': 0.17873920500278473, 'reward': 0.5681751370429993, 'reward_std': 0.17873920500278473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.118393175303936, 'sampling/sampling_logp_difference/max': 1.5386567115783691, 'sampling/importance_sampling_ratio/min': 0.21466928720474243, 'sampling/importance_sampling_ratio/mean': 0.9922995567321777, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6725741475820541, 'clip_ratio/low_mean': 0.05689032329246402, 'clip_ratio/low_min': 0.05689032329246402, 'clip_ratio/high_mean': 0.04010617733001709, 'clip_ratio/high_max': 0.04010617733001709, 'clip_ratio/region_mean': 0.09699650062248111, 'reward_total_mean': 0.5681751370429993, 'reward_meter_mean': 0.4962160289287567, 'reward_meter_std': 0.3201512396335602, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9090368747711182, 'reward_repeat_soft_std': 0.04927627369761467, 'reward_judge_quality_mean': 0.7112500071525574, 'reward_judge_quality_std': 0.24793073534965515, 'reward_total_composite_mean': 0.5681751370429993, 'reward_total_composite_std': 0.17873920500278473, 'epoch': 0.15}

 44%|████▍     | 1447/3300 [3:33:45<3:28:06,  6.74s/it]INFO 04-13 10:54:43 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1448/3300 [3:33:50<3:19:44,  6.47s/it]2026-04-13 10:54:48,721 | INFO | train_grpo_train | metrics_logged mode=train step=1448

                                                       
{'loss': 0.0022, 'grad_norm': 18.25330352783203, 'learning_rate': 5.615151515151516e-06, 'num_tokens': 2554090.0, 'completions/mean_length': 34.0, 'completions/min_length': 31.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7931654453277588, 'rewards/meter/std': 0.2793334424495697, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9400959014892578, 'rewards/repeat_soft/std': 0.05659913271665573, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5964663624763489, 'rewards/total_composite/std': 0.14883674681186676, 'reward': 0.5964663624763489, 'reward_std': 0.14883676171302795, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10333961993455887, 'sampling/sampling_logp_difference/max': 1.230234146118164, 'sampling/importance_sampling_ratio/min': 0.29222413897514343, 'sampling/importance_sampling_ratio/mean': 0.998220682144165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5057181008160114, 'clip_ratio/low_mean': 0.0633125938475132, 'clip_ratio/low_min': 0.0633125938475132, 'clip_ratio/high_mean': 0.04446381703019142, 'clip_ratio/high_max': 0.04446381703019142, 'clip_ratio/region_mean': 0.10777641087770462, 'reward_total_mean': 0.5964663624763489, 'reward_meter_mean': 0.7931654453277588, 'reward_meter_std': 0.2793334424495697, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9400959014892578, 'reward_repeat_soft_std': 0.05659913271665573, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5964663624763489, 'reward_total_composite_std': 0.14883674681186676, 'epoch': 0.15}

 44%|████▍     | 1448/3300 [3:33:50<3:19:44,  6.47s/it]INFO 04-13 10:54:48 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1449/3300 [3:33:57<3:20:26,  6.50s/it]2026-04-13 10:54:55,277 | INFO | train_grpo_train | metrics_logged mode=train step=1449

                                                       
{'loss': -0.0235, 'grad_norm': 11.540170669555664, 'learning_rate': 5.612121212121212e-06, 'num_tokens': 2556021.0, 'completions/mean_length': 67.375, 'completions/min_length': 55.0, 'completions/max_length': 87.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.375, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 87.0, 'rewards/meter/mean': 0.9395323991775513, 'rewards/meter/std': 0.05084110423922539, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8415687084197998, 'rewards/repeat_soft/std': 0.03868628293275833, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5630892515182495, 'rewards/total_composite/std': 0.04056117311120033, 'reward': 0.5630892515182495, 'reward_std': 0.04056116193532944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08219168335199356, 'sampling/sampling_logp_difference/max': 2.3244993686676025, 'sampling/importance_sampling_ratio/min': 0.09783240407705307, 'sampling/importance_sampling_ratio/mean': 1.0111548900604248, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.37682488933205605, 'clip_ratio/low_mean': 0.03030737955123186, 'clip_ratio/low_min': 0.03030737955123186, 'clip_ratio/high_mean': 0.04899826226755977, 'clip_ratio/high_max': 0.04899826226755977, 'clip_ratio/region_mean': 0.07930564181879163, 'reward_total_mean': 0.5630892515182495, 'reward_meter_mean': 0.9395323991775513, 'reward_meter_std': 0.05084110423922539, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8415687084197998, 'reward_repeat_soft_std': 0.03868628293275833, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5630892515182495, 'reward_total_composite_std': 0.04056117311120033, 'epoch': 0.15}

 44%|████▍     | 1449/3300 [3:33:57<3:20:26,  6.50s/it]INFO 04-13 10:54:55 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1450/3300 [3:34:04<3:26:08,  6.69s/it]2026-04-13 10:55:02,431 | INFO | train_grpo_train | metrics_logged mode=train step=1450

                                                       
{'loss': 0.0188, 'grad_norm': 11.66684627532959, 'learning_rate': 5.60909090909091e-06, 'num_tokens': 2557654.0, 'completions/mean_length': 42.125, 'completions/min_length': 36.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.3843640089035034, 'rewards/meter/std': 0.3152768909931183, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9392750263214111, 'rewards/repeat_soft/std': 0.03424127772450447, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.45239052176475525, 'rewards/total_composite/std': 0.09135284274816513, 'reward': 0.45239052176475525, 'reward_std': 0.09135284274816513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12498661130666733, 'sampling/sampling_logp_difference/max': 2.2886857986450195, 'sampling/importance_sampling_ratio/min': 0.10139963775873184, 'sampling/importance_sampling_ratio/mean': 1.0153913497924805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7969678267836571, 'clip_ratio/low_mean': 0.09065534360706806, 'clip_ratio/low_min': 0.09065534360706806, 'clip_ratio/high_mean': 0.03164640720933676, 'clip_ratio/high_max': 0.03164640720933676, 'clip_ratio/region_mean': 0.12230175081640482, 'reward_total_mean': 0.45239052176475525, 'reward_meter_mean': 0.3843640089035034, 'reward_meter_std': 0.3152768909931183, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9392750263214111, 'reward_repeat_soft_std': 0.03424127772450447, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.45239052176475525, 'reward_total_composite_std': 0.09135284274816513, 'epoch': 0.15}

 44%|████▍     | 1450/3300 [3:34:04<3:26:08,  6.69s/it]INFO 04-13 10:55:02 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.77s/it]

 30%|███       | 3/10 [00:08<00:22,  3.17s/it]

 40%|████      | 4/10 [00:12<00:19,  3.26s/it]

 50%|█████     | 5/10 [00:16<00:17,  3.52s/it]

 60%|██████    | 6/10 [00:19<00:14,  3.53s/it]

 70%|███████   | 7/10 [00:23<00:10,  3.63s/it]

 80%|████████  | 8/10 [00:26<00:07,  3.53s/it]

 90%|█████████ | 9/10 [00:30<00:03,  3.57s/it]

100%|██████████| 10/10 [00:33<00:00,  3.46s/it]2026-04-13 10:55:40,133 | INFO | train_grpo_train | metrics_logged mode=eval step=1450

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 37.6201, 'eval_samples_per_second': 2.127, 'eval_steps_per_second': 0.266, 'eval_num_tokens': 2557654.0, 'eval_completions/mean_length': 65.45, 'eval_completions/min_length': 30.7, 'eval_completions/max_length': 103.8, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 65.45, 'eval_completions/min_terminated_length': 30.7, 'eval_completions/max_terminated_length': 103.8, 'eval_rewards/meter/mean': 0.5202593326568603, 'eval_rewards/meter/std': 0.3611732065677643, 'eval_rewards/count_adherence/mean': 0.9629166543483734, 'eval_rewards/count_adherence/std': 0.06404209546744824, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/repeat_soft/mean': 0.9235949277877807, 'eval_rewards/repeat_soft/std': 0.06313590295612811, 'eval_rewards/judge_quality/mean': 0.4716249942779541, 'eval_rewards/judge_quality/std': 0.1344312877394259, 'eval_rewards/total_composite/mean': 0.4841545969247818, 'eval_rewards/total_composite/std': 0.11240602880716324, 'eval_reward': 0.4841545969247818, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06021547056734562, 'eval_sampling/sampling_logp_difference/max': 1.0568527698516845, 'eval_sampling/importance_sampling_ratio/min': 0.354967337846756, 'eval_sampling/importance_sampling_ratio/mean': 1.0121062040328979, 'eval_sampling/importance_sampling_ratio/max': 1.403704333305359, 'eval_entropy': 0.6505220174789429, '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': 0.4841545969247818, 'eval_reward_meter_mean': 0.5202593326568603, 'eval_reward_meter_std': 0.3611732065677643, 'eval_reward_count_adherence_mean': 0.9629166543483734, 'eval_reward_count_adherence_std': 0.06404209546744824, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_repeat_soft_mean': 0.9235949277877807, 'eval_reward_repeat_soft_std': 0.06313590295612811, 'eval_reward_judge_quality_mean': 0.4716249942779541, 'eval_reward_judge_quality_std': 0.1344312877394259, 'eval_reward_total_composite_mean': 0.4841545969247818, 'eval_reward_total_composite_std': 0.11240602880716324, 'epoch': 0.15}

 44%|████▍     | 1450/3300 [3:34:42<3:26:08,  6.69s/it]

100%|██████████| 10/10 [00:33<00:00,  3.46s/it]

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

 44%|████▍     | 1451/3300 [3:34:51<9:33:49, 18.62s/it]2026-04-13 10:55:48,887 | INFO | train_grpo_train | metrics_logged mode=train step=1451

                                                       
{'loss': -0.0034, 'grad_norm': 16.953227996826172, 'learning_rate': 5.606060606060606e-06, 'num_tokens': 2559106.0, 'completions/mean_length': 28.5, 'completions/min_length': 26.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6778678297996521, 'rewards/meter/std': 0.3919105529785156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9157227277755737, 'rewards/repeat_soft/std': 0.05926491692662239, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6333736777305603, 'rewards/total_composite/std': 0.22816839814186096, 'reward': 0.6333736777305603, 'reward_std': 0.22816839814186096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11051974445581436, 'sampling/sampling_logp_difference/max': 1.0624847412109375, 'sampling/importance_sampling_ratio/min': 0.3455960154533386, 'sampling/importance_sampling_ratio/mean': 1.0067046880722046, 'sampling/importance_sampling_ratio/max': 1.7910748720169067, 'entropy': 0.6300291679799557, 'clip_ratio/low_mean': 0.06631200667470694, 'clip_ratio/low_min': 0.06631200667470694, 'clip_ratio/high_mean': 0.0210628192871809, 'clip_ratio/high_max': 0.0210628192871809, 'clip_ratio/region_mean': 0.08737482596188784, 'reward_total_mean': 0.6333736777305603, 'reward_meter_mean': 0.6778678297996521, 'reward_meter_std': 0.3919105529785156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9157227277755737, 'reward_repeat_soft_std': 0.05926491692662239, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6333736777305603, 'reward_total_composite_std': 0.22816839814186096, 'epoch': 0.15}

 44%|████▍     | 1451/3300 [3:34:51<9:33:49, 18.62s/it]INFO 04-13 10:55:49 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1452/3300 [3:34:56<7:33:30, 14.72s/it]2026-04-13 10:55:54,506 | INFO | train_grpo_train | metrics_logged mode=train step=1452

                                                       
{'loss': 0.0033, 'grad_norm': 14.863313674926758, 'learning_rate': 5.603030303030303e-06, 'num_tokens': 2560431.0, 'completions/mean_length': 17.625, 'completions/min_length': 15.0, 'completions/max_length': 20.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 20.0, 'rewards/meter/mean': 0.9673625230789185, 'rewards/meter/std': 0.04691013693809509, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.8443201780319214, 'rewards/total_composite/std': 0.1482287347316742, 'reward': 0.8443201780319214, 'reward_std': 0.1482287347316742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1140875294804573, 'sampling/sampling_logp_difference/max': 1.420192003250122, 'sampling/importance_sampling_ratio/min': 0.24166762828826904, 'sampling/importance_sampling_ratio/mean': 1.0103696584701538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6248944774270058, 'clip_ratio/low_mean': 0.04411764815449715, 'clip_ratio/low_min': 0.04411764815449715, 'clip_ratio/high_mean': 0.08975694654509425, 'clip_ratio/high_max': 0.08975694654509425, 'clip_ratio/region_mean': 0.1338745946995914, 'reward_total_mean': 0.8443201780319214, 'reward_meter_mean': 0.9673625230789185, 'reward_meter_std': 0.04691013693809509, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.8443201780319214, 'reward_total_composite_std': 0.1482287347316742, 'epoch': 0.15}

 44%|████▍     | 1452/3300 [3:34:56<7:33:30, 14.72s/it]INFO 04-13 10:55:54 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1453/3300 [3:35:02<6:14:02, 12.15s/it]2026-04-13 10:56:00,670 | INFO | train_grpo_train | metrics_logged mode=train step=1453

                                                       
{'loss': 0.0488, 'grad_norm': 11.638545989990234, 'learning_rate': 5.600000000000001e-06, 'num_tokens': 2562004.0, 'completions/mean_length': 40.625, 'completions/min_length': 35.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.6810108423233032, 'rewards/meter/std': 0.2987150549888611, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9620885848999023, 'rewards/repeat_soft/std': 0.04251923784613609, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.19334925711154938, 'rewards/total_composite/mean': 0.5764018893241882, 'rewards/total_composite/std': 0.11793755739927292, 'reward': 0.5764018893241882, 'reward_std': 0.11793754994869232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11488058418035507, 'sampling/sampling_logp_difference/max': 1.1022356748580933, 'sampling/importance_sampling_ratio/min': 0.3321276903152466, 'sampling/importance_sampling_ratio/mean': 1.0061787366867065, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6907035112380981, 'clip_ratio/low_mean': 0.030543780885636806, 'clip_ratio/low_min': 0.030543780885636806, 'clip_ratio/high_mean': 0.07158117881044745, 'clip_ratio/high_max': 0.07158117881044745, 'clip_ratio/region_mean': 0.10212495969608426, 'reward_total_mean': 0.5764018893241882, 'reward_meter_mean': 0.6810108423233032, 'reward_meter_std': 0.2987150549888611, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9620885848999023, 'reward_repeat_soft_std': 0.04251923784613609, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.19334925711154938, 'reward_total_composite_mean': 0.5764018893241882, 'reward_total_composite_std': 0.11793755739927292, 'epoch': 0.15}

 44%|████▍     | 1453/3300 [3:35:02<6:14:02, 12.15s/it]INFO 04-13 10:56:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1454/3300 [3:35:10<5:30:07, 10.73s/it]2026-04-13 10:56:08,070 | INFO | train_grpo_train | metrics_logged mode=train step=1454

                                                       
{'loss': -0.0124, 'grad_norm': 7.159426689147949, 'learning_rate': 5.596969696969697e-06, 'num_tokens': 2564393.0, 'completions/mean_length': 115.625, 'completions/min_length': 109.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.625, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9090392589569092, 'rewards/meter/std': 0.18287572264671326, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9738054275512695, 'rewards/repeat_soft/std': 0.025494271889328957, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.626960039138794, 'rewards/total_composite/std': 0.055665262043476105, 'reward': 0.626960039138794, 'reward_std': 0.05566524714231491, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12418323010206223, 'sampling/sampling_logp_difference/max': 2.0376601219177246, 'sampling/importance_sampling_ratio/min': 0.1303333193063736, 'sampling/importance_sampling_ratio/mean': 1.0259209871292114, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.839865043759346, 'clip_ratio/low_mean': 0.11573239602148533, 'clip_ratio/low_min': 0.11573239602148533, 'clip_ratio/high_mean': 0.012096773833036423, 'clip_ratio/high_max': 0.012096773833036423, 'clip_ratio/region_mean': 0.12782916985452175, 'reward_total_mean': 0.626960039138794, 'reward_meter_mean': 0.9090392589569092, 'reward_meter_std': 0.18287572264671326, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9738054275512695, 'reward_repeat_soft_std': 0.025494271889328957, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.626960039138794, 'reward_total_composite_std': 0.055665262043476105, 'epoch': 0.15}

 44%|████▍     | 1454/3300 [3:35:10<5:30:07, 10.73s/it]INFO 04-13 10:56:08 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1455/3300 [3:35:16<4:51:36,  9.48s/it]2026-04-13 10:56:14,640 | INFO | train_grpo_train | metrics_logged mode=train step=1455

                                                       
{'loss': 0.0213, 'grad_norm': 10.604643821716309, 'learning_rate': 5.593939393939395e-06, 'num_tokens': 2566076.0, 'completions/mean_length': 47.375, 'completions/min_length': 40.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.4165652394294739, 'rewards/meter/std': 0.3665863275527954, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9656630754470825, 'rewards/repeat_soft/std': 0.025733020156621933, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.20057062804698944, 'rewards/total_composite/mean': 0.5180891752243042, 'rewards/total_composite/std': 0.19815339148044586, 'reward': 0.5180891752243042, 'reward_std': 0.19815339148044586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1462828516960144, 'sampling/sampling_logp_difference/max': 1.6169003248214722, 'sampling/importance_sampling_ratio/min': 0.19851306080818176, 'sampling/importance_sampling_ratio/mean': 1.005271077156067, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9341767653822899, 'clip_ratio/low_mean': 0.07179701700806618, 'clip_ratio/low_min': 0.07179701700806618, 'clip_ratio/high_mean': 0.07825229875743389, 'clip_ratio/high_max': 0.07825229875743389, 'clip_ratio/region_mean': 0.15004931576550007, 'reward_total_mean': 0.5180891752243042, 'reward_meter_mean': 0.4165652394294739, 'reward_meter_std': 0.3665863275527954, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9656630754470825, 'reward_repeat_soft_std': 0.025733020156621933, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.20057062804698944, 'reward_total_composite_mean': 0.5180891752243042, 'reward_total_composite_std': 0.19815339148044586, 'epoch': 0.15}

 44%|████▍     | 1455/3300 [3:35:16<4:51:36,  9.48s/it]INFO 04-13 10:56:14 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1456/3300 [3:35:23<4:23:30,  8.57s/it]2026-04-13 10:56:21,095 | INFO | train_grpo_train | metrics_logged mode=train step=1456

                                                       
{'loss': 0.0332, 'grad_norm': 11.490345001220703, 'learning_rate': 5.5909090909090915e-06, 'num_tokens': 2567633.0, 'completions/mean_length': 36.625, 'completions/min_length': 32.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.7870504856109619, 'rewards/meter/std': 0.2520248293876648, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9496994614601135, 'rewards/repeat_soft/std': 0.06635992974042892, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5600391626358032, 'rewards/total_composite/std': 0.06313219666481018, 'reward': 0.5600391626358032, 'reward_std': 0.06313218176364899, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08639819175004959, 'sampling/sampling_logp_difference/max': 1.2901191711425781, 'sampling/importance_sampling_ratio/min': 0.2752379775047302, 'sampling/importance_sampling_ratio/mean': 0.9957996010780334, 'sampling/importance_sampling_ratio/max': 1.5556946992874146, 'entropy': 0.4823068752884865, 'clip_ratio/low_mean': 0.044285462237894535, 'clip_ratio/low_min': 0.044285462237894535, 'clip_ratio/high_mean': 0.05146470386534929, 'clip_ratio/high_max': 0.05146470386534929, 'clip_ratio/region_mean': 0.09575016610324383, 'reward_total_mean': 0.5600391626358032, 'reward_meter_mean': 0.7870504856109619, 'reward_meter_std': 0.2520248293876648, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9496994614601135, 'reward_repeat_soft_std': 0.06635992974042892, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5600391626358032, 'reward_total_composite_std': 0.06313219666481018, 'epoch': 0.15}

 44%|████▍     | 1456/3300 [3:35:23<4:23:30,  8.57s/it]INFO 04-13 10:56:21 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1457/3300 [3:35:30<4:09:35,  8.13s/it]2026-04-13 10:56:28,174 | INFO | train_grpo_train | metrics_logged mode=train step=1457

                                                       
{'loss': 0.042, 'grad_norm': 11.235291481018066, 'learning_rate': 5.587878787878789e-06, 'num_tokens': 2569854.0, 'completions/mean_length': 98.625, 'completions/min_length': 87.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.625, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.7525693774223328, 'rewards/meter/std': 0.2684197425842285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9529146552085876, 'rewards/repeat_soft/std': 0.03146170824766159, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5393280982971191, 'rewards/total_composite/std': 0.0792858675122261, 'reward': 0.5393280982971191, 'reward_std': 0.07928586006164551, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1270921230316162, 'sampling/sampling_logp_difference/max': 1.6169508695602417, 'sampling/importance_sampling_ratio/min': 0.1985030472278595, 'sampling/importance_sampling_ratio/mean': 1.026190996170044, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7462729886174202, 'clip_ratio/low_mean': 0.043929059989750385, 'clip_ratio/low_min': 0.043929059989750385, 'clip_ratio/high_mean': 0.05161045212298632, 'clip_ratio/high_max': 0.05161045212298632, 'clip_ratio/region_mean': 0.0955395121127367, 'reward_total_mean': 0.5393280982971191, 'reward_meter_mean': 0.7525693774223328, 'reward_meter_std': 0.2684197425842285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9529146552085876, 'reward_repeat_soft_std': 0.03146170824766159, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5393280982971191, 'reward_total_composite_std': 0.0792858675122261, 'epoch': 0.15}

 44%|████▍     | 1457/3300 [3:35:30<4:09:35,  8.13s/it]INFO 04-13 10:56:28 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1458/3300 [3:35:36<3:54:59,  7.65s/it]2026-04-13 10:56:34,734 | INFO | train_grpo_train | metrics_logged mode=train step=1458

                                                       
{'loss': 0.043, 'grad_norm': 7.706801891326904, 'learning_rate': 5.584848484848485e-06, 'num_tokens': 2572174.0, 'completions/mean_length': 88.0, 'completions/min_length': 80.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.0, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.7622480392456055, 'rewards/meter/std': 0.2639716863632202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8169191479682922, 'rewards/repeat_soft/std': 0.05823111906647682, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5306315422058105, 'rewards/total_composite/std': 0.06933984160423279, 'reward': 0.5306315422058105, 'reward_std': 0.06933984160423279, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08421962708234787, 'sampling/sampling_logp_difference/max': 1.8180437088012695, 'sampling/importance_sampling_ratio/min': 0.16234302520751953, 'sampling/importance_sampling_ratio/mean': 1.0130640268325806, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43276629596948624, 'clip_ratio/low_mean': 0.030778031796216965, 'clip_ratio/low_min': 0.030778031796216965, 'clip_ratio/high_mean': 0.05446739564649761, 'clip_ratio/high_max': 0.05446739564649761, 'clip_ratio/region_mean': 0.08524542744271457, 'reward_total_mean': 0.5306315422058105, 'reward_meter_mean': 0.7622480392456055, 'reward_meter_std': 0.2639716863632202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8169191479682922, 'reward_repeat_soft_std': 0.05823111906647682, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5306315422058105, 'reward_total_composite_std': 0.06933984160423279, 'epoch': 0.15}

 44%|████▍     | 1458/3300 [3:35:36<3:54:59,  7.65s/it]INFO 04-13 10:56:34 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1459/3300 [3:35:42<3:39:08,  7.14s/it]2026-04-13 10:56:40,675 | INFO | train_grpo_train | metrics_logged mode=train step=1459

                                                       
{'loss': 0.0239, 'grad_norm': 12.355224609375, 'learning_rate': 5.5818181818181824e-06, 'num_tokens': 2573928.0, 'completions/mean_length': 44.25, 'completions/min_length': 39.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.6194192171096802, 'rewards/meter/std': 0.3770742416381836, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9371640682220459, 'rewards/repeat_soft/std': 0.05107033997774124, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5168610215187073, 'rewards/total_composite/std': 0.1104142963886261, 'reward': 0.5168610215187073, 'reward_std': 0.1104142889380455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10699944943189621, 'sampling/sampling_logp_difference/max': 1.3431096076965332, 'sampling/importance_sampling_ratio/min': 0.26103270053863525, 'sampling/importance_sampling_ratio/mean': 1.0050694942474365, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5965271331369877, 'clip_ratio/low_mean': 0.03899620100855827, 'clip_ratio/low_min': 0.03899620100855827, 'clip_ratio/high_mean': 0.056769791059195995, 'clip_ratio/high_max': 0.056769791059195995, 'clip_ratio/region_mean': 0.09576599206775427, 'reward_total_mean': 0.5168610215187073, 'reward_meter_mean': 0.6194192171096802, 'reward_meter_std': 0.3770742416381836, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9371640682220459, 'reward_repeat_soft_std': 0.05107033997774124, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5168610215187073, 'reward_total_composite_std': 0.1104142963886261, 'epoch': 0.15}

 44%|████▍     | 1459/3300 [3:35:42<3:39:08,  7.14s/it]INFO 04-13 10:56:40 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1460/3300 [3:35:49<3:37:43,  7.10s/it]2026-04-13 10:56:47,688 | INFO | train_grpo_train | metrics_logged mode=train step=1460

                                                       
{'loss': -0.0616, 'grad_norm': 11.842880249023438, 'learning_rate': 5.578787878787879e-06, 'num_tokens': 2576086.0, 'completions/mean_length': 92.75, 'completions/min_length': 66.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.75, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.05187898874282837, 'rewards/meter/std': 0.07328092306852341, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.966722846031189, 'rewards/repeat_soft/std': 0.021765653043985367, 'rewards/judge_quality/mean': 0.5187499523162842, 'rewards/judge_quality/std': 0.17908000946044922, 'rewards/total_composite/mean': 0.33348724246025085, 'rewards/total_composite/std': 0.04362621530890465, 'reward': 0.33348724246025085, 'reward_std': 0.04362621530890465, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15613049268722534, 'sampling/sampling_logp_difference/max': 3.2224905490875244, 'sampling/importance_sampling_ratio/min': 0.03985567018389702, 'sampling/importance_sampling_ratio/mean': 0.9895204305648804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5575118139386177, 'clip_ratio/low_mean': 0.10854641254991293, 'clip_ratio/low_min': 0.10854641254991293, 'clip_ratio/high_mean': 0.03530109953135252, 'clip_ratio/high_max': 0.03530109953135252, 'clip_ratio/region_mean': 0.14384751208126545, 'reward_total_mean': 0.33348724246025085, 'reward_meter_mean': 0.05187898874282837, 'reward_meter_std': 0.07328092306852341, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.966722846031189, 'reward_repeat_soft_std': 0.021765653043985367, 'reward_judge_quality_mean': 0.5187499523162842, 'reward_judge_quality_std': 0.17908000946044922, 'reward_total_composite_mean': 0.33348724246025085, 'reward_total_composite_std': 0.04362621530890465, 'epoch': 0.15}

 44%|████▍     | 1460/3300 [3:35:49<3:37:43,  7.10s/it]INFO 04-13 10:56:47 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1461/3300 [3:35:56<3:32:51,  6.94s/it]2026-04-13 10:56:54,267 | INFO | train_grpo_train | metrics_logged mode=train step=1461

                                                       
{'loss': 0.0691, 'grad_norm': 20.209810256958008, 'learning_rate': 5.575757575757577e-06, 'num_tokens': 2577949.0, 'completions/mean_length': 53.875, 'completions/min_length': 43.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9132522940635681, 'rewards/meter/std': 0.17970865964889526, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8632922172546387, 'rewards/repeat_soft/std': 0.07984157651662827, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5811365842819214, 'rewards/total_composite/std': 0.04944567382335663, 'reward': 0.5811365842819214, 'reward_std': 0.04944567382335663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11274424195289612, 'sampling/sampling_logp_difference/max': 1.6341705322265625, 'sampling/importance_sampling_ratio/min': 0.1951141357421875, 'sampling/importance_sampling_ratio/mean': 1.0093806982040405, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6105680763721466, 'clip_ratio/low_mean': 0.01848655566573143, 'clip_ratio/low_min': 0.01848655566573143, 'clip_ratio/high_mean': 0.06229339726269245, 'clip_ratio/high_max': 0.06229339726269245, 'clip_ratio/region_mean': 0.08077995292842388, 'reward_total_mean': 0.5811365842819214, 'reward_meter_mean': 0.9132522940635681, 'reward_meter_std': 0.17970865964889526, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8632922172546387, 'reward_repeat_soft_std': 0.07984157651662827, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5811365842819214, 'reward_total_composite_std': 0.04944567382335663, 'epoch': 0.15}

 44%|████▍     | 1461/3300 [3:35:56<3:32:51,  6.94s/it]INFO 04-13 10:56:54 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1462/3300 [3:36:02<3:24:30,  6.68s/it]2026-04-13 10:57:00,311 | INFO | train_grpo_train | metrics_logged mode=train step=1462

                                                       
{'loss': 0.0297, 'grad_norm': 11.300406455993652, 'learning_rate': 5.572727272727273e-06, 'num_tokens': 2579675.0, 'completions/mean_length': 59.75, 'completions/min_length': 51.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4085647165775299, 'rewards/meter/std': 0.404621422290802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.907824695110321, 'rewards/repeat_soft/std': 0.06312815845012665, 'rewards/judge_quality/mean': 0.45624998211860657, 'rewards/judge_quality/std': 0.18126046657562256, 'rewards/total_composite/mean': 0.46732097864151, 'rewards/total_composite/std': 0.16994957625865936, 'reward': 0.46732097864151, 'reward_std': 0.16994957625865936, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13315221667289734, 'sampling/sampling_logp_difference/max': 3.153635263442993, 'sampling/importance_sampling_ratio/min': 0.04269663244485855, 'sampling/importance_sampling_ratio/mean': 1.0035516023635864, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6467822268605232, 'clip_ratio/low_mean': 0.08823794219642878, 'clip_ratio/low_min': 0.08823794219642878, 'clip_ratio/high_mean': 0.03578192740678787, 'clip_ratio/high_max': 0.03578192740678787, 'clip_ratio/region_mean': 0.12401986960321665, 'reward_total_mean': 0.46732097864151, 'reward_meter_mean': 0.4085647165775299, 'reward_meter_std': 0.404621422290802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.907824695110321, 'reward_repeat_soft_std': 0.06312815845012665, 'reward_judge_quality_mean': 0.45624998211860657, 'reward_judge_quality_std': 0.18126046657562256, 'reward_total_composite_mean': 0.46732097864151, 'reward_total_composite_std': 0.16994957625865936, 'epoch': 0.15}

 44%|████▍     | 1462/3300 [3:36:02<3:24:30,  6.68s/it]INFO 04-13 10:57:00 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1463/3300 [3:36:08<3:21:59,  6.60s/it]2026-04-13 10:57:06,724 | INFO | train_grpo_train | metrics_logged mode=train step=1463

                                                       
{'loss': 0.0272, 'grad_norm': 13.791470527648926, 'learning_rate': 5.569696969696971e-06, 'num_tokens': 2581272.0, 'completions/mean_length': 29.625, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.7594460248947144, 'rewards/meter/std': 0.2380126267671585, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8953207731246948, 'rewards/repeat_soft/std': 0.11132125556468964, 'rewards/judge_quality/mean': 0.3675000071525574, 'rewards/judge_quality/std': 0.1348809003829956, 'rewards/total_composite/mean': 0.5211995840072632, 'rewards/total_composite/std': 0.10588626563549042, 'reward': 0.5211995840072632, 'reward_std': 0.10588628053665161, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12958385050296783, 'sampling/sampling_logp_difference/max': 1.413332223892212, 'sampling/importance_sampling_ratio/min': 0.24333108961582184, 'sampling/importance_sampling_ratio/mean': 0.9924678802490234, 'sampling/importance_sampling_ratio/max': 1.768438458442688, 'entropy': 0.7009607888758183, 'clip_ratio/low_mean': 0.058047713711857796, 'clip_ratio/low_min': 0.058047713711857796, 'clip_ratio/high_mean': 0.06673027109354734, 'clip_ratio/high_max': 0.06673027109354734, 'clip_ratio/region_mean': 0.12477798480540514, 'reward_total_mean': 0.5211995840072632, 'reward_meter_mean': 0.7594460248947144, 'reward_meter_std': 0.2380126267671585, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8953207731246948, 'reward_repeat_soft_std': 0.11132125556468964, 'reward_judge_quality_mean': 0.3675000071525574, 'reward_judge_quality_std': 0.1348809003829956, 'reward_total_composite_mean': 0.5211995840072632, 'reward_total_composite_std': 0.10588626563549042, 'epoch': 0.15}

 44%|████▍     | 1463/3300 [3:36:08<3:21:59,  6.60s/it]INFO 04-13 10:57:06 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1464/3300 [3:36:16<3:27:24,  6.78s/it]2026-04-13 10:57:13,931 | INFO | train_grpo_train | metrics_logged mode=train step=1464

                                                       
{'loss': 0.044, 'grad_norm': 9.621746063232422, 'learning_rate': 5.566666666666667e-06, 'num_tokens': 2582894.0, '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': 0.9511234164237976, 'rewards/meter/std': 0.10077424347400665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9086436033248901, 'rewards/repeat_soft/std': 0.08281069248914719, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.8650649785995483, 'rewards/total_composite/std': 0.12106508016586304, 'reward': 0.8650649785995483, 'reward_std': 0.12106508016586304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09151957184076309, 'sampling/sampling_logp_difference/max': 1.1939563751220703, 'sampling/importance_sampling_ratio/min': 0.30302003026008606, 'sampling/importance_sampling_ratio/mean': 1.0106608867645264, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5162233076989651, 'clip_ratio/low_mean': 0.02119965897873044, 'clip_ratio/low_min': 0.02119965897873044, 'clip_ratio/high_mean': 0.09480717871338129, 'clip_ratio/high_max': 0.09480717871338129, 'clip_ratio/region_mean': 0.11600683769211173, 'reward_total_mean': 0.8650649785995483, 'reward_meter_mean': 0.9511234164237976, 'reward_meter_std': 0.10077424347400665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9086436033248901, 'reward_repeat_soft_std': 0.08281069248914719, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.8650649785995483, 'reward_total_composite_std': 0.12106508016586304, 'epoch': 0.15}

 44%|████▍     | 1464/3300 [3:36:16<3:27:24,  6.78s/it]INFO 04-13 10:57:14 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1465/3300 [3:36:22<3:19:57,  6.54s/it]2026-04-13 10:57:19,909 | INFO | train_grpo_train | metrics_logged mode=train step=1465

                                                       
{'loss': -0.0256, 'grad_norm': 12.525416374206543, 'learning_rate': 5.563636363636364e-06, 'num_tokens': 2584391.0, 'completions/mean_length': 42.125, 'completions/min_length': 35.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.125, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9827646613121033, 'rewards/meter/std': 0.017621733248233795, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9358309507369995, 'rewards/repeat_soft/std': 0.02582893893122673, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6886136531829834, 'rewards/total_composite/std': 0.14731666445732117, 'reward': 0.6886136531829834, 'reward_std': 0.14731664955615997, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.105953149497509, 'sampling/sampling_logp_difference/max': 1.130956768989563, 'sampling/importance_sampling_ratio/min': 0.3227243423461914, 'sampling/importance_sampling_ratio/mean': 1.01067054271698, 'sampling/importance_sampling_ratio/max': 1.8950624465942383, 'entropy': 0.568340964615345, 'clip_ratio/low_mean': 0.07531766872853041, 'clip_ratio/low_min': 0.07531766872853041, 'clip_ratio/high_mean': 0.027173913083970547, 'clip_ratio/high_max': 0.027173913083970547, 'clip_ratio/region_mean': 0.10249158181250095, 'reward_total_mean': 0.6886136531829834, 'reward_meter_mean': 0.9827646613121033, 'reward_meter_std': 0.017621733248233795, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9358309507369995, 'reward_repeat_soft_std': 0.02582893893122673, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6886136531829834, 'reward_total_composite_std': 0.14731666445732117, 'epoch': 0.15}

 44%|████▍     | 1465/3300 [3:36:22<3:19:57,  6.54s/it]INFO 04-13 10:57:20 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1466/3300 [3:36:28<3:19:47,  6.54s/it]2026-04-13 10:57:26,481 | INFO | train_grpo_train | metrics_logged mode=train step=1466

                                                       
{'loss': 0.0033, 'grad_norm': 11.49193000793457, 'learning_rate': 5.560606060606061e-06, 'num_tokens': 2585914.0, 'completions/mean_length': 47.375, 'completions/min_length': 40.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.375, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.664898157119751, 'rewards/meter/std': 0.3572017252445221, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9430370330810547, 'rewards/repeat_soft/std': 0.09419349581003189, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.1865811049938202, 'rewards/total_composite/mean': 0.537501335144043, 'rewards/total_composite/std': 0.07693111896514893, 'reward': 0.537501335144043, 'reward_std': 0.07693111896514893, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12955878674983978, 'sampling/sampling_logp_difference/max': 3.5795273780822754, 'sampling/importance_sampling_ratio/min': 0.02788887731730938, 'sampling/importance_sampling_ratio/mean': 1.0106037855148315, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7453957945108414, 'clip_ratio/low_mean': 0.0381249999627471, 'clip_ratio/low_min': 0.0381249999627471, 'clip_ratio/high_mean': 0.07593381218612194, 'clip_ratio/high_max': 0.07593381218612194, 'clip_ratio/region_mean': 0.11405881214886904, 'reward_total_mean': 0.537501335144043, 'reward_meter_mean': 0.664898157119751, 'reward_meter_std': 0.3572017252445221, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9430370330810547, 'reward_repeat_soft_std': 0.09419349581003189, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.1865811049938202, 'reward_total_composite_mean': 0.537501335144043, 'reward_total_composite_std': 0.07693111896514893, 'epoch': 0.15}

 44%|████▍     | 1466/3300 [3:36:28<3:19:47,  6.54s/it]INFO 04-13 10:57:26 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1467/3300 [3:36:35<3:22:34,  6.63s/it]2026-04-13 10:57:33,285 | INFO | train_grpo_train | metrics_logged mode=train step=1467

                                                       
{'loss': -0.0289, 'grad_norm': 8.926210403442383, 'learning_rate': 5.557575757575758e-06, 'num_tokens': 2587774.0, 'completions/mean_length': 60.5, 'completions/min_length': 47.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.5, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7280633449554443, 'rewards/meter/std': 0.30291086435317993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8180413842201233, 'rewards/repeat_soft/std': 0.08221400529146194, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.25150617957115173, 'rewards/total_composite/mean': 0.5870155692100525, 'rewards/total_composite/std': 0.21065010130405426, 'reward': 0.5870155692100525, 'reward_std': 0.21065007150173187, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11216025054454803, 'sampling/sampling_logp_difference/max': 1.7116155624389648, 'sampling/importance_sampling_ratio/min': 0.18057382106781006, 'sampling/importance_sampling_ratio/mean': 1.0145072937011719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6791957169771194, 'clip_ratio/low_mean': 0.07617599982768297, 'clip_ratio/low_min': 0.07617599982768297, 'clip_ratio/high_mean': 0.02996880654245615, 'clip_ratio/high_max': 0.02996880654245615, 'clip_ratio/region_mean': 0.10614480637013912, 'reward_total_mean': 0.5870155692100525, 'reward_meter_mean': 0.7280633449554443, 'reward_meter_std': 0.30291086435317993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8180413842201233, 'reward_repeat_soft_std': 0.08221400529146194, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.25150617957115173, 'reward_total_composite_mean': 0.5870155692100525, 'reward_total_composite_std': 0.21065010130405426, 'epoch': 0.15}

 44%|████▍     | 1467/3300 [3:36:35<3:22:34,  6.63s/it]INFO 04-13 10:57:33 [block_pool.py:378] Successfully reset prefix cache

 44%|████▍     | 1468/3300 [3:36:41<3:19:00,  6.52s/it]2026-04-13 10:57:39,584 | INFO | train_grpo_train | metrics_logged mode=train step=1468

                                                       
{'loss': -0.0245, 'grad_norm': 16.370771408081055, 'learning_rate': 5.554545454545454e-06, 'num_tokens': 2589059.0, 'completions/mean_length': 23.625, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.625, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9890918731689453, 'rewards/meter/std': 0.00959152914583683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9175378680229187, 'rewards/repeat_soft/std': 0.11084415018558502, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.10336308926343918, 'rewards/total_composite/mean': 0.5601323843002319, 'rewards/total_composite/std': 0.061217617243528366, 'reward': 0.5601323843002319, 'reward_std': 0.06121760234236717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1455719918012619, 'sampling/sampling_logp_difference/max': 2.6551313400268555, 'sampling/importance_sampling_ratio/min': 0.07028961181640625, 'sampling/importance_sampling_ratio/mean': 0.9993245601654053, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.094538040459156, 'clip_ratio/low_mean': 0.06639610603451729, 'clip_ratio/low_min': 0.06639610603451729, 'clip_ratio/high_mean': 0.06223659124225378, 'clip_ratio/high_max': 0.06223659124225378, 'clip_ratio/region_mean': 0.12863269727677107, 'reward_total_mean': 0.5601323843002319, 'reward_meter_mean': 0.9890918731689453, 'reward_meter_std': 0.00959152914583683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9175378680229187, 'reward_repeat_soft_std': 0.11084415018558502, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.10336308926343918, 'reward_total_composite_mean': 0.5601323843002319, 'reward_total_composite_std': 0.061217617243528366, 'epoch': 0.15}

 44%|████▍     | 1468/3300 [3:36:41<3:19:00,  6.52s/it]INFO 04-13 10:57:39 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1469/3300 [3:36:49<3:28:40,  6.84s/it]2026-04-13 10:57:47,129 | INFO | train_grpo_train | metrics_logged mode=train step=1469

                                                       
{'loss': 0.0148, 'grad_norm': 6.228905200958252, 'learning_rate': 5.5515151515151524e-06, 'num_tokens': 2591513.0, 'completions/mean_length': 129.75, 'completions/min_length': 100.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.75, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.7403182983398438, 'rewards/meter/std': 0.1365850269794464, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7926737070083618, 'rewards/repeat_soft/std': 0.11876393854618073, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.4557223618030548, 'rewards/total_composite/std': 0.06416784226894379, 'reward': 0.4557223618030548, 'reward_std': 0.06416784226894379, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11837196350097656, 'sampling/sampling_logp_difference/max': 1.9244991540908813, 'sampling/importance_sampling_ratio/min': 0.1459488421678543, 'sampling/importance_sampling_ratio/mean': 1.000449538230896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6611907668411732, 'clip_ratio/low_mean': 0.06312037538737059, 'clip_ratio/low_min': 0.06312037538737059, 'clip_ratio/high_mean': 0.05303825717419386, 'clip_ratio/high_max': 0.05303825717419386, 'clip_ratio/region_mean': 0.11615863256156445, 'reward_total_mean': 0.4557223618030548, 'reward_meter_mean': 0.7403182983398438, 'reward_meter_std': 0.1365850269794464, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7926737070083618, 'reward_repeat_soft_std': 0.11876393854618073, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.4557223618030548, 'reward_total_composite_std': 0.06416784226894379, 'epoch': 0.15}

 45%|████▍     | 1469/3300 [3:36:49<3:28:40,  6.84s/it]INFO 04-13 10:57:47 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1470/3300 [3:36:56<3:31:04,  6.92s/it]2026-04-13 10:57:54,244 | INFO | train_grpo_train | metrics_logged mode=train step=1470

                                                       
{'loss': 0.0308, 'grad_norm': 7.634474277496338, 'learning_rate': 5.548484848484849e-06, 'num_tokens': 2594005.0, 'completions/mean_length': 116.5, 'completions/min_length': 90.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.5, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.7753080129623413, 'rewards/meter/std': 0.27682748436927795, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9754493236541748, 'rewards/repeat_soft/std': 0.017098110169172287, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.4989067316055298, 'rewards/total_composite/std': 0.08452694118022919, 'reward': 0.4989067316055298, 'reward_std': 0.08452694863080978, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14233583211898804, 'sampling/sampling_logp_difference/max': 1.5213804244995117, 'sampling/importance_sampling_ratio/min': 0.2184101939201355, 'sampling/importance_sampling_ratio/mean': 1.0129231214523315, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0378146320581436, 'clip_ratio/low_mean': 0.06215149536728859, 'clip_ratio/low_min': 0.06215149536728859, 'clip_ratio/high_mean': 0.07031365670263767, 'clip_ratio/high_max': 0.07031365670263767, 'clip_ratio/region_mean': 0.13246515206992626, 'reward_total_mean': 0.4989067316055298, 'reward_meter_mean': 0.7753080129623413, 'reward_meter_std': 0.27682748436927795, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9754493236541748, 'reward_repeat_soft_std': 0.017098110169172287, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.4989067316055298, 'reward_total_composite_std': 0.08452694118022919, 'epoch': 0.15}

 45%|████▍     | 1470/3300 [3:36:56<3:31:04,  6.92s/it]INFO 04-13 10:57:54 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1471/3300 [3:37:02<3:22:54,  6.66s/it]2026-04-13 10:58:00,285 | INFO | train_grpo_train | metrics_logged mode=train step=1471

                                                       
{'loss': 0.0236, 'grad_norm': 13.744756698608398, 'learning_rate': 5.545454545454546e-06, 'num_tokens': 2595614.0, 'completions/mean_length': 48.125, 'completions/min_length': 43.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6877424716949463, 'rewards/meter/std': 0.4287676513195038, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.892635703086853, 'rewards/repeat_soft/std': 0.08857487142086029, 'rewards/judge_quality/mean': 0.6987500190734863, 'rewards/judge_quality/std': 0.31651848554611206, 'rewards/total_composite/mean': 0.6591956615447998, 'rewards/total_composite/std': 0.28758805990219116, 'reward': 0.6591956615447998, 'reward_std': 0.28758805990219116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13303479552268982, 'sampling/sampling_logp_difference/max': 1.1878950595855713, 'sampling/importance_sampling_ratio/min': 0.30486229062080383, 'sampling/importance_sampling_ratio/mean': 1.0081233978271484, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8316401988267899, 'clip_ratio/low_mean': 0.05475694499909878, 'clip_ratio/low_min': 0.05475694499909878, 'clip_ratio/high_mean': 0.06690456345677376, 'clip_ratio/high_max': 0.06690456345677376, 'clip_ratio/region_mean': 0.12166150845587254, 'reward_total_mean': 0.6591956615447998, 'reward_meter_mean': 0.6877424716949463, 'reward_meter_std': 0.4287676513195038, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.892635703086853, 'reward_repeat_soft_std': 0.08857487142086029, 'reward_judge_quality_mean': 0.6987500190734863, 'reward_judge_quality_std': 0.31651848554611206, 'reward_total_composite_mean': 0.6591956615447998, 'reward_total_composite_std': 0.28758805990219116, 'epoch': 0.15}

 45%|████▍     | 1471/3300 [3:37:02<3:22:54,  6.66s/it]INFO 04-13 10:58:00 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1472/3300 [3:37:08<3:20:40,  6.59s/it]2026-04-13 10:58:06,701 | INFO | train_grpo_train | metrics_logged mode=train step=1472

                                                       
{'loss': -0.0372, 'grad_norm': 12.051742553710938, 'learning_rate': 5.5424242424242425e-06, 'num_tokens': 2597271.0, 'completions/mean_length': 43.125, 'completions/min_length': 34.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8499079346656799, 'rewards/meter/std': 0.3276419937610626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8506057858467102, 'rewards/repeat_soft/std': 0.047927580773830414, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.5555394887924194, 'rewards/total_composite/std': 0.09663747251033783, 'reward': 0.5555394887924194, 'reward_std': 0.09663745760917664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1002088412642479, 'sampling/sampling_logp_difference/max': 1.6085458993911743, 'sampling/importance_sampling_ratio/min': 0.20017847418785095, 'sampling/importance_sampling_ratio/mean': 1.0045076608657837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.625467523932457, 'clip_ratio/low_mean': 0.01907442230731249, 'clip_ratio/low_min': 0.01907442230731249, 'clip_ratio/high_mean': 0.06602906063199043, 'clip_ratio/high_max': 0.06602906063199043, 'clip_ratio/region_mean': 0.08510348293930292, 'reward_total_mean': 0.5555394887924194, 'reward_meter_mean': 0.8499079346656799, 'reward_meter_std': 0.3276419937610626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8506057858467102, 'reward_repeat_soft_std': 0.047927580773830414, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.5555394887924194, 'reward_total_composite_std': 0.09663747251033783, 'epoch': 0.15}

 45%|████▍     | 1472/3300 [3:37:08<3:20:40,  6.59s/it]INFO 04-13 10:58:06 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1473/3300 [3:37:14<3:15:22,  6.42s/it]2026-04-13 10:58:12,726 | INFO | train_grpo_train | metrics_logged mode=train step=1473

                                                       
{'loss': 0.078, 'grad_norm': 17.361730575561523, 'learning_rate': 5.53939393939394e-06, 'num_tokens': 2598815.0, 'completions/mean_length': 35.0, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.3028879761695862, 'rewards/meter/std': 0.32608795166015625, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9703353643417358, 'rewards/repeat_soft/std': 0.058450888842344284, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5166357755661011, 'rewards/total_composite/std': 0.20340381562709808, 'reward': 0.5166357755661011, 'reward_std': 0.2034038007259369, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11877524852752686, 'sampling/sampling_logp_difference/max': 1.2922561168670654, 'sampling/importance_sampling_ratio/min': 0.2746504247188568, 'sampling/importance_sampling_ratio/mean': 1.0062525272369385, 'sampling/importance_sampling_ratio/max': 1.9707258939743042, 'entropy': 0.5405756812542677, 'clip_ratio/low_mean': 0.0628345743753016, 'clip_ratio/low_min': 0.0628345743753016, 'clip_ratio/high_mean': 0.038352273404598236, 'clip_ratio/high_max': 0.038352273404598236, 'clip_ratio/region_mean': 0.10118684777989984, 'reward_total_mean': 0.5166357755661011, 'reward_meter_mean': 0.3028879761695862, 'reward_meter_std': 0.32608795166015625, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9703353643417358, 'reward_repeat_soft_std': 0.058450888842344284, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5166357755661011, 'reward_total_composite_std': 0.20340381562709808, 'epoch': 0.15}

 45%|████▍     | 1473/3300 [3:37:14<3:15:22,  6.42s/it]INFO 04-13 10:58:12 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1474/3300 [3:37:21<3:18:25,  6.52s/it]2026-04-13 10:58:19,481 | INFO | train_grpo_train | metrics_logged mode=train step=1474

                                                       
{'loss': 0.0647, 'grad_norm': 13.33869457244873, 'learning_rate': 5.536363636363636e-06, 'num_tokens': 2600649.0, 'completions/mean_length': 54.25, 'completions/min_length': 42.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.4765779674053192, 'rewards/meter/std': 0.3976920247077942, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9742038249969482, 'rewards/repeat_soft/std': 0.027118271216750145, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.19334925711154938, 'rewards/total_composite/mean': 0.4688124358654022, 'rewards/total_composite/std': 0.26618191599845886, 'reward': 0.4688124358654022, 'reward_std': 0.2661818861961365, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15565451979637146, 'sampling/sampling_logp_difference/max': 1.6274919509887695, 'sampling/importance_sampling_ratio/min': 0.1964215785264969, 'sampling/importance_sampling_ratio/mean': 1.0438464879989624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3133893460035324, 'clip_ratio/low_mean': 0.09236462041735649, 'clip_ratio/low_min': 0.09236462041735649, 'clip_ratio/high_mean': 0.045936762355268, 'clip_ratio/high_max': 0.045936762355268, 'clip_ratio/region_mean': 0.1383013827726245, 'reward_total_mean': 0.4688124358654022, 'reward_meter_mean': 0.4765779674053192, 'reward_meter_std': 0.3976920247077942, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9742038249969482, 'reward_repeat_soft_std': 0.027118271216750145, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.19334925711154938, 'reward_total_composite_mean': 0.4688124358654022, 'reward_total_composite_std': 0.26618191599845886, 'epoch': 0.15}

 45%|████▍     | 1474/3300 [3:37:21<3:18:25,  6.52s/it]INFO 04-13 10:58:19 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1475/3300 [3:37:29<3:28:29,  6.85s/it]2026-04-13 10:58:27,127 | INFO | train_grpo_train | metrics_logged mode=train step=1475

                                                       
{'loss': 0.0295, 'grad_norm': 10.713043212890625, 'learning_rate': 5.533333333333334e-06, 'num_tokens': 2603073.0, 'completions/mean_length': 104.0, 'completions/min_length': 81.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.0, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.648520827293396, 'rewards/meter/std': 0.2619701325893402, 'rewards/count_adherence/mean': 0.9249999523162842, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9017126560211182, 'rewards/repeat_soft/std': 0.07726913690567017, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955308735370636, 'rewards/total_composite/mean': 0.5715631246566772, 'rewards/total_composite/std': 0.18075372278690338, 'reward': 0.5715631246566772, 'reward_std': 0.18075372278690338, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09615106880664825, 'sampling/sampling_logp_difference/max': 3.520033359527588, 'sampling/importance_sampling_ratio/min': 0.029598446562886238, 'sampling/importance_sampling_ratio/mean': 1.006070852279663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3830401562154293, 'clip_ratio/low_mean': 0.06448802212253213, 'clip_ratio/low_min': 0.06448802212253213, 'clip_ratio/high_mean': 0.01771604921668768, 'clip_ratio/high_max': 0.01771604921668768, 'clip_ratio/region_mean': 0.08220407133921981, 'reward_total_mean': 0.5715631246566772, 'reward_meter_mean': 0.648520827293396, 'reward_meter_std': 0.2619701325893402, 'reward_count_adherence_mean': 0.9249999523162842, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9017126560211182, 'reward_repeat_soft_std': 0.07726913690567017, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955308735370636, 'reward_total_composite_mean': 0.5715631246566772, 'reward_total_composite_std': 0.18075372278690338, 'epoch': 0.15}

 45%|████▍     | 1475/3300 [3:37:29<3:28:29,  6.85s/it]INFO 04-13 10:58:27 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1476/3300 [3:37:41<4:13:20,  8.33s/it]2026-04-13 10:58:38,904 | INFO | train_grpo_train | metrics_logged mode=train step=1476

                                                       
{'loss': -0.1071, 'grad_norm': 2.849167585372925, 'learning_rate': 5.530303030303031e-06, 'num_tokens': 2604754.0, 'completions/mean_length': 109.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 51.57143020629883, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.7600843906402588, 'rewards/meter/std': 0.2595258355140686, 'rewards/count_adherence/mean': 0.8333333730697632, 'rewards/count_adherence/std': 0.2519763112068176, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8423864841461182, 'rewards/repeat_soft/std': 0.12611058354377747, 'rewards/judge_quality/mean': 0.3399999737739563, 'rewards/judge_quality/std': 0.15052290260791779, 'rewards/total_composite/mean': 0.44253116846084595, 'rewards/total_composite/std': 0.20070195198059082, 'reward': 0.44253116846084595, 'reward_std': 0.20070193707942963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10046947747468948, 'sampling/sampling_logp_difference/max': 1.9943890571594238, 'sampling/importance_sampling_ratio/min': 0.1360967755317688, 'sampling/importance_sampling_ratio/mean': 0.9980024099349976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3889933116734028, 'clip_ratio/low_mean': 0.005859375, 'clip_ratio/low_min': 0.005859375, 'clip_ratio/high_mean': 0.07746341777965426, 'clip_ratio/high_max': 0.07746341777965426, 'clip_ratio/region_mean': 0.08332279277965426, 'reward_total_mean': 0.44253116846084595, 'reward_meter_mean': 0.7600843906402588, 'reward_meter_std': 0.2595258355140686, 'reward_count_adherence_mean': 0.8333333730697632, 'reward_count_adherence_std': 0.2519763112068176, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8423864841461182, 'reward_repeat_soft_std': 0.12611058354377747, 'reward_judge_quality_mean': 0.3399999737739563, 'reward_judge_quality_std': 0.15052290260791779, 'reward_total_composite_mean': 0.44253116846084595, 'reward_total_composite_std': 0.20070195198059082, 'epoch': 0.15}

 45%|████▍     | 1476/3300 [3:37:41<4:13:20,  8.33s/it]INFO 04-13 10:58:39 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1477/3300 [3:37:48<4:06:05,  8.10s/it]2026-04-13 10:58:46,469 | INFO | train_grpo_train | metrics_logged mode=train step=1477

                                                       
{'loss': -0.0015, 'grad_norm': 6.90681791305542, 'learning_rate': 5.527272727272728e-06, 'num_tokens': 2607584.0, 'completions/mean_length': 128.75, 'completions/min_length': 101.0, 'completions/max_length': 169.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 128.75, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.33890193700790405, 'rewards/meter/std': 0.31740322709083557, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8658617734909058, 'rewards/repeat_soft/std': 0.06393294781446457, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.40142741799354553, 'rewards/total_composite/std': 0.08093750476837158, 'reward': 0.40142741799354553, 'reward_std': 0.08093750476837158, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13084861636161804, 'sampling/sampling_logp_difference/max': 4.246708869934082, 'sampling/importance_sampling_ratio/min': 0.014311256818473339, 'sampling/importance_sampling_ratio/mean': 0.9870463013648987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6351949870586395, 'clip_ratio/low_mean': 0.06428987253457308, 'clip_ratio/low_min': 0.06428987253457308, 'clip_ratio/high_mean': 0.05529332160949707, 'clip_ratio/high_max': 0.05529332160949707, 'clip_ratio/region_mean': 0.11958319414407015, 'reward_total_mean': 0.40142741799354553, 'reward_meter_mean': 0.33890193700790405, 'reward_meter_std': 0.31740322709083557, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8658617734909058, 'reward_repeat_soft_std': 0.06393294781446457, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.40142741799354553, 'reward_total_composite_std': 0.08093750476837158, 'epoch': 0.15}

 45%|████▍     | 1477/3300 [3:37:48<4:06:05,  8.10s/it]INFO 04-13 10:58:46 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1478/3300 [3:38:00<4:36:25,  9.10s/it]2026-04-13 10:58:57,899 | INFO | train_grpo_train | metrics_logged mode=train step=1478

                                                       
{'loss': -0.1887, 'grad_norm': 1.7883447408676147, 'learning_rate': 5.524242424242424e-06, 'num_tokens': 2609757.0, 'completions/mean_length': 144.625, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 92.14286041259766, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.8850111961364746, 'rewards/meter/std': 0.1782539188861847, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8292892575263977, 'rewards/repeat_soft/std': 0.0910106971859932, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4554700553417206, 'rewards/total_composite/std': 0.18945030868053436, 'reward': 0.4554700553417206, 'reward_std': 0.18945029377937317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10043580830097198, 'sampling/sampling_logp_difference/max': 2.3448104858398438, 'sampling/importance_sampling_ratio/min': 0.09586536884307861, 'sampling/importance_sampling_ratio/mean': 1.0046113729476929, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5433511883020401, 'clip_ratio/low_mean': 0.006793478038161993, 'clip_ratio/low_min': 0.006793478038161993, 'clip_ratio/high_mean': 0.10017213178798556, 'clip_ratio/high_max': 0.10017213178798556, 'clip_ratio/region_mean': 0.10696560982614756, 'reward_total_mean': 0.4554700553417206, 'reward_meter_mean': 0.8850111961364746, 'reward_meter_std': 0.1782539188861847, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8292892575263977, 'reward_repeat_soft_std': 0.0910106971859932, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4554700553417206, 'reward_total_composite_std': 0.18945030868053436, 'epoch': 0.15}

 45%|████▍     | 1478/3300 [3:38:00<4:36:25,  9.10s/it]INFO 04-13 10:58:58 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1479/3300 [3:38:06<4:09:07,  8.21s/it]2026-04-13 10:59:04,028 | INFO | train_grpo_train | metrics_logged mode=train step=1479

                                                       
{'loss': 0.1153, 'grad_norm': 14.996219635009766, 'learning_rate': 5.521212121212122e-06, 'num_tokens': 2611292.0, 'completions/mean_length': 44.875, 'completions/min_length': 35.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8259141445159912, 'rewards/meter/std': 0.26048386096954346, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9079709053039551, 'rewards/repeat_soft/std': 0.06270492821931839, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.6216350197792053, 'rewards/total_composite/std': 0.1506887674331665, 'reward': 0.6216350197792053, 'reward_std': 0.1506887674331665, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12402523308992386, 'sampling/sampling_logp_difference/max': 2.003143310546875, 'sampling/importance_sampling_ratio/min': 0.13491055369377136, 'sampling/importance_sampling_ratio/mean': 0.9949265718460083, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7351835444569588, 'clip_ratio/low_mean': 0.06681718723848462, 'clip_ratio/low_min': 0.06681718723848462, 'clip_ratio/high_mean': 0.031944445334374905, 'clip_ratio/high_max': 0.031944445334374905, 'clip_ratio/region_mean': 0.09876163257285953, 'reward_total_mean': 0.6216350197792053, 'reward_meter_mean': 0.8259141445159912, 'reward_meter_std': 0.26048386096954346, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9079709053039551, 'reward_repeat_soft_std': 0.06270492821931839, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.6216350197792053, 'reward_total_composite_std': 0.1506887674331665, 'epoch': 0.15}

 45%|████▍     | 1479/3300 [3:38:06<4:09:07,  8.21s/it]INFO 04-13 10:59:04 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1480/3300 [3:38:13<4:00:25,  7.93s/it]2026-04-13 10:59:11,286 | INFO | train_grpo_train | metrics_logged mode=train step=1480

                                                       
{'loss': 0.004, 'grad_norm': 8.53366470336914, 'learning_rate': 5.518181818181818e-06, 'num_tokens': 2613336.0, 'completions/mean_length': 94.5, 'completions/min_length': 81.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.5, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8894447088241577, 'rewards/meter/std': 0.139001727104187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8424639701843262, 'rewards/repeat_soft/std': 0.075047567486763, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5691879987716675, 'rewards/total_composite/std': 0.03418510779738426, 'reward': 0.5691879987716675, 'reward_std': 0.03418509662151337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09812415391206741, 'sampling/sampling_logp_difference/max': 2.183708667755127, 'sampling/importance_sampling_ratio/min': 0.11262307316064835, 'sampling/importance_sampling_ratio/mean': 1.0124567747116089, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5634704940021038, 'clip_ratio/low_mean': 0.031258889473974705, 'clip_ratio/low_min': 0.031258889473974705, 'clip_ratio/high_mean': 0.062191106379032135, 'clip_ratio/high_max': 0.062191106379032135, 'clip_ratio/region_mean': 0.09344999585300684, 'reward_total_mean': 0.5691879987716675, 'reward_meter_mean': 0.8894447088241577, 'reward_meter_std': 0.139001727104187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8424639701843262, 'reward_repeat_soft_std': 0.075047567486763, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5691879987716675, 'reward_total_composite_std': 0.03418510779738426, 'epoch': 0.15}

 45%|████▍     | 1480/3300 [3:38:13<4:00:25,  7.93s/it]INFO 04-13 10:59:11 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1481/3300 [3:38:20<3:56:24,  7.80s/it]2026-04-13 10:59:18,787 | INFO | train_grpo_train | metrics_logged mode=train step=1481

                                                       
{'loss': 0.0185, 'grad_norm': 5.042724132537842, 'learning_rate': 5.515151515151515e-06, 'num_tokens': 2615841.0, 'completions/mean_length': 120.125, 'completions/min_length': 91.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.125, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.9654009342193604, 'rewards/meter/std': 0.02672361209988594, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7379403114318848, 'rewards/repeat_soft/std': 0.07007953524589539, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.5003182888031006, 'rewards/total_composite/std': 0.07372362166643143, 'reward': 0.5003182888031006, 'reward_std': 0.07372362166643143, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08715896308422089, 'sampling/sampling_logp_difference/max': 2.041287660598755, 'sampling/importance_sampling_ratio/min': 0.12986138463020325, 'sampling/importance_sampling_ratio/mean': 1.0069060325622559, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5354104712605476, 'clip_ratio/low_mean': 0.018736762925982475, 'clip_ratio/low_min': 0.018736762925982475, 'clip_ratio/high_mean': 0.05698342062532902, 'clip_ratio/high_max': 0.05698342062532902, 'clip_ratio/region_mean': 0.07572018355131149, 'reward_total_mean': 0.5003182888031006, 'reward_meter_mean': 0.9654009342193604, 'reward_meter_std': 0.02672361209988594, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7379403114318848, 'reward_repeat_soft_std': 0.07007953524589539, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.5003182888031006, 'reward_total_composite_std': 0.07372362166643143, 'epoch': 0.15}

 45%|████▍     | 1481/3300 [3:38:21<3:56:24,  7.80s/it]INFO 04-13 10:59:19 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1482/3300 [3:38:27<3:41:08,  7.30s/it]2026-04-13 10:59:24,926 | INFO | train_grpo_train | metrics_logged mode=train step=1482

                                                       
{'loss': 0.0199, 'grad_norm': 13.904169082641602, 'learning_rate': 5.512121212121213e-06, 'num_tokens': 2617555.0, 'completions/mean_length': 51.25, 'completions/min_length': 46.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.7591448426246643, 'rewards/meter/std': 0.43144935369491577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9451721906661987, 'rewards/repeat_soft/std': 0.06980152428150177, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.549022376537323, 'rewards/total_composite/std': 0.11372621357440948, 'reward': 0.549022376537323, 'reward_std': 0.11372620612382889, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11626243591308594, 'sampling/sampling_logp_difference/max': 2.5900814533233643, 'sampling/importance_sampling_ratio/min': 0.07501393556594849, 'sampling/importance_sampling_ratio/mean': 0.9951414465904236, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6460620984435081, 'clip_ratio/low_mean': 0.04374057427048683, 'clip_ratio/low_min': 0.04374057427048683, 'clip_ratio/high_mean': 0.07585607282817364, 'clip_ratio/high_max': 0.07585607282817364, 'clip_ratio/region_mean': 0.11959664709866047, 'reward_total_mean': 0.549022376537323, 'reward_meter_mean': 0.7591448426246643, 'reward_meter_std': 0.43144935369491577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9451721906661987, 'reward_repeat_soft_std': 0.06980152428150177, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.549022376537323, 'reward_total_composite_std': 0.11372621357440948, 'epoch': 0.15}

 45%|████▍     | 1482/3300 [3:38:27<3:41:08,  7.30s/it]INFO 04-13 10:59:25 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1483/3300 [3:38:32<3:27:01,  6.84s/it]2026-04-13 10:59:30,681 | INFO | train_grpo_train | metrics_logged mode=train step=1483

                                                       
{'loss': -0.0115, 'grad_norm': 16.531383514404297, 'learning_rate': 5.50909090909091e-06, 'num_tokens': 2618989.0, 'completions/mean_length': 26.25, 'completions/min_length': 21.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6387969255447388, 'rewards/meter/std': 0.4885605573654175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9475217461585999, 'rewards/repeat_soft/std': 0.042364854365587234, 'rewards/judge_quality/mean': 0.4975000023841858, 'rewards/judge_quality/std': 0.17136012017726898, 'rewards/total_composite/mean': 0.5319289565086365, 'rewards/total_composite/std': 0.1341337114572525, 'reward': 0.5319289565086365, 'reward_std': 0.1341337114572525, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12180305272340775, 'sampling/sampling_logp_difference/max': 0.840731143951416, 'sampling/importance_sampling_ratio/min': 0.431395024061203, 'sampling/importance_sampling_ratio/mean': 1.0240724086761475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9381069242954254, 'clip_ratio/low_mean': 0.05408653989434242, 'clip_ratio/low_min': 0.05408653989434242, 'clip_ratio/high_mean': 0.07852909667417407, 'clip_ratio/high_max': 0.07852909667417407, 'clip_ratio/region_mean': 0.1326156365685165, 'reward_total_mean': 0.5319289565086365, 'reward_meter_mean': 0.6387969255447388, 'reward_meter_std': 0.4885605573654175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9475217461585999, 'reward_repeat_soft_std': 0.042364854365587234, 'reward_judge_quality_mean': 0.4975000023841858, 'reward_judge_quality_std': 0.17136012017726898, 'reward_total_composite_mean': 0.5319289565086365, 'reward_total_composite_std': 0.1341337114572525, 'epoch': 0.15}

 45%|████▍     | 1483/3300 [3:38:32<3:27:01,  6.84s/it]INFO 04-13 10:59:30 [block_pool.py:378] Successfully reset prefix cache

 45%|████▍     | 1484/3300 [3:38:39<3:25:12,  6.78s/it]2026-04-13 10:59:37,335 | INFO | train_grpo_train | metrics_logged mode=train step=1484

                                                       
{'loss': -0.0041, 'grad_norm': 8.879104614257812, 'learning_rate': 5.506060606060607e-06, 'num_tokens': 2620724.0, 'completions/mean_length': 59.875, 'completions/min_length': 48.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9281506538391113, 'rewards/meter/std': 0.12714247405529022, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9748289585113525, 'rewards/repeat_soft/std': 0.026928458362817764, 'rewards/judge_quality/mean': 0.5012500286102295, 'rewards/judge_quality/std': 0.16974246501922607, 'rewards/total_composite/mean': 0.6508640050888062, 'rewards/total_composite/std': 0.1229616180062294, 'reward': 0.6508640050888062, 'reward_std': 0.1229616105556488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1533975899219513, 'sampling/sampling_logp_difference/max': 1.600853443145752, 'sampling/importance_sampling_ratio/min': 0.20172427594661713, 'sampling/importance_sampling_ratio/mean': 1.0146669149398804, 'sampling/importance_sampling_ratio/max': 1.929390549659729, 'entropy': 1.1840372160077095, 'clip_ratio/low_mean': 0.1217160364612937, 'clip_ratio/low_min': 0.1217160364612937, 'clip_ratio/high_mean': 0.012295082211494446, 'clip_ratio/high_max': 0.012295082211494446, 'clip_ratio/region_mean': 0.13401111867278814, 'reward_total_mean': 0.6508640050888062, 'reward_meter_mean': 0.9281506538391113, 'reward_meter_std': 0.12714247405529022, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9748289585113525, 'reward_repeat_soft_std': 0.026928458362817764, 'reward_judge_quality_mean': 0.5012500286102295, 'reward_judge_quality_std': 0.16974246501922607, 'reward_total_composite_mean': 0.6508640050888062, 'reward_total_composite_std': 0.1229616180062294, 'epoch': 0.15}

 45%|████▍     | 1484/3300 [3:38:39<3:25:12,  6.78s/it]INFO 04-13 10:59:37 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1485/3300 [3:38:46<3:27:16,  6.85s/it]2026-04-13 10:59:44,368 | INFO | train_grpo_train | metrics_logged mode=train step=1485

                                                       
{'loss': 0.0026, 'grad_norm': 8.277298927307129, 'learning_rate': 5.5030303030303034e-06, 'num_tokens': 2623061.0, 'completions/mean_length': 108.125, 'completions/min_length': 99.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.125, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9773924350738525, 'rewards/meter/std': 0.013282565400004387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9117889404296875, 'rewards/repeat_soft/std': 0.04147614538669586, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6035964488983154, 'rewards/total_composite/std': 0.00719041284173727, 'reward': 0.6035964488983154, 'reward_std': 0.007190412376075983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11500288546085358, 'sampling/sampling_logp_difference/max': 1.3996589183807373, 'sampling/importance_sampling_ratio/min': 0.2466810792684555, 'sampling/importance_sampling_ratio/mean': 1.0076630115509033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6722720637917519, 'clip_ratio/low_mean': 0.06540508940815926, 'clip_ratio/low_min': 0.06540508940815926, 'clip_ratio/high_mean': 0.038259128108620644, 'clip_ratio/high_max': 0.038259128108620644, 'clip_ratio/region_mean': 0.1036642175167799, 'reward_total_mean': 0.6035964488983154, 'reward_meter_mean': 0.9773924350738525, 'reward_meter_std': 0.013282565400004387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9117889404296875, 'reward_repeat_soft_std': 0.04147614538669586, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6035964488983154, 'reward_total_composite_std': 0.00719041284173727, 'epoch': 0.15}

 45%|████▌     | 1485/3300 [3:38:46<3:27:16,  6.85s/it]INFO 04-13 10:59:44 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1486/3300 [3:38:53<3:24:53,  6.78s/it]2026-04-13 10:59:50,954 | INFO | train_grpo_train | metrics_logged mode=train step=1486

                                                       
{'loss': -0.0171, 'grad_norm': 10.497859954833984, 'learning_rate': 5.500000000000001e-06, 'num_tokens': 2624788.0, 'completions/mean_length': 65.875, 'completions/min_length': 47.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.807790994644165, 'rewards/meter/std': 0.20373433828353882, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.970165491104126, 'rewards/repeat_soft/std': 0.036199379712343216, 'rewards/judge_quality/mean': 0.35499998927116394, 'rewards/judge_quality/std': 0.11928357183933258, 'rewards/total_composite/mean': 0.48034781217575073, 'rewards/total_composite/std': 0.20813879370689392, 'reward': 0.48034781217575073, 'reward_std': 0.20813877880573273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13982558250427246, 'sampling/sampling_logp_difference/max': 1.3635053634643555, 'sampling/importance_sampling_ratio/min': 0.25576266646385193, 'sampling/importance_sampling_ratio/mean': 0.996860921382904, 'sampling/importance_sampling_ratio/max': 1.9622875452041626, 'entropy': 1.1020059809088707, 'clip_ratio/low_mean': 0.03239944111555815, 'clip_ratio/low_min': 0.03239944111555815, 'clip_ratio/high_mean': 0.09075272921472788, 'clip_ratio/high_max': 0.09075272921472788, 'clip_ratio/region_mean': 0.12315217033028603, 'reward_total_mean': 0.48034781217575073, 'reward_meter_mean': 0.807790994644165, 'reward_meter_std': 0.20373433828353882, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.970165491104126, 'reward_repeat_soft_std': 0.036199379712343216, 'reward_judge_quality_mean': 0.35499998927116394, 'reward_judge_quality_std': 0.11928357183933258, 'reward_total_composite_mean': 0.48034781217575073, 'reward_total_composite_std': 0.20813879370689392, 'epoch': 0.15}

 45%|████▌     | 1486/3300 [3:38:53<3:24:53,  6.78s/it]INFO 04-13 10:59:51 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1487/3300 [3:39:00<3:32:55,  7.05s/it]2026-04-13 10:59:58,626 | INFO | train_grpo_train | metrics_logged mode=train step=1487

                                                       
{'loss': 0.0495, 'grad_norm': 5.990231990814209, 'learning_rate': 5.496969696969697e-06, 'num_tokens': 2627478.0, 'completions/mean_length': 146.25, 'completions/min_length': 131.0, 'completions/max_length': 159.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.25, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 159.0, 'rewards/meter/mean': 0.9197298288345337, 'rewards/meter/std': 0.11296083778142929, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.875533401966095, 'rewards/repeat_soft/std': 0.05981547012925148, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6126554012298584, 'rewards/total_composite/std': 0.09052237123250961, 'reward': 0.6126554012298584, 'reward_std': 0.09052237123250961, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11949143558740616, 'sampling/sampling_logp_difference/max': 1.4177286624908447, 'sampling/importance_sampling_ratio/min': 0.24701178073883057, 'sampling/importance_sampling_ratio/mean': 1.0083190202713013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7267304733395576, 'clip_ratio/low_mean': 0.06900496501475573, 'clip_ratio/low_min': 0.06900496501475573, 'clip_ratio/high_mean': 0.05193363409489393, 'clip_ratio/high_max': 0.05193363409489393, 'clip_ratio/region_mean': 0.12093859910964966, 'reward_total_mean': 0.6126554012298584, 'reward_meter_mean': 0.9197298288345337, 'reward_meter_std': 0.11296083778142929, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.875533401966095, 'reward_repeat_soft_std': 0.05981547012925148, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6126554012298584, 'reward_total_composite_std': 0.09052237123250961, 'epoch': 0.15}

 45%|████▌     | 1487/3300 [3:39:00<3:32:55,  7.05s/it]INFO 04-13 10:59:58 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1488/3300 [3:39:07<3:29:16,  6.93s/it]2026-04-13 11:00:05,280 | INFO | train_grpo_train | metrics_logged mode=train step=1488

                                                       
{'loss': -0.0701, 'grad_norm': 15.173059463500977, 'learning_rate': 5.493939393939395e-06, 'num_tokens': 2628993.0, 'completions/mean_length': 30.375, 'completions/min_length': 28.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8411796689033508, 'rewards/meter/std': 0.26568955183029175, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8814701437950134, 'rewards/repeat_soft/std': 0.09990499913692474, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5878442525863647, 'rewards/total_composite/std': 0.11131254583597183, 'reward': 0.5878442525863647, 'reward_std': 0.11131253838539124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09718216955661774, 'sampling/sampling_logp_difference/max': 2.024062156677246, 'sampling/importance_sampling_ratio/min': 0.13211768865585327, 'sampling/importance_sampling_ratio/mean': 0.9881425499916077, 'sampling/importance_sampling_ratio/max': 1.7737125158309937, 'entropy': 0.5191093683242798, 'clip_ratio/low_mean': 0.03064450016245246, 'clip_ratio/low_min': 0.03064450016245246, 'clip_ratio/high_mean': 0.04625057056546211, 'clip_ratio/high_max': 0.04625057056546211, 'clip_ratio/region_mean': 0.07689507072791457, 'reward_total_mean': 0.5878442525863647, 'reward_meter_mean': 0.8411796689033508, 'reward_meter_std': 0.26568955183029175, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8814701437950134, 'reward_repeat_soft_std': 0.09990499913692474, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5878442525863647, 'reward_total_composite_std': 0.11131254583597183, 'epoch': 0.15}

 45%|████▌     | 1488/3300 [3:39:07<3:29:16,  6.93s/it]INFO 04-13 11:00:05 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1489/3300 [3:39:14<3:28:02,  6.89s/it]2026-04-13 11:00:12,098 | INFO | train_grpo_train | metrics_logged mode=train step=1489

                                                       
{'loss': 0.0822, 'grad_norm': 9.129810333251953, 'learning_rate': 5.490909090909091e-06, 'num_tokens': 2630972.0, 'completions/mean_length': 82.375, 'completions/min_length': 67.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.375, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.8473856449127197, 'rewards/meter/std': 0.31352749466896057, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9352773427963257, 'rewards/repeat_soft/std': 0.04949399083852768, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22403764724731445, 'rewards/total_composite/mean': 0.5794332027435303, 'rewards/total_composite/std': 0.13291306793689728, 'reward': 0.5794332027435303, 'reward_std': 0.13291305303573608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11948537081480026, 'sampling/sampling_logp_difference/max': 1.7878718376159668, 'sampling/importance_sampling_ratio/min': 0.16731585562229156, 'sampling/importance_sampling_ratio/mean': 1.0016365051269531, 'sampling/importance_sampling_ratio/max': 1.9319500923156738, 'entropy': 0.756651908159256, 'clip_ratio/low_mean': 0.029689370654523373, 'clip_ratio/low_min': 0.029689370654523373, 'clip_ratio/high_mean': 0.10294715873897076, 'clip_ratio/high_max': 0.10294715873897076, 'clip_ratio/region_mean': 0.13263652939349413, 'reward_total_mean': 0.5794332027435303, 'reward_meter_mean': 0.8473856449127197, 'reward_meter_std': 0.31352749466896057, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9352773427963257, 'reward_repeat_soft_std': 0.04949399083852768, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22403764724731445, 'reward_total_composite_mean': 0.5794332027435303, 'reward_total_composite_std': 0.13291306793689728, 'epoch': 0.15}

 45%|████▌     | 1489/3300 [3:39:14<3:28:02,  6.89s/it]INFO 04-13 11:00:12 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1490/3300 [3:39:20<3:22:41,  6.72s/it]2026-04-13 11:00:18,401 | INFO | train_grpo_train | metrics_logged mode=train step=1490

                                                       
{'loss': 0.0197, 'grad_norm': 8.857263565063477, 'learning_rate': 5.487878787878789e-06, 'num_tokens': 2632388.0, 'completions/mean_length': 23.0, 'completions/min_length': 23.0, 'completions/max_length': 23.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.0, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 23.0, 'rewards/meter/mean': 0.9874359369277954, 'rewards/meter/std': 0.006660922896116972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9256622195243835, 'rewards/repeat_soft/std': 0.049346212297677994, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.6204668283462524, 'rewards/total_composite/std': 0.009920109063386917, 'reward': 0.6204668283462524, 'reward_std': 0.009920092299580574, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07062021642923355, 'sampling/sampling_logp_difference/max': 1.7588739395141602, 'sampling/importance_sampling_ratio/min': 0.17223870754241943, 'sampling/importance_sampling_ratio/mean': 1.0116403102874756, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3152022659778595, 'clip_ratio/low_mean': 0.027173913549631834, 'clip_ratio/low_min': 0.027173913549631834, 'clip_ratio/high_mean': 0.021739130839705467, 'clip_ratio/high_max': 0.021739130839705467, 'clip_ratio/region_mean': 0.0489130443893373, 'reward_total_mean': 0.6204668283462524, 'reward_meter_mean': 0.9874359369277954, 'reward_meter_std': 0.006660922896116972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9256622195243835, 'reward_repeat_soft_std': 0.049346212297677994, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.6204668283462524, 'reward_total_composite_std': 0.009920109063386917, 'epoch': 0.15}

 45%|████▌     | 1490/3300 [3:39:20<3:22:41,  6.72s/it]INFO 04-13 11:00:18 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1491/3300 [3:39:26<3:14:12,  6.44s/it]2026-04-13 11:00:24,200 | INFO | train_grpo_train | metrics_logged mode=train step=1491

                                                       
{'loss': 0.005, 'grad_norm': 21.160472869873047, 'learning_rate': 5.484848484848485e-06, 'num_tokens': 2633915.0, 'completions/mean_length': 22.875, 'completions/min_length': 18.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9593546390533447, 'rewards/meter/std': 0.03311125934123993, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9424046277999878, 'rewards/repeat_soft/std': 0.049576517194509506, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.0843462198972702, 'rewards/total_composite/mean': 0.5809230804443359, 'rewards/total_composite/std': 0.056494951248168945, 'reward': 0.5809230804443359, 'reward_std': 0.05649494379758835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13785071671009064, 'sampling/sampling_logp_difference/max': 1.004812240600586, 'sampling/importance_sampling_ratio/min': 0.3661133646965027, 'sampling/importance_sampling_ratio/mean': 1.0086857080459595, 'sampling/importance_sampling_ratio/max': 1.6974269151687622, 'entropy': 0.9706518948078156, 'clip_ratio/low_mean': 0.015567766036838293, 'clip_ratio/low_min': 0.015567766036838293, 'clip_ratio/high_mean': 0.08649774361401796, 'clip_ratio/high_max': 0.08649774361401796, 'clip_ratio/region_mean': 0.10206550965085626, 'reward_total_mean': 0.5809230804443359, 'reward_meter_mean': 0.9593546390533447, 'reward_meter_std': 0.03311125934123993, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9424046277999878, 'reward_repeat_soft_std': 0.049576517194509506, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.0843462198972702, 'reward_total_composite_mean': 0.5809230804443359, 'reward_total_composite_std': 0.056494951248168945, 'epoch': 0.15}

 45%|████▌     | 1491/3300 [3:39:26<3:14:12,  6.44s/it]INFO 04-13 11:00:24 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1492/3300 [3:39:37<3:57:18,  7.88s/it]2026-04-13 11:00:35,418 | INFO | train_grpo_train | metrics_logged mode=train step=1492

                                                       
{'loss': -0.1434, 'grad_norm': 2.424750804901123, 'learning_rate': 5.4818181818181825e-06, 'num_tokens': 2635684.0, 'completions/mean_length': 113.125, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 56.142860412597656, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9021384716033936, 'rewards/meter/std': 0.15267567336559296, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9634636640548706, 'rewards/repeat_soft/std': 0.030924834311008453, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.13767844438552856, 'rewards/total_composite/mean': 0.5216336250305176, 'rewards/total_composite/std': 0.21433673799037933, 'reward': 0.5216336250305176, 'reward_std': 0.21433670818805695, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12702292203903198, 'sampling/sampling_logp_difference/max': 1.3553638458251953, 'sampling/importance_sampling_ratio/min': 0.2578534483909607, 'sampling/importance_sampling_ratio/mean': 0.9867607355117798, 'sampling/importance_sampling_ratio/max': 1.7102025747299194, 'entropy': 0.7838902771472931, 'clip_ratio/low_mean': 0.004464285913854837, 'clip_ratio/low_min': 0.004464285913854837, 'clip_ratio/high_mean': 0.10617516562342644, 'clip_ratio/high_max': 0.10617516562342644, 'clip_ratio/region_mean': 0.11063945153728127, 'reward_total_mean': 0.5216336250305176, 'reward_meter_mean': 0.9021384716033936, 'reward_meter_std': 0.15267567336559296, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9634636640548706, 'reward_repeat_soft_std': 0.030924834311008453, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.13767844438552856, 'reward_total_composite_mean': 0.5216336250305176, 'reward_total_composite_std': 0.21433673799037933, 'epoch': 0.15}

 45%|████▌     | 1492/3300 [3:39:37<3:57:18,  7.88s/it]INFO 04-13 11:00:35 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1493/3300 [3:39:44<3:51:43,  7.69s/it]2026-04-13 11:00:42,688 | INFO | train_grpo_train | metrics_logged mode=train step=1493

                                                       
{'loss': -0.0612, 'grad_norm': 8.406460762023926, 'learning_rate': 5.478787878787879e-06, 'num_tokens': 2637724.0, 'completions/mean_length': 87.0, 'completions/min_length': 78.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.0, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9535136222839355, 'rewards/meter/std': 0.03741809353232384, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8257770538330078, 'rewards/repeat_soft/std': 0.12242306768894196, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.634614884853363, 'rewards/total_composite/std': 0.1459803432226181, 'reward': 0.634614884853363, 'reward_std': 0.1459803283214569, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10829455405473709, 'sampling/sampling_logp_difference/max': 1.737403392791748, 'sampling/importance_sampling_ratio/min': 0.17597675323486328, 'sampling/importance_sampling_ratio/mean': 1.0054004192352295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6760761961340904, 'clip_ratio/low_mean': 0.07716662669554353, 'clip_ratio/low_min': 0.07716662669554353, 'clip_ratio/high_mean': 0.028583532199263573, 'clip_ratio/high_max': 0.028583532199263573, 'clip_ratio/region_mean': 0.1057501588948071, 'reward_total_mean': 0.634614884853363, 'reward_meter_mean': 0.9535136222839355, 'reward_meter_std': 0.03741809353232384, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8257770538330078, 'reward_repeat_soft_std': 0.12242306768894196, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.634614884853363, 'reward_total_composite_std': 0.1459803432226181, 'epoch': 0.15}

 45%|████▌     | 1493/3300 [3:39:44<3:51:43,  7.69s/it]INFO 04-13 11:00:42 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1494/3300 [3:39:51<3:39:18,  7.29s/it]2026-04-13 11:00:49,023 | INFO | train_grpo_train | metrics_logged mode=train step=1494

                                                       
{'loss': 0.0655, 'grad_norm': 11.62306022644043, 'learning_rate': 5.475757575757576e-06, 'num_tokens': 2639318.0, 'completions/mean_length': 46.25, 'completions/min_length': 36.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7219693660736084, 'rewards/meter/std': 0.33736079931259155, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877862930297852, 'rewards/repeat_soft/std': 0.009149615652859211, 'rewards/judge_quality/mean': 0.5724999904632568, 'rewards/judge_quality/std': 0.194770947098732, 'rewards/total_composite/mean': 0.6372451186180115, 'rewards/total_composite/std': 0.19359150528907776, 'reward': 0.6372451186180115, 'reward_std': 0.19359152019023895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12142278254032135, 'sampling/sampling_logp_difference/max': 1.2413525581359863, 'sampling/importance_sampling_ratio/min': 0.28899309039115906, 'sampling/importance_sampling_ratio/mean': 0.9877916574478149, 'sampling/importance_sampling_ratio/max': 1.7015717029571533, 'entropy': 0.7238209135830402, 'clip_ratio/low_mean': 0.095611322671175, 'clip_ratio/low_min': 0.095611322671175, 'clip_ratio/high_mean': 0.03800287377089262, 'clip_ratio/high_max': 0.03800287377089262, 'clip_ratio/region_mean': 0.13361419644206762, 'reward_total_mean': 0.6372451186180115, 'reward_meter_mean': 0.7219693660736084, 'reward_meter_std': 0.33736079931259155, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877862930297852, 'reward_repeat_soft_std': 0.009149615652859211, 'reward_judge_quality_mean': 0.5724999904632568, 'reward_judge_quality_std': 0.194770947098732, 'reward_total_composite_mean': 0.6372451186180115, 'reward_total_composite_std': 0.19359150528907776, 'epoch': 0.15}

 45%|████▌     | 1494/3300 [3:39:51<3:39:18,  7.29s/it]INFO 04-13 11:00:49 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1495/3300 [3:39:57<3:29:56,  6.98s/it]2026-04-13 11:00:55,287 | INFO | train_grpo_train | metrics_logged mode=train step=1495

                                                       
{'loss': 0.0637, 'grad_norm': 9.856406211853027, 'learning_rate': 5.472727272727273e-06, 'num_tokens': 2641005.0, 'completions/mean_length': 58.875, 'completions/min_length': 53.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6126219034194946, 'rewards/meter/std': 0.4486890733242035, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9383866786956787, 'rewards/repeat_soft/std': 0.04639485850930214, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.4990963935852051, 'rewards/total_composite/std': 0.11686215549707413, 'reward': 0.4990963935852051, 'reward_std': 0.11686215549707413, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11863676458597183, 'sampling/sampling_logp_difference/max': 1.8329658508300781, 'sampling/importance_sampling_ratio/min': 0.1599385142326355, 'sampling/importance_sampling_ratio/mean': 1.023113489151001, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7165866494178772, 'clip_ratio/low_mean': 0.05935907084494829, 'clip_ratio/low_min': 0.05935907084494829, 'clip_ratio/high_mean': 0.041931509505957365, 'clip_ratio/high_max': 0.041931509505957365, 'clip_ratio/region_mean': 0.10129058035090566, 'reward_total_mean': 0.4990963935852051, 'reward_meter_mean': 0.6126219034194946, 'reward_meter_std': 0.4486890733242035, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9383866786956787, 'reward_repeat_soft_std': 0.04639485850930214, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.4990963935852051, 'reward_total_composite_std': 0.11686215549707413, 'epoch': 0.15}

 45%|████▌     | 1495/3300 [3:39:57<3:29:56,  6.98s/it]INFO 04-13 11:00:55 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1496/3300 [3:40:10<4:28:29,  8.93s/it]2026-04-13 11:01:08,770 | INFO | train_grpo_train | metrics_logged mode=train step=1496

                                                       
{'loss': 0.0202, 'grad_norm': 13.381893157958984, 'learning_rate': 5.469696969696971e-06, 'num_tokens': 2642740.0, 'completions/mean_length': 44.875, 'completions/min_length': 35.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8217130899429321, 'rewards/meter/std': 0.16472452878952026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9239823818206787, 'rewards/repeat_soft/std': 0.07076884806156158, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5703728199005127, 'rewards/total_composite/std': 0.05117390304803848, 'reward': 0.5703728199005127, 'reward_std': 0.05117391049861908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1393130123615265, 'sampling/sampling_logp_difference/max': 1.482508659362793, 'sampling/importance_sampling_ratio/min': 0.22706735134124756, 'sampling/importance_sampling_ratio/mean': 0.9985095262527466, 'sampling/importance_sampling_ratio/max': 1.9565730094909668, 'entropy': 1.0107583776116371, 'clip_ratio/low_mean': 0.03506616735830903, 'clip_ratio/low_min': 0.03506616735830903, 'clip_ratio/high_mean': 0.11435185372829437, 'clip_ratio/high_max': 0.11435185372829437, 'clip_ratio/region_mean': 0.1494180210866034, 'reward_total_mean': 0.5703728199005127, 'reward_meter_mean': 0.8217130899429321, 'reward_meter_std': 0.16472452878952026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9239823818206787, 'reward_repeat_soft_std': 0.07076884806156158, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5703728199005127, 'reward_total_composite_std': 0.05117390304803848, 'epoch': 0.15}

 45%|████▌     | 1496/3300 [3:40:10<4:28:29,  8.93s/it]INFO 04-13 11:01:09 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1497/3300 [3:40:16<4:00:35,  8.01s/it]2026-04-13 11:01:14,620 | INFO | train_grpo_train | metrics_logged mode=train step=1497

                                                       
{'loss': -0.0586, 'grad_norm': 15.628218650817871, 'learning_rate': 5.466666666666667e-06, 'num_tokens': 2644152.0, 'completions/mean_length': 27.5, 'completions/min_length': 22.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.5, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8665649890899658, 'rewards/meter/std': 0.3190095126628876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9560876488685608, 'rewards/repeat_soft/std': 0.018136821687221527, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5825812816619873, 'rewards/total_composite/std': 0.09054209291934967, 'reward': 0.5825812816619873, 'reward_std': 0.09054208546876907, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14445458352565765, 'sampling/sampling_logp_difference/max': 1.2789459228515625, 'sampling/importance_sampling_ratio/min': 0.278330534696579, 'sampling/importance_sampling_ratio/mean': 1.0274255275726318, 'sampling/importance_sampling_ratio/max': 1.9164459705352783, 'entropy': 1.5635968893766403, 'clip_ratio/low_mean': 0.03804347664117813, 'clip_ratio/low_min': 0.03804347664117813, 'clip_ratio/high_mean': 0.14897077903151512, 'clip_ratio/high_max': 0.14897077903151512, 'clip_ratio/region_mean': 0.18701425567269325, 'reward_total_mean': 0.5825812816619873, 'reward_meter_mean': 0.8665649890899658, 'reward_meter_std': 0.3190095126628876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9560876488685608, 'reward_repeat_soft_std': 0.018136821687221527, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5825812816619873, 'reward_total_composite_std': 0.09054209291934967, 'epoch': 0.15}

 45%|████▌     | 1497/3300 [3:40:16<4:00:35,  8.01s/it]INFO 04-13 11:01:14 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1498/3300 [3:40:22<3:42:49,  7.42s/it]2026-04-13 11:01:20,668 | INFO | train_grpo_train | metrics_logged mode=train step=1498

                                                       
{'loss': 0.0737, 'grad_norm': 9.467653274536133, 'learning_rate': 5.463636363636364e-06, 'num_tokens': 2645644.0, 'completions/mean_length': 44.5, 'completions/min_length': 37.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.8815276622772217, 'rewards/meter/std': 0.28615087270736694, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9526903629302979, 'rewards/repeat_soft/std': 0.04133959859609604, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.6274303197860718, 'rewards/total_composite/std': 0.1451360434293747, 'reward': 0.6274303197860718, 'reward_std': 0.1451360434293747, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14966563880443573, 'sampling/sampling_logp_difference/max': 2.6682653427124023, 'sampling/importance_sampling_ratio/min': 0.06937245279550552, 'sampling/importance_sampling_ratio/mean': 0.9882627725601196, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.726091668009758, 'clip_ratio/low_mean': 0.07268185913562775, 'clip_ratio/low_min': 0.07268185913562775, 'clip_ratio/high_mean': 0.0625273548066616, 'clip_ratio/high_max': 0.0625273548066616, 'clip_ratio/region_mean': 0.13520921394228935, 'reward_total_mean': 0.6274303197860718, 'reward_meter_mean': 0.8815276622772217, 'reward_meter_std': 0.28615087270736694, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9526903629302979, 'reward_repeat_soft_std': 0.04133959859609604, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.6274303197860718, 'reward_total_composite_std': 0.1451360434293747, 'epoch': 0.15}

 45%|████▌     | 1498/3300 [3:40:22<3:42:49,  7.42s/it]INFO 04-13 11:01:20 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1499/3300 [3:40:29<3:38:28,  7.28s/it]2026-04-13 11:01:27,618 | INFO | train_grpo_train | metrics_logged mode=train step=1499

                                                       
{'loss': -0.0227, 'grad_norm': 8.826386451721191, 'learning_rate': 5.460606060606061e-06, 'num_tokens': 2647574.0, 'completions/mean_length': 78.25, 'completions/min_length': 60.0, 'completions/max_length': 103.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 103.0, 'rewards/meter/mean': 0.5142233371734619, 'rewards/meter/std': 0.3521144390106201, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9376263618469238, 'rewards/repeat_soft/std': 0.07676321268081665, 'rewards/judge_quality/mean': 0.41874998807907104, 'rewards/judge_quality/std': 0.14574317634105682, 'rewards/total_composite/mean': 0.47821107506752014, 'rewards/total_composite/std': 0.0897725373506546, 'reward': 0.47821107506752014, 'reward_std': 0.089772529900074, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1335970014333725, 'sampling/sampling_logp_difference/max': 1.471923828125, 'sampling/importance_sampling_ratio/min': 0.22948358952999115, 'sampling/importance_sampling_ratio/mean': 1.028091549873352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.117735967040062, 'clip_ratio/low_mean': 0.08763296157121658, 'clip_ratio/low_min': 0.08763296157121658, 'clip_ratio/high_mean': 0.04597425367683172, 'clip_ratio/high_max': 0.04597425367683172, 'clip_ratio/region_mean': 0.1336072152480483, 'reward_total_mean': 0.47821107506752014, 'reward_meter_mean': 0.5142233371734619, 'reward_meter_std': 0.3521144390106201, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9376263618469238, 'reward_repeat_soft_std': 0.07676321268081665, 'reward_judge_quality_mean': 0.41874998807907104, 'reward_judge_quality_std': 0.14574317634105682, 'reward_total_composite_mean': 0.47821107506752014, 'reward_total_composite_std': 0.0897725373506546, 'epoch': 0.15}

 45%|████▌     | 1499/3300 [3:40:29<3:38:28,  7.28s/it]INFO 04-13 11:01:27 [block_pool.py:378] Successfully reset prefix cache

 45%|████▌     | 1500/3300 [3:40:36<3:32:42,  7.09s/it]2026-04-13 11:01:34,270 | INFO | train_grpo_train | metrics_logged mode=train step=1500

                                                       
{'loss': 0.0307, 'grad_norm': 10.32536506652832, 'learning_rate': 5.457575757575758e-06, 'num_tokens': 2649204.0, 'completions/mean_length': 51.75, 'completions/min_length': 45.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.7600724697113037, 'rewards/meter/std': 0.37609532475471497, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9694873690605164, 'rewards/repeat_soft/std': 0.04620639607310295, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6760584115982056, 'rewards/total_composite/std': 0.216745987534523, 'reward': 0.6760584115982056, 'reward_std': 0.216745987534523, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15095128118991852, 'sampling/sampling_logp_difference/max': 4.3736796379089355, 'sampling/importance_sampling_ratio/min': 0.012604773975908756, 'sampling/importance_sampling_ratio/mean': 1.0109680891036987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8784070312976837, 'clip_ratio/low_mean': 0.0693869749084115, 'clip_ratio/low_min': 0.0693869749084115, 'clip_ratio/high_mean': 0.0614222576841712, 'clip_ratio/high_max': 0.0614222576841712, 'clip_ratio/region_mean': 0.1308092325925827, 'reward_total_mean': 0.6760584115982056, 'reward_meter_mean': 0.7600724697113037, 'reward_meter_std': 0.37609532475471497, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9694873690605164, 'reward_repeat_soft_std': 0.04620639607310295, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6760584115982056, 'reward_total_composite_std': 0.216745987534523, 'epoch': 0.15}

 45%|████▌     | 1500/3300 [3:40:36<3:32:42,  7.09s/it]INFO 04-13 11:01:34 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:18,  2.31s/it]

 30%|███       | 3/10 [00:13<00:35,  5.04s/it]

 40%|████      | 4/10 [00:16<00:26,  4.44s/it]

 50%|█████     | 5/10 [00:21<00:21,  4.38s/it]

 60%|██████    | 6/10 [00:25<00:16,  4.20s/it]

 70%|███████   | 7/10 [00:29<00:12,  4.33s/it]

 80%|████████  | 8/10 [00:33<00:08,  4.17s/it]

 90%|█████████ | 9/10 [00:37<00:04,  4.01s/it]

100%|██████████| 10/10 [00:40<00:00,  3.86s/it]2026-04-13 11:02:19,335 | INFO | train_grpo_train | metrics_logged mode=eval step=1500

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 44.9802, 'eval_samples_per_second': 1.779, 'eval_steps_per_second': 0.222, 'eval_num_tokens': 2649204.0, 'eval_completions/mean_length': 85.5875, 'eval_completions/min_length': 32.7, 'eval_completions/max_length': 162.7, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 80.42500076293945, 'eval_completions/min_terminated_length': 32.7, 'eval_completions/max_terminated_length': 127.9, 'eval_rewards/meter/mean': 0.8153466045856476, 'eval_rewards/meter/std': 0.2554396173916757, 'eval_rewards/count_adherence/mean': 0.9520833134651184, 'eval_rewards/count_adherence/std': 0.07691999115049838, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.9100002408027649, 'eval_rewards/repeat_soft/std': 0.07414132617413997, 'eval_rewards/judge_quality/mean': 0.49362499713897706, 'eval_rewards/judge_quality/std': 0.16515217600390314, 'eval_rewards/total_composite/mean': 0.582668250799179, 'eval_rewards/total_composite/std': 0.15843873098492622, 'eval_reward': 0.582668250799179, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06454889848828316, 'eval_sampling/sampling_logp_difference/max': 1.154489517211914, 'eval_sampling/importance_sampling_ratio/min': 0.3234517902135849, 'eval_sampling/importance_sampling_ratio/mean': 1.0130776166915894, 'eval_sampling/importance_sampling_ratio/max': 1.4355666995048524, 'eval_entropy': 0.7258538126945495, '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': 0.582668250799179, 'eval_reward_meter_mean': 0.8153466045856476, 'eval_reward_meter_std': 0.2554396173916757, 'eval_reward_count_adherence_mean': 0.9520833134651184, 'eval_reward_count_adherence_std': 0.07691999115049838, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.9100002408027649, 'eval_reward_repeat_soft_std': 0.07414132617413997, 'eval_reward_judge_quality_mean': 0.49362499713897706, 'eval_reward_judge_quality_std': 0.16515217600390314, 'eval_reward_total_composite_mean': 0.582668250799179, 'eval_reward_total_composite_std': 0.15843873098492622, 'epoch': 0.15}

 45%|████▌     | 1500/3300 [3:41:21<3:32:42,  7.09s/it]

100%|██████████| 10/10 [00:40<00:00,  3.86s/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(

 45%|████▌     | 1501/3300 [3:41:36<11:31:14, 23.05s/it]2026-04-13 11:02:34,577 | INFO | train_grpo_train | metrics_logged mode=train step=1501

                                                        
{'loss': -0.2061, 'grad_norm': 2.157881498336792, 'learning_rate': 5.4545454545454545e-06, 'num_tokens': 2651603.0, 'completions/mean_length': 162.875, 'completions/min_length': 101.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.00000762939453, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8726204633712769, 'rewards/meter/std': 0.1572498232126236, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.1414213627576828, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8686163425445557, 'rewards/repeat_soft/std': 0.08269593119621277, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.44510364532470703, 'rewards/total_composite/std': 0.18498866260051727, 'reward': 0.44510364532470703, 'reward_std': 0.18498864769935608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10576401650905609, 'sampling/sampling_logp_difference/max': 1.4901682138442993, 'sampling/importance_sampling_ratio/min': 0.2253347635269165, 'sampling/importance_sampling_ratio/mean': 1.0134966373443604, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6739145219326019, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09370322432368994, 'clip_ratio/high_max': 0.09370322432368994, 'clip_ratio/region_mean': 0.09370322432368994, 'reward_total_mean': 0.44510364532470703, 'reward_meter_mean': 0.8726204633712769, 'reward_meter_std': 0.1572498232126236, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.1414213627576828, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8686163425445557, 'reward_repeat_soft_std': 0.08269593119621277, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.44510364532470703, 'reward_total_composite_std': 0.18498866260051727, 'epoch': 0.15}

 45%|████▌     | 1501/3300 [3:41:36<11:31:14, 23.05s/it]INFO 04-13 11:02:34 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1502/3300 [3:41:43<9:01:45, 18.08s/it] 2026-04-13 11:02:41,046 | INFO | train_grpo_train | metrics_logged mode=train step=1502

                                                       
{'loss': 0.0589, 'grad_norm': 19.415096282958984, 'learning_rate': 5.451515151515152e-06, 'num_tokens': 2653263.0, 'completions/mean_length': 49.5, 'completions/min_length': 42.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8572096824645996, 'rewards/meter/std': 0.30716466903686523, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8570389151573181, 'rewards/repeat_soft/std': 0.11212202161550522, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5695780515670776, 'rewards/total_composite/std': 0.08213376253843307, 'reward': 0.5695780515670776, 'reward_std': 0.08213376253843307, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10910096019506454, 'sampling/sampling_logp_difference/max': 1.8397150039672852, 'sampling/importance_sampling_ratio/min': 0.15886269509792328, 'sampling/importance_sampling_ratio/mean': 1.006670355796814, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6812138892710209, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.10409970721229911, 'clip_ratio/high_max': 0.10409970721229911, 'clip_ratio/region_mean': 0.11972470721229911, 'reward_total_mean': 0.5695780515670776, 'reward_meter_mean': 0.8572096824645996, 'reward_meter_std': 0.30716466903686523, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8570389151573181, 'reward_repeat_soft_std': 0.11212202161550522, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5695780515670776, 'reward_total_composite_std': 0.08213376253843307, 'epoch': 0.15}

 46%|████▌     | 1502/3300 [3:41:43<9:01:45, 18.08s/it]INFO 04-13 11:02:41 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1503/3300 [3:41:50<7:20:59, 14.72s/it]2026-04-13 11:02:47,939 | INFO | train_grpo_train | metrics_logged mode=train step=1503

                                                       
{'loss': -0.0259, 'grad_norm': 8.051602363586426, 'learning_rate': 5.448484848484848e-06, 'num_tokens': 2655185.0, 'completions/mean_length': 60.25, 'completions/min_length': 52.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.25, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9878210425376892, 'rewards/meter/std': 0.007200425956398249, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9534200429916382, 'rewards/repeat_soft/std': 0.04252530634403229, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.6198652386665344, 'rewards/total_composite/std': 0.008003455586731434, 'reward': 0.6198652386665344, 'reward_std': 0.008003444410860538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1118922159075737, 'sampling/sampling_logp_difference/max': 1.7507295608520508, 'sampling/importance_sampling_ratio/min': 0.1736472100019455, 'sampling/importance_sampling_ratio/mean': 1.0283818244934082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7226510271430016, 'clip_ratio/low_mean': 0.053972672671079636, 'clip_ratio/low_min': 0.053972672671079636, 'clip_ratio/high_mean': 0.06846752669662237, 'clip_ratio/high_max': 0.06846752669662237, 'clip_ratio/region_mean': 0.12244019936770201, 'reward_total_mean': 0.6198652386665344, 'reward_meter_mean': 0.9878210425376892, 'reward_meter_std': 0.007200425956398249, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9534200429916382, 'reward_repeat_soft_std': 0.04252530634403229, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.6198652386665344, 'reward_total_composite_std': 0.008003455586731434, 'epoch': 0.15}

 46%|████▌     | 1503/3300 [3:41:50<7:20:59, 14.72s/it]INFO 04-13 11:02:48 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1504/3300 [3:41:57<6:12:45, 12.45s/it]2026-04-13 11:02:55,096 | INFO | train_grpo_train | metrics_logged mode=train step=1504

                                                       
{'loss': 0.0935, 'grad_norm': 5.836754322052002, 'learning_rate': 5.445454545454546e-06, 'num_tokens': 2657649.0, 'completions/mean_length': 137.0, 'completions/min_length': 96.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 137.0, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.8068217039108276, 'rewards/meter/std': 0.33071011304855347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8825508952140808, 'rewards/repeat_soft/std': 0.0859769731760025, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.6047463417053223, 'rewards/total_composite/std': 0.14661374688148499, 'reward': 0.6047463417053223, 'reward_std': 0.14661374688148499, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1181572675704956, 'sampling/sampling_logp_difference/max': 2.094663143157959, 'sampling/importance_sampling_ratio/min': 0.12311170995235443, 'sampling/importance_sampling_ratio/mean': 1.0147782564163208, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7575431764125824, 'clip_ratio/low_mean': 0.051041049882769585, 'clip_ratio/low_min': 0.051041049882769585, 'clip_ratio/high_mean': 0.07136560790240765, 'clip_ratio/high_max': 0.07136560790240765, 'clip_ratio/region_mean': 0.12240665778517723, 'reward_total_mean': 0.6047463417053223, 'reward_meter_mean': 0.8068217039108276, 'reward_meter_std': 0.33071011304855347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8825508952140808, 'reward_repeat_soft_std': 0.0859769731760025, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.6047463417053223, 'reward_total_composite_std': 0.14661374688148499, 'epoch': 0.15}

 46%|████▌     | 1504/3300 [3:41:57<6:12:45, 12.45s/it]INFO 04-13 11:02:55 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1505/3300 [3:42:03<5:16:25, 10.58s/it]2026-04-13 11:03:01,296 | INFO | train_grpo_train | metrics_logged mode=train step=1505

                                                       
{'loss': 0.0829, 'grad_norm': 16.15242576599121, 'learning_rate': 5.442424242424243e-06, 'num_tokens': 2659213.0, 'completions/mean_length': 34.5, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9436606764793396, 'rewards/meter/std': 0.03392787277698517, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9545820951461792, 'rewards/repeat_soft/std': 0.04496331885457039, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7173162698745728, 'rewards/total_composite/std': 0.16378594934940338, 'reward': 0.7173162698745728, 'reward_std': 0.16378596425056458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08711379766464233, 'sampling/sampling_logp_difference/max': 1.1202850341796875, 'sampling/importance_sampling_ratio/min': 0.32618680596351624, 'sampling/importance_sampling_ratio/mean': 1.0169613361358643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.553403202444315, 'clip_ratio/low_mean': 0.05818144977092743, 'clip_ratio/low_min': 0.05818144977092743, 'clip_ratio/high_mean': 0.04139785002917051, 'clip_ratio/high_max': 0.04139785002917051, 'clip_ratio/region_mean': 0.09957929980009794, 'reward_total_mean': 0.7173162698745728, 'reward_meter_mean': 0.9436606764793396, 'reward_meter_std': 0.03392787277698517, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9545820951461792, 'reward_repeat_soft_std': 0.04496331885457039, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7173162698745728, 'reward_total_composite_std': 0.16378594934940338, 'epoch': 0.15}

 46%|████▌     | 1505/3300 [3:42:03<5:16:25, 10.58s/it]INFO 04-13 11:03:01 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1506/3300 [3:42:10<4:46:27,  9.58s/it]2026-04-13 11:03:08,547 | INFO | train_grpo_train | metrics_logged mode=train step=1506

                                                       
{'loss': 0.0089, 'grad_norm': 7.340456008911133, 'learning_rate': 5.43939393939394e-06, 'num_tokens': 2661594.0, 'completions/mean_length': 113.625, 'completions/min_length': 102.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.8746345043182373, 'rewards/meter/std': 0.2627345025539398, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8636871576309204, 'rewards/repeat_soft/std': 0.07790645956993103, 'rewards/judge_quality/mean': 0.44874998927116394, 'rewards/judge_quality/std': 0.16137246787548065, 'rewards/total_composite/mean': 0.5864561200141907, 'rewards/total_composite/std': 0.13236787915229797, 'reward': 0.5864561200141907, 'reward_std': 0.13236786425113678, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09377837926149368, 'sampling/sampling_logp_difference/max': 2.0795183181762695, 'sampling/importance_sampling_ratio/min': 0.12499039620161057, 'sampling/importance_sampling_ratio/mean': 1.0091928243637085, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5727298632264137, 'clip_ratio/low_mean': 0.04311327030882239, 'clip_ratio/low_min': 0.04311327030882239, 'clip_ratio/high_mean': 0.046685791574418545, 'clip_ratio/high_max': 0.046685791574418545, 'clip_ratio/region_mean': 0.08979906188324094, 'reward_total_mean': 0.5864561200141907, 'reward_meter_mean': 0.8746345043182373, 'reward_meter_std': 0.2627345025539398, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8636871576309204, 'reward_repeat_soft_std': 0.07790645956993103, 'reward_judge_quality_mean': 0.44874998927116394, 'reward_judge_quality_std': 0.16137246787548065, 'reward_total_composite_mean': 0.5864561200141907, 'reward_total_composite_std': 0.13236787915229797, 'epoch': 0.15}

 46%|████▌     | 1506/3300 [3:42:10<4:46:27,  9.58s/it]INFO 04-13 11:03:08 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1507/3300 [3:42:19<4:34:51,  9.20s/it]2026-04-13 11:03:16,852 | INFO | train_grpo_train | metrics_logged mode=train step=1507

                                                       
{'loss': -0.0042, 'grad_norm': 7.900362968444824, 'learning_rate': 5.436363636363636e-06, 'num_tokens': 2663816.0, 'completions/mean_length': 99.75, 'completions/min_length': 85.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.75, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.7890487909317017, 'rewards/meter/std': 0.29200175404548645, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9465805888175964, 'rewards/repeat_soft/std': 0.04801928624510765, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5079265236854553, 'rewards/total_composite/std': 0.2088804692029953, 'reward': 0.5079265236854553, 'reward_std': 0.2088804692029953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1373182237148285, 'sampling/sampling_logp_difference/max': 2.233384132385254, 'sampling/importance_sampling_ratio/min': 0.1071651503443718, 'sampling/importance_sampling_ratio/mean': 0.9960684776306152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7693092972040176, 'clip_ratio/low_mean': 0.016129031777381897, 'clip_ratio/low_min': 0.016129031777381897, 'clip_ratio/high_mean': 0.11910470807924867, 'clip_ratio/high_max': 0.11910470807924867, 'clip_ratio/region_mean': 0.13523373985663056, 'reward_total_mean': 0.5079265236854553, 'reward_meter_mean': 0.7890487909317017, 'reward_meter_std': 0.29200175404548645, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9465805888175964, 'reward_repeat_soft_std': 0.04801928624510765, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5079265236854553, 'reward_total_composite_std': 0.2088804692029953, 'epoch': 0.15}

 46%|████▌     | 1507/3300 [3:42:19<4:34:51,  9.20s/it]INFO 04-13 11:03:17 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1508/3300 [3:42:25<4:06:28,  8.25s/it]2026-04-13 11:03:22,903 | INFO | train_grpo_train | metrics_logged mode=train step=1508

                                                       
{'loss': 0.0014, 'grad_norm': 11.292783737182617, 'learning_rate': 5.4333333333333335e-06, 'num_tokens': 2665539.0, 'completions/mean_length': 46.375, 'completions/min_length': 42.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.948706865310669, 'rewards/meter/std': 0.10278742760419846, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8677626848220825, 'rewards/repeat_soft/std': 0.03913799673318863, 'rewards/judge_quality/mean': 0.48124998807907104, 'rewards/judge_quality/std': 0.2294986993074417, 'rewards/total_composite/mean': 0.6284198760986328, 'rewards/total_composite/std': 0.15320169925689697, 'reward': 0.6284198760986328, 'reward_std': 0.15320169925689697, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09286518394947052, 'sampling/sampling_logp_difference/max': 0.9282870292663574, 'sampling/importance_sampling_ratio/min': 0.39523017406463623, 'sampling/importance_sampling_ratio/mean': 1.0193724632263184, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6417422741651535, 'clip_ratio/low_mean': 0.06929143890738487, 'clip_ratio/low_min': 0.06929143890738487, 'clip_ratio/high_mean': 0.03787878900766373, 'clip_ratio/high_max': 0.03787878900766373, 'clip_ratio/region_mean': 0.1071702279150486, 'reward_total_mean': 0.6284198760986328, 'reward_meter_mean': 0.948706865310669, 'reward_meter_std': 0.10278742760419846, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8677626848220825, 'reward_repeat_soft_std': 0.03913799673318863, 'reward_judge_quality_mean': 0.48124998807907104, 'reward_judge_quality_std': 0.2294986993074417, 'reward_total_composite_mean': 0.6284198760986328, 'reward_total_composite_std': 0.15320169925689697, 'epoch': 0.15}

 46%|████▌     | 1508/3300 [3:42:25<4:06:28,  8.25s/it]INFO 04-13 11:03:23 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1509/3300 [3:42:31<3:52:33,  7.79s/it]2026-04-13 11:03:29,616 | INFO | train_grpo_train | metrics_logged mode=train step=1509

                                                       
{'loss': 0.0009, 'grad_norm': 8.47575855255127, 'learning_rate': 5.430303030303032e-06, 'num_tokens': 2667852.0, 'completions/mean_length': 98.125, 'completions/min_length': 92.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.125, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.6732974052429199, 'rewards/meter/std': 0.3062664270401001, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9173442721366882, 'rewards/repeat_soft/std': 0.08115822076797485, 'rewards/judge_quality/mean': 0.5525000095367432, 'rewards/judge_quality/std': 0.18941642343997955, 'rewards/total_composite/mean': 0.5861427187919617, 'rewards/total_composite/std': 0.1619904786348343, 'reward': 0.5861427187919617, 'reward_std': 0.1619904637336731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12370063364505768, 'sampling/sampling_logp_difference/max': 2.313128709793091, 'sampling/importance_sampling_ratio/min': 0.09895117580890656, 'sampling/importance_sampling_ratio/mean': 1.005765438079834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7182970643043518, 'clip_ratio/low_mean': 0.07507901545614004, 'clip_ratio/low_min': 0.07507901545614004, 'clip_ratio/high_mean': 0.059414125978946686, 'clip_ratio/high_max': 0.059414125978946686, 'clip_ratio/region_mean': 0.13449314143508673, 'reward_total_mean': 0.5861427187919617, 'reward_meter_mean': 0.6732974052429199, 'reward_meter_std': 0.3062664270401001, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9173442721366882, 'reward_repeat_soft_std': 0.08115822076797485, 'reward_judge_quality_mean': 0.5525000095367432, 'reward_judge_quality_std': 0.18941642343997955, 'reward_total_composite_mean': 0.5861427187919617, 'reward_total_composite_std': 0.1619904786348343, 'epoch': 0.15}

 46%|████▌     | 1509/3300 [3:42:31<3:52:33,  7.79s/it]INFO 04-13 11:03:29 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1510/3300 [3:42:37<3:37:20,  7.29s/it]2026-04-13 11:03:35,722 | INFO | train_grpo_train | metrics_logged mode=train step=1510

                                                       
{'loss': 0.0284, 'grad_norm': 20.081600189208984, 'learning_rate': 5.427272727272728e-06, 'num_tokens': 2669263.0, 'completions/mean_length': 22.375, 'completions/min_length': 19.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6335141658782959, 'rewards/meter/std': 0.4142892062664032, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9450810551643372, 'rewards/repeat_soft/std': 0.02819715067744255, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5171016454696655, 'rewards/total_composite/std': 0.114821657538414, 'reward': 0.5171016454696655, 'reward_std': 0.11482164263725281, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14239293336868286, 'sampling/sampling_logp_difference/max': 1.3075666427612305, 'sampling/importance_sampling_ratio/min': 0.27047741413116455, 'sampling/importance_sampling_ratio/mean': 1.0207654237747192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9001251459121704, 'clip_ratio/low_mean': 0.0656578941270709, 'clip_ratio/low_min': 0.0656578941270709, 'clip_ratio/high_mean': 0.09755183570086956, 'clip_ratio/high_max': 0.09755183570086956, 'clip_ratio/region_mean': 0.16320972982794046, 'reward_total_mean': 0.5171016454696655, 'reward_meter_mean': 0.6335141658782959, 'reward_meter_std': 0.4142892062664032, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9450810551643372, 'reward_repeat_soft_std': 0.02819715067744255, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5171016454696655, 'reward_total_composite_std': 0.114821657538414, 'epoch': 0.15}

 46%|████▌     | 1510/3300 [3:42:37<3:37:20,  7.29s/it]INFO 04-13 11:03:35 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1511/3300 [3:42:48<4:10:26,  8.40s/it]2026-04-13 11:03:46,718 | INFO | train_grpo_train | metrics_logged mode=train step=1511

                                                       
{'loss': -0.1045, 'grad_norm': 2.0721373558044434, 'learning_rate': 5.424242424242425e-06, 'num_tokens': 2670635.0, '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': 55.0, 'rewards/meter/mean': 0.7495286464691162, 'rewards/meter/std': 0.22441412508487701, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9604498147964478, 'rewards/repeat_soft/std': 0.04676219820976257, 'rewards/judge_quality/mean': 0.33375000953674316, 'rewards/judge_quality/std': 0.16159361600875854, 'rewards/total_composite/mean': 0.47083914279937744, 'rewards/total_composite/std': 0.201918825507164, 'reward': 0.47083914279937744, 'reward_std': 0.2019188106060028, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13865235447883606, 'sampling/sampling_logp_difference/max': 1.438107967376709, 'sampling/importance_sampling_ratio/min': 0.23737645149230957, 'sampling/importance_sampling_ratio/mean': 1.026513934135437, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7795788422226906, 'clip_ratio/low_mean': 0.04449760727584362, 'clip_ratio/low_min': 0.04449760727584362, 'clip_ratio/high_mean': 0.0813118212390691, 'clip_ratio/high_max': 0.0813118212390691, 'clip_ratio/region_mean': 0.12580942851491272, 'reward_total_mean': 0.47083914279937744, 'reward_meter_mean': 0.7495286464691162, 'reward_meter_std': 0.22441412508487701, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9604498147964478, 'reward_repeat_soft_std': 0.04676219820976257, 'reward_judge_quality_mean': 0.33375000953674316, 'reward_judge_quality_std': 0.16159361600875854, 'reward_total_composite_mean': 0.47083914279937744, 'reward_total_composite_std': 0.201918825507164, 'epoch': 0.15}

 46%|████▌     | 1511/3300 [3:42:48<4:10:26,  8.40s/it]INFO 04-13 11:03:46 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1512/3300 [3:42:54<3:49:10,  7.69s/it]2026-04-13 11:03:52,755 | INFO | train_grpo_train | metrics_logged mode=train step=1512

                                                       
{'loss': 0.0172, 'grad_norm': 13.740234375, 'learning_rate': 5.421212121212122e-06, 'num_tokens': 2672290.0, 'completions/mean_length': 41.875, 'completions/min_length': 36.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.449136346578598, 'rewards/meter/std': 0.31989529728889465, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9736288785934448, 'rewards/repeat_soft/std': 0.027965646237134933, 'rewards/judge_quality/mean': 0.6050000190734863, 'rewards/judge_quality/std': 0.22025960683822632, 'rewards/total_composite/mean': 0.5165993571281433, 'rewards/total_composite/std': 0.12132763862609863, 'reward': 0.5165993571281433, 'reward_std': 0.12132763862609863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14902806282043457, 'sampling/sampling_logp_difference/max': 1.2676098346710205, 'sampling/importance_sampling_ratio/min': 0.28150367736816406, 'sampling/importance_sampling_ratio/mean': 1.0072979927062988, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8480989933013916, 'clip_ratio/low_mean': 0.059883545152843, 'clip_ratio/low_min': 0.059883545152843, 'clip_ratio/high_mean': 0.07934731431305408, 'clip_ratio/high_max': 0.07934731431305408, 'clip_ratio/region_mean': 0.13923085946589708, 'reward_total_mean': 0.5165993571281433, 'reward_meter_mean': 0.449136346578598, 'reward_meter_std': 0.31989529728889465, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9736288785934448, 'reward_repeat_soft_std': 0.027965646237134933, 'reward_judge_quality_mean': 0.6050000190734863, 'reward_judge_quality_std': 0.22025960683822632, 'reward_total_composite_mean': 0.5165993571281433, 'reward_total_composite_std': 0.12132763862609863, 'epoch': 0.15}

 46%|████▌     | 1512/3300 [3:42:54<3:49:10,  7.69s/it]INFO 04-13 11:03:52 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1513/3300 [3:43:06<4:19:59,  8.73s/it]2026-04-13 11:04:03,909 | INFO | train_grpo_train | metrics_logged mode=train step=1513

                                                       
{'loss': -0.189, 'grad_norm': 1.9482091665267944, 'learning_rate': 5.418181818181819e-06, 'num_tokens': 2674183.0, 'completions/mean_length': 207.625, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 106.16667175292969, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.6227824687957764, 'rewards/meter/std': 0.3599480390548706, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9174457788467407, 'rewards/repeat_soft/std': 0.03720393404364586, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.4073261320590973, 'rewards/total_composite/std': 0.2590354084968567, 'reward': 0.4073261320590973, 'reward_std': 0.2590354084968567, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12539781630039215, 'sampling/sampling_logp_difference/max': 2.0272881984710693, 'sampling/importance_sampling_ratio/min': 0.13169217109680176, 'sampling/importance_sampling_ratio/mean': 1.0209314823150635, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4624735414981842, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09595593344420195, 'clip_ratio/high_max': 0.09595593344420195, 'clip_ratio/region_mean': 0.09595593344420195, 'reward_total_mean': 0.4073261320590973, 'reward_meter_mean': 0.6227824687957764, 'reward_meter_std': 0.3599480390548706, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9174457788467407, 'reward_repeat_soft_std': 0.03720393404364586, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.4073261320590973, 'reward_total_composite_std': 0.2590354084968567, 'epoch': 0.15}

 46%|████▌     | 1513/3300 [3:43:06<4:19:59,  8.73s/it]INFO 04-13 11:04:04 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1514/3300 [3:43:12<3:57:47,  7.99s/it]2026-04-13 11:04:10,171 | INFO | train_grpo_train | metrics_logged mode=train step=1514

                                                       
{'loss': 0.0397, 'grad_norm': 14.658851623535156, 'learning_rate': 5.415151515151515e-06, 'num_tokens': 2675842.0, 'completions/mean_length': 42.375, 'completions/min_length': 34.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.40207529067993164, 'rewards/meter/std': 0.38992664217948914, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9891808032989502, 'rewards/repeat_soft/std': 0.009330120868980885, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4599241614341736, 'rewards/total_composite/std': 0.10620740801095963, 'reward': 0.4599241614341736, 'reward_std': 0.10620740801095963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15452642738819122, 'sampling/sampling_logp_difference/max': 1.1953868865966797, 'sampling/importance_sampling_ratio/min': 0.3025868833065033, 'sampling/importance_sampling_ratio/mean': 1.004881739616394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.845390260219574, 'clip_ratio/low_mean': 0.11816397495567799, 'clip_ratio/low_min': 0.11816397495567799, 'clip_ratio/high_mean': 0.0432773120701313, 'clip_ratio/high_max': 0.0432773120701313, 'clip_ratio/region_mean': 0.1614412870258093, 'reward_total_mean': 0.4599241614341736, 'reward_meter_mean': 0.40207529067993164, 'reward_meter_std': 0.38992664217948914, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9891808032989502, 'reward_repeat_soft_std': 0.009330120868980885, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4599241614341736, 'reward_total_composite_std': 0.10620740801095963, 'epoch': 0.15}

 46%|████▌     | 1514/3300 [3:43:12<3:57:47,  7.99s/it]INFO 04-13 11:04:10 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1515/3300 [3:43:18<3:39:26,  7.38s/it]2026-04-13 11:04:16,117 | INFO | train_grpo_train | metrics_logged mode=train step=1515

                                                       
{'loss': -0.0282, 'grad_norm': 10.259291648864746, 'learning_rate': 5.412121212121213e-06, 'num_tokens': 2677438.0, 'completions/mean_length': 32.5, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.7878463864326477, 'rewards/meter/std': 0.2908877432346344, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9157747030258179, 'rewards/repeat_soft/std': 0.08889061957597733, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5524482727050781, 'rewards/total_composite/std': 0.07280430197715759, 'reward': 0.5524482727050781, 'reward_std': 0.07280430942773819, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08970874547958374, 'sampling/sampling_logp_difference/max': 1.055037260055542, 'sampling/importance_sampling_ratio/min': 0.348179429769516, 'sampling/importance_sampling_ratio/mean': 0.9857891201972961, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34806801564991474, 'clip_ratio/low_mean': 0.01293103490024805, 'clip_ratio/low_min': 0.01293103490024805, 'clip_ratio/high_mean': 0.07328375522047281, 'clip_ratio/high_max': 0.07328375522047281, 'clip_ratio/region_mean': 0.08621479012072086, 'reward_total_mean': 0.5524482727050781, 'reward_meter_mean': 0.7878463864326477, 'reward_meter_std': 0.2908877432346344, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9157747030258179, 'reward_repeat_soft_std': 0.08889061957597733, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5524482727050781, 'reward_total_composite_std': 0.07280430197715759, 'epoch': 0.15}

 46%|████▌     | 1515/3300 [3:43:18<3:39:26,  7.38s/it]INFO 04-13 11:04:16 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1516/3300 [3:43:24<3:25:48,  6.92s/it]2026-04-13 11:04:21,978 | INFO | train_grpo_train | metrics_logged mode=train step=1516

                                                       
{'loss': 0.0945, 'grad_norm': 10.050907135009766, 'learning_rate': 5.409090909090909e-06, 'num_tokens': 2678942.0, 'completions/mean_length': 45.0, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.9644712805747986, 'rewards/meter/std': 0.04411641135811806, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8823840618133545, 'rewards/repeat_soft/std': 0.07079489529132843, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.27994900941848755, 'rewards/total_composite/mean': 0.701180100440979, 'rewards/total_composite/std': 0.1814952939748764, 'reward': 0.701180100440979, 'reward_std': 0.1814952939748764, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10215334594249725, 'sampling/sampling_logp_difference/max': 2.1007111072540283, 'sampling/importance_sampling_ratio/min': 0.12236937880516052, 'sampling/importance_sampling_ratio/mean': 1.007930874824524, 'sampling/importance_sampling_ratio/max': 1.9660403728485107, 'entropy': 0.5422038696706295, 'clip_ratio/low_mean': 0.04473423445597291, 'clip_ratio/low_min': 0.04473423445597291, 'clip_ratio/high_mean': 0.04288395494222641, 'clip_ratio/high_max': 0.04288395494222641, 'clip_ratio/region_mean': 0.08761818939819932, 'reward_total_mean': 0.701180100440979, 'reward_meter_mean': 0.9644712805747986, 'reward_meter_std': 0.04411641135811806, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8823840618133545, 'reward_repeat_soft_std': 0.07079489529132843, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.27994900941848755, 'reward_total_composite_mean': 0.701180100440979, 'reward_total_composite_std': 0.1814952939748764, 'epoch': 0.15}

 46%|████▌     | 1516/3300 [3:43:24<3:25:48,  6.92s/it]INFO 04-13 11:04:22 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1517/3300 [3:43:29<3:15:23,  6.58s/it]2026-04-13 11:04:27,747 | INFO | train_grpo_train | metrics_logged mode=train step=1517

                                                       
{'loss': 0.1147, 'grad_norm': 20.944910049438477, 'learning_rate': 5.406060606060607e-06, 'num_tokens': 2680412.0, 'completions/mean_length': 23.75, 'completions/min_length': 20.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.7645701169967651, 'rewards/meter/std': 0.3575473725795746, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9476451277732849, 'rewards/repeat_soft/std': 0.021861007437109947, 'rewards/judge_quality/mean': 0.5975000262260437, 'rewards/judge_quality/std': 0.2750454545021057, 'rewards/total_composite/mean': 0.6685272455215454, 'rewards/total_composite/std': 0.2298787236213684, 'reward': 0.6685272455215454, 'reward_std': 0.2298787236213684, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1604284942150116, 'sampling/sampling_logp_difference/max': 1.1559343338012695, 'sampling/importance_sampling_ratio/min': 0.31476327776908875, 'sampling/importance_sampling_ratio/mean': 1.0079511404037476, 'sampling/importance_sampling_ratio/max': 1.9789609909057617, 'entropy': 1.0180640891194344, 'clip_ratio/low_mean': 0.09042960871011019, 'clip_ratio/low_min': 0.09042960871011019, 'clip_ratio/high_mean': 0.038977273274213076, 'clip_ratio/high_max': 0.038977273274213076, 'clip_ratio/region_mean': 0.12940688198432326, 'reward_total_mean': 0.6685272455215454, 'reward_meter_mean': 0.7645701169967651, 'reward_meter_std': 0.3575473725795746, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9476451277732849, 'reward_repeat_soft_std': 0.021861007437109947, 'reward_judge_quality_mean': 0.5975000262260437, 'reward_judge_quality_std': 0.2750454545021057, 'reward_total_composite_mean': 0.6685272455215454, 'reward_total_composite_std': 0.2298787236213684, 'epoch': 0.15}

 46%|████▌     | 1517/3300 [3:43:29<3:15:23,  6.58s/it]INFO 04-13 11:04:27 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1518/3300 [3:43:36<3:17:50,  6.66s/it]2026-04-13 11:04:34,608 | INFO | train_grpo_train | metrics_logged mode=train step=1518

                                                       
{'loss': -0.0043, 'grad_norm': 7.887856960296631, 'learning_rate': 5.4030303030303036e-06, 'num_tokens': 2682797.0, 'completions/mean_length': 106.125, 'completions/min_length': 78.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.125, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.9521242380142212, 'rewards/meter/std': 0.03605160117149353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.928634762763977, 'rewards/repeat_soft/std': 0.04874228313565254, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.6227474212646484, 'rewards/total_composite/std': 0.0639571100473404, 'reward': 0.6227474212646484, 'reward_std': 0.0639571025967598, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13802793622016907, 'sampling/sampling_logp_difference/max': 2.396497964859009, 'sampling/importance_sampling_ratio/min': 0.09103620797395706, 'sampling/importance_sampling_ratio/mean': 1.0033305883407593, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6701874136924744, 'clip_ratio/low_mean': 0.10489683970808983, 'clip_ratio/low_min': 0.10489683970808983, 'clip_ratio/high_mean': 0.013392857275903225, 'clip_ratio/high_max': 0.013392857275903225, 'clip_ratio/region_mean': 0.11828969698399305, 'reward_total_mean': 0.6227474212646484, 'reward_meter_mean': 0.9521242380142212, 'reward_meter_std': 0.03605160117149353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.928634762763977, 'reward_repeat_soft_std': 0.04874228313565254, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.6227474212646484, 'reward_total_composite_std': 0.0639571100473404, 'epoch': 0.15}

 46%|████▌     | 1518/3300 [3:43:36<3:17:50,  6.66s/it]INFO 04-13 11:04:34 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1519/3300 [3:43:43<3:18:18,  6.68s/it]2026-04-13 11:04:41,332 | INFO | train_grpo_train | metrics_logged mode=train step=1519

                                                       
{'loss': 0.1012, 'grad_norm': 10.09461498260498, 'learning_rate': 5.400000000000001e-06, 'num_tokens': 2685120.0, 'completions/mean_length': 106.375, 'completions/min_length': 85.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.375, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.7603713870048523, 'rewards/meter/std': 0.3669488728046417, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9532291889190674, 'rewards/repeat_soft/std': 0.042330704629421234, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6824182271957397, 'rewards/total_composite/std': 0.20744413137435913, 'reward': 0.6824182271957397, 'reward_std': 0.20744411647319794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15000297129154205, 'sampling/sampling_logp_difference/max': 2.788607120513916, 'sampling/importance_sampling_ratio/min': 0.13063034415245056, 'sampling/importance_sampling_ratio/mean': 1.01371431350708, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9093747064471245, 'clip_ratio/low_mean': 0.08507308643311262, 'clip_ratio/low_min': 0.08507308643311262, 'clip_ratio/high_mean': 0.049013751558959484, 'clip_ratio/high_max': 0.049013751558959484, 'clip_ratio/region_mean': 0.1340868379920721, 'reward_total_mean': 0.6824182271957397, 'reward_meter_mean': 0.7603713870048523, 'reward_meter_std': 0.3669488728046417, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9532291889190674, 'reward_repeat_soft_std': 0.042330704629421234, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6824182271957397, 'reward_total_composite_std': 0.20744413137435913, 'epoch': 0.15}

 46%|████▌     | 1519/3300 [3:43:43<3:18:18,  6.68s/it]INFO 04-13 11:04:41 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1520/3300 [3:43:50<3:25:16,  6.92s/it]2026-04-13 11:04:48,807 | INFO | train_grpo_train | metrics_logged mode=train step=1520

                                                       
{'loss': -0.0249, 'grad_norm': 7.378472328186035, 'learning_rate': 5.396969696969697e-06, 'num_tokens': 2687601.0, 'completions/mean_length': 121.125, 'completions/min_length': 93.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.125, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.926666259765625, 'rewards/meter/std': 0.16308124363422394, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9598362445831299, 'rewards/repeat_soft/std': 0.024580257013440132, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6049239039421082, 'rewards/total_composite/std': 0.10893219709396362, 'reward': 0.6049239039421082, 'reward_std': 0.10893220454454422, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12834319472312927, 'sampling/sampling_logp_difference/max': 1.499781608581543, 'sampling/importance_sampling_ratio/min': 0.2231789082288742, 'sampling/importance_sampling_ratio/mean': 1.0122907161712646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7687502652406693, 'clip_ratio/low_mean': 0.09747287072241306, 'clip_ratio/low_min': 0.09747287072241306, 'clip_ratio/high_mean': 0.031536294147372246, 'clip_ratio/high_max': 0.031536294147372246, 'clip_ratio/region_mean': 0.1290091648697853, 'reward_total_mean': 0.6049239039421082, 'reward_meter_mean': 0.926666259765625, 'reward_meter_std': 0.16308124363422394, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9598362445831299, 'reward_repeat_soft_std': 0.024580257013440132, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6049239039421082, 'reward_total_composite_std': 0.10893219709396362, 'epoch': 0.15}

 46%|████▌     | 1520/3300 [3:43:51<3:25:16,  6.92s/it]INFO 04-13 11:04:49 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1521/3300 [3:43:57<3:18:14,  6.69s/it]2026-04-13 11:04:54,951 | INFO | train_grpo_train | metrics_logged mode=train step=1521

                                                       
{'loss': -0.0609, 'grad_norm': 11.856502532958984, 'learning_rate': 5.3939393939393945e-06, 'num_tokens': 2689306.0, 'completions/mean_length': 53.125, 'completions/min_length': 41.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.9026252627372742, 'rewards/meter/std': 0.233467236161232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9118351936340332, 'rewards/repeat_soft/std': 0.05899510160088539, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.6094926595687866, 'rewards/total_composite/std': 0.09445804357528687, 'reward': 0.6094926595687866, 'reward_std': 0.09445804357528687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11438848823308945, 'sampling/sampling_logp_difference/max': 2.0489282608032227, 'sampling/importance_sampling_ratio/min': 0.12887294590473175, 'sampling/importance_sampling_ratio/mean': 1.0063875913619995, 'sampling/importance_sampling_ratio/max': 1.8470425605773926, 'entropy': 0.7874886989593506, 'clip_ratio/low_mean': 0.05083679733797908, 'clip_ratio/low_min': 0.05083679733797908, 'clip_ratio/high_mean': 0.04340798268094659, 'clip_ratio/high_max': 0.04340798268094659, 'clip_ratio/region_mean': 0.09424478001892567, 'reward_total_mean': 0.6094926595687866, 'reward_meter_mean': 0.9026252627372742, 'reward_meter_std': 0.233467236161232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9118351936340332, 'reward_repeat_soft_std': 0.05899510160088539, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.6094926595687866, 'reward_total_composite_std': 0.09445804357528687, 'epoch': 0.15}

 46%|████▌     | 1521/3300 [3:43:57<3:18:14,  6.69s/it]INFO 04-13 11:04:55 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1522/3300 [3:44:03<3:11:46,  6.47s/it]2026-04-13 11:05:00,925 | INFO | train_grpo_train | metrics_logged mode=train step=1522

                                                       
{'loss': 0.0047, 'grad_norm': 20.712505340576172, 'learning_rate': 5.390909090909091e-06, 'num_tokens': 2690716.0, 'completions/mean_length': 26.25, 'completions/min_length': 21.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.25, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.565726637840271, 'rewards/meter/std': 0.45630767941474915, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9610613584518433, 'rewards/repeat_soft/std': 0.004068940877914429, 'rewards/judge_quality/mean': 0.8237500190734863, 'rewards/judge_quality/std': 0.2722361385822296, 'rewards/total_composite/mean': 0.6389440894126892, 'rewards/total_composite/std': 0.3503834903240204, 'reward': 0.6389440894126892, 'reward_std': 0.350383460521698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17062167823314667, 'sampling/sampling_logp_difference/max': 1.8133325576782227, 'sampling/importance_sampling_ratio/min': 0.1631096601486206, 'sampling/importance_sampling_ratio/mean': 1.0215020179748535, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.5077793076634407, 'clip_ratio/low_mean': 0.029363354202359915, 'clip_ratio/low_min': 0.029363354202359915, 'clip_ratio/high_mean': 0.08101852051913738, 'clip_ratio/high_max': 0.08101852051913738, 'clip_ratio/region_mean': 0.1103818747214973, 'reward_total_mean': 0.6389440894126892, 'reward_meter_mean': 0.565726637840271, 'reward_meter_std': 0.45630767941474915, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9610613584518433, 'reward_repeat_soft_std': 0.004068940877914429, 'reward_judge_quality_mean': 0.8237500190734863, 'reward_judge_quality_std': 0.2722361385822296, 'reward_total_composite_mean': 0.6389440894126892, 'reward_total_composite_std': 0.3503834903240204, 'epoch': 0.15}

 46%|████▌     | 1522/3300 [3:44:03<3:11:46,  6.47s/it]INFO 04-13 11:05:01 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1523/3300 [3:44:09<3:06:58,  6.31s/it]2026-04-13 11:05:06,864 | INFO | train_grpo_train | metrics_logged mode=train step=1523

                                                       
{'loss': 0.0921, 'grad_norm': 11.37076473236084, 'learning_rate': 5.387878787878789e-06, 'num_tokens': 2692426.0, 'completions/mean_length': 50.75, 'completions/min_length': 36.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7785061597824097, 'rewards/meter/std': 0.3467750549316406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9427660703659058, 'rewards/repeat_soft/std': 0.044391099363565445, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.648284912109375, 'rewards/total_composite/std': 0.1843511015176773, 'reward': 0.648284912109375, 'reward_std': 0.1843511164188385, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14299741387367249, 'sampling/sampling_logp_difference/max': 3.153827428817749, 'sampling/importance_sampling_ratio/min': 0.04268842563033104, 'sampling/importance_sampling_ratio/mean': 0.9996667504310608, 'sampling/importance_sampling_ratio/max': 1.9937615394592285, 'entropy': 0.9337623193860054, 'clip_ratio/low_mean': 0.07621855894103646, 'clip_ratio/low_min': 0.07621855894103646, 'clip_ratio/high_mean': 0.031250000931322575, 'clip_ratio/high_max': 0.031250000931322575, 'clip_ratio/region_mean': 0.10746855987235904, 'reward_total_mean': 0.648284912109375, 'reward_meter_mean': 0.7785061597824097, 'reward_meter_std': 0.3467750549316406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9427660703659058, 'reward_repeat_soft_std': 0.044391099363565445, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.648284912109375, 'reward_total_composite_std': 0.1843511015176773, 'epoch': 0.15}

 46%|████▌     | 1523/3300 [3:44:09<3:06:58,  6.31s/it]INFO 04-13 11:05:07 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1524/3300 [3:44:16<3:15:27,  6.60s/it]2026-04-13 11:05:14,145 | INFO | train_grpo_train | metrics_logged mode=train step=1524

                                                       
{'loss': 0.1018, 'grad_norm': 7.0124335289001465, 'learning_rate': 5.384848484848485e-06, 'num_tokens': 2694722.0, 'completions/mean_length': 111.0, 'completions/min_length': 88.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.0, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.7280535101890564, 'rewards/meter/std': 0.3529597818851471, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7642220854759216, 'rewards/repeat_soft/std': 0.08997632563114166, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.49593979120254517, 'rewards/total_composite/std': 0.09022112190723419, 'reward': 0.49593979120254517, 'reward_std': 0.090221107006073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08519119769334793, 'sampling/sampling_logp_difference/max': 1.982034683227539, 'sampling/importance_sampling_ratio/min': 0.1377885937690735, 'sampling/importance_sampling_ratio/mean': 0.9952489733695984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4152562692761421, 'clip_ratio/low_mean': 0.03927290253341198, 'clip_ratio/low_min': 0.03927290253341198, 'clip_ratio/high_mean': 0.045991506427526474, 'clip_ratio/high_max': 0.045991506427526474, 'clip_ratio/region_mean': 0.08526440896093845, 'reward_total_mean': 0.49593979120254517, 'reward_meter_mean': 0.7280535101890564, 'reward_meter_std': 0.3529597818851471, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7642220854759216, 'reward_repeat_soft_std': 0.08997632563114166, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.49593979120254517, 'reward_total_composite_std': 0.09022112190723419, 'epoch': 0.15}

 46%|████▌     | 1524/3300 [3:44:16<3:15:27,  6.60s/it]INFO 04-13 11:05:14 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1525/3300 [3:44:22<3:10:51,  6.45s/it]2026-04-13 11:05:20,242 | INFO | train_grpo_train | metrics_logged mode=train step=1525

                                                       
{'loss': 0.0811, 'grad_norm': 12.178094863891602, 'learning_rate': 5.381818181818183e-06, 'num_tokens': 2696332.0, 'completions/mean_length': 51.25, 'completions/min_length': 44.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8399039506912231, 'rewards/meter/std': 0.21824374794960022, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9909389019012451, 'rewards/repeat_soft/std': 0.00940681528300047, 'rewards/judge_quality/mean': 0.7437500357627869, 'rewards/judge_quality/std': 0.2083909660577774, 'rewards/total_composite/mean': 0.752814531326294, 'rewards/total_composite/std': 0.15112841129302979, 'reward': 0.752814531326294, 'reward_std': 0.15112842619419098, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11794614046812057, 'sampling/sampling_logp_difference/max': 1.3004674911499023, 'sampling/importance_sampling_ratio/min': 0.27240443229675293, 'sampling/importance_sampling_ratio/mean': 1.02642023563385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7635732442140579, 'clip_ratio/low_mean': 0.06484845839440823, 'clip_ratio/low_min': 0.06484845839440823, 'clip_ratio/high_mean': 0.07267182040959597, 'clip_ratio/high_max': 0.07267182040959597, 'clip_ratio/region_mean': 0.1375202788040042, 'reward_total_mean': 0.752814531326294, 'reward_meter_mean': 0.8399039506912231, 'reward_meter_std': 0.21824374794960022, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9909389019012451, 'reward_repeat_soft_std': 0.00940681528300047, 'reward_judge_quality_mean': 0.7437500357627869, 'reward_judge_quality_std': 0.2083909660577774, 'reward_total_composite_mean': 0.752814531326294, 'reward_total_composite_std': 0.15112841129302979, 'epoch': 0.15}

 46%|████▌     | 1525/3300 [3:44:22<3:10:51,  6.45s/it]INFO 04-13 11:05:20 [block_pool.py:378] Successfully reset prefix cache

 46%|████▌     | 1526/3300 [3:44:29<3:13:18,  6.54s/it]2026-04-13 11:05:26,999 | INFO | train_grpo_train | metrics_logged mode=train step=1526

                                                       
{'loss': 0.0942, 'grad_norm': 9.61211109161377, 'learning_rate': 5.378787878787879e-06, 'num_tokens': 2698420.0, 'completions/mean_length': 84.0, 'completions/min_length': 68.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.6344290375709534, 'rewards/meter/std': 0.37753942608833313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9780125021934509, 'rewards/repeat_soft/std': 0.034678518772125244, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.49886345863342285, 'rewards/total_composite/std': 0.09963525831699371, 'reward': 0.49886345863342285, 'reward_std': 0.09963525831699371, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16740047931671143, 'sampling/sampling_logp_difference/max': 3.027803897857666, 'sampling/importance_sampling_ratio/min': 0.04842185974121094, 'sampling/importance_sampling_ratio/mean': 1.015329122543335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4986537620425224, 'clip_ratio/low_mean': 0.07367149833589792, 'clip_ratio/low_min': 0.07367149833589792, 'clip_ratio/high_mean': 0.07234195433557034, 'clip_ratio/high_max': 0.07234195433557034, 'clip_ratio/region_mean': 0.14601345267146826, 'reward_total_mean': 0.49886345863342285, 'reward_meter_mean': 0.6344290375709534, 'reward_meter_std': 0.37753942608833313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9780125021934509, 'reward_repeat_soft_std': 0.034678518772125244, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.49886345863342285, 'reward_total_composite_std': 0.09963525831699371, 'epoch': 0.15}

 46%|████▌     | 1526/3300 [3:44:29<3:13:18,  6.54s/it]INFO 04-13 11:05:27 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1527/3300 [3:44:35<3:15:54,  6.63s/it]2026-04-13 11:05:33,826 | INFO | train_grpo_train | metrics_logged mode=train step=1527

                                                       
{'loss': 0.0509, 'grad_norm': 9.31325626373291, 'learning_rate': 5.375757575757576e-06, 'num_tokens': 2700394.0, 'completions/mean_length': 83.75, 'completions/min_length': 64.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.8043165802955627, 'rewards/meter/std': 0.2480888068675995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8991732001304626, 'rewards/repeat_soft/std': 0.083254374563694, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.1891333907842636, 'rewards/total_composite/mean': 0.6492887139320374, 'rewards/total_composite/std': 0.14161844551563263, 'reward': 0.6492887139320374, 'reward_std': 0.14161844551563263, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11522864550352097, 'sampling/sampling_logp_difference/max': 1.8459806442260742, 'sampling/importance_sampling_ratio/min': 0.15787042677402496, 'sampling/importance_sampling_ratio/mean': 1.0108704566955566, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.574681032449007, 'clip_ratio/low_mean': 0.06570903304964304, 'clip_ratio/low_min': 0.06570903304964304, 'clip_ratio/high_mean': 0.05417239014059305, 'clip_ratio/high_max': 0.05417239014059305, 'clip_ratio/region_mean': 0.11988142319023609, 'reward_total_mean': 0.6492887139320374, 'reward_meter_mean': 0.8043165802955627, 'reward_meter_std': 0.2480888068675995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8991732001304626, 'reward_repeat_soft_std': 0.083254374563694, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.1891333907842636, 'reward_total_composite_mean': 0.6492887139320374, 'reward_total_composite_std': 0.14161844551563263, 'epoch': 0.15}

 46%|████▋     | 1527/3300 [3:44:36<3:15:54,  6.63s/it]INFO 04-13 11:05:34 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1528/3300 [3:44:42<3:18:22,  6.72s/it]2026-04-13 11:05:40,751 | INFO | train_grpo_train | metrics_logged mode=train step=1528

                                                       
{'loss': 0.0588, 'grad_norm': 7.919285774230957, 'learning_rate': 5.372727272727273e-06, 'num_tokens': 2702601.0, 'completions/mean_length': 104.875, 'completions/min_length': 85.0, 'completions/max_length': 137.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.875, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9438287019729614, 'rewards/meter/std': 0.11375102400779724, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8729192018508911, 'rewards/repeat_soft/std': 0.12133605778217316, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5657995939254761, 'rewards/total_composite/std': 0.06492535769939423, 'reward': 0.5657995939254761, 'reward_std': 0.06492535769939423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11691422015428543, 'sampling/sampling_logp_difference/max': 1.281205177307129, 'sampling/importance_sampling_ratio/min': 0.2777024209499359, 'sampling/importance_sampling_ratio/mean': 1.0154949426651, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8299001902341843, 'clip_ratio/low_mean': 0.04468996450304985, 'clip_ratio/low_min': 0.04468996450304985, 'clip_ratio/high_mean': 0.0835497984662652, 'clip_ratio/high_max': 0.0835497984662652, 'clip_ratio/region_mean': 0.12823976296931505, 'reward_total_mean': 0.5657995939254761, 'reward_meter_mean': 0.9438287019729614, 'reward_meter_std': 0.11375102400779724, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8729192018508911, 'reward_repeat_soft_std': 0.12133605778217316, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5657995939254761, 'reward_total_composite_std': 0.06492535769939423, 'epoch': 0.15}

 46%|████▋     | 1528/3300 [3:44:42<3:18:22,  6.72s/it]INFO 04-13 11:05:40 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1529/3300 [3:44:50<3:24:59,  6.95s/it]2026-04-13 11:05:48,233 | INFO | train_grpo_train | metrics_logged mode=train step=1529

                                                       
{'loss': 0.0283, 'grad_norm': 6.661654949188232, 'learning_rate': 5.36969696969697e-06, 'num_tokens': 2704969.0, 'completions/mean_length': 122.0, 'completions/min_length': 111.0, 'completions/max_length': 136.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.0, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 136.0, 'rewards/meter/mean': 0.562111496925354, 'rewards/meter/std': 0.3219490349292755, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8385282754898071, 'rewards/repeat_soft/std': 0.077992282807827, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.45416495203971863, 'rewards/total_composite/std': 0.08038046211004257, 'reward': 0.45416495203971863, 'reward_std': 0.08038046956062317, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09357380121946335, 'sampling/sampling_logp_difference/max': 2.4090735912323, 'sampling/importance_sampling_ratio/min': 0.08989854902029037, 'sampling/importance_sampling_ratio/mean': 1.0042166709899902, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5388602800667286, 'clip_ratio/low_mean': 0.026751196943223476, 'clip_ratio/low_min': 0.026751196943223476, 'clip_ratio/high_mean': 0.058976574800908566, 'clip_ratio/high_max': 0.058976574800908566, 'clip_ratio/region_mean': 0.08572777174413204, 'reward_total_mean': 0.45416495203971863, 'reward_meter_mean': 0.562111496925354, 'reward_meter_std': 0.3219490349292755, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8385282754898071, 'reward_repeat_soft_std': 0.077992282807827, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.45416495203971863, 'reward_total_composite_std': 0.08038046211004257, 'epoch': 0.15}

 46%|████▋     | 1529/3300 [3:44:50<3:24:59,  6.95s/it]INFO 04-13 11:05:48 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1530/3300 [3:45:02<4:06:47,  8.37s/it]2026-04-13 11:05:59,918 | INFO | train_grpo_train | metrics_logged mode=train step=1530

                                                       
{'loss': -0.0947, 'grad_norm': 3.3263089656829834, 'learning_rate': 5.366666666666666e-06, 'num_tokens': 2706594.0, 'completions/mean_length': 105.125, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 47.000003814697266, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5275822281837463, 'rewards/meter/std': 0.42082855105400085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9427675008773804, 'rewards/repeat_soft/std': 0.09034770727157593, 'rewards/judge_quality/mean': 0.6312500238418579, 'rewards/judge_quality/std': 0.3341701030731201, 'rewards/total_composite/mean': 0.5328185558319092, 'rewards/total_composite/std': 0.2762130796909332, 'reward': 0.5328185558319092, 'reward_std': 0.27621304988861084, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11439567804336548, 'sampling/sampling_logp_difference/max': 1.2095632553100586, 'sampling/importance_sampling_ratio/min': 0.2983275353908539, 'sampling/importance_sampling_ratio/mean': 1.0066003799438477, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.531870748847723, 'clip_ratio/low_mean': 0.04078907240182161, 'clip_ratio/low_min': 0.04078907240182161, 'clip_ratio/high_mean': 0.0648478849325329, 'clip_ratio/high_max': 0.0648478849325329, 'clip_ratio/region_mean': 0.10563695733435452, 'reward_total_mean': 0.5328185558319092, 'reward_meter_mean': 0.5275822281837463, 'reward_meter_std': 0.42082855105400085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9427675008773804, 'reward_repeat_soft_std': 0.09034770727157593, 'reward_judge_quality_mean': 0.6312500238418579, 'reward_judge_quality_std': 0.3341701030731201, 'reward_total_composite_mean': 0.5328185558319092, 'reward_total_composite_std': 0.2762130796909332, 'epoch': 0.15}

 46%|████▋     | 1530/3300 [3:45:02<4:06:47,  8.37s/it]INFO 04-13 11:06:00 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1531/3300 [3:45:08<3:49:40,  7.79s/it]2026-04-13 11:06:06,359 | INFO | train_grpo_train | metrics_logged mode=train step=1531

                                                       
{'loss': 0.0017, 'grad_norm': 13.402848243713379, 'learning_rate': 5.3636363636363645e-06, 'num_tokens': 2708116.0, 'completions/mean_length': 40.25, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5361714363098145, 'rewards/meter/std': 0.3404063284397125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9781786799430847, 'rewards/repeat_soft/std': 0.028630316257476807, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4972688853740692, 'rewards/total_composite/std': 0.09261411428451538, 'reward': 0.4972688853740692, 'reward_std': 0.09261411428451538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11025577038526535, 'sampling/sampling_logp_difference/max': 2.0586910247802734, 'sampling/importance_sampling_ratio/min': 0.1276209056377411, 'sampling/importance_sampling_ratio/mean': 1.0056194067001343, 'sampling/importance_sampling_ratio/max': 1.8145413398742676, 'entropy': 0.7311753779649734, 'clip_ratio/low_mean': 0.035480367951095104, 'clip_ratio/low_min': 0.035480367951095104, 'clip_ratio/high_mean': 0.0829462818801403, 'clip_ratio/high_max': 0.0829462818801403, 'clip_ratio/region_mean': 0.11842664983123541, 'reward_total_mean': 0.4972688853740692, 'reward_meter_mean': 0.5361714363098145, 'reward_meter_std': 0.3404063284397125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9781786799430847, 'reward_repeat_soft_std': 0.028630316257476807, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4972688853740692, 'reward_total_composite_std': 0.09261411428451538, 'epoch': 0.15}

 46%|████▋     | 1531/3300 [3:45:08<3:49:40,  7.79s/it]INFO 04-13 11:06:06 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1532/3300 [3:45:14<3:35:58,  7.33s/it]2026-04-13 11:06:12,622 | INFO | train_grpo_train | metrics_logged mode=train step=1532

                                                       
{'loss': -0.0447, 'grad_norm': 14.792496681213379, 'learning_rate': 5.360606060606061e-06, 'num_tokens': 2709704.0, 'completions/mean_length': 37.5, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.5, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9494877457618713, 'rewards/meter/std': 0.009832732379436493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9202165007591248, 'rewards/repeat_soft/std': 0.09014234691858292, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6734165549278259, 'rewards/total_composite/std': 0.146859809756279, 'reward': 0.6734165549278259, 'reward_std': 0.146859809756279, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10582664608955383, 'sampling/sampling_logp_difference/max': 1.6556751728057861, 'sampling/importance_sampling_ratio/min': 0.19096307456493378, 'sampling/importance_sampling_ratio/mean': 1.0053645372390747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5843740478157997, 'clip_ratio/low_mean': 0.0696247685700655, 'clip_ratio/low_min': 0.0696247685700655, 'clip_ratio/high_mean': 0.025600685738027096, 'clip_ratio/high_max': 0.025600685738027096, 'clip_ratio/region_mean': 0.0952254543080926, 'reward_total_mean': 0.6734165549278259, 'reward_meter_mean': 0.9494877457618713, 'reward_meter_std': 0.009832732379436493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9202165007591248, 'reward_repeat_soft_std': 0.09014234691858292, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6734165549278259, 'reward_total_composite_std': 0.146859809756279, 'epoch': 0.15}

 46%|████▋     | 1532/3300 [3:45:14<3:35:58,  7.33s/it]INFO 04-13 11:06:12 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1533/3300 [3:45:21<3:27:59,  7.06s/it]2026-04-13 11:06:19,063 | INFO | train_grpo_train | metrics_logged mode=train step=1533

                                                       
{'loss': 0.0066, 'grad_norm': 11.148467063903809, 'learning_rate': 5.357575757575758e-06, 'num_tokens': 2711471.0, 'completions/mean_length': 46.875, 'completions/min_length': 41.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9281042814254761, 'rewards/meter/std': 0.14827781915664673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9227937459945679, 'rewards/repeat_soft/std': 0.053246546536684036, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5917915105819702, 'rewards/total_composite/std': 0.03999544680118561, 'reward': 0.5917915105819702, 'reward_std': 0.0399954617023468, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14617781341075897, 'sampling/sampling_logp_difference/max': 4.057716369628906, 'sampling/importance_sampling_ratio/min': 0.0172884538769722, 'sampling/importance_sampling_ratio/mean': 0.996834933757782, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8533198460936546, 'clip_ratio/low_mean': 0.016735858283936977, 'clip_ratio/low_min': 0.016735858283936977, 'clip_ratio/high_mean': 0.07233824674040079, 'clip_ratio/high_max': 0.07233824674040079, 'clip_ratio/region_mean': 0.08907410502433777, 'reward_total_mean': 0.5917915105819702, 'reward_meter_mean': 0.9281042814254761, 'reward_meter_std': 0.14827781915664673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9227937459945679, 'reward_repeat_soft_std': 0.053246546536684036, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5917915105819702, 'reward_total_composite_std': 0.03999544680118561, 'epoch': 0.15}

 46%|████▋     | 1533/3300 [3:45:21<3:27:59,  7.06s/it]INFO 04-13 11:06:19 [block_pool.py:378] Successfully reset prefix cache

 46%|████▋     | 1534/3300 [3:45:27<3:23:33,  6.92s/it]2026-04-13 11:06:25,639 | INFO | train_grpo_train | metrics_logged mode=train step=1534

                                                       
{'loss': 0.0335, 'grad_norm': 11.622137069702148, 'learning_rate': 5.3545454545454546e-06, 'num_tokens': 2713067.0, 'completions/mean_length': 46.5, 'completions/min_length': 36.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7989070415496826, 'rewards/meter/std': 0.30848050117492676, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9807281494140625, 'rewards/repeat_soft/std': 0.02614726312458515, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5687021613121033, 'rewards/total_composite/std': 0.07431824505329132, 'reward': 0.5687021613121033, 'reward_std': 0.07431824505329132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15608374774456024, 'sampling/sampling_logp_difference/max': 1.595898151397705, 'sampling/importance_sampling_ratio/min': 0.20841604471206665, 'sampling/importance_sampling_ratio/mean': 0.9950181841850281, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.86784628033638, 'clip_ratio/low_mean': 0.024980408139526844, 'clip_ratio/low_min': 0.024980408139526844, 'clip_ratio/high_mean': 0.11696935445070267, 'clip_ratio/high_max': 0.11696935445070267, 'clip_ratio/region_mean': 0.1419497625902295, 'reward_total_mean': 0.5687021613121033, 'reward_meter_mean': 0.7989070415496826, 'reward_meter_std': 0.30848050117492676, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9807281494140625, 'reward_repeat_soft_std': 0.02614726312458515, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5687021613121033, 'reward_total_composite_std': 0.07431824505329132, 'epoch': 0.15}

 46%|████▋     | 1534/3300 [3:45:27<3:23:33,  6.92s/it]INFO 04-13 11:06:25 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1535/3300 [3:45:34<3:19:38,  6.79s/it]2026-04-13 11:06:32,121 | INFO | train_grpo_train | metrics_logged mode=train step=1535

                                                       
{'loss': 0.0702, 'grad_norm': 13.182768821716309, 'learning_rate': 5.351515151515152e-06, 'num_tokens': 2714639.0, 'completions/mean_length': 38.5, 'completions/min_length': 28.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7237865924835205, 'rewards/meter/std': 0.2100200653076172, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9861932992935181, 'rewards/repeat_soft/std': 0.038746047765016556, 'rewards/judge_quality/mean': 0.6200000047683716, 'rewards/judge_quality/std': 0.22677870094776154, 'rewards/total_composite/mean': 0.6452229022979736, 'rewards/total_composite/std': 0.15417109429836273, 'reward': 0.6452229022979736, 'reward_std': 0.15417106449604034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14780864119529724, 'sampling/sampling_logp_difference/max': 1.209925889968872, 'sampling/importance_sampling_ratio/min': 0.29821938276290894, 'sampling/importance_sampling_ratio/mean': 1.0319126844406128, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8952219486236572, 'clip_ratio/low_mean': 0.07706950325518847, 'clip_ratio/low_min': 0.07706950325518847, 'clip_ratio/high_mean': 0.037288233172148466, 'clip_ratio/high_max': 0.037288233172148466, 'clip_ratio/region_mean': 0.11435773642733693, 'reward_total_mean': 0.6452229022979736, 'reward_meter_mean': 0.7237865924835205, 'reward_meter_std': 0.2100200653076172, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9861932992935181, 'reward_repeat_soft_std': 0.038746047765016556, 'reward_judge_quality_mean': 0.6200000047683716, 'reward_judge_quality_std': 0.22677870094776154, 'reward_total_composite_mean': 0.6452229022979736, 'reward_total_composite_std': 0.15417109429836273, 'epoch': 0.15}

 47%|████▋     | 1535/3300 [3:45:34<3:19:38,  6.79s/it]INFO 04-13 11:06:32 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1536/3300 [3:45:41<3:19:19,  6.78s/it]2026-04-13 11:06:38,876 | INFO | train_grpo_train | metrics_logged mode=train step=1536

                                                       
{'loss': 0.0485, 'grad_norm': 8.649129867553711, 'learning_rate': 5.348484848484848e-06, 'num_tokens': 2716736.0, 'completions/mean_length': 90.125, 'completions/min_length': 81.0, 'completions/max_length': 102.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.125, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 102.0, 'rewards/meter/mean': 0.7510343790054321, 'rewards/meter/std': 0.2199547290802002, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9506515264511108, 'rewards/repeat_soft/std': 0.047333285212516785, 'rewards/judge_quality/mean': 0.5149999856948853, 'rewards/judge_quality/std': 0.1810288280248642, 'rewards/total_composite/mean': 0.5881261825561523, 'rewards/total_composite/std': 0.08761488646268845, 'reward': 0.5881261825561523, 'reward_std': 0.08761487901210785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13279852271080017, 'sampling/sampling_logp_difference/max': 2.762014865875244, 'sampling/importance_sampling_ratio/min': 0.0631643682718277, 'sampling/importance_sampling_ratio/mean': 1.008737325668335, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5726853087544441, 'clip_ratio/low_mean': 0.043127247132360935, 'clip_ratio/low_min': 0.043127247132360935, 'clip_ratio/high_mean': 0.06832017423585057, 'clip_ratio/high_max': 0.06832017423585057, 'clip_ratio/region_mean': 0.11144742136821151, 'reward_total_mean': 0.5881261825561523, 'reward_meter_mean': 0.7510343790054321, 'reward_meter_std': 0.2199547290802002, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9506515264511108, 'reward_repeat_soft_std': 0.047333285212516785, 'reward_judge_quality_mean': 0.5149999856948853, 'reward_judge_quality_std': 0.1810288280248642, 'reward_total_composite_mean': 0.5881261825561523, 'reward_total_composite_std': 0.08761488646268845, 'epoch': 0.15}

 47%|████▋     | 1536/3300 [3:45:41<3:19:19,  6.78s/it]INFO 04-13 11:06:39 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1537/3300 [3:45:48<3:24:32,  6.96s/it]2026-04-13 11:06:46,255 | INFO | train_grpo_train | metrics_logged mode=train step=1537

                                                       
{'loss': 0.0057, 'grad_norm': 6.371781826019287, 'learning_rate': 5.3454545454545455e-06, 'num_tokens': 2719387.0, 'completions/mean_length': 142.375, 'completions/min_length': 118.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.375, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.7671759128570557, 'rewards/meter/std': 0.2969895601272583, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8740166425704956, 'rewards/repeat_soft/std': 0.04647092521190643, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.20860078930854797, 'rewards/total_composite/mean': 0.5356192588806152, 'rewards/total_composite/std': 0.16517037153244019, 'reward': 0.5356192588806152, 'reward_std': 0.16517037153244019, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12230098247528076, 'sampling/sampling_logp_difference/max': 2.7857043743133545, 'sampling/importance_sampling_ratio/min': 0.06168562173843384, 'sampling/importance_sampling_ratio/mean': 1.016041874885559, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8719235211610794, 'clip_ratio/low_mean': 0.037878320552408695, 'clip_ratio/low_min': 0.037878320552408695, 'clip_ratio/high_mean': 0.07058449182659388, 'clip_ratio/high_max': 0.07058449182659388, 'clip_ratio/region_mean': 0.10846281237900257, 'reward_total_mean': 0.5356192588806152, 'reward_meter_mean': 0.7671759128570557, 'reward_meter_std': 0.2969895601272583, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8740166425704956, 'reward_repeat_soft_std': 0.04647092521190643, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.20860078930854797, 'reward_total_composite_mean': 0.5356192588806152, 'reward_total_composite_std': 0.16517037153244019, 'epoch': 0.15}

 47%|████▋     | 1537/3300 [3:45:48<3:24:32,  6.96s/it]INFO 04-13 11:06:46 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1538/3300 [3:45:55<3:22:22,  6.89s/it]2026-04-13 11:06:52,986 | INFO | train_grpo_train | metrics_logged mode=train step=1538

                                                       
{'loss': 0.036, 'grad_norm': 6.371514797210693, 'learning_rate': 5.342424242424244e-06, 'num_tokens': 2721639.0, 'completions/mean_length': 96.5, 'completions/min_length': 81.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.5, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9475618600845337, 'rewards/meter/std': 0.10585737228393555, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8539997935295105, 'rewards/repeat_soft/std': 0.04982974752783775, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6270068287849426, 'rewards/total_composite/std': 0.1197318509221077, 'reward': 0.6270068287849426, 'reward_std': 0.1197318509221077, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10939937829971313, 'sampling/sampling_logp_difference/max': 1.5694153308868408, 'sampling/importance_sampling_ratio/min': 0.20816683769226074, 'sampling/importance_sampling_ratio/mean': 1.0134103298187256, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6863407343626022, 'clip_ratio/low_mean': 0.08676877291873097, 'clip_ratio/low_min': 0.08676877291873097, 'clip_ratio/high_mean': 0.010174418799579144, 'clip_ratio/high_max': 0.010174418799579144, 'clip_ratio/region_mean': 0.09694319171831012, 'reward_total_mean': 0.6270068287849426, 'reward_meter_mean': 0.9475618600845337, 'reward_meter_std': 0.10585737228393555, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8539997935295105, 'reward_repeat_soft_std': 0.04982974752783775, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6270068287849426, 'reward_total_composite_std': 0.1197318509221077, 'epoch': 0.15}

 47%|████▋     | 1538/3300 [3:45:55<3:22:22,  6.89s/it]INFO 04-13 11:06:53 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1539/3300 [3:46:01<3:17:24,  6.73s/it]2026-04-13 11:06:59,333 | INFO | train_grpo_train | metrics_logged mode=train step=1539

                                                       
{'loss': 0.0437, 'grad_norm': 14.528196334838867, 'learning_rate': 5.33939393939394e-06, 'num_tokens': 2723288.0, 'completions/mean_length': 46.125, 'completions/min_length': 41.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8299557566642761, 'rewards/meter/std': 0.3053044378757477, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9574869275093079, 'rewards/repeat_soft/std': 0.04177146032452583, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.1527603268623352, 'rewards/total_composite/mean': 0.5731433629989624, 'rewards/total_composite/std': 0.12193405628204346, 'reward': 0.5731433629989624, 'reward_std': 0.12193406373262405, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1316116452217102, 'sampling/sampling_logp_difference/max': 1.407602310180664, 'sampling/importance_sampling_ratio/min': 0.24472936987876892, 'sampling/importance_sampling_ratio/mean': 1.0131338834762573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.930720329284668, 'clip_ratio/low_mean': 0.04696767684072256, 'clip_ratio/low_min': 0.04696767684072256, 'clip_ratio/high_mean': 0.06994033884257078, 'clip_ratio/high_max': 0.06994033884257078, 'clip_ratio/region_mean': 0.11690801568329334, 'reward_total_mean': 0.5731433629989624, 'reward_meter_mean': 0.8299557566642761, 'reward_meter_std': 0.3053044378757477, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9574869275093079, 'reward_repeat_soft_std': 0.04177146032452583, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.1527603268623352, 'reward_total_composite_mean': 0.5731433629989624, 'reward_total_composite_std': 0.12193405628204346, 'epoch': 0.15}

 47%|████▋     | 1539/3300 [3:46:01<3:17:24,  6.73s/it]INFO 04-13 11:06:59 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1540/3300 [3:46:09<3:28:37,  7.11s/it]2026-04-13 11:07:07,342 | INFO | train_grpo_train | metrics_logged mode=train step=1540

                                                       
{'loss': 0.0371, 'grad_norm': 8.206191062927246, 'learning_rate': 5.336363636363637e-06, 'num_tokens': 2725521.0, 'completions/mean_length': 110.125, 'completions/min_length': 77.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.125, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.8557904958724976, 'rewards/meter/std': 0.24603213369846344, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8825615048408508, 'rewards/repeat_soft/std': 0.1511286348104477, 'rewards/judge_quality/mean': 0.4362499713897705, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5596288442611694, 'rewards/total_composite/std': 0.07174958288669586, 'reward': 0.5596288442611694, 'reward_std': 0.07174960523843765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11753819137811661, 'sampling/sampling_logp_difference/max': 2.8877174854278564, 'sampling/importance_sampling_ratio/min': 0.055703211575746536, 'sampling/importance_sampling_ratio/mean': 1.0019290447235107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7809856422245502, 'clip_ratio/low_mean': 0.017171660671010613, 'clip_ratio/low_min': 0.017171660671010613, 'clip_ratio/high_mean': 0.10712305922061205, 'clip_ratio/high_max': 0.10712305922061205, 'clip_ratio/region_mean': 0.12429471989162266, 'reward_total_mean': 0.5596288442611694, 'reward_meter_mean': 0.8557904958724976, 'reward_meter_std': 0.24603213369846344, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8825615048408508, 'reward_repeat_soft_std': 0.1511286348104477, 'reward_judge_quality_mean': 0.4362499713897705, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5596288442611694, 'reward_total_composite_std': 0.07174958288669586, 'epoch': 0.15}

 47%|████▋     | 1540/3300 [3:46:09<3:28:37,  7.11s/it]INFO 04-13 11:07:07 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1541/3300 [3:46:20<4:06:26,  8.41s/it]2026-04-13 11:07:18,769 | INFO | train_grpo_train | metrics_logged mode=train step=1541

                                                       
{'loss': -0.1553, 'grad_norm': 3.9842867851257324, 'learning_rate': 5.333333333333334e-06, 'num_tokens': 2727819.0, 'completions/mean_length': 163.25, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.42857360839844, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.6968671679496765, 'rewards/meter/std': 0.3204905390739441, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9399411678314209, 'rewards/repeat_soft/std': 0.05086035653948784, 'rewards/judge_quality/mean': 0.6687500476837158, 'rewards/judge_quality/std': 0.29844537377357483, 'rewards/total_composite/mean': 0.5980728268623352, 'rewards/total_composite/std': 0.30698275566101074, 'reward': 0.5980728268623352, 'reward_std': 0.30698272585868835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10841288417577744, 'sampling/sampling_logp_difference/max': 1.840155839920044, 'sampling/importance_sampling_ratio/min': 0.1587926745414734, 'sampling/importance_sampling_ratio/mean': 1.000960350036621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5811756551265717, 'clip_ratio/low_mean': 0.03184491116553545, 'clip_ratio/low_min': 0.03184491116553545, 'clip_ratio/high_mean': 0.05904944706708193, 'clip_ratio/high_max': 0.05904944706708193, 'clip_ratio/region_mean': 0.09089435823261738, 'reward_total_mean': 0.5980728268623352, 'reward_meter_mean': 0.6968671679496765, 'reward_meter_std': 0.3204905390739441, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9399411678314209, 'reward_repeat_soft_std': 0.05086035653948784, 'reward_judge_quality_mean': 0.6687500476837158, 'reward_judge_quality_std': 0.29844537377357483, 'reward_total_composite_mean': 0.5980728268623352, 'reward_total_composite_std': 0.30698275566101074, 'epoch': 0.15}

 47%|████▋     | 1541/3300 [3:46:20<4:06:26,  8.41s/it]INFO 04-13 11:07:19 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1542/3300 [3:46:27<3:47:07,  7.75s/it]2026-04-13 11:07:24,988 | INFO | train_grpo_train | metrics_logged mode=train step=1542

                                                       
{'loss': 0.0187, 'grad_norm': 9.514369010925293, 'learning_rate': 5.330303030303031e-06, 'num_tokens': 2729353.0, 'completions/mean_length': 48.75, 'completions/min_length': 42.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.5638735890388489, 'rewards/meter/std': 0.3950325548648834, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9393631219863892, 'rewards/repeat_soft/std': 0.046571388840675354, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.6202842593193054, 'rewards/total_composite/std': 0.23236724734306335, 'reward': 0.6202842593193054, 'reward_std': 0.23236723244190216, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11959965527057648, 'sampling/sampling_logp_difference/max': 1.501955509185791, 'sampling/importance_sampling_ratio/min': 0.2226942628622055, 'sampling/importance_sampling_ratio/mean': 1.0103263854980469, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7916103079915047, 'clip_ratio/low_mean': 0.06776149990037084, 'clip_ratio/low_min': 0.06776149990037084, 'clip_ratio/high_mean': 0.047608437947928905, 'clip_ratio/high_max': 0.047608437947928905, 'clip_ratio/region_mean': 0.11536993784829974, 'reward_total_mean': 0.6202842593193054, 'reward_meter_mean': 0.5638735890388489, 'reward_meter_std': 0.3950325548648834, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9393631219863892, 'reward_repeat_soft_std': 0.046571388840675354, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.6202842593193054, 'reward_total_composite_std': 0.23236724734306335, 'epoch': 0.15}

 47%|████▋     | 1542/3300 [3:46:27<3:47:07,  7.75s/it]INFO 04-13 11:07:25 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1543/3300 [3:46:35<3:48:27,  7.80s/it]2026-04-13 11:07:32,909 | INFO | train_grpo_train | metrics_logged mode=train step=1543

                                                       
{'loss': -0.0023, 'grad_norm': 7.811085224151611, 'learning_rate': 5.327272727272727e-06, 'num_tokens': 2731860.0, 'completions/mean_length': 127.375, 'completions/min_length': 100.0, 'completions/max_length': 163.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.375, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 163.0, 'rewards/meter/mean': 0.937529444694519, 'rewards/meter/std': 0.07273654639720917, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8321306705474854, 'rewards/repeat_soft/std': 0.10501664131879807, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5340712666511536, 'rewards/total_composite/std': 0.03933795914053917, 'reward': 0.5340712666511536, 'reward_std': 0.03933796286582947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11424104124307632, 'sampling/sampling_logp_difference/max': 1.647552490234375, 'sampling/importance_sampling_ratio/min': 0.19252052903175354, 'sampling/importance_sampling_ratio/mean': 1.0067178010940552, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.728199802339077, 'clip_ratio/low_mean': 0.030051613226532936, 'clip_ratio/low_min': 0.030051613226532936, 'clip_ratio/high_mean': 0.06726025650277734, 'clip_ratio/high_max': 0.06726025650277734, 'clip_ratio/region_mean': 0.09731186972931027, 'reward_total_mean': 0.5340712666511536, 'reward_meter_mean': 0.937529444694519, 'reward_meter_std': 0.07273654639720917, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8321306705474854, 'reward_repeat_soft_std': 0.10501664131879807, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5340712666511536, 'reward_total_composite_std': 0.03933795914053917, 'epoch': 0.15}

 47%|████▋     | 1543/3300 [3:46:35<3:48:27,  7.80s/it]INFO 04-13 11:07:33 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1544/3300 [3:46:41<3:40:34,  7.54s/it]2026-04-13 11:07:39,827 | INFO | train_grpo_train | metrics_logged mode=train step=1544

                                                       
{'loss': -0.0121, 'grad_norm': 10.400113105773926, 'learning_rate': 5.324242424242425e-06, 'num_tokens': 2733700.0, 'completions/mean_length': 55.0, 'completions/min_length': 45.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.979071855545044, 'rewards/meter/std': 0.01928120292723179, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9689692854881287, 'rewards/repeat_soft/std': 0.03453250229358673, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.6551904082298279, 'rewards/total_composite/std': 0.11491691321134567, 'reward': 0.6551904082298279, 'reward_std': 0.11491690576076508, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1051703691482544, 'sampling/sampling_logp_difference/max': 0.999671459197998, 'sampling/importance_sampling_ratio/min': 0.43396344780921936, 'sampling/importance_sampling_ratio/mean': 1.0276440382003784, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8735250607132912, 'clip_ratio/low_mean': 0.07392397476360202, 'clip_ratio/low_min': 0.07392397476360202, 'clip_ratio/high_mean': 0.008196720853447914, 'clip_ratio/high_max': 0.008196720853447914, 'clip_ratio/region_mean': 0.08212069561704993, 'reward_total_mean': 0.6551904082298279, 'reward_meter_mean': 0.979071855545044, 'reward_meter_std': 0.01928120292723179, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9689692854881287, 'reward_repeat_soft_std': 0.03453250229358673, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.6551904082298279, 'reward_total_composite_std': 0.11491691321134567, 'epoch': 0.16}

 47%|████▋     | 1544/3300 [3:46:42<3:40:34,  7.54s/it]INFO 04-13 11:07:40 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1545/3300 [3:46:48<3:32:34,  7.27s/it]2026-04-13 11:07:46,467 | INFO | train_grpo_train | metrics_logged mode=train step=1545

                                                       
{'loss': 0.008, 'grad_norm': 14.160773277282715, 'learning_rate': 5.321212121212122e-06, 'num_tokens': 2735461.0, 'completions/mean_length': 51.125, 'completions/min_length': 44.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8905959129333496, 'rewards/meter/std': 0.2683301270008087, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9754196405410767, 'rewards/repeat_soft/std': 0.03474640101194382, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.601094126701355, 'rewards/total_composite/std': 0.04728719964623451, 'reward': 0.601094126701355, 'reward_std': 0.04728719964623451, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1266704797744751, 'sampling/sampling_logp_difference/max': 1.1628007888793945, 'sampling/importance_sampling_ratio/min': 0.31260940432548523, 'sampling/importance_sampling_ratio/mean': 1.0115855932235718, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.953790083527565, 'clip_ratio/low_mean': 0.017500000074505806, 'clip_ratio/low_min': 0.017500000074505806, 'clip_ratio/high_mean': 0.11213839892297983, 'clip_ratio/high_max': 0.11213839892297983, 'clip_ratio/region_mean': 0.12963839899748564, 'reward_total_mean': 0.601094126701355, 'reward_meter_mean': 0.8905959129333496, 'reward_meter_std': 0.2683301270008087, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9754196405410767, 'reward_repeat_soft_std': 0.03474640101194382, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.601094126701355, 'reward_total_composite_std': 0.04728719964623451, 'epoch': 0.16}

 47%|████▋     | 1545/3300 [3:46:48<3:32:34,  7.27s/it]INFO 04-13 11:07:46 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1546/3300 [3:46:59<4:07:06,  8.45s/it]2026-04-13 11:07:57,683 | INFO | train_grpo_train | metrics_logged mode=train step=1546

                                                       
{'loss': -0.1924, 'grad_norm': 1.910956859588623, 'learning_rate': 5.318181818181819e-06, 'num_tokens': 2737383.0, 'completions/mean_length': 209.25, 'completions/min_length': 94.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 108.33333587646484, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9042237997055054, 'rewards/meter/std': 0.13206376135349274, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9281355142593384, 'rewards/repeat_soft/std': 0.059728119522333145, 'rewards/judge_quality/mean': 0.45500001311302185, 'rewards/judge_quality/std': 0.2919393479824066, 'rewards/total_composite/mean': 0.5200342535972595, 'rewards/total_composite/std': 0.3319006562232971, 'reward': 0.5200342535972595, 'reward_std': 0.33190062642097473, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11505308002233505, 'sampling/sampling_logp_difference/max': 1.2959656715393066, 'sampling/importance_sampling_ratio/min': 0.27363350987434387, 'sampling/importance_sampling_ratio/mean': 1.0099626779556274, 'sampling/importance_sampling_ratio/max': 1.8460681438446045, 'entropy': 0.6150470823049545, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10359776020050049, 'clip_ratio/high_max': 0.10359776020050049, 'clip_ratio/region_mean': 0.10359776020050049, 'reward_total_mean': 0.5200342535972595, 'reward_meter_mean': 0.9042237997055054, 'reward_meter_std': 0.13206376135349274, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9281355142593384, 'reward_repeat_soft_std': 0.059728119522333145, 'reward_judge_quality_mean': 0.45500001311302185, 'reward_judge_quality_std': 0.2919393479824066, 'reward_total_composite_mean': 0.5200342535972595, 'reward_total_composite_std': 0.3319006562232971, 'epoch': 0.16}

 47%|████▋     | 1546/3300 [3:46:59<4:07:06,  8.45s/it]INFO 04-13 11:07:57 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1547/3300 [3:47:07<4:01:38,  8.27s/it]2026-04-13 11:08:05,528 | INFO | train_grpo_train | metrics_logged mode=train step=1547

                                                       
{'loss': 0.0779, 'grad_norm': 9.501943588256836, 'learning_rate': 5.3151515151515155e-06, 'num_tokens': 2739383.0, 'completions/mean_length': 75.0, 'completions/min_length': 64.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 75.0, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9446517825126648, 'rewards/meter/std': 0.02080017700791359, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8228054046630859, 'rewards/repeat_soft/std': 0.11904467642307281, 'rewards/judge_quality/mean': 0.47749999165534973, 'rewards/judge_quality/std': 0.2303258776664734, 'rewards/total_composite/mean': 0.6034237146377563, 'rewards/total_composite/std': 0.15840618312358856, 'reward': 0.6034237146377563, 'reward_std': 0.15840616822242737, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09276086091995239, 'sampling/sampling_logp_difference/max': 2.794069528579712, 'sampling/importance_sampling_ratio/min': 0.061171770095825195, 'sampling/importance_sampling_ratio/mean': 0.99891197681427, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36848473735153675, 'clip_ratio/low_mean': 0.06293998192995787, 'clip_ratio/low_min': 0.06293998192995787, 'clip_ratio/high_mean': 0.01844532322138548, 'clip_ratio/high_max': 0.01844532322138548, 'clip_ratio/region_mean': 0.08138530515134335, 'reward_total_mean': 0.6034237146377563, 'reward_meter_mean': 0.9446517825126648, 'reward_meter_std': 0.02080017700791359, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8228054046630859, 'reward_repeat_soft_std': 0.11904467642307281, 'reward_judge_quality_mean': 0.47749999165534973, 'reward_judge_quality_std': 0.2303258776664734, 'reward_total_composite_mean': 0.6034237146377563, 'reward_total_composite_std': 0.15840618312358856, 'epoch': 0.16}

 47%|████▋     | 1547/3300 [3:47:07<4:01:38,  8.27s/it]INFO 04-13 11:08:05 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1548/3300 [3:47:15<3:55:33,  8.07s/it]2026-04-13 11:08:13,123 | INFO | train_grpo_train | metrics_logged mode=train step=1548

                                                       
{'loss': -0.1042, 'grad_norm': 7.731827735900879, 'learning_rate': 5.312121212121213e-06, 'num_tokens': 2741814.0, 'completions/mean_length': 112.875, 'completions/min_length': 81.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.875, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9604154229164124, 'rewards/meter/std': 0.015933865681290627, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9142602682113647, 'rewards/repeat_soft/std': 0.03433630242943764, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5805824398994446, 'rewards/total_composite/std': 0.02216915413737297, 'reward': 0.5805824398994446, 'reward_std': 0.02216913178563118, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09675724804401398, 'sampling/sampling_logp_difference/max': 1.280354380607605, 'sampling/importance_sampling_ratio/min': 0.2779387831687927, 'sampling/importance_sampling_ratio/mean': 1.0128365755081177, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7276768982410431, 'clip_ratio/low_mean': 0.03882122738286853, 'clip_ratio/low_min': 0.03882122738286853, 'clip_ratio/high_mean': 0.06372740771621466, 'clip_ratio/high_max': 0.06372740771621466, 'clip_ratio/region_mean': 0.10254863509908319, 'reward_total_mean': 0.5805824398994446, 'reward_meter_mean': 0.9604154229164124, 'reward_meter_std': 0.015933865681290627, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9142602682113647, 'reward_repeat_soft_std': 0.03433630242943764, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5805824398994446, 'reward_total_composite_std': 0.02216915413737297, 'epoch': 0.16}

 47%|████▋     | 1548/3300 [3:47:15<3:55:33,  8.07s/it]INFO 04-13 11:08:13 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1549/3300 [3:47:21<3:38:06,  7.47s/it]2026-04-13 11:08:19,215 | INFO | train_grpo_train | metrics_logged mode=train step=1549

                                                       
{'loss': -0.0294, 'grad_norm': 15.26572322845459, 'learning_rate': 5.309090909090909e-06, 'num_tokens': 2743498.0, 'completions/mean_length': 39.5, 'completions/min_length': 35.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8108031749725342, 'rewards/meter/std': 0.30466896295547485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9867522716522217, 'rewards/repeat_soft/std': 0.010853597894310951, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.7405623197555542, 'rewards/total_composite/std': 0.2211495190858841, 'reward': 0.7405623197555542, 'reward_std': 0.2211495041847229, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1325124204158783, 'sampling/sampling_logp_difference/max': 1.735328197479248, 'sampling/importance_sampling_ratio/min': 0.17634230852127075, 'sampling/importance_sampling_ratio/mean': 0.9997811913490295, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.604987595230341, 'clip_ratio/low_mean': 0.053365386091172695, 'clip_ratio/low_min': 0.053365386091172695, 'clip_ratio/high_mean': 0.05798851721920073, 'clip_ratio/high_max': 0.05798851721920073, 'clip_ratio/region_mean': 0.11135390331037343, 'reward_total_mean': 0.7405623197555542, 'reward_meter_mean': 0.8108031749725342, 'reward_meter_std': 0.30466896295547485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9867522716522217, 'reward_repeat_soft_std': 0.010853597894310951, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.7405623197555542, 'reward_total_composite_std': 0.2211495190858841, 'epoch': 0.16}

 47%|████▋     | 1549/3300 [3:47:21<3:38:06,  7.47s/it]INFO 04-13 11:08:19 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1550/3300 [3:47:27<3:25:32,  7.05s/it]2026-04-13 11:08:25,268 | INFO | train_grpo_train | metrics_logged mode=train step=1550

                                                       
{'loss': 0.0702, 'grad_norm': 17.046966552734375, 'learning_rate': 5.306060606060606e-06, 'num_tokens': 2744953.0, 'completions/mean_length': 18.875, 'completions/min_length': 16.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9664780497550964, 'rewards/meter/std': 0.018068116158246994, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9465569853782654, 'rewards/repeat_soft/std': 0.026359153911471367, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.72638338804245, 'rewards/total_composite/std': 0.16048485040664673, 'reward': 0.72638338804245, 'reward_std': 0.16048482060432434, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11172085255384445, 'sampling/sampling_logp_difference/max': 1.3375508785247803, 'sampling/importance_sampling_ratio/min': 0.4185125529766083, 'sampling/importance_sampling_ratio/mean': 1.0193119049072266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6745321527123451, 'clip_ratio/low_mean': 0.07676024967804551, 'clip_ratio/low_min': 0.07676024967804551, 'clip_ratio/high_mean': 0.026610644534230232, 'clip_ratio/high_max': 0.026610644534230232, 'clip_ratio/region_mean': 0.10337089421227574, 'reward_total_mean': 0.72638338804245, 'reward_meter_mean': 0.9664780497550964, 'reward_meter_std': 0.018068116158246994, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9465569853782654, 'reward_repeat_soft_std': 0.026359153911471367, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.72638338804245, 'reward_total_composite_std': 0.16048485040664673, 'epoch': 0.16}

 47%|████▋     | 1550/3300 [3:47:27<3:25:32,  7.05s/it]INFO 04-13 11:08:25 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:15,  1.88s/it]

 30%|███       | 3/10 [00:12<00:33,  4.82s/it]

 40%|████      | 4/10 [00:16<00:26,  4.40s/it]

 50%|█████     | 5/10 [00:25<00:29,  5.90s/it]

 60%|██████    | 6/10 [00:29<00:21,  5.29s/it]

 70%|███████   | 7/10 [00:34<00:15,  5.15s/it]

 80%|████████  | 8/10 [00:38<00:09,  4.80s/it]

 90%|█████████ | 9/10 [00:47<00:06,  6.18s/it]

100%|██████████| 10/10 [00:50<00:00,  5.38s/it]2026-04-13 11:09:20,704 | INFO | train_grpo_train | metrics_logged mode=eval step=1550

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 55.3483, 'eval_samples_per_second': 1.445, 'eval_steps_per_second': 0.181, 'eval_num_tokens': 2744953.0, 'eval_completions/mean_length': 97.5375, 'eval_completions/min_length': 36.7, 'eval_completions/max_length': 238.9, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 75.92916793823242, 'eval_completions/min_terminated_length': 36.7, 'eval_completions/max_terminated_length': 123.3, 'eval_rewards/meter/mean': 0.7568637251853942, 'eval_rewards/meter/std': 0.29639622159302237, 'eval_rewards/count_adherence/mean': 0.9468749761581421, 'eval_rewards/count_adherence/std': 0.09409432746469974, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.15235702097415924, 'eval_rewards/repeat_soft/mean': 0.923405921459198, 'eval_rewards/repeat_soft/std': 0.06663752272725106, 'eval_rewards/judge_quality/mean': 0.5013749897480011, 'eval_rewards/judge_quality/std': 0.19865378523245453, 'eval_rewards/total_composite/mean': 0.5569140017032623, 'eval_rewards/total_composite/std': 0.19040723517537117, 'eval_reward': 0.5569140017032623, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06511233486235142, 'eval_sampling/sampling_logp_difference/max': 1.0043030738830567, 'eval_sampling/importance_sampling_ratio/min': 0.37784353643655777, 'eval_sampling/importance_sampling_ratio/mean': 1.0135634899139405, 'eval_sampling/importance_sampling_ratio/max': 1.3928049325942993, 'eval_entropy': 0.7222438454627991, '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': 0.5569140017032623, 'eval_reward_meter_mean': 0.7568637251853942, 'eval_reward_meter_std': 0.29639622159302237, 'eval_reward_count_adherence_mean': 0.9468749761581421, 'eval_reward_count_adherence_std': 0.09409432746469974, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.15235702097415924, 'eval_reward_repeat_soft_mean': 0.923405921459198, 'eval_reward_repeat_soft_std': 0.06663752272725106, 'eval_reward_judge_quality_mean': 0.5013749897480011, 'eval_reward_judge_quality_std': 0.19865378523245453, 'eval_reward_total_composite_mean': 0.5569140017032623, 'eval_reward_total_composite_std': 0.19040723517537117, 'epoch': 0.16}

 47%|████▋     | 1550/3300 [3:48:22<3:25:32,  7.05s/it]

100%|██████████| 10/10 [00:50<00:00,  5.38s/it]

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

 47%|████▋     | 1551/3300 [3:48:33<12:00:42, 24.72s/it]2026-04-13 11:09:31,235 | INFO | train_grpo_train | metrics_logged mode=train step=1551

                                                        
{'loss': -0.066, 'grad_norm': 6.368042469024658, 'learning_rate': 5.303030303030303e-06, 'num_tokens': 2747835.0, 'completions/mean_length': 152.25, 'completions/min_length': 118.0, 'completions/max_length': 183.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 152.25, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 183.0, 'rewards/meter/mean': 0.9868429899215698, 'rewards/meter/std': 0.008497558534145355, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9355205297470093, 'rewards/repeat_soft/std': 0.07157598435878754, 'rewards/judge_quality/mean': 0.30124998092651367, 'rewards/judge_quality/std': 0.10398317128419876, 'rewards/total_composite/mean': 0.4304884374141693, 'rewards/total_composite/std': 0.18392300605773926, 'reward': 0.4304884374141693, 'reward_std': 0.18392299115657806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1302221268415451, 'sampling/sampling_logp_difference/max': 1.7571523189544678, 'sampling/importance_sampling_ratio/min': 0.1725354939699173, 'sampling/importance_sampling_ratio/mean': 1.0175302028656006, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1224376186728477, 'clip_ratio/low_mean': 0.026544908992946148, 'clip_ratio/low_min': 0.026544908992946148, 'clip_ratio/high_mean': 0.099517279304564, 'clip_ratio/high_max': 0.099517279304564, 'clip_ratio/region_mean': 0.12606218829751015, 'reward_total_mean': 0.4304884374141693, 'reward_meter_mean': 0.9868429899215698, 'reward_meter_std': 0.008497558534145355, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9355205297470093, 'reward_repeat_soft_std': 0.07157598435878754, 'reward_judge_quality_mean': 0.30124998092651367, 'reward_judge_quality_std': 0.10398317128419876, 'reward_total_composite_mean': 0.4304884374141693, 'reward_total_composite_std': 0.18392300605773926, 'epoch': 0.16}

 47%|████▋     | 1551/3300 [3:48:33<12:00:42, 24.72s/it]INFO 04-13 11:09:31 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1552/3300 [3:48:39<9:17:38, 19.14s/it] 2026-04-13 11:09:37,352 | INFO | train_grpo_train | metrics_logged mode=train step=1552

                                                       
{'loss': -0.0647, 'grad_norm': 15.391485214233398, 'learning_rate': 5.300000000000001e-06, 'num_tokens': 2749242.0, 'completions/mean_length': 22.875, 'completions/min_length': 15.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.971284031867981, 'rewards/meter/std': 0.01080433838069439, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9125404357910156, 'rewards/repeat_soft/std': 0.03231460601091385, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7617834806442261, 'rewards/total_composite/std': 0.16494220495224, 'reward': 0.7617834806442261, 'reward_std': 0.16494223475456238, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1157231256365776, 'sampling/sampling_logp_difference/max': 1.5222516059875488, 'sampling/importance_sampling_ratio/min': 0.21821999549865723, 'sampling/importance_sampling_ratio/mean': 0.9835056662559509, 'sampling/importance_sampling_ratio/max': 1.7103513479232788, 'entropy': 0.5968308188021183, 'clip_ratio/low_mean': 0.06343860179185867, 'clip_ratio/low_min': 0.06343860179185867, 'clip_ratio/high_mean': 0.049401901196688414, 'clip_ratio/high_max': 0.049401901196688414, 'clip_ratio/region_mean': 0.11284050298854709, 'reward_total_mean': 0.7617834806442261, 'reward_meter_mean': 0.971284031867981, 'reward_meter_std': 0.01080433838069439, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9125404357910156, 'reward_repeat_soft_std': 0.03231460601091385, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7617834806442261, 'reward_total_composite_std': 0.16494220495224, 'epoch': 0.16}

 47%|████▋     | 1552/3300 [3:48:39<9:17:38, 19.14s/it]INFO 04-13 11:09:37 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1553/3300 [3:48:46<7:29:41, 15.44s/it]2026-04-13 11:09:44,177 | INFO | train_grpo_train | metrics_logged mode=train step=1553

                                                       
{'loss': 0.0853, 'grad_norm': 8.848579406738281, 'learning_rate': 5.296969696969697e-06, 'num_tokens': 2750795.0, 'completions/mean_length': 44.125, 'completions/min_length': 34.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8539493083953857, 'rewards/meter/std': 0.15820306539535522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9636582136154175, 'rewards/repeat_soft/std': 0.04516100138425827, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.6176408529281616, 'rewards/total_composite/std': 0.12416830658912659, 'reward': 0.6176408529281616, 'reward_std': 0.12416831403970718, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12705570459365845, 'sampling/sampling_logp_difference/max': 1.7956466674804688, 'sampling/importance_sampling_ratio/min': 0.3185770809650421, 'sampling/importance_sampling_ratio/mean': 1.0181162357330322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7664505019783974, 'clip_ratio/low_mean': 0.08077190211042762, 'clip_ratio/low_min': 0.08077190211042762, 'clip_ratio/high_mean': 0.04592503793537617, 'clip_ratio/high_max': 0.04592503793537617, 'clip_ratio/region_mean': 0.12669694004580379, 'reward_total_mean': 0.6176408529281616, 'reward_meter_mean': 0.8539493083953857, 'reward_meter_std': 0.15820306539535522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9636582136154175, 'reward_repeat_soft_std': 0.04516100138425827, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.6176408529281616, 'reward_total_composite_std': 0.12416830658912659, 'epoch': 0.16}

 47%|████▋     | 1553/3300 [3:48:46<7:29:41, 15.44s/it]INFO 04-13 11:09:44 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1554/3300 [3:48:52<6:06:35, 12.60s/it]2026-04-13 11:09:50,124 | INFO | train_grpo_train | metrics_logged mode=train step=1554

                                                       
{'loss': 0.0926, 'grad_norm': 18.151506423950195, 'learning_rate': 5.293939393939395e-06, 'num_tokens': 2752138.0, 'completions/mean_length': 23.875, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.893831729888916, 'rewards/meter/std': 0.23648913204669952, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9596153497695923, 'rewards/repeat_soft/std': 0.008158913813531399, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.7982645630836487, 'rewards/total_composite/std': 0.17805226147174835, 'reward': 0.7982645630836487, 'reward_std': 0.17805226147174835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14744354784488678, 'sampling/sampling_logp_difference/max': 1.1399688720703125, 'sampling/importance_sampling_ratio/min': 0.31982898712158203, 'sampling/importance_sampling_ratio/mean': 1.0191422700881958, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8114384114742279, 'clip_ratio/low_mean': 0.029230769723653793, 'clip_ratio/low_min': 0.029230769723653793, 'clip_ratio/high_mean': 0.0935286795720458, 'clip_ratio/high_max': 0.0935286795720458, 'clip_ratio/region_mean': 0.1227594492956996, 'reward_total_mean': 0.7982645630836487, 'reward_meter_mean': 0.893831729888916, 'reward_meter_std': 0.23648913204669952, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9596153497695923, 'reward_repeat_soft_std': 0.008158913813531399, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.7982645630836487, 'reward_total_composite_std': 0.17805226147174835, 'epoch': 0.16}

 47%|████▋     | 1554/3300 [3:48:52<6:06:35, 12.60s/it]INFO 04-13 11:09:50 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1555/3300 [3:48:59<5:17:04, 10.90s/it]2026-04-13 11:09:57,074 | INFO | train_grpo_train | metrics_logged mode=train step=1555

                                                       
{'loss': 0.0074, 'grad_norm': 10.675823211669922, 'learning_rate': 5.290909090909091e-06, 'num_tokens': 2754057.0, 'completions/mean_length': 73.875, 'completions/min_length': 66.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.875, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.9280035495758057, 'rewards/meter/std': 0.035113625228405, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8458590507507324, 'rewards/repeat_soft/std': 0.03096371702849865, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5825026035308838, 'rewards/total_composite/std': 0.013317407108843327, 'reward': 0.5825026035308838, 'reward_std': 0.01331739965826273, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0960083156824112, 'sampling/sampling_logp_difference/max': 1.5728020668029785, 'sampling/importance_sampling_ratio/min': 0.20746304094791412, 'sampling/importance_sampling_ratio/mean': 0.9991890788078308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4552772305905819, 'clip_ratio/low_mean': 0.033438229002058506, 'clip_ratio/low_min': 0.033438229002058506, 'clip_ratio/high_mean': 0.0584720391780138, 'clip_ratio/high_max': 0.0584720391780138, 'clip_ratio/region_mean': 0.09191026818007231, 'reward_total_mean': 0.5825026035308838, 'reward_meter_mean': 0.9280035495758057, 'reward_meter_std': 0.035113625228405, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8458590507507324, 'reward_repeat_soft_std': 0.03096371702849865, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5825026035308838, 'reward_total_composite_std': 0.013317407108843327, 'epoch': 0.16}

 47%|████▋     | 1555/3300 [3:48:59<5:17:04, 10.90s/it]INFO 04-13 11:09:57 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1556/3300 [3:49:06<4:47:47,  9.90s/it]2026-04-13 11:10:04,639 | INFO | train_grpo_train | metrics_logged mode=train step=1556

                                                       
{'loss': 0.0166, 'grad_norm': 8.030050277709961, 'learning_rate': 5.287878787878788e-06, 'num_tokens': 2756457.0, 'completions/mean_length': 94.0, 'completions/min_length': 76.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.0, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.8477905988693237, 'rewards/meter/std': 0.2628907859325409, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9841978549957275, 'rewards/repeat_soft/std': 0.015399984084069729, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5814843773841858, 'rewards/total_composite/std': 0.07323091477155685, 'reward': 0.5814843773841858, 'reward_std': 0.07323091477155685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.148122638463974, 'sampling/sampling_logp_difference/max': 1.7706518173217773, 'sampling/importance_sampling_ratio/min': 0.26092320680618286, 'sampling/importance_sampling_ratio/mean': 1.0171198844909668, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1991672441363335, 'clip_ratio/low_mean': 0.03827519528567791, 'clip_ratio/low_min': 0.03827519528567791, 'clip_ratio/high_mean': 0.1132746646180749, 'clip_ratio/high_max': 0.1132746646180749, 'clip_ratio/region_mean': 0.1515498599037528, 'reward_total_mean': 0.5814843773841858, 'reward_meter_mean': 0.8477905988693237, 'reward_meter_std': 0.2628907859325409, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9841978549957275, 'reward_repeat_soft_std': 0.015399984084069729, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5814843773841858, 'reward_total_composite_std': 0.07323091477155685, 'epoch': 0.16}

 47%|████▋     | 1556/3300 [3:49:06<4:47:47,  9.90s/it]INFO 04-13 11:10:04 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1557/3300 [3:49:18<5:03:07, 10.43s/it]2026-04-13 11:10:16,369 | INFO | train_grpo_train | metrics_logged mode=train step=1557

                                                       
{'loss': -0.1633, 'grad_norm': 2.751214027404785, 'learning_rate': 5.284848484848485e-06, 'num_tokens': 2758656.0, 'completions/mean_length': 166.875, 'completions/min_length': 97.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 117.5714340209961, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.8048547506332397, 'rewards/meter/std': 0.33105307817459106, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9741945862770081, 'rewards/repeat_soft/std': 0.017214907333254814, 'rewards/judge_quality/mean': 0.39375001192092896, 'rewards/judge_quality/std': 0.25048166513442993, 'rewards/total_composite/mean': 0.5263358354568481, 'rewards/total_composite/std': 0.2428174614906311, 'reward': 0.5263358354568481, 'reward_std': 0.2428174614906311, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13197213411331177, 'sampling/sampling_logp_difference/max': 1.4400482177734375, 'sampling/importance_sampling_ratio/min': 0.23691634833812714, 'sampling/importance_sampling_ratio/mean': 1.0212165117263794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8606486544013023, 'clip_ratio/low_mean': 0.03507152758538723, 'clip_ratio/low_min': 0.03507152758538723, 'clip_ratio/high_mean': 0.0820219162851572, 'clip_ratio/high_max': 0.0820219162851572, 'clip_ratio/region_mean': 0.11709344387054443, 'reward_total_mean': 0.5263358354568481, 'reward_meter_mean': 0.8048547506332397, 'reward_meter_std': 0.33105307817459106, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9741945862770081, 'reward_repeat_soft_std': 0.017214907333254814, 'reward_judge_quality_mean': 0.39375001192092896, 'reward_judge_quality_std': 0.25048166513442993, 'reward_total_composite_mean': 0.5263358354568481, 'reward_total_composite_std': 0.2428174614906311, 'epoch': 0.16}

 47%|████▋     | 1557/3300 [3:49:18<5:03:07, 10.43s/it]INFO 04-13 11:10:16 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1558/3300 [3:49:29<5:10:57, 10.71s/it]2026-04-13 11:10:27,675 | INFO | train_grpo_train | metrics_logged mode=train step=1558

                                                       
{'loss': -0.1534, 'grad_norm': 2.0736265182495117, 'learning_rate': 5.281818181818183e-06, 'num_tokens': 2760392.0, 'completions/mean_length': 116.0, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.42857360839844, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.758922815322876, 'rewards/meter/std': 0.28375113010406494, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9984496831893921, 'rewards/repeat_soft/std': 0.001939329202286899, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.13767844438552856, 'rewards/total_composite/mean': 0.5132649540901184, 'rewards/total_composite/std': 0.213321253657341, 'reward': 0.5132649540901184, 'reward_std': 0.213321253657341, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15871377289295197, 'sampling/sampling_logp_difference/max': 1.6210627555847168, 'sampling/importance_sampling_ratio/min': 0.197688490152359, 'sampling/importance_sampling_ratio/mean': 1.0191385746002197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.068819299340248, 'clip_ratio/low_mean': 0.01515151560306549, 'clip_ratio/low_min': 0.01515151560306549, 'clip_ratio/high_mean': 0.09964854456484318, 'clip_ratio/high_max': 0.09964854456484318, 'clip_ratio/region_mean': 0.11480006016790867, 'reward_total_mean': 0.5132649540901184, 'reward_meter_mean': 0.758922815322876, 'reward_meter_std': 0.28375113010406494, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9984496831893921, 'reward_repeat_soft_std': 0.001939329202286899, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.13767844438552856, 'reward_total_composite_mean': 0.5132649540901184, 'reward_total_composite_std': 0.213321253657341, 'epoch': 0.16}

 47%|████▋     | 1558/3300 [3:49:29<5:10:57, 10.71s/it]INFO 04-13 11:10:27 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1559/3300 [3:49:36<4:32:53,  9.40s/it]2026-04-13 11:10:34,024 | INFO | train_grpo_train | metrics_logged mode=train step=1559

                                                       
{'loss': -0.0118, 'grad_norm': 14.156771659851074, 'learning_rate': 5.278787878787879e-06, 'num_tokens': 2761903.0, 'completions/mean_length': 29.875, 'completions/min_length': 26.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9599217176437378, 'rewards/meter/std': 0.06778952479362488, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.958297610282898, 'rewards/repeat_soft/std': 0.011176233179867268, 'rewards/judge_quality/mean': 0.5637500286102295, 'rewards/judge_quality/std': 0.22012579441070557, 'rewards/total_composite/mean': 0.6975878477096558, 'rewards/total_composite/std': 0.14779505133628845, 'reward': 0.6975878477096558, 'reward_std': 0.14779506623744965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09784701466560364, 'sampling/sampling_logp_difference/max': 0.8822650909423828, 'sampling/importance_sampling_ratio/min': 0.413844496011734, 'sampling/importance_sampling_ratio/mean': 1.0062772035598755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7347541488707066, 'clip_ratio/low_mean': 0.059528898447752, 'clip_ratio/low_min': 0.059528898447752, 'clip_ratio/high_mean': 0.011742424685508013, 'clip_ratio/high_max': 0.011742424685508013, 'clip_ratio/region_mean': 0.07127132313326001, 'reward_total_mean': 0.6975878477096558, 'reward_meter_mean': 0.9599217176437378, 'reward_meter_std': 0.06778952479362488, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.958297610282898, 'reward_repeat_soft_std': 0.011176233179867268, 'reward_judge_quality_mean': 0.5637500286102295, 'reward_judge_quality_std': 0.22012579441070557, 'reward_total_composite_mean': 0.6975878477096558, 'reward_total_composite_std': 0.14779505133628845, 'epoch': 0.16}

 47%|████▋     | 1559/3300 [3:49:36<4:32:53,  9.40s/it]INFO 04-13 11:10:34 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1560/3300 [3:49:47<4:52:13, 10.08s/it]2026-04-13 11:10:45,672 | INFO | train_grpo_train | metrics_logged mode=train step=1560

                                                       
{'loss': -0.1116, 'grad_norm': 1.5735626220703125, 'learning_rate': 5.2757575757575764e-06, 'num_tokens': 2763398.0, 'completions/mean_length': 163.875, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 47.833335876464844, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.812476396560669, 'rewards/meter/std': 0.35773009061813354, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9805666208267212, 'rewards/repeat_soft/std': 0.021628832444548607, 'rewards/judge_quality/mean': 0.33125001192092896, 'rewards/judge_quality/std': 0.1738995909690857, 'rewards/total_composite/mean': 0.4497215151786804, 'rewards/total_composite/std': 0.28081437945365906, 'reward': 0.4497215151786804, 'reward_std': 0.28081434965133667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13447198271751404, 'sampling/sampling_logp_difference/max': 2.093869686126709, 'sampling/importance_sampling_ratio/min': 0.12320942431688309, 'sampling/importance_sampling_ratio/mean': 1.0144320726394653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6690042838454247, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.079138049390167, 'clip_ratio/high_max': 0.079138049390167, 'clip_ratio/region_mean': 0.079138049390167, 'reward_total_mean': 0.4497215151786804, 'reward_meter_mean': 0.812476396560669, 'reward_meter_std': 0.35773009061813354, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9805666208267212, 'reward_repeat_soft_std': 0.021628832444548607, 'reward_judge_quality_mean': 0.33125001192092896, 'reward_judge_quality_std': 0.1738995909690857, 'reward_total_composite_mean': 0.4497215151786804, 'reward_total_composite_std': 0.28081437945365906, 'epoch': 0.16}

 47%|████▋     | 1560/3300 [3:49:47<4:52:13, 10.08s/it]INFO 04-13 11:10:45 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1561/3300 [3:49:54<4:20:07,  8.97s/it]2026-04-13 11:10:52,079 | INFO | train_grpo_train | metrics_logged mode=train step=1561

                                                       
{'loss': 0.0944, 'grad_norm': 12.714040756225586, 'learning_rate': 5.272727272727273e-06, 'num_tokens': 2765074.0, 'completions/mean_length': 49.5, 'completions/min_length': 41.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.6791974306106567, 'rewards/meter/std': 0.40748533606529236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9711135625839233, 'rewards/repeat_soft/std': 0.032192979007959366, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.5897715091705322, 'rewards/total_composite/std': 0.15912891924381256, 'reward': 0.5897715091705322, 'reward_std': 0.15912891924381256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1287555992603302, 'sampling/sampling_logp_difference/max': 1.287729263305664, 'sampling/importance_sampling_ratio/min': 0.2758965492248535, 'sampling/importance_sampling_ratio/mean': 1.0143460035324097, 'sampling/importance_sampling_ratio/max': 1.9852994680404663, 'entropy': 1.0453985407948494, 'clip_ratio/low_mean': 0.031547619961202145, 'clip_ratio/low_min': 0.031547619961202145, 'clip_ratio/high_mean': 0.07998319528996944, 'clip_ratio/high_max': 0.07998319528996944, 'clip_ratio/region_mean': 0.11153081525117159, 'reward_total_mean': 0.5897715091705322, 'reward_meter_mean': 0.6791974306106567, 'reward_meter_std': 0.40748533606529236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9711135625839233, 'reward_repeat_soft_std': 0.032192979007959366, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5897715091705322, 'reward_total_composite_std': 0.15912891924381256, 'epoch': 0.16}

 47%|████▋     | 1561/3300 [3:49:54<4:20:07,  8.97s/it]INFO 04-13 11:10:52 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1562/3300 [3:50:01<4:04:05,  8.43s/it]2026-04-13 11:10:59,241 | INFO | train_grpo_train | metrics_logged mode=train step=1562

                                                       
{'loss': 0.0353, 'grad_norm': 9.678753852844238, 'learning_rate': 5.26969696969697e-06, 'num_tokens': 2767338.0, 'completions/mean_length': 92.0, 'completions/min_length': 82.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 92.0, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.845755934715271, 'rewards/meter/std': 0.2165696918964386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8743439316749573, 'rewards/repeat_soft/std': 0.13308818638324738, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5560852289199829, 'rewards/total_composite/std': 0.06631311029195786, 'reward': 0.5560852289199829, 'reward_std': 0.06631311029195786, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13134834170341492, 'sampling/sampling_logp_difference/max': 1.8900089263916016, 'sampling/importance_sampling_ratio/min': 0.1510704606771469, 'sampling/importance_sampling_ratio/mean': 1.0071284770965576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.544047586619854, 'clip_ratio/low_mean': 0.051638985984027386, 'clip_ratio/low_min': 0.051638985984027386, 'clip_ratio/high_mean': 0.07158447615802288, 'clip_ratio/high_max': 0.07158447615802288, 'clip_ratio/region_mean': 0.12322346214205027, 'reward_total_mean': 0.5560852289199829, 'reward_meter_mean': 0.845755934715271, 'reward_meter_std': 0.2165696918964386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8743439316749573, 'reward_repeat_soft_std': 0.13308818638324738, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5560852289199829, 'reward_total_composite_std': 0.06631311029195786, 'epoch': 0.16}

 47%|████▋     | 1562/3300 [3:50:01<4:04:05,  8.43s/it]INFO 04-13 11:10:59 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1563/3300 [3:50:08<3:54:04,  8.09s/it]2026-04-13 11:11:06,538 | INFO | train_grpo_train | metrics_logged mode=train step=1563

                                                       
{'loss': -0.0562, 'grad_norm': 6.637635231018066, 'learning_rate': 5.2666666666666665e-06, 'num_tokens': 2769672.0, 'completions/mean_length': 108.75, 'completions/min_length': 73.0, 'completions/max_length': 129.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.9818673133850098, 'rewards/meter/std': 0.013576094061136246, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8238779306411743, 'rewards/repeat_soft/std': 0.09231465309858322, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.1866959184408188, 'rewards/total_composite/mean': 0.6786043047904968, 'rewards/total_composite/std': 0.1297779232263565, 'reward': 0.6786043047904968, 'reward_std': 0.1297779232263565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08684387803077698, 'sampling/sampling_logp_difference/max': 1.6086792945861816, 'sampling/importance_sampling_ratio/min': 0.20015178620815277, 'sampling/importance_sampling_ratio/mean': 1.0073845386505127, 'sampling/importance_sampling_ratio/max': 1.6928097009658813, 'entropy': 0.6217380240559578, 'clip_ratio/low_mean': 0.03960631275549531, 'clip_ratio/low_min': 0.03960631275549531, 'clip_ratio/high_mean': 0.02663089195266366, 'clip_ratio/high_max': 0.02663089195266366, 'clip_ratio/region_mean': 0.06623720470815897, 'reward_total_mean': 0.6786043047904968, 'reward_meter_mean': 0.9818673133850098, 'reward_meter_std': 0.013576094061136246, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8238779306411743, 'reward_repeat_soft_std': 0.09231465309858322, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.1866959184408188, 'reward_total_composite_mean': 0.6786043047904968, 'reward_total_composite_std': 0.1297779232263565, 'epoch': 0.16}

 47%|████▋     | 1563/3300 [3:50:08<3:54:04,  8.09s/it]INFO 04-13 11:11:06 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1564/3300 [3:50:14<3:36:19,  7.48s/it]2026-04-13 11:11:12,584 | INFO | train_grpo_train | metrics_logged mode=train step=1564

                                                       
{'loss': 0.0165, 'grad_norm': 17.923070907592773, 'learning_rate': 5.263636363636364e-06, 'num_tokens': 2771042.0, 'completions/mean_length': 23.25, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9900604486465454, 'rewards/meter/std': 0.005437163170427084, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6170778274536133, 'rewards/total_composite/std': 0.007128218654543161, 'reward': 0.6170778274536133, 'reward_std': 0.007128225173801184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1269725263118744, 'sampling/sampling_logp_difference/max': 1.31215238571167, 'sampling/importance_sampling_ratio/min': 0.2692399322986603, 'sampling/importance_sampling_ratio/mean': 1.0162959098815918, 'sampling/importance_sampling_ratio/max': 1.8745906352996826, 'entropy': 0.8843695893883705, 'clip_ratio/low_mean': 0.07846467662602663, 'clip_ratio/low_min': 0.07846467662602663, 'clip_ratio/high_mean': 0.0, 'clip_ratio/high_max': 0.0, 'clip_ratio/region_mean': 0.07846467662602663, 'reward_total_mean': 0.6170778274536133, 'reward_meter_mean': 0.9900604486465454, 'reward_meter_std': 0.005437163170427084, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6170778274536133, 'reward_total_composite_std': 0.007128218654543161, 'epoch': 0.16}

 47%|████▋     | 1564/3300 [3:50:14<3:36:19,  7.48s/it]INFO 04-13 11:11:12 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1565/3300 [3:50:21<3:31:47,  7.32s/it]2026-04-13 11:11:19,551 | INFO | train_grpo_train | metrics_logged mode=train step=1565

                                                       
{'loss': 0.0515, 'grad_norm': 11.058951377868652, 'learning_rate': 5.26060606060606e-06, 'num_tokens': 2772582.0, 'completions/mean_length': 42.5, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.5421852469444275, 'rewards/meter/std': 0.4336164891719818, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9924154281616211, 'rewards/repeat_soft/std': 0.012711554765701294, 'rewards/judge_quality/mean': 0.5600000023841858, 'rewards/judge_quality/std': 0.22258226573467255, 'rewards/total_composite/mean': 0.5454724431037903, 'rewards/total_composite/std': 0.19200962781906128, 'reward': 0.5454724431037903, 'reward_std': 0.19200964272022247, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13946400582790375, 'sampling/sampling_logp_difference/max': 2.6357040405273438, 'sampling/importance_sampling_ratio/min': 0.07166849821805954, 'sampling/importance_sampling_ratio/mean': 0.9839051365852356, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7409085147082806, 'clip_ratio/low_mean': 0.056764266453683376, 'clip_ratio/low_min': 0.056764266453683376, 'clip_ratio/high_mean': 0.08312970027327538, 'clip_ratio/high_max': 0.08312970027327538, 'clip_ratio/region_mean': 0.13989396672695875, 'reward_total_mean': 0.5454724431037903, 'reward_meter_mean': 0.5421852469444275, 'reward_meter_std': 0.4336164891719818, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9924154281616211, 'reward_repeat_soft_std': 0.012711554765701294, 'reward_judge_quality_mean': 0.5600000023841858, 'reward_judge_quality_std': 0.22258226573467255, 'reward_total_composite_mean': 0.5454724431037903, 'reward_total_composite_std': 0.19200962781906128, 'epoch': 0.16}

 47%|████▋     | 1565/3300 [3:50:21<3:31:47,  7.32s/it]INFO 04-13 11:11:19 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1566/3300 [3:50:28<3:23:32,  7.04s/it]2026-04-13 11:11:25,937 | INFO | train_grpo_train | metrics_logged mode=train step=1566

                                                       
{'loss': 0.0025, 'grad_norm': 13.1821870803833, 'learning_rate': 5.257575757575758e-06, 'num_tokens': 2774094.0, 'completions/mean_length': 36.0, 'completions/min_length': 32.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.12457047402858734, 'rewards/meter/std': 0.33992356061935425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8819153308868408, 'rewards/repeat_soft/std': 0.07343287765979767, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3662950396537781, 'rewards/total_composite/std': 0.0971856340765953, 'reward': 0.3662950396537781, 'reward_std': 0.0971856340765953, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11318037658929825, 'sampling/sampling_logp_difference/max': 1.9890861511230469, 'sampling/importance_sampling_ratio/min': 0.13682040572166443, 'sampling/importance_sampling_ratio/mean': 0.9863535761833191, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3793261554092169, 'clip_ratio/low_mean': 0.0840162024833262, 'clip_ratio/low_min': 0.0840162024833262, 'clip_ratio/high_mean': 0.01315789483487606, 'clip_ratio/high_max': 0.01315789483487606, 'clip_ratio/region_mean': 0.09717409731820226, 'reward_total_mean': 0.3662950396537781, 'reward_meter_mean': 0.12457047402858734, 'reward_meter_std': 0.33992356061935425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8819153308868408, 'reward_repeat_soft_std': 0.07343287765979767, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3662950396537781, 'reward_total_composite_std': 0.0971856340765953, 'epoch': 0.16}

 47%|████▋     | 1566/3300 [3:50:28<3:23:32,  7.04s/it]INFO 04-13 11:11:26 [block_pool.py:378] Successfully reset prefix cache

 47%|████▋     | 1567/3300 [3:50:34<3:19:52,  6.92s/it]2026-04-13 11:11:32,568 | INFO | train_grpo_train | metrics_logged mode=train step=1567

                                                       
{'loss': 0.021, 'grad_norm': 8.928327560424805, 'learning_rate': 5.2545454545454555e-06, 'num_tokens': 2775783.0, 'completions/mean_length': 59.125, 'completions/min_length': 50.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8751605749130249, 'rewards/meter/std': 0.2470269799232483, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8825733065605164, 'rewards/repeat_soft/std': 0.07122129946947098, 'rewards/judge_quality/mean': 0.6100000143051147, 'rewards/judge_quality/std': 0.24628673493862152, 'rewards/total_composite/mean': 0.666850209236145, 'rewards/total_composite/std': 0.16365578770637512, 'reward': 0.666850209236145, 'reward_std': 0.16365578770637512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11339724808931351, 'sampling/sampling_logp_difference/max': 1.3507094383239746, 'sampling/importance_sampling_ratio/min': 0.2590563893318176, 'sampling/importance_sampling_ratio/mean': 1.0102709531784058, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7823491543531418, 'clip_ratio/low_mean': 0.0842600530013442, 'clip_ratio/low_min': 0.0842600530013442, 'clip_ratio/high_mean': 0.019696970470249653, 'clip_ratio/high_max': 0.019696970470249653, 'clip_ratio/region_mean': 0.10395702347159386, 'reward_total_mean': 0.666850209236145, 'reward_meter_mean': 0.8751605749130249, 'reward_meter_std': 0.2470269799232483, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8825733065605164, 'reward_repeat_soft_std': 0.07122129946947098, 'reward_judge_quality_mean': 0.6100000143051147, 'reward_judge_quality_std': 0.24628673493862152, 'reward_total_composite_mean': 0.666850209236145, 'reward_total_composite_std': 0.16365578770637512, 'epoch': 0.16}

 47%|████▋     | 1567/3300 [3:50:34<3:19:52,  6.92s/it]INFO 04-13 11:11:32 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1568/3300 [3:50:41<3:14:37,  6.74s/it]2026-04-13 11:11:38,881 | INFO | train_grpo_train | metrics_logged mode=train step=1568

                                                       
{'loss': -0.0343, 'grad_norm': 13.269099235534668, 'learning_rate': 5.251515151515152e-06, 'num_tokens': 2777320.0, 'completions/mean_length': 47.125, 'completions/min_length': 38.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7809305191040039, 'rewards/meter/std': 0.3151097893714905, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9559616446495056, 'rewards/repeat_soft/std': 0.03860804811120033, 'rewards/judge_quality/mean': 0.6449999809265137, 'rewards/judge_quality/std': 0.24928471446037292, 'rewards/total_composite/mean': 0.6395946741104126, 'rewards/total_composite/std': 0.13574348390102386, 'reward': 0.6395946741104126, 'reward_std': 0.13574348390102386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11881216615438461, 'sampling/sampling_logp_difference/max': 1.609349012374878, 'sampling/importance_sampling_ratio/min': 0.2000177800655365, 'sampling/importance_sampling_ratio/mean': 1.0256962776184082, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6737695634365082, 'clip_ratio/low_mean': 0.08250083262100816, 'clip_ratio/low_min': 0.08250083262100816, 'clip_ratio/high_mean': 0.014433962292969227, 'clip_ratio/high_max': 0.014433962292969227, 'clip_ratio/region_mean': 0.09693479491397738, 'reward_total_mean': 0.6395946741104126, 'reward_meter_mean': 0.7809305191040039, 'reward_meter_std': 0.3151097893714905, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9559616446495056, 'reward_repeat_soft_std': 0.03860804811120033, 'reward_judge_quality_mean': 0.6449999809265137, 'reward_judge_quality_std': 0.24928471446037292, 'reward_total_composite_mean': 0.6395946741104126, 'reward_total_composite_std': 0.13574348390102386, 'epoch': 0.16}

 48%|████▊     | 1568/3300 [3:50:41<3:14:37,  6.74s/it]INFO 04-13 11:11:39 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1569/3300 [3:50:48<3:20:20,  6.94s/it]2026-04-13 11:11:46,297 | INFO | train_grpo_train | metrics_logged mode=train step=1569

                                                       
{'loss': 0.0921, 'grad_norm': 19.56976890563965, 'learning_rate': 5.248484848484849e-06, 'num_tokens': 2779015.0, 'completions/mean_length': 38.875, 'completions/min_length': 33.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8211154341697693, 'rewards/meter/std': 0.22690525650978088, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9778926372528076, 'rewards/repeat_soft/std': 0.03570332005620003, 'rewards/judge_quality/mean': 0.7699999809265137, 'rewards/judge_quality/std': 0.22677870094776154, 'rewards/total_composite/mean': 0.7675187587738037, 'rewards/total_composite/std': 0.17436037957668304, 'reward': 0.7675187587738037, 'reward_std': 0.17436039447784424, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10905809700489044, 'sampling/sampling_logp_difference/max': 1.2931843996047974, 'sampling/importance_sampling_ratio/min': 0.27439558506011963, 'sampling/importance_sampling_ratio/mean': 1.0121809244155884, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4464311897754669, 'clip_ratio/low_mean': 0.024127907119691372, 'clip_ratio/low_min': 0.024127907119691372, 'clip_ratio/high_mean': 0.05762811517342925, 'clip_ratio/high_max': 0.05762811517342925, 'clip_ratio/region_mean': 0.08175602229312062, 'reward_total_mean': 0.7675187587738037, 'reward_meter_mean': 0.8211154341697693, 'reward_meter_std': 0.22690525650978088, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9778926372528076, 'reward_repeat_soft_std': 0.03570332005620003, 'reward_judge_quality_mean': 0.7699999809265137, 'reward_judge_quality_std': 0.22677870094776154, 'reward_total_composite_mean': 0.7675187587738037, 'reward_total_composite_std': 0.17436037957668304, 'epoch': 0.16}

 48%|████▊     | 1569/3300 [3:50:48<3:20:20,  6.94s/it]INFO 04-13 11:11:46 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1570/3300 [3:50:56<3:25:33,  7.13s/it]2026-04-13 11:11:53,874 | INFO | train_grpo_train | metrics_logged mode=train step=1570

                                                       
{'loss': 0.0308, 'grad_norm': 5.472067832946777, 'learning_rate': 5.245454545454546e-06, 'num_tokens': 2781743.0, 'completions/mean_length': 153.0, 'completions/min_length': 146.0, 'completions/max_length': 166.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.0, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 166.0, 'rewards/meter/mean': 0.9894278645515442, 'rewards/meter/std': 0.005916774272918701, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7560806274414062, 'rewards/repeat_soft/std': 0.0717843696475029, 'rewards/judge_quality/mean': 0.29750001430511475, 'rewards/judge_quality/std': 0.1349867582321167, 'rewards/total_composite/mean': 0.4711803197860718, 'rewards/total_composite/std': 0.08733160048723221, 'reward': 0.4711803197860718, 'reward_std': 0.08733160048723221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08227641135454178, 'sampling/sampling_logp_difference/max': 1.9980177879333496, 'sampling/importance_sampling_ratio/min': 0.13560381531715393, 'sampling/importance_sampling_ratio/mean': 1.00179123878479, 'sampling/importance_sampling_ratio/max': 1.942197322845459, 'entropy': 0.4608866088092327, 'clip_ratio/low_mean': 0.028170245699584484, 'clip_ratio/low_min': 0.028170245699584484, 'clip_ratio/high_mean': 0.05055607762187719, 'clip_ratio/high_max': 0.05055607762187719, 'clip_ratio/region_mean': 0.07872632332146168, 'reward_total_mean': 0.4711803197860718, 'reward_meter_mean': 0.9894278645515442, 'reward_meter_std': 0.005916774272918701, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7560806274414062, 'reward_repeat_soft_std': 0.0717843696475029, 'reward_judge_quality_mean': 0.29750001430511475, 'reward_judge_quality_std': 0.1349867582321167, 'reward_total_composite_mean': 0.4711803197860718, 'reward_total_composite_std': 0.08733160048723221, 'epoch': 0.16}

 48%|████▊     | 1570/3300 [3:50:56<3:25:33,  7.13s/it]INFO 04-13 11:11:54 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1571/3300 [3:51:02<3:19:35,  6.93s/it]2026-04-13 11:12:00,365 | INFO | train_grpo_train | metrics_logged mode=train step=1571

                                                       
{'loss': 0.0842, 'grad_norm': 9.9489164352417, 'learning_rate': 5.242424242424244e-06, 'num_tokens': 2783489.0, 'completions/mean_length': 56.25, 'completions/min_length': 49.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9855920672416687, 'rewards/meter/std': 0.015676230192184448, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9529132843017578, 'rewards/repeat_soft/std': 0.04825044795870781, 'rewards/judge_quality/mean': 0.5487499833106995, 'rewards/judge_quality/std': 0.22937417030334473, 'rewards/total_composite/mean': 0.6934705376625061, 'rewards/total_composite/std': 0.14567831158638, 'reward': 0.6934705376625061, 'reward_std': 0.14567831158638, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1211368516087532, 'sampling/sampling_logp_difference/max': 1.8575502634048462, 'sampling/importance_sampling_ratio/min': 0.15605445206165314, 'sampling/importance_sampling_ratio/mean': 1.0127365589141846, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6561549454927444, 'clip_ratio/low_mean': 0.08774976572021842, 'clip_ratio/low_min': 0.08774976572021842, 'clip_ratio/high_mean': 0.028686946723610163, 'clip_ratio/high_max': 0.028686946723610163, 'clip_ratio/region_mean': 0.11643671244382858, 'reward_total_mean': 0.6934705376625061, 'reward_meter_mean': 0.9855920672416687, 'reward_meter_std': 0.015676230192184448, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9529132843017578, 'reward_repeat_soft_std': 0.04825044795870781, 'reward_judge_quality_mean': 0.5487499833106995, 'reward_judge_quality_std': 0.22937417030334473, 'reward_total_composite_mean': 0.6934705376625061, 'reward_total_composite_std': 0.14567831158638, 'epoch': 0.16}

 48%|████▊     | 1571/3300 [3:51:02<3:19:35,  6.93s/it]INFO 04-13 11:12:00 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1572/3300 [3:51:09<3:16:56,  6.84s/it]2026-04-13 11:12:06,953 | INFO | train_grpo_train | metrics_logged mode=train step=1572

                                                       
{'loss': 0.0262, 'grad_norm': 10.350753784179688, 'learning_rate': 5.23939393939394e-06, 'num_tokens': 2785038.0, 'completions/mean_length': 35.625, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.9386126399040222, 'rewards/meter/std': 0.03209441527724266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9973484873771667, 'rewards/repeat_soft/std': 0.0028345861937850714, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.8387596011161804, 'rewards/total_composite/std': 0.14296871423721313, 'reward': 0.8387596011161804, 'reward_std': 0.14296871423721313, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0853315219283104, 'sampling/sampling_logp_difference/max': 1.6540765762329102, 'sampling/importance_sampling_ratio/min': 0.19126859307289124, 'sampling/importance_sampling_ratio/mean': 1.0000499486923218, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3170384429395199, 'clip_ratio/low_mean': 0.035556891933083534, 'clip_ratio/low_min': 0.035556891933083534, 'clip_ratio/high_mean': 0.04899440938606858, 'clip_ratio/high_max': 0.04899440938606858, 'clip_ratio/region_mean': 0.08455130131915212, 'reward_total_mean': 0.8387596011161804, 'reward_meter_mean': 0.9386126399040222, 'reward_meter_std': 0.03209441527724266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9973484873771667, 'reward_repeat_soft_std': 0.0028345861937850714, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.8387596011161804, 'reward_total_composite_std': 0.14296871423721313, 'epoch': 0.16}

 48%|████▊     | 1572/3300 [3:51:09<3:16:56,  6.84s/it]INFO 04-13 11:12:07 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1573/3300 [3:51:15<3:16:44,  6.84s/it]2026-04-13 11:12:13,780 | INFO | train_grpo_train | metrics_logged mode=train step=1573

                                                       
{'loss': 0.0591, 'grad_norm': 11.370959281921387, 'learning_rate': 5.236363636363637e-06, 'num_tokens': 2786685.0, 'completions/mean_length': 55.875, 'completions/min_length': 41.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8713705539703369, 'rewards/meter/std': 0.3254017233848572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9093286991119385, 'rewards/repeat_soft/std': 0.055463191121816635, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5767042636871338, 'rewards/total_composite/std': 0.08499462157487869, 'reward': 0.5767042636871338, 'reward_std': 0.08499462157487869, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1327207386493683, 'sampling/sampling_logp_difference/max': 1.3430163860321045, 'sampling/importance_sampling_ratio/min': 0.2610570192337036, 'sampling/importance_sampling_ratio/mean': 1.010684609413147, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8427535966038704, 'clip_ratio/low_mean': 0.016393441706895828, 'clip_ratio/low_min': 0.016393441706895828, 'clip_ratio/high_mean': 0.11803002236410975, 'clip_ratio/high_max': 0.11803002236410975, 'clip_ratio/region_mean': 0.13442346407100558, 'reward_total_mean': 0.5767042636871338, 'reward_meter_mean': 0.8713705539703369, 'reward_meter_std': 0.3254017233848572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9093286991119385, 'reward_repeat_soft_std': 0.055463191121816635, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5767042636871338, 'reward_total_composite_std': 0.08499462157487869, 'epoch': 0.16}

 48%|████▊     | 1573/3300 [3:51:16<3:16:44,  6.84s/it]INFO 04-13 11:12:14 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1574/3300 [3:51:22<3:16:17,  6.82s/it]2026-04-13 11:12:20,576 | INFO | train_grpo_train | metrics_logged mode=train step=1574

                                                       
{'loss': -0.0634, 'grad_norm': 7.636651992797852, 'learning_rate': 5.233333333333334e-06, 'num_tokens': 2788583.0, 'completions/mean_length': 76.25, 'completions/min_length': 61.0, 'completions/max_length': 98.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 98.0, 'rewards/meter/mean': 0.9867457151412964, 'rewards/meter/std': 0.004665898624807596, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8288445472717285, 'rewards/repeat_soft/std': 0.08534655719995499, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5937082767486572, 'rewards/total_composite/std': 0.011892632581293583, 'reward': 0.5937082767486572, 'reward_std': 0.011892640963196754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11635272949934006, 'sampling/sampling_logp_difference/max': 1.3306999206542969, 'sampling/importance_sampling_ratio/min': 0.2642922103404999, 'sampling/importance_sampling_ratio/mean': 1.0000841617584229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6212470978498459, 'clip_ratio/low_mean': 0.019583360757678747, 'clip_ratio/low_min': 0.019583360757678747, 'clip_ratio/high_mean': 0.07040644437074661, 'clip_ratio/high_max': 0.07040644437074661, 'clip_ratio/region_mean': 0.08998980512842536, 'reward_total_mean': 0.5937082767486572, 'reward_meter_mean': 0.9867457151412964, 'reward_meter_std': 0.004665898624807596, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8288445472717285, 'reward_repeat_soft_std': 0.08534655719995499, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5937082767486572, 'reward_total_composite_std': 0.011892632581293583, 'epoch': 0.16}

 48%|████▊     | 1574/3300 [3:51:22<3:16:17,  6.82s/it]INFO 04-13 11:12:20 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1575/3300 [3:51:34<3:56:00,  8.21s/it]2026-04-13 11:12:32,024 | INFO | train_grpo_train | metrics_logged mode=train step=1575

                                                       
{'loss': -0.1517, 'grad_norm': 3.322378635406494, 'learning_rate': 5.230303030303031e-06, 'num_tokens': 2790237.0, 'completions/mean_length': 120.75, 'completions/min_length': 44.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 64.85714721679688, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.5820126533508301, 'rewards/meter/std': 0.3395563066005707, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9731205701828003, 'rewards/repeat_soft/std': 0.0636206567287445, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.47881633043289185, 'rewards/total_composite/std': 0.22543832659721375, 'reward': 0.47881633043289185, 'reward_std': 0.22543832659721375, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13704617321491241, 'sampling/sampling_logp_difference/max': 1.5153605937957764, 'sampling/importance_sampling_ratio/min': 0.21972894668579102, 'sampling/importance_sampling_ratio/mean': 1.012619137763977, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5075528435409069, 'clip_ratio/low_mean': 0.027435066178441048, 'clip_ratio/low_min': 0.027435066178441048, 'clip_ratio/high_mean': 0.06846528127789497, 'clip_ratio/high_max': 0.06846528127789497, 'clip_ratio/region_mean': 0.09590034745633602, 'reward_total_mean': 0.47881633043289185, 'reward_meter_mean': 0.5820126533508301, 'reward_meter_std': 0.3395563066005707, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9731205701828003, 'reward_repeat_soft_std': 0.0636206567287445, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.47881633043289185, 'reward_total_composite_std': 0.22543832659721375, 'epoch': 0.16}

 48%|████▊     | 1575/3300 [3:51:34<3:56:00,  8.21s/it]INFO 04-13 11:12:32 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1576/3300 [3:51:40<3:43:41,  7.79s/it]2026-04-13 11:12:38,813 | INFO | train_grpo_train | metrics_logged mode=train step=1576

                                                       
{'loss': 0.0672, 'grad_norm': 8.657578468322754, 'learning_rate': 5.2272727272727274e-06, 'num_tokens': 2792093.0, 'completions/mean_length': 71.0, 'completions/min_length': 56.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.0, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.8741507530212402, 'rewards/meter/std': 0.13478398323059082, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8293139934539795, 'rewards/repeat_soft/std': 0.07133729010820389, 'rewards/judge_quality/mean': 0.5612499713897705, 'rewards/judge_quality/std': 0.25614938139915466, 'rewards/total_composite/mean': 0.6353857517242432, 'rewards/total_composite/std': 0.13864047825336456, 'reward': 0.6353857517242432, 'reward_std': 0.13864046335220337, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0998210459947586, 'sampling/sampling_logp_difference/max': 1.2027044296264648, 'sampling/importance_sampling_ratio/min': 0.30038073658943176, 'sampling/importance_sampling_ratio/mean': 1.0102102756500244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5485082268714905, 'clip_ratio/low_mean': 0.05175823159515858, 'clip_ratio/low_min': 0.05175823159515858, 'clip_ratio/high_mean': 0.04950087238103151, 'clip_ratio/high_max': 0.04950087238103151, 'clip_ratio/region_mean': 0.10125910397619009, 'reward_total_mean': 0.6353857517242432, 'reward_meter_mean': 0.8741507530212402, 'reward_meter_std': 0.13478398323059082, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8293139934539795, 'reward_repeat_soft_std': 0.07133729010820389, 'reward_judge_quality_mean': 0.5612499713897705, 'reward_judge_quality_std': 0.25614938139915466, 'reward_total_composite_mean': 0.6353857517242432, 'reward_total_composite_std': 0.13864047825336456, 'epoch': 0.16}

 48%|████▊     | 1576/3300 [3:51:41<3:43:41,  7.79s/it]INFO 04-13 11:12:39 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1577/3300 [3:51:47<3:35:27,  7.50s/it]2026-04-13 11:12:45,661 | INFO | train_grpo_train | metrics_logged mode=train step=1577

                                                       
{'loss': 0.0788, 'grad_norm': 14.405783653259277, 'learning_rate': 5.224242424242425e-06, 'num_tokens': 2793539.0, 'completions/mean_length': 33.75, 'completions/min_length': 28.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8508832454681396, 'rewards/meter/std': 0.2590540051460266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9896299839019775, 'rewards/repeat_soft/std': 0.010065720416605473, 'rewards/judge_quality/mean': 0.6487500667572021, 'rewards/judge_quality/std': 0.2507951855659485, 'rewards/total_composite/mean': 0.6965362429618835, 'rewards/total_composite/std': 0.1784001588821411, 'reward': 0.6965362429618835, 'reward_std': 0.1784001588821411, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11630921065807343, 'sampling/sampling_logp_difference/max': 1.6127252578735352, 'sampling/importance_sampling_ratio/min': 0.2402961254119873, 'sampling/importance_sampling_ratio/mean': 0.9971534609794617, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5220210142433643, 'clip_ratio/low_mean': 0.08428848814219236, 'clip_ratio/low_min': 0.08428848814219236, 'clip_ratio/high_mean': 0.045522186905145645, 'clip_ratio/high_max': 0.045522186905145645, 'clip_ratio/region_mean': 0.129810675047338, 'reward_total_mean': 0.6965362429618835, 'reward_meter_mean': 0.8508832454681396, 'reward_meter_std': 0.2590540051460266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9896299839019775, 'reward_repeat_soft_std': 0.010065720416605473, 'reward_judge_quality_mean': 0.6487500667572021, 'reward_judge_quality_std': 0.2507951855659485, 'reward_total_composite_mean': 0.6965362429618835, 'reward_total_composite_std': 0.1784001588821411, 'epoch': 0.16}

 48%|████▊     | 1577/3300 [3:51:47<3:35:27,  7.50s/it]INFO 04-13 11:12:45 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1578/3300 [3:51:59<4:12:44,  8.81s/it]2026-04-13 11:12:57,560 | INFO | train_grpo_train | metrics_logged mode=train step=1578

                                                       
{'loss': -0.0652, 'grad_norm': 1.7189676761627197, 'learning_rate': 5.221212121212121e-06, 'num_tokens': 2795070.0, 'completions/mean_length': 210.375, 'completions/min_length': 21.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 29.399999618530273, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.19838173687458038, 'rewards/meter/std': 0.3413299024105072, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.4319060742855072, 'rewards/total_composite/mean': 0.28746530413627625, 'rewards/total_composite/std': 0.2562018632888794, 'reward': 0.28746530413627625, 'reward_std': 0.2562018632888794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16784901916980743, 'sampling/sampling_logp_difference/max': 1.1486377716064453, 'sampling/importance_sampling_ratio/min': 0.3170683979988098, 'sampling/importance_sampling_ratio/mean': 1.0461905002593994, 'sampling/importance_sampling_ratio/max': 1.9723833799362183, 'entropy': 1.0352696031332016, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.12192688602954149, 'clip_ratio/high_max': 0.12192688602954149, 'clip_ratio/region_mean': 0.12192688602954149, 'reward_total_mean': 0.28746530413627625, 'reward_meter_mean': 0.19838173687458038, 'reward_meter_std': 0.3413299024105072, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.4319060742855072, 'reward_total_composite_mean': 0.28746530413627625, 'reward_total_composite_std': 0.2562018632888794, 'epoch': 0.16}

 48%|████▊     | 1578/3300 [3:51:59<4:12:44,  8.81s/it]INFO 04-13 11:12:57 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1579/3300 [3:52:07<4:00:08,  8.37s/it]2026-04-13 11:13:04,866 | INFO | train_grpo_train | metrics_logged mode=train step=1579

                                                       
{'loss': -0.0589, 'grad_norm': 11.315073013305664, 'learning_rate': 5.218181818181819e-06, 'num_tokens': 2796880.0, 'completions/mean_length': 51.25, 'completions/min_length': 41.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8555561304092407, 'rewards/meter/std': 0.2454449087381363, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.916056752204895, 'rewards/repeat_soft/std': 0.06853128224611282, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5733703374862671, 'rewards/total_composite/std': 0.06686627119779587, 'reward': 0.5733703374862671, 'reward_std': 0.06686627119779587, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1285472810268402, 'sampling/sampling_logp_difference/max': 1.6143684387207031, 'sampling/importance_sampling_ratio/min': 0.19901631772518158, 'sampling/importance_sampling_ratio/mean': 1.0262603759765625, 'sampling/importance_sampling_ratio/max': 1.9754222631454468, 'entropy': 0.8895671665668488, 'clip_ratio/low_mean': 0.029666979797184467, 'clip_ratio/low_min': 0.029666979797184467, 'clip_ratio/high_mean': 0.09269695077091455, 'clip_ratio/high_max': 0.09269695077091455, 'clip_ratio/region_mean': 0.12236393056809902, 'reward_total_mean': 0.5733703374862671, 'reward_meter_mean': 0.8555561304092407, 'reward_meter_std': 0.2454449087381363, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.916056752204895, 'reward_repeat_soft_std': 0.06853128224611282, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5733703374862671, 'reward_total_composite_std': 0.06686627119779587, 'epoch': 0.16}

 48%|████▊     | 1579/3300 [3:52:07<4:00:08,  8.37s/it]INFO 04-13 11:13:05 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1580/3300 [3:52:14<3:50:58,  8.06s/it]2026-04-13 11:13:12,188 | INFO | train_grpo_train | metrics_logged mode=train step=1580

                                                       
{'loss': 0.0251, 'grad_norm': 7.267446517944336, 'learning_rate': 5.215151515151516e-06, 'num_tokens': 2799157.0, 'completions/mean_length': 117.625, 'completions/min_length': 95.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.625, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.9608144760131836, 'rewards/meter/std': 0.05136120691895485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8403679132461548, 'rewards/repeat_soft/std': 0.13496080040931702, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.1891333907842636, 'rewards/total_composite/mean': 0.6941373348236084, 'rewards/total_composite/std': 0.12480224668979645, 'reward': 0.6941373348236084, 'reward_std': 0.12480224668979645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1125214546918869, 'sampling/sampling_logp_difference/max': 2.707551956176758, 'sampling/importance_sampling_ratio/min': 0.06669989228248596, 'sampling/importance_sampling_ratio/mean': 1.0147680044174194, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.761542446911335, 'clip_ratio/low_mean': 0.039915972389280796, 'clip_ratio/low_min': 0.039915972389280796, 'clip_ratio/high_mean': 0.047846470028162, 'clip_ratio/high_max': 0.047846470028162, 'clip_ratio/region_mean': 0.0877624424174428, 'reward_total_mean': 0.6941373348236084, 'reward_meter_mean': 0.9608144760131836, 'reward_meter_std': 0.05136120691895485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8403679132461548, 'reward_repeat_soft_std': 0.13496080040931702, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.1891333907842636, 'reward_total_composite_mean': 0.6941373348236084, 'reward_total_composite_std': 0.12480224668979645, 'epoch': 0.16}

 48%|████▊     | 1580/3300 [3:52:14<3:50:58,  8.06s/it]INFO 04-13 11:13:12 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1581/3300 [3:52:20<3:38:44,  7.63s/it]2026-04-13 11:13:18,893 | INFO | train_grpo_train | metrics_logged mode=train step=1581

                                                       
{'loss': 0.0349, 'grad_norm': 12.470683097839355, 'learning_rate': 5.212121212121213e-06, 'num_tokens': 2800876.0, 'completions/mean_length': 51.875, 'completions/min_length': 46.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8088762760162354, 'rewards/meter/std': 0.33293840289115906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9293208122253418, 'rewards/repeat_soft/std': 0.01859375834465027, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.5894694328308105, 'rewards/total_composite/std': 0.12230564653873444, 'reward': 0.5894694328308105, 'reward_std': 0.12230563163757324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12058217823505402, 'sampling/sampling_logp_difference/max': 1.4676432609558105, 'sampling/importance_sampling_ratio/min': 0.2304680049419403, 'sampling/importance_sampling_ratio/mean': 0.9969631433486938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7287614792585373, 'clip_ratio/low_mean': 0.0363562093116343, 'clip_ratio/low_min': 0.0363562093116343, 'clip_ratio/high_mean': 0.05789553700014949, 'clip_ratio/high_max': 0.05789553700014949, 'clip_ratio/region_mean': 0.09425174631178379, 'reward_total_mean': 0.5894694328308105, 'reward_meter_mean': 0.8088762760162354, 'reward_meter_std': 0.33293840289115906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9293208122253418, 'reward_repeat_soft_std': 0.01859375834465027, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.5894694328308105, 'reward_total_composite_std': 0.12230564653873444, 'epoch': 0.16}

 48%|████▊     | 1581/3300 [3:52:21<3:38:44,  7.63s/it]INFO 04-13 11:13:19 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1582/3300 [3:52:28<3:34:32,  7.49s/it]2026-04-13 11:13:25,993 | INFO | train_grpo_train | metrics_logged mode=train step=1582

                                                       
{'loss': 0.0159, 'grad_norm': 17.356124877929688, 'learning_rate': 5.209090909090909e-06, 'num_tokens': 2802668.0, 'completions/mean_length': 63.0, 'completions/min_length': 59.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.7036300301551819, 'rewards/meter/std': 0.23615922033786774, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9461240768432617, 'rewards/repeat_soft/std': 0.02267386019229889, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.11667262762784958, 'rewards/total_composite/mean': 0.5510309338569641, 'rewards/total_composite/std': 0.07714087516069412, 'reward': 0.5510309338569641, 'reward_std': 0.07714088261127472, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08768946677446365, 'sampling/sampling_logp_difference/max': 3.278637647628784, 'sampling/importance_sampling_ratio/min': 0.037679556757211685, 'sampling/importance_sampling_ratio/mean': 0.9985697269439697, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42118654400110245, 'clip_ratio/low_mean': 0.017747920006513596, 'clip_ratio/low_min': 0.017747920006513596, 'clip_ratio/high_mean': 0.06158871715888381, 'clip_ratio/high_max': 0.06158871715888381, 'clip_ratio/region_mean': 0.0793366371653974, 'reward_total_mean': 0.5510309338569641, 'reward_meter_mean': 0.7036300301551819, 'reward_meter_std': 0.23615922033786774, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9461240768432617, 'reward_repeat_soft_std': 0.02267386019229889, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.11667262762784958, 'reward_total_composite_mean': 0.5510309338569641, 'reward_total_composite_std': 0.07714087516069412, 'epoch': 0.16}

 48%|████▊     | 1582/3300 [3:52:28<3:34:32,  7.49s/it]INFO 04-13 11:13:26 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1583/3300 [3:52:34<3:23:44,  7.12s/it]2026-04-13 11:13:32,244 | INFO | train_grpo_train | metrics_logged mode=train step=1583

                                                       
{'loss': -0.0114, 'grad_norm': 14.827067375183105, 'learning_rate': 5.2060606060606065e-06, 'num_tokens': 2804195.0, 'completions/mean_length': 35.875, 'completions/min_length': 33.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7313417792320251, 'rewards/meter/std': 0.16144664585590363, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9974443912506104, 'rewards/repeat_soft/std': 0.004507945850491524, 'rewards/judge_quality/mean': 0.7150000333786011, 'rewards/judge_quality/std': 0.24307554960250854, 'rewards/total_composite/mean': 0.6837367415428162, 'rewards/total_composite/std': 0.12577074766159058, 'reward': 0.6837367415428162, 'reward_std': 0.12577074766159058, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.144041508436203, 'sampling/sampling_logp_difference/max': 3.5102028846740723, 'sampling/importance_sampling_ratio/min': 0.029890848323702812, 'sampling/importance_sampling_ratio/mean': 0.9953535795211792, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5109865702688694, 'clip_ratio/low_mean': 0.047726212767884135, 'clip_ratio/low_min': 0.047726212767884135, 'clip_ratio/high_mean': 0.057368865702301264, 'clip_ratio/high_max': 0.057368865702301264, 'clip_ratio/region_mean': 0.1050950784701854, 'reward_total_mean': 0.6837367415428162, 'reward_meter_mean': 0.7313417792320251, 'reward_meter_std': 0.16144664585590363, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9974443912506104, 'reward_repeat_soft_std': 0.004507945850491524, 'reward_judge_quality_mean': 0.7150000333786011, 'reward_judge_quality_std': 0.24307554960250854, 'reward_total_composite_mean': 0.6837367415428162, 'reward_total_composite_std': 0.12577074766159058, 'epoch': 0.16}

 48%|████▊     | 1583/3300 [3:52:34<3:23:44,  7.12s/it]INFO 04-13 11:13:32 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1584/3300 [3:52:47<4:11:48,  8.80s/it]2026-04-13 11:13:44,979 | INFO | train_grpo_train | metrics_logged mode=train step=1584

                                                       
{'loss': -0.109, 'grad_norm': 3.57348370552063, 'learning_rate': 5.203030303030303e-06, 'num_tokens': 2806091.0, 'completions/mean_length': 142.0, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 89.14286041259766, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.7479578256607056, 'rewards/meter/std': 0.26813703775405884, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9292437434196472, 'rewards/repeat_soft/std': 0.049904964864254, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.33907175064086914, 'rewards/total_composite/mean': 0.6143718957901001, 'rewards/total_composite/std': 0.3079480528831482, 'reward': 0.6143718957901001, 'reward_std': 0.3079480528831482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11741182208061218, 'sampling/sampling_logp_difference/max': 1.9511277675628662, 'sampling/importance_sampling_ratio/min': 0.14211371541023254, 'sampling/importance_sampling_ratio/mean': 0.9921196103096008, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.519912701100111, 'clip_ratio/low_mean': 0.02894775429740548, 'clip_ratio/low_min': 0.02894775429740548, 'clip_ratio/high_mean': 0.051714906468987465, 'clip_ratio/high_max': 0.051714906468987465, 'clip_ratio/region_mean': 0.08066266076639295, 'reward_total_mean': 0.6143718957901001, 'reward_meter_mean': 0.7479578256607056, 'reward_meter_std': 0.26813703775405884, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9292437434196472, 'reward_repeat_soft_std': 0.049904964864254, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.33907175064086914, 'reward_total_composite_mean': 0.6143718957901001, 'reward_total_composite_std': 0.3079480528831482, 'epoch': 0.16}

 48%|████▊     | 1584/3300 [3:52:47<4:11:48,  8.80s/it]INFO 04-13 11:13:45 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1585/3300 [3:52:54<3:56:44,  8.28s/it]2026-04-13 11:13:52,047 | INFO | train_grpo_train | metrics_logged mode=train step=1585

                                                       
{'loss': 0.1246, 'grad_norm': 10.92084789276123, 'learning_rate': 5.2e-06, 'num_tokens': 2807954.0, 'completions/mean_length': 55.875, 'completions/min_length': 43.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.7135234475135803, 'rewards/meter/std': 0.3027102053165436, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.930819034576416, 'rewards/repeat_soft/std': 0.031038982793688774, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5591105222702026, 'rewards/total_composite/std': 0.09256841987371445, 'reward': 0.5591105222702026, 'reward_std': 0.09256841987371445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10486295819282532, 'sampling/sampling_logp_difference/max': 1.120133399963379, 'sampling/importance_sampling_ratio/min': 0.3262362778186798, 'sampling/importance_sampling_ratio/mean': 1.009124517440796, 'sampling/importance_sampling_ratio/max': 1.8682174682617188, 'entropy': 0.8714019283652306, 'clip_ratio/low_mean': 0.0363570312038064, 'clip_ratio/low_min': 0.0363570312038064, 'clip_ratio/high_mean': 0.06640502624213696, 'clip_ratio/high_max': 0.06640502624213696, 'clip_ratio/region_mean': 0.10276205744594336, 'reward_total_mean': 0.5591105222702026, 'reward_meter_mean': 0.7135234475135803, 'reward_meter_std': 0.3027102053165436, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.930819034576416, 'reward_repeat_soft_std': 0.031038982793688774, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5591105222702026, 'reward_total_composite_std': 0.09256841987371445, 'epoch': 0.16}

 48%|████▊     | 1585/3300 [3:52:54<3:56:44,  8.28s/it]INFO 04-13 11:13:52 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1586/3300 [3:53:00<3:39:43,  7.69s/it]2026-04-13 11:13:58,353 | INFO | train_grpo_train | metrics_logged mode=train step=1586

                                                       
{'loss': 0.0281, 'grad_norm': 14.093966484069824, 'learning_rate': 5.196969696969697e-06, 'num_tokens': 2809459.0, 'completions/mean_length': 45.125, 'completions/min_length': 38.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.7918084859848022, 'rewards/meter/std': 0.1963483691215515, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9725749492645264, 'rewards/repeat_soft/std': 0.03801165893673897, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.19108712673187256, 'rewards/total_composite/mean': 0.5859659910202026, 'rewards/total_composite/std': 0.1032167300581932, 'reward': 0.5859659910202026, 'reward_std': 0.1032167449593544, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1340153068304062, 'sampling/sampling_logp_difference/max': 1.883195400238037, 'sampling/importance_sampling_ratio/min': 0.15210328996181488, 'sampling/importance_sampling_ratio/mean': 1.0184165239334106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6959806382656097, 'clip_ratio/low_mean': 0.020072502084076405, 'clip_ratio/low_min': 0.020072502084076405, 'clip_ratio/high_mean': 0.08884621039032936, 'clip_ratio/high_max': 0.08884621039032936, 'clip_ratio/region_mean': 0.10891871247440577, 'reward_total_mean': 0.5859659910202026, 'reward_meter_mean': 0.7918084859848022, 'reward_meter_std': 0.1963483691215515, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9725749492645264, 'reward_repeat_soft_std': 0.03801165893673897, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.19108712673187256, 'reward_total_composite_mean': 0.5859659910202026, 'reward_total_composite_std': 0.1032167300581932, 'epoch': 0.16}

 48%|████▊     | 1586/3300 [3:53:00<3:39:43,  7.69s/it]INFO 04-13 11:13:58 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1587/3300 [3:53:06<3:23:54,  7.14s/it]2026-04-13 11:14:04,224 | INFO | train_grpo_train | metrics_logged mode=train step=1587

                                                       
{'loss': 0.0598, 'grad_norm': 20.771202087402344, 'learning_rate': 5.193939393939395e-06, 'num_tokens': 2810919.0, 'completions/mean_length': 31.5, 'completions/min_length': 27.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.48673421144485474, 'rewards/meter/std': 0.358243852853775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9822884202003479, 'rewards/repeat_soft/std': 0.009964286349713802, 'rewards/judge_quality/mean': 0.7312500476837158, 'rewards/judge_quality/std': 0.20131267607212067, 'rewards/total_composite/mean': 0.5827430486679077, 'rewards/total_composite/std': 0.19626817107200623, 'reward': 0.5827430486679077, 'reward_std': 0.19626817107200623, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12723451852798462, 'sampling/sampling_logp_difference/max': 2.145318031311035, 'sampling/importance_sampling_ratio/min': 0.1170308068394661, 'sampling/importance_sampling_ratio/mean': 1.0058783292770386, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4881987236440182, 'clip_ratio/low_mean': 0.05622329097241163, 'clip_ratio/low_min': 0.05622329097241163, 'clip_ratio/high_mean': 0.049179146997630596, 'clip_ratio/high_max': 0.049179146997630596, 'clip_ratio/region_mean': 0.10540243797004223, 'reward_total_mean': 0.5827430486679077, 'reward_meter_mean': 0.48673421144485474, 'reward_meter_std': 0.358243852853775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9822884202003479, 'reward_repeat_soft_std': 0.009964286349713802, 'reward_judge_quality_mean': 0.7312500476837158, 'reward_judge_quality_std': 0.20131267607212067, 'reward_total_composite_mean': 0.5827430486679077, 'reward_total_composite_std': 0.19626817107200623, 'epoch': 0.16}

 48%|████▊     | 1587/3300 [3:53:06<3:23:54,  7.14s/it]INFO 04-13 11:14:04 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1588/3300 [3:53:17<3:58:30,  8.36s/it]2026-04-13 11:14:15,433 | INFO | train_grpo_train | metrics_logged mode=train step=1588

                                                       
{'loss': -0.0514, 'grad_norm': 3.0134294033050537, 'learning_rate': 5.190909090909091e-06, 'num_tokens': 2812318.0, 'completions/mean_length': 83.875, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.71428680419922, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.701058030128479, 'rewards/meter/std': 0.34651628136634827, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9576764702796936, 'rewards/repeat_soft/std': 0.013642913661897182, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.3147788345813751, 'rewards/total_composite/mean': 0.5050708055496216, 'rewards/total_composite/std': 0.26179876923561096, 'reward': 0.5050708055496216, 'reward_std': 0.2617987394332886, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1175791397690773, 'sampling/sampling_logp_difference/max': 1.7134544849395752, 'sampling/importance_sampling_ratio/min': 0.18024207651615143, 'sampling/importance_sampling_ratio/mean': 0.9964831471443176, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5598728880286217, 'clip_ratio/low_mean': 0.054387480951845646, 'clip_ratio/low_min': 0.054387480951845646, 'clip_ratio/high_mean': 0.06264881044626236, 'clip_ratio/high_max': 0.06264881044626236, 'clip_ratio/region_mean': 0.117036291398108, 'reward_total_mean': 0.5050708055496216, 'reward_meter_mean': 0.701058030128479, 'reward_meter_std': 0.34651628136634827, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9576764702796936, 'reward_repeat_soft_std': 0.013642913661897182, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.3147788345813751, 'reward_total_composite_mean': 0.5050708055496216, 'reward_total_composite_std': 0.26179876923561096, 'epoch': 0.16}

 48%|████▊     | 1588/3300 [3:53:17<3:58:30,  8.36s/it]INFO 04-13 11:14:15 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1589/3300 [3:53:23<3:41:48,  7.78s/it]2026-04-13 11:14:21,865 | INFO | train_grpo_train | metrics_logged mode=train step=1589

                                                       
{'loss': 0.0049, 'grad_norm': 10.57528305053711, 'learning_rate': 5.187878787878788e-06, 'num_tokens': 2813869.0, 'completions/mean_length': 46.875, 'completions/min_length': 35.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.92644202709198, 'rewards/meter/std': 0.16375716030597687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.908367395401001, 'rewards/repeat_soft/std': 0.08913388103246689, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5940006971359253, 'rewards/total_composite/std': 0.05687683820724487, 'reward': 0.5940006971359253, 'reward_std': 0.05687684565782547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10283184051513672, 'sampling/sampling_logp_difference/max': 1.772202491760254, 'sampling/importance_sampling_ratio/min': 0.16995824873447418, 'sampling/importance_sampling_ratio/mean': 1.022234320640564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.686737447977066, 'clip_ratio/low_mean': 0.02113095298409462, 'clip_ratio/low_min': 0.02113095298409462, 'clip_ratio/high_mean': 0.09037395427003503, 'clip_ratio/high_max': 0.09037395427003503, 'clip_ratio/region_mean': 0.11150490725412965, 'reward_total_mean': 0.5940006971359253, 'reward_meter_mean': 0.92644202709198, 'reward_meter_std': 0.16375716030597687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.908367395401001, 'reward_repeat_soft_std': 0.08913388103246689, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5940006971359253, 'reward_total_composite_std': 0.05687683820724487, 'epoch': 0.16}

 48%|████▊     | 1589/3300 [3:53:24<3:41:48,  7.78s/it]INFO 04-13 11:14:22 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1590/3300 [3:53:31<3:35:51,  7.57s/it]2026-04-13 11:14:28,937 | INFO | train_grpo_train | metrics_logged mode=train step=1590

                                                       
{'loss': 0.0177, 'grad_norm': 6.9513840675354, 'learning_rate': 5.184848484848485e-06, 'num_tokens': 2815914.0, 'completions/mean_length': 96.625, 'completions/min_length': 73.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 96.625, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.9399340152740479, 'rewards/meter/std': 0.0858582854270935, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9564548134803772, 'rewards/repeat_soft/std': 0.034047335386276245, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6397402286529541, 'rewards/total_composite/std': 0.1189008429646492, 'reward': 0.6397402286529541, 'reward_std': 0.118900828063488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12704916298389435, 'sampling/sampling_logp_difference/max': 3.7976233959198, 'sampling/importance_sampling_ratio/min': 0.0224240031093359, 'sampling/importance_sampling_ratio/mean': 1.0120757818222046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8992115184664726, 'clip_ratio/low_mean': 0.1037961496040225, 'clip_ratio/low_min': 0.1037961496040225, 'clip_ratio/high_mean': 0.013020833022892475, 'clip_ratio/high_max': 0.013020833022892475, 'clip_ratio/region_mean': 0.11681698262691498, 'reward_total_mean': 0.6397402286529541, 'reward_meter_mean': 0.9399340152740479, 'reward_meter_std': 0.0858582854270935, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9564548134803772, 'reward_repeat_soft_std': 0.034047335386276245, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6397402286529541, 'reward_total_composite_std': 0.1189008429646492, 'epoch': 0.16}

 48%|████▊     | 1590/3300 [3:53:31<3:35:51,  7.57s/it]INFO 04-13 11:14:29 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1591/3300 [3:53:36<3:19:27,  7.00s/it]2026-04-13 11:14:34,611 | INFO | train_grpo_train | metrics_logged mode=train step=1591

                                                       
{'loss': 0.1216, 'grad_norm': 23.791715621948242, 'learning_rate': 5.181818181818182e-06, 'num_tokens': 2817297.0, 'completions/mean_length': 21.875, 'completions/min_length': 18.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8926088809967041, 'rewards/meter/std': 0.1404341459274292, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.947945773601532, 'rewards/repeat_soft/std': 0.04116548225283623, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.511832058429718, 'rewards/total_composite/std': 0.05234541743993759, 'reward': 0.511832058429718, 'reward_std': 0.05234541371464729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18677814304828644, 'sampling/sampling_logp_difference/max': 1.87799072265625, 'sampling/importance_sampling_ratio/min': 0.15289701521396637, 'sampling/importance_sampling_ratio/mean': 1.034908413887024, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.168588176369667, 'clip_ratio/low_mean': 0.12529785558581352, 'clip_ratio/low_min': 0.12529785558581352, 'clip_ratio/high_mean': 0.054459065198898315, 'clip_ratio/high_max': 0.054459065198898315, 'clip_ratio/region_mean': 0.17975692078471184, 'reward_total_mean': 0.511832058429718, 'reward_meter_mean': 0.8926088809967041, 'reward_meter_std': 0.1404341459274292, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.947945773601532, 'reward_repeat_soft_std': 0.04116548225283623, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.511832058429718, 'reward_total_composite_std': 0.05234541743993759, 'epoch': 0.16}

 48%|████▊     | 1591/3300 [3:53:36<3:19:27,  7.00s/it]INFO 04-13 11:14:34 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1592/3300 [3:53:42<3:12:27,  6.76s/it]2026-04-13 11:14:40,808 | INFO | train_grpo_train | metrics_logged mode=train step=1592

                                                       
{'loss': 0.0895, 'grad_norm': 15.207016944885254, 'learning_rate': 5.1787878787878784e-06, 'num_tokens': 2818966.0, 'completions/mean_length': 43.625, 'completions/min_length': 34.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6591077446937561, 'rewards/meter/std': 0.35364237427711487, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9206712245941162, 'rewards/repeat_soft/std': 0.0638064593076706, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5571283102035522, 'rewards/total_composite/std': 0.17710694670677185, 'reward': 0.5571283102035522, 'reward_std': 0.17710696160793304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14776752889156342, 'sampling/sampling_logp_difference/max': 2.2035584449768066, 'sampling/importance_sampling_ratio/min': 0.11040957272052765, 'sampling/importance_sampling_ratio/mean': 0.9959150552749634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7373897694051266, 'clip_ratio/low_mean': 0.04277499159798026, 'clip_ratio/low_min': 0.04277499159798026, 'clip_ratio/high_mean': 0.07668900489807129, 'clip_ratio/high_max': 0.07668900489807129, 'clip_ratio/region_mean': 0.11946399649605155, 'reward_total_mean': 0.5571283102035522, 'reward_meter_mean': 0.6591077446937561, 'reward_meter_std': 0.35364237427711487, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9206712245941162, 'reward_repeat_soft_std': 0.0638064593076706, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5571283102035522, 'reward_total_composite_std': 0.17710694670677185, 'epoch': 0.16}

 48%|████▊     | 1592/3300 [3:53:43<3:12:27,  6.76s/it]INFO 04-13 11:14:41 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1593/3300 [3:53:49<3:11:52,  6.74s/it]2026-04-13 11:14:47,510 | INFO | train_grpo_train | metrics_logged mode=train step=1593

                                                       
{'loss': 0.0399, 'grad_norm': 10.931285858154297, 'learning_rate': 5.1757575757575765e-06, 'num_tokens': 2820738.0, 'completions/mean_length': 61.5, 'completions/min_length': 55.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.5, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.8601760268211365, 'rewards/meter/std': 0.2870243191719055, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9424629211425781, 'rewards/repeat_soft/std': 0.04623568803071976, 'rewards/judge_quality/mean': 0.367499977350235, 'rewards/judge_quality/std': 0.09808888286352158, 'rewards/total_composite/mean': 0.554111123085022, 'rewards/total_composite/std': 0.08999910205602646, 'reward': 0.554111123085022, 'reward_std': 0.08999911695718765, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12144916504621506, 'sampling/sampling_logp_difference/max': 1.7190871238708496, 'sampling/importance_sampling_ratio/min': 0.17922969162464142, 'sampling/importance_sampling_ratio/mean': 1.0130892992019653, 'sampling/importance_sampling_ratio/max': 1.9167232513427734, 'entropy': 1.0045983046293259, 'clip_ratio/low_mean': 0.061697943136096, 'clip_ratio/low_min': 0.061697943136096, 'clip_ratio/high_mean': 0.08305913768708706, 'clip_ratio/high_max': 0.08305913768708706, 'clip_ratio/region_mean': 0.14475708082318306, 'reward_total_mean': 0.554111123085022, 'reward_meter_mean': 0.8601760268211365, 'reward_meter_std': 0.2870243191719055, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9424629211425781, 'reward_repeat_soft_std': 0.04623568803071976, 'reward_judge_quality_mean': 0.367499977350235, 'reward_judge_quality_std': 0.09808888286352158, 'reward_total_composite_mean': 0.554111123085022, 'reward_total_composite_std': 0.08999910205602646, 'epoch': 0.16}

 48%|████▊     | 1593/3300 [3:53:49<3:11:52,  6.74s/it]INFO 04-13 11:14:47 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1594/3300 [3:53:58<3:29:58,  7.38s/it]2026-04-13 11:14:56,387 | INFO | train_grpo_train | metrics_logged mode=train step=1594

                                                       
{'loss': 0.0554, 'grad_norm': 7.387360572814941, 'learning_rate': 5.172727272727273e-06, 'num_tokens': 2823131.0, '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': 0.8398451209068298, 'rewards/meter/std': 0.19838230311870575, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.809610903263092, 'rewards/repeat_soft/std': 0.05944375321269035, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5107194185256958, 'rewards/total_composite/std': 0.058343540877103806, 'reward': 0.5107194185256958, 'reward_std': 0.058343540877103806, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09610215574502945, 'sampling/sampling_logp_difference/max': 2.1601510047912598, 'sampling/importance_sampling_ratio/min': 0.11530770361423492, 'sampling/importance_sampling_ratio/mean': 0.9995241761207581, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5531741492450237, 'clip_ratio/low_mean': 0.015402622055262327, 'clip_ratio/low_min': 0.015402622055262327, 'clip_ratio/high_mean': 0.055950433015823364, 'clip_ratio/high_max': 0.055950433015823364, 'clip_ratio/region_mean': 0.07135305507108569, 'reward_total_mean': 0.5107194185256958, 'reward_meter_mean': 0.8398451209068298, 'reward_meter_std': 0.19838230311870575, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.809610903263092, 'reward_repeat_soft_std': 0.05944375321269035, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5107194185256958, 'reward_total_composite_std': 0.058343540877103806, 'epoch': 0.16}

 48%|████▊     | 1594/3300 [3:53:58<3:29:58,  7.38s/it]INFO 04-13 11:14:56 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1595/3300 [3:54:13<4:30:57,  9.54s/it]2026-04-13 11:15:10,940 | INFO | train_grpo_train | metrics_logged mode=train step=1595

                                                       
{'loss': -0.0829, 'grad_norm': 1.6510740518569946, 'learning_rate': 5.16969696969697e-06, 'num_tokens': 2824713.0, 'completions/mean_length': 220.75, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.21364955604076385, 'rewards/meter/std': 0.2237122654914856, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9871094822883606, 'rewards/repeat_soft/std': 0.013966692611575127, 'rewards/judge_quality/mean': 0.32249999046325684, 'rewards/judge_quality/std': 0.29605743288993835, 'rewards/total_composite/mean': 0.26387202739715576, 'rewards/total_composite/std': 0.2245648354291916, 'reward': 0.26387202739715576, 'reward_std': 0.2245648354291916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1586504429578781, 'sampling/sampling_logp_difference/max': 2.5145673751831055, 'sampling/importance_sampling_ratio/min': 0.08089790493249893, 'sampling/importance_sampling_ratio/mean': 1.002648949623108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6664703115820885, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08654541755095124, 'clip_ratio/high_max': 0.08654541755095124, 'clip_ratio/region_mean': 0.08654541755095124, 'reward_total_mean': 0.26387202739715576, 'reward_meter_mean': 0.21364955604076385, 'reward_meter_std': 0.2237122654914856, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9871094822883606, 'reward_repeat_soft_std': 0.013966692611575127, 'reward_judge_quality_mean': 0.32249999046325684, 'reward_judge_quality_std': 0.29605743288993835, 'reward_total_composite_mean': 0.26387202739715576, 'reward_total_composite_std': 0.2245648354291916, 'epoch': 0.16}

 48%|████▊     | 1595/3300 [3:54:13<4:30:57,  9.54s/it]INFO 04-13 11:15:11 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1596/3300 [3:54:25<4:53:57, 10.35s/it]2026-04-13 11:15:23,206 | INFO | train_grpo_train | metrics_logged mode=train step=1596

                                                       
{'loss': -0.2111, 'grad_norm': 2.712610960006714, 'learning_rate': 5.1666666666666675e-06, 'num_tokens': 2827335.0, 'completions/mean_length': 186.75, 'completions/min_length': 123.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 140.2857208251953, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.6589198112487793, 'rewards/meter/std': 0.3036969006061554, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.2357022613286972, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.883997917175293, 'rewards/repeat_soft/std': 0.05677222087979317, 'rewards/judge_quality/mean': 0.33124998211860657, 'rewards/judge_quality/std': 0.13715866208076477, 'rewards/total_composite/mean': 0.41331344842910767, 'rewards/total_composite/std': 0.1814599186182022, 'reward': 0.41331344842910767, 'reward_std': 0.18145990371704102, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10325267165899277, 'sampling/sampling_logp_difference/max': 1.4933323860168457, 'sampling/importance_sampling_ratio/min': 0.22462287545204163, 'sampling/importance_sampling_ratio/mean': 1.0112303495407104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5912267565727234, 'clip_ratio/low_mean': 0.01158059504814446, 'clip_ratio/low_min': 0.01158059504814446, 'clip_ratio/high_mean': 0.07480679452419281, 'clip_ratio/high_max': 0.07480679452419281, 'clip_ratio/region_mean': 0.08638738957233727, 'reward_total_mean': 0.41331344842910767, 'reward_meter_mean': 0.6589198112487793, 'reward_meter_std': 0.3036969006061554, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.2357022613286972, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.883997917175293, 'reward_repeat_soft_std': 0.05677222087979317, 'reward_judge_quality_mean': 0.33124998211860657, 'reward_judge_quality_std': 0.13715866208076477, 'reward_total_composite_mean': 0.41331344842910767, 'reward_total_composite_std': 0.1814599186182022, 'epoch': 0.16}

 48%|████▊     | 1596/3300 [3:54:25<4:53:57, 10.35s/it]INFO 04-13 11:15:23 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1597/3300 [3:54:34<4:39:57,  9.86s/it]2026-04-13 11:15:31,980 | INFO | train_grpo_train | metrics_logged mode=train step=1597

                                                       
{'loss': 0.0247, 'grad_norm': 7.101766586303711, 'learning_rate': 5.163636363636364e-06, 'num_tokens': 2829633.0, 'completions/mean_length': 130.25, 'completions/min_length': 116.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.25, 'completions/min_terminated_length': 116.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.8372750282287598, 'rewards/meter/std': 0.21894672513008118, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9086341857910156, 'rewards/repeat_soft/std': 0.07240957766771317, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.49496954679489136, 'rewards/total_composite/std': 0.07301773875951767, 'reward': 0.49496954679489136, 'reward_std': 0.07301773130893707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11287858337163925, 'sampling/sampling_logp_difference/max': 1.9383482933044434, 'sampling/importance_sampling_ratio/min': 0.1439415067434311, 'sampling/importance_sampling_ratio/mean': 1.019770622253418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9102844372391701, 'clip_ratio/low_mean': 0.03527601668611169, 'clip_ratio/low_min': 0.03527601668611169, 'clip_ratio/high_mean': 0.060892132110893726, 'clip_ratio/high_max': 0.060892132110893726, 'clip_ratio/region_mean': 0.09616814879700541, 'reward_total_mean': 0.49496954679489136, 'reward_meter_mean': 0.8372750282287598, 'reward_meter_std': 0.21894672513008118, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9086341857910156, 'reward_repeat_soft_std': 0.07240957766771317, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.49496954679489136, 'reward_total_composite_std': 0.07301773875951767, 'epoch': 0.16}

 48%|████▊     | 1597/3300 [3:54:34<4:39:57,  9.86s/it]INFO 04-13 11:15:32 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1598/3300 [3:54:41<4:20:48,  9.19s/it]2026-04-13 11:15:39,552 | INFO | train_grpo_train | metrics_logged mode=train step=1598

                                                       
{'loss': -0.0178, 'grad_norm': 10.958575248718262, 'learning_rate': 5.160606060606061e-06, 'num_tokens': 2831096.0, 'completions/mean_length': 41.875, 'completions/min_length': 34.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.5208902359008789, 'rewards/meter/std': 0.401324987411499, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9967353343963623, 'rewards/repeat_soft/std': 0.00640062615275383, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.17880557477474213, 'rewards/total_composite/mean': 0.5438495874404907, 'rewards/total_composite/std': 0.14783990383148193, 'reward': 0.5438495874404907, 'reward_std': 0.14783990383148193, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14300960302352905, 'sampling/sampling_logp_difference/max': 3.099897861480713, 'sampling/importance_sampling_ratio/min': 0.04505380615592003, 'sampling/importance_sampling_ratio/mean': 1.0001353025436401, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5781061500310898, 'clip_ratio/low_mean': 0.05817630235105753, 'clip_ratio/low_min': 0.05817630235105753, 'clip_ratio/high_mean': 0.06277035549283028, 'clip_ratio/high_max': 0.06277035549283028, 'clip_ratio/region_mean': 0.1209466578438878, 'reward_total_mean': 0.5438495874404907, 'reward_meter_mean': 0.5208902359008789, 'reward_meter_std': 0.401324987411499, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9967353343963623, 'reward_repeat_soft_std': 0.00640062615275383, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.17880557477474213, 'reward_total_composite_mean': 0.5438495874404907, 'reward_total_composite_std': 0.14783990383148193, 'epoch': 0.16}

 48%|████▊     | 1598/3300 [3:54:41<4:20:48,  9.19s/it]INFO 04-13 11:15:39 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1599/3300 [3:54:49<4:07:04,  8.72s/it]2026-04-13 11:15:47,149 | INFO | train_grpo_train | metrics_logged mode=train step=1599

                                                       
{'loss': 0.0411, 'grad_norm': 11.115921020507812, 'learning_rate': 5.1575757575757575e-06, 'num_tokens': 2832890.0, 'completions/mean_length': 71.25, 'completions/min_length': 51.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.8588839769363403, 'rewards/meter/std': 0.20465929806232452, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8969908952713013, 'rewards/repeat_soft/std': 0.07345731556415558, 'rewards/judge_quality/mean': 0.6362500190734863, 'rewards/judge_quality/std': 0.18392062187194824, 'rewards/total_composite/mean': 0.6836203336715698, 'rewards/total_composite/std': 0.1292240470647812, 'reward': 0.6836203336715698, 'reward_std': 0.1292240470647812, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12643003463745117, 'sampling/sampling_logp_difference/max': 1.3608005046844482, 'sampling/importance_sampling_ratio/min': 0.2564553916454315, 'sampling/importance_sampling_ratio/mean': 1.0082217454910278, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7264133766293526, 'clip_ratio/low_mean': 0.08147330209612846, 'clip_ratio/low_min': 0.08147330209612846, 'clip_ratio/high_mean': 0.04042944964021444, 'clip_ratio/high_max': 0.04042944964021444, 'clip_ratio/region_mean': 0.12190275173634291, 'reward_total_mean': 0.6836203336715698, 'reward_meter_mean': 0.8588839769363403, 'reward_meter_std': 0.20465929806232452, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8969908952713013, 'reward_repeat_soft_std': 0.07345731556415558, 'reward_judge_quality_mean': 0.6362500190734863, 'reward_judge_quality_std': 0.18392062187194824, 'reward_total_composite_mean': 0.6836203336715698, 'reward_total_composite_std': 0.1292240470647812, 'epoch': 0.16}

 48%|████▊     | 1599/3300 [3:54:49<4:07:04,  8.72s/it]INFO 04-13 11:15:47 [block_pool.py:378] Successfully reset prefix cache

 48%|████▊     | 1600/3300 [3:55:01<4:34:13,  9.68s/it]2026-04-13 11:15:59,082 | INFO | train_grpo_train | metrics_logged mode=train step=1600

                                                       
{'loss': -0.1727, 'grad_norm': 2.6750028133392334, 'learning_rate': 5.154545454545456e-06, 'num_tokens': 2835263.0, 'completions/mean_length': 164.625, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 115.00000762939453, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.812462329864502, 'rewards/meter/std': 0.2528509795665741, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9228575229644775, 'rewards/repeat_soft/std': 0.06156541034579277, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.2768702805042267, 'rewards/total_composite/mean': 0.5253216028213501, 'rewards/total_composite/std': 0.23880189657211304, 'reward': 0.5253216028213501, 'reward_std': 0.23880188167095184, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1308581382036209, 'sampling/sampling_logp_difference/max': 2.108630657196045, 'sampling/importance_sampling_ratio/min': 0.12140409648418427, 'sampling/importance_sampling_ratio/mean': 1.0056959390640259, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7257308885455132, 'clip_ratio/low_mean': 0.027334056794643402, 'clip_ratio/low_min': 0.027334056794643402, 'clip_ratio/high_mean': 0.0635377811267972, 'clip_ratio/high_max': 0.0635377811267972, 'clip_ratio/region_mean': 0.0908718379214406, 'reward_total_mean': 0.5253216028213501, 'reward_meter_mean': 0.812462329864502, 'reward_meter_std': 0.2528509795665741, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9228575229644775, 'reward_repeat_soft_std': 0.06156541034579277, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.2768702805042267, 'reward_total_composite_mean': 0.5253216028213501, 'reward_total_composite_std': 0.23880189657211304, 'epoch': 0.16}

 48%|████▊     | 1600/3300 [3:55:01<4:34:13,  9.68s/it]INFO 04-13 11:15:59 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:08<00:34,  4.35s/it]

 30%|███       | 3/10 [00:13<00:31,  4.45s/it]

 40%|████      | 4/10 [00:17<00:26,  4.43s/it]

 50%|█████     | 5/10 [00:22<00:23,  4.64s/it]

 60%|██████    | 6/10 [00:32<00:24,  6.22s/it]

 70%|███████   | 7/10 [00:36<00:16,  5.51s/it]

 80%|████████  | 8/10 [00:39<00:09,  4.95s/it]

 90%|█████████ | 9/10 [00:48<00:06,  6.14s/it]

100%|██████████| 10/10 [00:52<00:00,  5.30s/it]2026-04-13 11:16:55,181 | INFO | train_grpo_train | metrics_logged mode=eval step=1600

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 56.006, 'eval_samples_per_second': 1.428, 'eval_steps_per_second': 0.179, 'eval_num_tokens': 2835263.0, 'eval_completions/mean_length': 96.5125, 'eval_completions/min_length': 39.7, 'eval_completions/max_length': 245.6, 'eval_completions/clipped_ratio': 0.0375, 'eval_completions/mean_terminated_length': 80.67321548461913, 'eval_completions/min_terminated_length': 39.7, 'eval_completions/max_terminated_length': 132.1, 'eval_rewards/meter/mean': 0.7673405587673188, 'eval_rewards/meter/std': 0.268173411488533, 'eval_rewards/count_adherence/mean': 0.9447916507720947, 'eval_rewards/count_adherence/std': 0.09196658246219158, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.1414213538169861, 'eval_rewards/repeat_soft/mean': 0.8982382655143738, 'eval_rewards/repeat_soft/std': 0.08441178202629089, 'eval_rewards/judge_quality/mean': 0.46950000524520874, 'eval_rewards/judge_quality/std': 0.1970653548836708, 'eval_rewards/total_composite/mean': 0.5493904024362564, 'eval_rewards/total_composite/std': 0.1925140403211117, 'eval_reward': 0.5493904024362564, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.062535485252738, 'eval_sampling/sampling_logp_difference/max': 1.031826877593994, 'eval_sampling/importance_sampling_ratio/min': 0.36239324510097504, 'eval_sampling/importance_sampling_ratio/mean': 1.0141837477684021, 'eval_sampling/importance_sampling_ratio/max': 1.4294534564018249, 'eval_entropy': 0.6990857660770416, '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': 0.5493904024362564, 'eval_reward_meter_mean': 0.7673405587673188, 'eval_reward_meter_std': 0.268173411488533, 'eval_reward_count_adherence_mean': 0.9447916507720947, 'eval_reward_count_adherence_std': 0.09196658246219158, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.1414213538169861, 'eval_reward_repeat_soft_mean': 0.8982382655143738, 'eval_reward_repeat_soft_std': 0.08441178202629089, 'eval_reward_judge_quality_mean': 0.46950000524520874, 'eval_reward_judge_quality_std': 0.1970653548836708, 'eval_reward_total_composite_mean': 0.5493904024362564, 'eval_reward_total_composite_std': 0.1925140403211117, 'epoch': 0.16}

 48%|████▊     | 1600/3300 [3:55:57<4:34:13,  9.68s/it]

100%|██████████| 10/10 [00:52<00:00,  5.30s/it]

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

 49%|████▊     | 1601/3300 [3:56:11<13:10:33, 27.92s/it]2026-04-13 11:17:09,559 | INFO | train_grpo_train | metrics_logged mode=train step=1601

                                                        
{'loss': -0.17, 'grad_norm': 2.645550012588501, 'learning_rate': 5.151515151515152e-06, 'num_tokens': 2837297.0, 'completions/mean_length': 148.25, 'completions/min_length': 80.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 96.28572082519531, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.44858697056770325, 'rewards/meter/std': 0.30310821533203125, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8310996294021606, 'rewards/repeat_soft/std': 0.048754941672086716, 'rewards/judge_quality/mean': 0.4775000214576721, 'rewards/judge_quality/std': 0.2539263069629669, 'rewards/total_composite/mean': 0.4004678726196289, 'rewards/total_composite/std': 0.18112020194530487, 'reward': 0.4004678726196289, 'reward_std': 0.18112018704414368, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10312364995479584, 'sampling/sampling_logp_difference/max': 1.91813325881958, 'sampling/importance_sampling_ratio/min': 0.1468808948993683, 'sampling/importance_sampling_ratio/mean': 0.9989551305770874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4168057069182396, 'clip_ratio/low_mean': 0.027486841194331646, 'clip_ratio/low_min': 0.027486841194331646, 'clip_ratio/high_mean': 0.05989459529519081, 'clip_ratio/high_max': 0.05989459529519081, 'clip_ratio/region_mean': 0.08738143648952246, 'reward_total_mean': 0.4004678726196289, 'reward_meter_mean': 0.44858697056770325, 'reward_meter_std': 0.30310821533203125, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8310996294021606, 'reward_repeat_soft_std': 0.048754941672086716, 'reward_judge_quality_mean': 0.4775000214576721, 'reward_judge_quality_std': 0.2539263069629669, 'reward_total_composite_mean': 0.4004678726196289, 'reward_total_composite_std': 0.18112020194530487, 'epoch': 0.16}

 49%|████▊     | 1601/3300 [3:56:11<13:10:33, 27.92s/it]INFO 04-13 11:17:09 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1602/3300 [3:56:23<10:51:59, 23.04s/it]2026-04-13 11:17:21,333 | INFO | train_grpo_train | metrics_logged mode=train step=1602

                                                        
{'loss': -0.1398, 'grad_norm': 2.3615663051605225, 'learning_rate': 5.148484848484849e-06, 'num_tokens': 2839115.0, 'completions/mean_length': 126.25, 'completions/min_length': 62.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 71.14286041259766, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8981894254684448, 'rewards/meter/std': 0.12483187019824982, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9875437021255493, 'rewards/repeat_soft/std': 0.012164885178208351, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22211645543575287, 'rewards/total_composite/mean': 0.5541506409645081, 'rewards/total_composite/std': 0.2510529160499573, 'reward': 0.5541506409645081, 'reward_std': 0.2510529160499573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13300660252571106, 'sampling/sampling_logp_difference/max': 2.366598129272461, 'sampling/importance_sampling_ratio/min': 0.09379927068948746, 'sampling/importance_sampling_ratio/mean': 0.9997671842575073, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8854727745056152, 'clip_ratio/low_mean': 0.01689189113676548, 'clip_ratio/low_min': 0.01689189113676548, 'clip_ratio/high_mean': 0.11515548266470432, 'clip_ratio/high_max': 0.11515548266470432, 'clip_ratio/region_mean': 0.1320473738014698, 'reward_total_mean': 0.5541506409645081, 'reward_meter_mean': 0.8981894254684448, 'reward_meter_std': 0.12483187019824982, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9875437021255493, 'reward_repeat_soft_std': 0.012164885178208351, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22211645543575287, 'reward_total_composite_mean': 0.5541506409645081, 'reward_total_composite_std': 0.2510529160499573, 'epoch': 0.16}

 49%|████▊     | 1602/3300 [3:56:23<10:51:59, 23.04s/it]INFO 04-13 11:17:21 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1603/3300 [3:56:29<8:31:58, 18.10s/it] 2026-04-13 11:17:27,844 | INFO | train_grpo_train | metrics_logged mode=train step=1603

                                                       
{'loss': 0.0289, 'grad_norm': 11.913492202758789, 'learning_rate': 5.145454545454546e-06, 'num_tokens': 2840919.0, 'completions/mean_length': 43.5, 'completions/min_length': 36.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8211958408355713, 'rewards/meter/std': 0.2598826587200165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9115462899208069, 'rewards/repeat_soft/std': 0.05229616537690163, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22403764724731445, 'rewards/total_composite/mean': 0.5429308414459229, 'rewards/total_composite/std': 0.07537403702735901, 'reward': 0.5429308414459229, 'reward_std': 0.07537402212619781, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11811105161905289, 'sampling/sampling_logp_difference/max': 1.1230762004852295, 'sampling/importance_sampling_ratio/min': 0.3252776265144348, 'sampling/importance_sampling_ratio/mean': 1.0160083770751953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8078811019659042, 'clip_ratio/low_mean': 0.05214684922248125, 'clip_ratio/low_min': 0.05214684922248125, 'clip_ratio/high_mean': 0.08267963957041502, 'clip_ratio/high_max': 0.08267963957041502, 'clip_ratio/region_mean': 0.13482648879289627, 'reward_total_mean': 0.5429308414459229, 'reward_meter_mean': 0.8211958408355713, 'reward_meter_std': 0.2598826587200165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9115462899208069, 'reward_repeat_soft_std': 0.05229616537690163, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22403764724731445, 'reward_total_composite_mean': 0.5429308414459229, 'reward_total_composite_std': 0.07537403702735901, 'epoch': 0.16}

 49%|████▊     | 1603/3300 [3:56:30<8:31:58, 18.10s/it]INFO 04-13 11:17:28 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1604/3300 [3:56:36<6:52:08, 14.58s/it]2026-04-13 11:17:34,211 | INFO | train_grpo_train | metrics_logged mode=train step=1604

                                                       
{'loss': 0.0867, 'grad_norm': 18.628082275390625, 'learning_rate': 5.142424242424243e-06, 'num_tokens': 2842277.0, 'completions/mean_length': 24.75, 'completions/min_length': 22.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8188531398773193, 'rewards/meter/std': 0.34643879532814026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.53125, 'rewards/judge_quality/std': 0.24833375215530396, 'rewards/total_composite/mean': 0.6387529373168945, 'rewards/total_composite/std': 0.20481234788894653, 'reward': 0.6387529373168945, 'reward_std': 0.20481234788894653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12999440729618073, 'sampling/sampling_logp_difference/max': 1.6062629222869873, 'sampling/importance_sampling_ratio/min': 0.20063599944114685, 'sampling/importance_sampling_ratio/mean': 1.0326086282730103, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9076575338840485, 'clip_ratio/low_mean': 0.08458688016980886, 'clip_ratio/low_min': 0.08458688016980886, 'clip_ratio/high_mean': 0.03840909153223038, 'clip_ratio/high_max': 0.03840909153223038, 'clip_ratio/region_mean': 0.12299597170203924, 'reward_total_mean': 0.6387529373168945, 'reward_meter_mean': 0.8188531398773193, 'reward_meter_std': 0.34643879532814026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.53125, 'reward_judge_quality_std': 0.24833375215530396, 'reward_total_composite_mean': 0.6387529373168945, 'reward_total_composite_std': 0.20481234788894653, 'epoch': 0.16}

 49%|████▊     | 1604/3300 [3:56:36<6:52:08, 14.58s/it]INFO 04-13 11:17:34 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1605/3300 [3:56:43<5:49:54, 12.39s/it]2026-04-13 11:17:41,428 | INFO | train_grpo_train | metrics_logged mode=train step=1605

                                                       
{'loss': 0.0535, 'grad_norm': 16.2254638671875, 'learning_rate': 5.139393939393939e-06, 'num_tokens': 2843690.0, 'completions/mean_length': 22.625, 'completions/min_length': 18.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9037623405456543, 'rewards/meter/std': 0.17803955078125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.3462499976158142, 'rewards/judge_quality/std': 0.10336308926343918, 'rewards/total_composite/mean': 0.5511980056762695, 'rewards/total_composite/std': 0.07901346683502197, 'reward': 0.5511980056762695, 'reward_std': 0.07901346683502197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13475042581558228, 'sampling/sampling_logp_difference/max': 1.7557687759399414, 'sampling/importance_sampling_ratio/min': 0.17277435958385468, 'sampling/importance_sampling_ratio/mean': 1.0087201595306396, 'sampling/importance_sampling_ratio/max': 1.7569578886032104, 'entropy': 0.8755036517977715, 'clip_ratio/low_mean': 0.022361111361533403, 'clip_ratio/low_min': 0.022361111361533403, 'clip_ratio/high_mean': 0.06772394385188818, 'clip_ratio/high_max': 0.06772394385188818, 'clip_ratio/region_mean': 0.09008505521342158, 'reward_total_mean': 0.5511980056762695, 'reward_meter_mean': 0.9037623405456543, 'reward_meter_std': 0.17803955078125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.3462499976158142, 'reward_judge_quality_std': 0.10336308926343918, 'reward_total_composite_mean': 0.5511980056762695, 'reward_total_composite_std': 0.07901346683502197, 'epoch': 0.16}

 49%|████▊     | 1605/3300 [3:56:43<5:49:54, 12.39s/it]INFO 04-13 11:17:41 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1606/3300 [3:56:49<4:58:39, 10.58s/it]2026-04-13 11:17:47,786 | INFO | train_grpo_train | metrics_logged mode=train step=1606

                                                       
{'loss': -0.0313, 'grad_norm': 14.100306510925293, 'learning_rate': 5.1363636363636375e-06, 'num_tokens': 2845067.0, 'completions/mean_length': 22.125, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.9127197265625, 'rewards/meter/std': 0.09622737020254135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9573202729225159, 'rewards/repeat_soft/std': 0.011277626268565655, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.21185828745365143, 'rewards/total_composite/mean': 0.6148068904876709, 'rewards/total_composite/std': 0.12625357508659363, 'reward': 0.6148068904876709, 'reward_std': 0.12625358998775482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1196405440568924, 'sampling/sampling_logp_difference/max': 1.4739208221435547, 'sampling/importance_sampling_ratio/min': 0.22902576625347137, 'sampling/importance_sampling_ratio/mean': 1.0227696895599365, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7830843478441238, 'clip_ratio/low_mean': 0.08167270477861166, 'clip_ratio/low_min': 0.08167270477861166, 'clip_ratio/high_mean': 0.03645833395421505, 'clip_ratio/high_max': 0.03645833395421505, 'clip_ratio/region_mean': 0.11813103873282671, 'reward_total_mean': 0.6148068904876709, 'reward_meter_mean': 0.9127197265625, 'reward_meter_std': 0.09622737020254135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9573202729225159, 'reward_repeat_soft_std': 0.011277626268565655, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.21185828745365143, 'reward_total_composite_mean': 0.6148068904876709, 'reward_total_composite_std': 0.12625357508659363, 'epoch': 0.16}

 49%|████▊     | 1606/3300 [3:56:50<4:58:39, 10.58s/it]INFO 04-13 11:17:48 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1607/3300 [3:56:56<4:21:24,  9.26s/it]2026-04-13 11:17:53,982 | INFO | train_grpo_train | metrics_logged mode=train step=1607

                                                       
{'loss': 0.0546, 'grad_norm': 13.587140083312988, 'learning_rate': 5.133333333333334e-06, 'num_tokens': 2846618.0, 'completions/mean_length': 40.875, 'completions/min_length': 37.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.875, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6240348815917969, 'rewards/meter/std': 0.4144648611545563, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9673728942871094, 'rewards/repeat_soft/std': 0.04772808402776718, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.605128288269043, 'rewards/total_composite/std': 0.2106853872537613, 'reward': 0.605128288269043, 'reward_std': 0.2106853872537613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10679075866937637, 'sampling/sampling_logp_difference/max': 3.874640464782715, 'sampling/importance_sampling_ratio/min': 0.020761799067258835, 'sampling/importance_sampling_ratio/mean': 1.0191521644592285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5109239295125008, 'clip_ratio/low_mean': 0.07677730452269316, 'clip_ratio/low_min': 0.07677730452269316, 'clip_ratio/high_mean': 0.03319597104564309, 'clip_ratio/high_max': 0.03319597104564309, 'clip_ratio/region_mean': 0.10997327556833625, 'reward_total_mean': 0.605128288269043, 'reward_meter_mean': 0.6240348815917969, 'reward_meter_std': 0.4144648611545563, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9673728942871094, 'reward_repeat_soft_std': 0.04772808402776718, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.605128288269043, 'reward_total_composite_std': 0.2106853872537613, 'epoch': 0.16}

 49%|████▊     | 1607/3300 [3:56:56<4:21:24,  9.26s/it]INFO 04-13 11:17:54 [block_pool.py:378] Successfully reset prefix cache

 49%|████▊     | 1608/3300 [3:57:03<4:06:05,  8.73s/it]2026-04-13 11:18:01,452 | INFO | train_grpo_train | metrics_logged mode=train step=1608

                                                       
{'loss': -0.0536, 'grad_norm': 7.547384262084961, 'learning_rate': 5.130303030303031e-06, 'num_tokens': 2849068.0, 'completions/mean_length': 109.25, 'completions/min_length': 92.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.25, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9808225631713867, 'rewards/meter/std': 0.0065976474434137344, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.841166615486145, 'rewards/repeat_soft/std': 0.06930014491081238, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5404136776924133, 'rewards/total_composite/std': 0.0432850681245327, 'reward': 0.5404136776924133, 'reward_std': 0.043285053223371506, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10502967983484268, 'sampling/sampling_logp_difference/max': 1.924008846282959, 'sampling/importance_sampling_ratio/min': 0.14602041244506836, 'sampling/importance_sampling_ratio/mean': 1.0074819326400757, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7072211876511574, 'clip_ratio/low_mean': 0.020380434580147266, 'clip_ratio/low_min': 0.020380434580147266, 'clip_ratio/high_mean': 0.07942765299230814, 'clip_ratio/high_max': 0.07942765299230814, 'clip_ratio/region_mean': 0.0998080875724554, 'reward_total_mean': 0.5404136776924133, 'reward_meter_mean': 0.9808225631713867, 'reward_meter_std': 0.0065976474434137344, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.841166615486145, 'reward_repeat_soft_std': 0.06930014491081238, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5404136776924133, 'reward_total_composite_std': 0.0432850681245327, 'epoch': 0.16}

 49%|████▊     | 1608/3300 [3:57:03<4:06:05,  8.73s/it]INFO 04-13 11:18:01 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1609/3300 [3:57:14<4:23:38,  9.35s/it]2026-04-13 11:18:12,270 | INFO | train_grpo_train | metrics_logged mode=train step=1609

                                                       
{'loss': -0.0007, 'grad_norm': 16.270160675048828, 'learning_rate': 5.1272727272727275e-06, 'num_tokens': 2850776.0, 'completions/mean_length': 47.5, 'completions/min_length': 37.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.847195029258728, 'rewards/meter/std': 0.323220819234848, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9017714262008667, 'rewards/repeat_soft/std': 0.030768852680921555, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5402442216873169, 'rewards/total_composite/std': 0.09381219744682312, 'reward': 0.5402442216873169, 'reward_std': 0.09381220489740372, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10080137848854065, 'sampling/sampling_logp_difference/max': 1.389388084411621, 'sampling/importance_sampling_ratio/min': 0.24922776222229004, 'sampling/importance_sampling_ratio/mean': 0.9930405616760254, 'sampling/importance_sampling_ratio/max': 1.710292935371399, 'entropy': 0.6024736911058426, 'clip_ratio/low_mean': 0.039260704070329666, 'clip_ratio/low_min': 0.039260704070329666, 'clip_ratio/high_mean': 0.03794161765836179, 'clip_ratio/high_max': 0.03794161765836179, 'clip_ratio/region_mean': 0.07720232172869146, 'reward_total_mean': 0.5402442216873169, 'reward_meter_mean': 0.847195029258728, 'reward_meter_std': 0.323220819234848, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9017714262008667, 'reward_repeat_soft_std': 0.030768852680921555, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5402442216873169, 'reward_total_composite_std': 0.09381219744682312, 'epoch': 0.16}

 49%|████▉     | 1609/3300 [3:57:14<4:23:38,  9.35s/it]INFO 04-13 11:18:12 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1610/3300 [3:57:25<4:41:49, 10.01s/it]2026-04-13 11:18:23,798 | INFO | train_grpo_train | metrics_logged mode=train step=1610

                                                       
{'loss': -0.1903, 'grad_norm': 2.445298433303833, 'learning_rate': 5.124242424242425e-06, 'num_tokens': 2853113.0, 'completions/mean_length': 186.125, 'completions/min_length': 127.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 139.57144165039062, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.8654612302780151, 'rewards/meter/std': 0.17720764875411987, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.0589255727827549, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8106361031532288, 'rewards/repeat_soft/std': 0.029637936502695084, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.21633309125900269, 'rewards/total_composite/mean': 0.4939092993736267, 'rewards/total_composite/std': 0.229624405503273, 'reward': 0.4939092993736267, 'reward_std': 0.22962439060211182, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09812058508396149, 'sampling/sampling_logp_difference/max': 1.952284812927246, 'sampling/importance_sampling_ratio/min': 0.1419493705034256, 'sampling/importance_sampling_ratio/mean': 1.0083082914352417, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.561351902782917, 'clip_ratio/low_mean': 0.02542989421635866, 'clip_ratio/low_min': 0.02542989421635866, 'clip_ratio/high_mean': 0.05852258764207363, 'clip_ratio/high_max': 0.05852258764207363, 'clip_ratio/region_mean': 0.08395248185843229, 'reward_total_mean': 0.4939092993736267, 'reward_meter_mean': 0.8654612302780151, 'reward_meter_std': 0.17720764875411987, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.0589255727827549, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8106361031532288, 'reward_repeat_soft_std': 0.029637936502695084, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.21633309125900269, 'reward_total_composite_mean': 0.4939092993736267, 'reward_total_composite_std': 0.229624405503273, 'epoch': 0.16}

 49%|████▉     | 1610/3300 [3:57:26<4:41:49, 10.01s/it]INFO 04-13 11:18:24 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1611/3300 [3:57:31<4:06:52,  8.77s/it]2026-04-13 11:18:29,682 | INFO | train_grpo_train | metrics_logged mode=train step=1611

                                                       
{'loss': 0.0264, 'grad_norm': 12.777203559875488, 'learning_rate': 5.121212121212121e-06, 'num_tokens': 2854664.0, 'completions/mean_length': 42.875, 'completions/min_length': 39.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6671066284179688, 'rewards/meter/std': 0.336922287940979, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9815101623535156, 'rewards/repeat_soft/std': 0.018132537603378296, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.6651303768157959, 'rewards/total_composite/std': 0.18651041388511658, 'reward': 0.6651303768157959, 'reward_std': 0.18651039898395538, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10066421329975128, 'sampling/sampling_logp_difference/max': 1.44620943069458, 'sampling/importance_sampling_ratio/min': 0.23546114563941956, 'sampling/importance_sampling_ratio/mean': 0.9972047209739685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44500717893242836, 'clip_ratio/low_mean': 0.03842564998194575, 'clip_ratio/low_min': 0.03842564998194575, 'clip_ratio/high_mean': 0.04087729286402464, 'clip_ratio/high_max': 0.04087729286402464, 'clip_ratio/region_mean': 0.07930294284597039, 'reward_total_mean': 0.6651303768157959, 'reward_meter_mean': 0.6671066284179688, 'reward_meter_std': 0.336922287940979, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9815101623535156, 'reward_repeat_soft_std': 0.018132537603378296, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.6651303768157959, 'reward_total_composite_std': 0.18651041388511658, 'epoch': 0.16}

 49%|████▉     | 1611/3300 [3:57:31<4:06:52,  8.77s/it]INFO 04-13 11:18:29 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1612/3300 [3:57:39<3:53:27,  8.30s/it]2026-04-13 11:18:36,877 | INFO | train_grpo_train | metrics_logged mode=train step=1612

                                                       
{'loss': 0.0258, 'grad_norm': 7.783813953399658, 'learning_rate': 5.1181818181818185e-06, 'num_tokens': 2857219.0, 'completions/mean_length': 114.375, 'completions/min_length': 104.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.375, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.9686040282249451, 'rewards/meter/std': 0.022422077134251595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8075989484786987, 'rewards/repeat_soft/std': 0.07894428819417953, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6254061460494995, 'rewards/total_composite/std': 0.11612702906131744, 'reward': 0.6254061460494995, 'reward_std': 0.11612702161073685, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09219850599765778, 'sampling/sampling_logp_difference/max': 1.4588334560394287, 'sampling/importance_sampling_ratio/min': 0.2325073480606079, 'sampling/importance_sampling_ratio/mean': 1.004338264465332, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6646230891346931, 'clip_ratio/low_mean': 0.06814689747989178, 'clip_ratio/low_min': 0.06814689747989178, 'clip_ratio/high_mean': 0.009259259328246117, 'clip_ratio/high_max': 0.009259259328246117, 'clip_ratio/region_mean': 0.0774061568081379, 'reward_total_mean': 0.6254061460494995, 'reward_meter_mean': 0.9686040282249451, 'reward_meter_std': 0.022422077134251595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8075989484786987, 'reward_repeat_soft_std': 0.07894428819417953, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6254061460494995, 'reward_total_composite_std': 0.11612702906131744, 'epoch': 0.16}

 49%|████▉     | 1612/3300 [3:57:39<3:53:27,  8.30s/it]INFO 04-13 11:18:37 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1613/3300 [3:57:51<4:28:11,  9.54s/it]2026-04-13 11:18:49,319 | INFO | train_grpo_train | metrics_logged mode=train step=1613

                                                       
{'loss': -0.2401, 'grad_norm': 1.867802381515503, 'learning_rate': 5.115151515151515e-06, 'num_tokens': 2859937.0, 'completions/mean_length': 199.75, 'completions/min_length': 127.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 155.1428680419922, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.9049348831176758, 'rewards/meter/std': 0.14540007710456848, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8620929718017578, 'rewards/repeat_soft/std': 0.07246813923120499, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.48327773809432983, 'rewards/total_composite/std': 0.20073343813419342, 'reward': 0.48327773809432983, 'reward_std': 0.20073342323303223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10677609592676163, 'sampling/sampling_logp_difference/max': 2.683267593383789, 'sampling/importance_sampling_ratio/min': 0.06833948194980621, 'sampling/importance_sampling_ratio/mean': 1.0043482780456543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5648591257631779, 'clip_ratio/low_mean': 0.012867647223174572, 'clip_ratio/low_min': 0.012867647223174572, 'clip_ratio/high_mean': 0.06916146818548441, 'clip_ratio/high_max': 0.06916146818548441, 'clip_ratio/region_mean': 0.08202911540865898, 'reward_total_mean': 0.48327773809432983, 'reward_meter_mean': 0.9049348831176758, 'reward_meter_std': 0.14540007710456848, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.1414213478565216, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8620929718017578, 'reward_repeat_soft_std': 0.07246813923120499, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.48327773809432983, 'reward_total_composite_std': 0.20073343813419342, 'epoch': 0.16}

 49%|████▉     | 1613/3300 [3:57:51<4:28:11,  9.54s/it]INFO 04-13 11:18:49 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1614/3300 [3:58:02<4:42:07, 10.04s/it]2026-04-13 11:19:00,532 | INFO | train_grpo_train | metrics_logged mode=train step=1614

                                                       
{'loss': -0.1467, 'grad_norm': 2.9070706367492676, 'learning_rate': 5.112121212121213e-06, 'num_tokens': 2861901.0, 'completions/mean_length': 130.5, 'completions/min_length': 68.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 76.0, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 85.0, 'rewards/meter/mean': 0.8117458820343018, 'rewards/meter/std': 0.31275635957717896, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7944717407226562, 'rewards/repeat_soft/std': 0.10366769880056381, 'rewards/judge_quality/mean': 0.33125001192092896, 'rewards/judge_quality/std': 0.13715866208076477, 'rewards/total_composite/mean': 0.46274441480636597, 'rewards/total_composite/std': 0.21137595176696777, 'reward': 0.46274441480636597, 'reward_std': 0.21137595176696777, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08003520220518112, 'sampling/sampling_logp_difference/max': 1.8128626346588135, 'sampling/importance_sampling_ratio/min': 0.16318632662296295, 'sampling/importance_sampling_ratio/mean': 1.011260986328125, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34130437672138214, 'clip_ratio/low_mean': 0.004687500186264515, 'clip_ratio/low_min': 0.004687500186264515, 'clip_ratio/high_mean': 0.05632254155352712, 'clip_ratio/high_max': 0.05632254155352712, 'clip_ratio/region_mean': 0.06101004173979163, 'reward_total_mean': 0.46274441480636597, 'reward_meter_mean': 0.8117458820343018, 'reward_meter_std': 0.31275635957717896, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7944717407226562, 'reward_repeat_soft_std': 0.10366769880056381, 'reward_judge_quality_mean': 0.33125001192092896, 'reward_judge_quality_std': 0.13715866208076477, 'reward_total_composite_mean': 0.46274441480636597, 'reward_total_composite_std': 0.21137595176696777, 'epoch': 0.16}

 49%|████▉     | 1614/3300 [3:58:02<4:42:07, 10.04s/it]INFO 04-13 11:19:00 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1615/3300 [3:58:09<4:11:50,  8.97s/it]2026-04-13 11:19:06,993 | INFO | train_grpo_train | metrics_logged mode=train step=1615

                                                       
{'loss': 0.0854, 'grad_norm': 10.08071517944336, 'learning_rate': 5.109090909090909e-06, 'num_tokens': 2863527.0, 'completions/mean_length': 50.25, 'completions/min_length': 43.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.25, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8818612694740295, 'rewards/meter/std': 0.27754804491996765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8982758522033691, 'rewards/repeat_soft/std': 0.07960784435272217, 'rewards/judge_quality/mean': 0.606249988079071, 'rewards/judge_quality/std': 0.2492811232805252, 'rewards/total_composite/mean': 0.693257212638855, 'rewards/total_composite/std': 0.19573436677455902, 'reward': 0.693257212638855, 'reward_std': 0.19573435187339783, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11168952286243439, 'sampling/sampling_logp_difference/max': 1.2561886310577393, 'sampling/importance_sampling_ratio/min': 0.2847371995449066, 'sampling/importance_sampling_ratio/mean': 1.0139323472976685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7719178572297096, 'clip_ratio/low_mean': 0.053172151325270534, 'clip_ratio/low_min': 0.053172151325270534, 'clip_ratio/high_mean': 0.045388901606202126, 'clip_ratio/high_max': 0.045388901606202126, 'clip_ratio/region_mean': 0.09856105293147266, 'reward_total_mean': 0.693257212638855, 'reward_meter_mean': 0.8818612694740295, 'reward_meter_std': 0.27754804491996765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8982758522033691, 'reward_repeat_soft_std': 0.07960784435272217, 'reward_judge_quality_mean': 0.606249988079071, 'reward_judge_quality_std': 0.2492811232805252, 'reward_total_composite_mean': 0.693257212638855, 'reward_total_composite_std': 0.19573436677455902, 'epoch': 0.16}

 49%|████▉     | 1615/3300 [3:58:09<4:11:50,  8.97s/it]INFO 04-13 11:19:07 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1616/3300 [3:58:15<3:45:59,  8.05s/it]2026-04-13 11:19:12,910 | INFO | train_grpo_train | metrics_logged mode=train step=1616

                                                       
{'loss': 0.095, 'grad_norm': 19.3624210357666, 'learning_rate': 5.106060606060607e-06, 'num_tokens': 2864983.0, 'completions/mean_length': 25.0, 'completions/min_length': 17.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9213307499885559, 'rewards/meter/std': 0.14470230042934418, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9494390487670898, 'rewards/repeat_soft/std': 0.03694179654121399, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5803236961364746, 'rewards/total_composite/std': 0.04910682886838913, 'reward': 0.5803236961364746, 'reward_std': 0.04910682141780853, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14111049473285675, 'sampling/sampling_logp_difference/max': 1.9327747821807861, 'sampling/importance_sampling_ratio/min': 0.14474600553512573, 'sampling/importance_sampling_ratio/mean': 1.0063116550445557, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9546639770269394, 'clip_ratio/low_mean': 0.047727273777127266, 'clip_ratio/low_min': 0.047727273777127266, 'clip_ratio/high_mean': 0.1064573572948575, 'clip_ratio/high_max': 0.1064573572948575, 'clip_ratio/region_mean': 0.15418463107198477, 'reward_total_mean': 0.5803236961364746, 'reward_meter_mean': 0.9213307499885559, 'reward_meter_std': 0.14470230042934418, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9494390487670898, 'reward_repeat_soft_std': 0.03694179654121399, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5803236961364746, 'reward_total_composite_std': 0.04910682886838913, 'epoch': 0.16}

 49%|████▉     | 1616/3300 [3:58:15<3:45:59,  8.05s/it]INFO 04-13 11:19:13 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1617/3300 [3:58:21<3:32:47,  7.59s/it]2026-04-13 11:19:19,413 | INFO | train_grpo_train | metrics_logged mode=train step=1617

                                                       
{'loss': 0.0395, 'grad_norm': 9.595121383666992, 'learning_rate': 5.103030303030303e-06, 'num_tokens': 2866961.0, 'completions/mean_length': 71.25, 'completions/min_length': 60.0, 'completions/max_length': 82.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 71.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 82.0, 'rewards/meter/mean': 0.9378806352615356, 'rewards/meter/std': 0.08389707654714584, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.956716775894165, 'rewards/repeat_soft/std': 0.005239794030785561, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.15645630657672882, 'rewards/total_composite/mean': 0.7527035474777222, 'rewards/total_composite/std': 0.08143587410449982, 'reward': 0.7527035474777222, 'reward_std': 0.08143586665391922, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10567019134759903, 'sampling/sampling_logp_difference/max': 1.5408134460449219, 'sampling/importance_sampling_ratio/min': 0.21420679986476898, 'sampling/importance_sampling_ratio/mean': 1.0093913078308105, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5745541453361511, 'clip_ratio/low_mean': 0.01658957451581955, 'clip_ratio/low_min': 0.01658957451581955, 'clip_ratio/high_mean': 0.07150601642206311, 'clip_ratio/high_max': 0.07150601642206311, 'clip_ratio/region_mean': 0.08809559093788266, 'reward_total_mean': 0.7527035474777222, 'reward_meter_mean': 0.9378806352615356, 'reward_meter_std': 0.08389707654714584, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.956716775894165, 'reward_repeat_soft_std': 0.005239794030785561, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.15645630657672882, 'reward_total_composite_mean': 0.7527035474777222, 'reward_total_composite_std': 0.08143587410449982, 'epoch': 0.16}

 49%|████▉     | 1617/3300 [3:58:21<3:32:47,  7.59s/it]INFO 04-13 11:19:19 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1618/3300 [3:58:27<3:17:19,  7.04s/it]2026-04-13 11:19:25,159 | INFO | train_grpo_train | metrics_logged mode=train step=1618

                                                       
{'loss': 0.0556, 'grad_norm': 13.132983207702637, 'learning_rate': 5.1e-06, 'num_tokens': 2868520.0, 'completions/mean_length': 35.875, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7930529713630676, 'rewards/meter/std': 0.27568158507347107, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9989027976989746, 'rewards/repeat_soft/std': 0.002089538611471653, 'rewards/judge_quality/mean': 0.6112500429153442, 'rewards/judge_quality/std': 0.25587037205696106, 'rewards/total_composite/mean': 0.6624631285667419, 'rewards/total_composite/std': 0.17252354323863983, 'reward': 0.6624631285667419, 'reward_std': 0.17252352833747864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08907654136419296, 'sampling/sampling_logp_difference/max': 1.5596189498901367, 'sampling/importance_sampling_ratio/min': 0.4622865915298462, 'sampling/importance_sampling_ratio/mean': 1.020411491394043, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3689921572804451, 'clip_ratio/low_mean': 0.06342792743816972, 'clip_ratio/low_min': 0.06342792743816972, 'clip_ratio/high_mean': 0.029411765281111002, 'clip_ratio/high_max': 0.029411765281111002, 'clip_ratio/region_mean': 0.09283969271928072, 'reward_total_mean': 0.6624631285667419, 'reward_meter_mean': 0.7930529713630676, 'reward_meter_std': 0.27568158507347107, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9989027976989746, 'reward_repeat_soft_std': 0.002089538611471653, 'reward_judge_quality_mean': 0.6112500429153442, 'reward_judge_quality_std': 0.25587037205696106, 'reward_total_composite_mean': 0.6624631285667419, 'reward_total_composite_std': 0.17252354323863983, 'epoch': 0.16}

 49%|████▉     | 1618/3300 [3:58:27<3:17:19,  7.04s/it]INFO 04-13 11:19:25 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1619/3300 [3:58:38<3:53:29,  8.33s/it]2026-04-13 11:19:36,516 | INFO | train_grpo_train | metrics_logged mode=train step=1619

                                                       
{'loss': -0.1018, 'grad_norm': 3.055025815963745, 'learning_rate': 5.096969696969697e-06, 'num_tokens': 2870943.0, 'completions/mean_length': 188.875, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 142.71429443359375, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.3641549348831177, 'rewards/meter/std': 0.39495429396629333, 'rewards/count_adherence/mean': 0.8541666269302368, 'rewards/count_adherence/std': 0.2878147065639496, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8099063038825989, 'rewards/repeat_soft/std': 0.1071389839053154, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.34355688095092773, 'rewards/total_composite/std': 0.17706337571144104, 'reward': 0.34355688095092773, 'reward_std': 0.17706336081027985, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1074066087603569, 'sampling/sampling_logp_difference/max': 1.994025707244873, 'sampling/importance_sampling_ratio/min': 0.13614623248577118, 'sampling/importance_sampling_ratio/mean': 1.004380464553833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4460967816412449, 'clip_ratio/low_mean': 0.046347555704414845, 'clip_ratio/low_min': 0.046347555704414845, 'clip_ratio/high_mean': 0.029849925078451633, 'clip_ratio/high_max': 0.029849925078451633, 'clip_ratio/region_mean': 0.07619748078286648, 'reward_total_mean': 0.34355688095092773, 'reward_meter_mean': 0.3641549348831177, 'reward_meter_std': 0.39495429396629333, 'reward_count_adherence_mean': 0.8541666269302368, 'reward_count_adherence_std': 0.2878147065639496, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8099063038825989, 'reward_repeat_soft_std': 0.1071389839053154, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.34355688095092773, 'reward_total_composite_std': 0.17706337571144104, 'epoch': 0.16}

 49%|████▉     | 1619/3300 [3:58:38<3:53:29,  8.33s/it]INFO 04-13 11:19:36 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1620/3300 [3:58:46<3:50:26,  8.23s/it]2026-04-13 11:19:44,518 | INFO | train_grpo_train | metrics_logged mode=train step=1620

                                                       
{'loss': 0.0153, 'grad_norm': 7.13424825668335, 'learning_rate': 5.093939393939395e-06, 'num_tokens': 2873349.0, 'completions/mean_length': 121.75, 'completions/min_length': 105.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.75, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.7869986295700073, 'rewards/meter/std': 0.22102504968643188, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8619891405105591, 'rewards/repeat_soft/std': 0.059215281158685684, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5751908421516418, 'rewards/total_composite/std': 0.13902568817138672, 'reward': 0.5751908421516418, 'reward_std': 0.13902568817138672, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12312452495098114, 'sampling/sampling_logp_difference/max': 2.1598167419433594, 'sampling/importance_sampling_ratio/min': 0.115346260368824, 'sampling/importance_sampling_ratio/mean': 1.0205926895141602, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0005771592259407, 'clip_ratio/low_mean': 0.04885080549865961, 'clip_ratio/low_min': 0.04885080549865961, 'clip_ratio/high_mean': 0.06665956694632769, 'clip_ratio/high_max': 0.06665956694632769, 'clip_ratio/region_mean': 0.1155103724449873, 'reward_total_mean': 0.5751908421516418, 'reward_meter_mean': 0.7869986295700073, 'reward_meter_std': 0.22102504968643188, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8619891405105591, 'reward_repeat_soft_std': 0.059215281158685684, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5751908421516418, 'reward_total_composite_std': 0.13902568817138672, 'epoch': 0.16}

 49%|████▉     | 1620/3300 [3:58:46<3:50:26,  8.23s/it]INFO 04-13 11:19:44 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1621/3300 [3:58:59<4:30:40,  9.67s/it]2026-04-13 11:19:57,548 | INFO | train_grpo_train | metrics_logged mode=train step=1621

                                                       
{'loss': -0.0467, 'grad_norm': 4.196609973907471, 'learning_rate': 5.090909090909091e-06, 'num_tokens': 2874984.0, 'completions/mean_length': 104.375, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 46.142860412597656, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.4812016189098358, 'rewards/meter/std': 0.37489452958106995, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9685986042022705, 'rewards/repeat_soft/std': 0.06977908313274384, 'rewards/judge_quality/mean': 0.5612500309944153, 'rewards/judge_quality/std': 0.3223324716091156, 'rewards/total_composite/mean': 0.43485894799232483, 'rewards/total_composite/std': 0.23783420026302338, 'reward': 0.43485894799232483, 'reward_std': 0.23783420026302338, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12444452941417694, 'sampling/sampling_logp_difference/max': 2.1760754585266113, 'sampling/importance_sampling_ratio/min': 0.11348603665828705, 'sampling/importance_sampling_ratio/mean': 1.0035669803619385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4616983085870743, 'clip_ratio/low_mean': 0.037504821084439754, 'clip_ratio/low_min': 0.037504821084439754, 'clip_ratio/high_mean': 0.06014752900227904, 'clip_ratio/high_max': 0.06014752900227904, 'clip_ratio/region_mean': 0.0976523500867188, 'reward_total_mean': 0.43485894799232483, 'reward_meter_mean': 0.4812016189098358, 'reward_meter_std': 0.37489452958106995, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9685986042022705, 'reward_repeat_soft_std': 0.06977908313274384, 'reward_judge_quality_mean': 0.5612500309944153, 'reward_judge_quality_std': 0.3223324716091156, 'reward_total_composite_mean': 0.43485894799232483, 'reward_total_composite_std': 0.23783420026302338, 'epoch': 0.16}

 49%|████▉     | 1621/3300 [3:58:59<4:30:40,  9.67s/it]INFO 04-13 11:19:57 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1622/3300 [3:59:05<3:59:36,  8.57s/it]2026-04-13 11:20:03,534 | INFO | train_grpo_train | metrics_logged mode=train step=1622

                                                       
{'loss': -0.0033, 'grad_norm': 12.808685302734375, 'learning_rate': 5.0878787878787885e-06, 'num_tokens': 2876610.0, 'completions/mean_length': 42.25, 'completions/min_length': 37.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.761579155921936, 'rewards/meter/std': 0.2876737117767334, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9501906633377075, 'rewards/repeat_soft/std': 0.05366018787026405, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5522565841674805, 'rewards/total_composite/std': 0.07785069197416306, 'reward': 0.5522565841674805, 'reward_std': 0.07785066962242126, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11096014082431793, 'sampling/sampling_logp_difference/max': 1.0162148475646973, 'sampling/importance_sampling_ratio/min': 0.3619624376296997, 'sampling/importance_sampling_ratio/mean': 0.996402382850647, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5768076628446579, 'clip_ratio/low_mean': 0.0435085401404649, 'clip_ratio/low_min': 0.0435085401404649, 'clip_ratio/high_mean': 0.0699709472246468, 'clip_ratio/high_max': 0.0699709472246468, 'clip_ratio/region_mean': 0.11347948736511171, 'reward_total_mean': 0.5522565841674805, 'reward_meter_mean': 0.761579155921936, 'reward_meter_std': 0.2876737117767334, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9501906633377075, 'reward_repeat_soft_std': 0.05366018787026405, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5522565841674805, 'reward_total_composite_std': 0.07785069197416306, 'epoch': 0.16}

 49%|████▉     | 1622/3300 [3:59:05<3:59:36,  8.57s/it]INFO 04-13 11:20:03 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1623/3300 [3:59:11<3:36:58,  7.76s/it]2026-04-13 11:20:09,424 | INFO | train_grpo_train | metrics_logged mode=train step=1623

                                                       
{'loss': 0.0678, 'grad_norm': 17.057676315307617, 'learning_rate': 5.084848484848486e-06, 'num_tokens': 2878072.0, 'completions/mean_length': 35.75, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5909867882728577, 'rewards/meter/std': 0.37435105443000793, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9783687591552734, 'rewards/repeat_soft/std': 0.03977498412132263, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.542860746383667, 'rewards/total_composite/std': 0.14172787964344025, 'reward': 0.542860746383667, 'reward_std': 0.14172787964344025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1391218900680542, 'sampling/sampling_logp_difference/max': 2.116826057434082, 'sampling/importance_sampling_ratio/min': 0.12041321396827698, 'sampling/importance_sampling_ratio/mean': 0.9980522990226746, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47563890740275383, 'clip_ratio/low_mean': 0.04050394566729665, 'clip_ratio/low_min': 0.04050394566729665, 'clip_ratio/high_mean': 0.09049043618142605, 'clip_ratio/high_max': 0.09049043618142605, 'clip_ratio/region_mean': 0.1309943818487227, 'reward_total_mean': 0.542860746383667, 'reward_meter_mean': 0.5909867882728577, 'reward_meter_std': 0.37435105443000793, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9783687591552734, 'reward_repeat_soft_std': 0.03977498412132263, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.542860746383667, 'reward_total_composite_std': 0.14172787964344025, 'epoch': 0.16}

 49%|████▉     | 1623/3300 [3:59:11<3:36:58,  7.76s/it]INFO 04-13 11:20:09 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1624/3300 [3:59:18<3:29:13,  7.49s/it]2026-04-13 11:20:16,284 | INFO | train_grpo_train | metrics_logged mode=train step=1624

                                                       
{'loss': -0.0129, 'grad_norm': 8.933954238891602, 'learning_rate': 5.081818181818182e-06, 'num_tokens': 2880331.0, 'completions/mean_length': 87.375, 'completions/min_length': 69.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.8871557712554932, 'rewards/meter/std': 0.16150468587875366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.828206479549408, 'rewards/repeat_soft/std': 0.04936762526631355, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.5923798084259033, 'rewards/total_composite/std': 0.08415987342596054, 'reward': 0.5923798084259033, 'reward_std': 0.08415987342596054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1125749796628952, 'sampling/sampling_logp_difference/max': 1.4348459243774414, 'sampling/importance_sampling_ratio/min': 0.23815205693244934, 'sampling/importance_sampling_ratio/mean': 1.0075618028640747, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7525682523846626, 'clip_ratio/low_mean': 0.03880079975351691, 'clip_ratio/low_min': 0.03880079975351691, 'clip_ratio/high_mean': 0.05439908429980278, 'clip_ratio/high_max': 0.05439908429980278, 'clip_ratio/region_mean': 0.09319988405331969, 'reward_total_mean': 0.5923798084259033, 'reward_meter_mean': 0.8871557712554932, 'reward_meter_std': 0.16150468587875366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.828206479549408, 'reward_repeat_soft_std': 0.04936762526631355, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.5923798084259033, 'reward_total_composite_std': 0.08415987342596054, 'epoch': 0.16}

 49%|████▉     | 1624/3300 [3:59:18<3:29:13,  7.49s/it]INFO 04-13 11:20:16 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1625/3300 [3:59:24<3:19:48,  7.16s/it]2026-04-13 11:20:22,660 | INFO | train_grpo_train | metrics_logged mode=train step=1625

                                                       
{'loss': 0.0353, 'grad_norm': 16.366573333740234, 'learning_rate': 5.078787878787879e-06, 'num_tokens': 2881690.0, 'completions/mean_length': 21.875, 'completions/min_length': 19.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.7903445959091187, 'rewards/meter/std': 0.32402628660202026, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.937935471534729, 'rewards/repeat_soft/std': 0.040868453681468964, 'rewards/judge_quality/mean': 0.42124998569488525, 'rewards/judge_quality/std': 0.06998724490404129, 'rewards/total_composite/mean': 0.5541468262672424, 'rewards/total_composite/std': 0.09281747043132782, 'reward': 0.5541468262672424, 'reward_std': 0.09281745553016663, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.144802987575531, 'sampling/sampling_logp_difference/max': 1.0754953622817993, 'sampling/importance_sampling_ratio/min': 0.3917113244533539, 'sampling/importance_sampling_ratio/mean': 1.0188369750976562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7806518524885178, 'clip_ratio/low_mean': 0.05853242985904217, 'clip_ratio/low_min': 0.05853242985904217, 'clip_ratio/high_mean': 0.08434872049838305, 'clip_ratio/high_max': 0.08434872049838305, 'clip_ratio/region_mean': 0.1428811503574252, 'reward_total_mean': 0.5541468262672424, 'reward_meter_mean': 0.7903445959091187, 'reward_meter_std': 0.32402628660202026, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.937935471534729, 'reward_repeat_soft_std': 0.040868453681468964, 'reward_judge_quality_mean': 0.42124998569488525, 'reward_judge_quality_std': 0.06998724490404129, 'reward_total_composite_mean': 0.5541468262672424, 'reward_total_composite_std': 0.09281747043132782, 'epoch': 0.16}

 49%|████▉     | 1625/3300 [3:59:24<3:19:48,  7.16s/it]INFO 04-13 11:20:22 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1626/3300 [3:59:32<3:20:58,  7.20s/it]2026-04-13 11:20:29,968 | INFO | train_grpo_train | metrics_logged mode=train step=1626

                                                       
{'loss': -0.014, 'grad_norm': 9.989459991455078, 'learning_rate': 5.075757575757576e-06, 'num_tokens': 2883911.0, 'completions/mean_length': 103.625, 'completions/min_length': 84.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.625, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.778885543346405, 'rewards/meter/std': 0.38644787669181824, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6627370119094849, 'rewards/repeat_soft/std': 0.16214919090270996, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.49848833680152893, 'rewards/total_composite/std': 0.09143819659948349, 'reward': 0.49848833680152893, 'reward_std': 0.0914381891489029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08699291944503784, 'sampling/sampling_logp_difference/max': 1.5512089729309082, 'sampling/importance_sampling_ratio/min': 0.21199151873588562, 'sampling/importance_sampling_ratio/mean': 0.9883591532707214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42546793818473816, 'clip_ratio/low_mean': 0.03153377631679177, 'clip_ratio/low_min': 0.03153377631679177, 'clip_ratio/high_mean': 0.05089415982365608, 'clip_ratio/high_max': 0.05089415982365608, 'clip_ratio/region_mean': 0.08242793614044785, 'reward_total_mean': 0.49848833680152893, 'reward_meter_mean': 0.778885543346405, 'reward_meter_std': 0.38644787669181824, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6627370119094849, 'reward_repeat_soft_std': 0.16214919090270996, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.49848833680152893, 'reward_total_composite_std': 0.09143819659948349, 'epoch': 0.16}

 49%|████▉     | 1626/3300 [3:59:32<3:20:58,  7.20s/it]INFO 04-13 11:20:30 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1627/3300 [3:59:38<3:13:59,  6.96s/it]2026-04-13 11:20:36,351 | INFO | train_grpo_train | metrics_logged mode=train step=1627

                                                       
{'loss': 0.0291, 'grad_norm': 11.057270050048828, 'learning_rate': 5.072727272727274e-06, 'num_tokens': 2885676.0, 'completions/mean_length': 50.625, 'completions/min_length': 40.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9232890605926514, 'rewards/meter/std': 0.0828256607055664, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8278992176055908, 'rewards/repeat_soft/std': 0.11741218715906143, 'rewards/judge_quality/mean': 0.5349999666213989, 'rewards/judge_quality/std': 0.24663449823856354, 'rewards/total_composite/mean': 0.6471283435821533, 'rewards/total_composite/std': 0.16611239314079285, 'reward': 0.6471283435821533, 'reward_std': 0.16611239314079285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10562446713447571, 'sampling/sampling_logp_difference/max': 1.053572654724121, 'sampling/importance_sampling_ratio/min': 0.3486897647380829, 'sampling/importance_sampling_ratio/mean': 1.019388198852539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7362676560878754, 'clip_ratio/low_mean': 0.07240128144621849, 'clip_ratio/low_min': 0.07240128144621849, 'clip_ratio/high_mean': 0.02590811997652054, 'clip_ratio/high_max': 0.02590811997652054, 'clip_ratio/region_mean': 0.09830940142273903, 'reward_total_mean': 0.6471283435821533, 'reward_meter_mean': 0.9232890605926514, 'reward_meter_std': 0.0828256607055664, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8278992176055908, 'reward_repeat_soft_std': 0.11741218715906143, 'reward_judge_quality_mean': 0.5349999666213989, 'reward_judge_quality_std': 0.24663449823856354, 'reward_total_composite_mean': 0.6471283435821533, 'reward_total_composite_std': 0.16611239314079285, 'epoch': 0.16}

 49%|████▉     | 1627/3300 [3:59:38<3:13:59,  6.96s/it]INFO 04-13 11:20:36 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1628/3300 [3:59:44<3:07:34,  6.73s/it]2026-04-13 11:20:42,550 | INFO | train_grpo_train | metrics_logged mode=train step=1628

                                                       
{'loss': -0.0076, 'grad_norm': 11.232089042663574, 'learning_rate': 5.06969696969697e-06, 'num_tokens': 2887311.0, 'completions/mean_length': 58.375, 'completions/min_length': 47.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9384300708770752, 'rewards/meter/std': 0.12097222357988358, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9826440811157227, 'rewards/repeat_soft/std': 0.014948301948606968, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.13695022463798523, 'rewards/total_composite/mean': 0.6538193225860596, 'rewards/total_composite/std': 0.101405069231987, 'reward': 0.6538193225860596, 'reward_std': 0.1014050617814064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12784811854362488, 'sampling/sampling_logp_difference/max': 1.5097036361694336, 'sampling/importance_sampling_ratio/min': 0.22097545862197876, 'sampling/importance_sampling_ratio/mean': 1.0171589851379395, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8741703927516937, 'clip_ratio/low_mean': 0.09202880971133709, 'clip_ratio/low_min': 0.09202880971133709, 'clip_ratio/high_mean': 0.02007575798779726, 'clip_ratio/high_max': 0.02007575798779726, 'clip_ratio/region_mean': 0.11210456769913435, 'reward_total_mean': 0.6538193225860596, 'reward_meter_mean': 0.9384300708770752, 'reward_meter_std': 0.12097222357988358, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9826440811157227, 'reward_repeat_soft_std': 0.014948301948606968, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.13695022463798523, 'reward_total_composite_mean': 0.6538193225860596, 'reward_total_composite_std': 0.101405069231987, 'epoch': 0.16}

 49%|████▉     | 1628/3300 [3:59:44<3:07:34,  6.73s/it]INFO 04-13 11:20:42 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1629/3300 [3:59:51<3:06:39,  6.70s/it]2026-04-13 11:20:49,191 | INFO | train_grpo_train | metrics_logged mode=train step=1629

                                                       
{'loss': 0.0336, 'grad_norm': 12.065169334411621, 'learning_rate': 5.0666666666666676e-06, 'num_tokens': 2888937.0, 'completions/mean_length': 52.25, 'completions/min_length': 48.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9526774883270264, 'rewards/meter/std': 0.07780593633651733, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9728760123252869, 'rewards/repeat_soft/std': 0.0253890473395586, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6060123443603516, 'rewards/total_composite/std': 0.022973021492362022, 'reward': 0.6060123443603516, 'reward_std': 0.022973012179136276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13202910125255585, 'sampling/sampling_logp_difference/max': 1.6118357181549072, 'sampling/importance_sampling_ratio/min': 0.199521005153656, 'sampling/importance_sampling_ratio/mean': 1.0097827911376953, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.658332109451294, 'clip_ratio/low_mean': 0.028846154920756817, 'clip_ratio/low_min': 0.028846154920756817, 'clip_ratio/high_mean': 0.09457175945863128, 'clip_ratio/high_max': 0.09457175945863128, 'clip_ratio/region_mean': 0.1234179143793881, 'reward_total_mean': 0.6060123443603516, 'reward_meter_mean': 0.9526774883270264, 'reward_meter_std': 0.07780593633651733, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9728760123252869, 'reward_repeat_soft_std': 0.0253890473395586, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6060123443603516, 'reward_total_composite_std': 0.022973021492362022, 'epoch': 0.16}

 49%|████▉     | 1629/3300 [3:59:51<3:06:39,  6.70s/it]INFO 04-13 11:20:49 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1630/3300 [3:59:59<3:19:36,  7.17s/it]2026-04-13 11:20:57,460 | INFO | train_grpo_train | metrics_logged mode=train step=1630

                                                       
{'loss': -0.061, 'grad_norm': 17.314697265625, 'learning_rate': 5.063636363636364e-06, 'num_tokens': 2890344.0, 'completions/mean_length': 31.875, 'completions/min_length': 23.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.8703351616859436, 'rewards/meter/std': 0.1738702803850174, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.7996700406074524, 'rewards/total_composite/std': 0.1645873337984085, 'reward': 0.7996700406074524, 'reward_std': 0.16458731889724731, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12682412564754486, 'sampling/sampling_logp_difference/max': 1.3500020503997803, 'sampling/importance_sampling_ratio/min': 0.25923970341682434, 'sampling/importance_sampling_ratio/mean': 1.004669189453125, 'sampling/importance_sampling_ratio/max': 1.9404537677764893, 'entropy': 0.7230124175548553, 'clip_ratio/low_mean': 0.035490778274834156, 'clip_ratio/low_min': 0.035490778274834156, 'clip_ratio/high_mean': 0.0786974485963583, 'clip_ratio/high_max': 0.0786974485963583, 'clip_ratio/region_mean': 0.11418822687119246, 'reward_total_mean': 0.7996700406074524, 'reward_meter_mean': 0.8703351616859436, 'reward_meter_std': 0.1738702803850174, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.7996700406074524, 'reward_total_composite_std': 0.1645873337984085, 'epoch': 0.16}

 49%|████▉     | 1630/3300 [3:59:59<3:19:36,  7.17s/it]INFO 04-13 11:20:57 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1631/3300 [4:00:05<3:10:38,  6.85s/it]2026-04-13 11:21:03,567 | INFO | train_grpo_train | metrics_logged mode=train step=1631

                                                       
{'loss': 0.0586, 'grad_norm': 12.814247131347656, 'learning_rate': 5.060606060606061e-06, 'num_tokens': 2892114.0, 'completions/mean_length': 39.25, 'completions/min_length': 34.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.7549862861633301, 'rewards/meter/std': 0.29110273718833923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9077358245849609, 'rewards/repeat_soft/std': 0.044362038373947144, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5813066959381104, 'rewards/total_composite/std': 0.153243288397789, 'reward': 0.5813066959381104, 'reward_std': 0.1532432734966278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09214208275079727, 'sampling/sampling_logp_difference/max': 1.1270325183868408, 'sampling/importance_sampling_ratio/min': 0.3239932656288147, 'sampling/importance_sampling_ratio/mean': 1.006922721862793, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5208986103534698, 'clip_ratio/low_mean': 0.02512002200819552, 'clip_ratio/low_min': 0.02512002200819552, 'clip_ratio/high_mean': 0.04087544255889952, 'clip_ratio/high_max': 0.04087544255889952, 'clip_ratio/region_mean': 0.06599546456709504, 'reward_total_mean': 0.5813066959381104, 'reward_meter_mean': 0.7549862861633301, 'reward_meter_std': 0.29110273718833923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9077358245849609, 'reward_repeat_soft_std': 0.044362038373947144, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5813066959381104, 'reward_total_composite_std': 0.153243288397789, 'epoch': 0.16}

 49%|████▉     | 1631/3300 [4:00:05<3:10:38,  6.85s/it]INFO 04-13 11:21:03 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1632/3300 [4:00:11<3:04:47,  6.65s/it]2026-04-13 11:21:09,741 | INFO | train_grpo_train | metrics_logged mode=train step=1632

                                                       
{'loss': 0.0718, 'grad_norm': 12.439215660095215, 'learning_rate': 5.057575757575758e-06, 'num_tokens': 2893812.0, 'completions/mean_length': 40.25, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.25, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6020994186401367, 'rewards/meter/std': 0.3678845763206482, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9884793758392334, 'rewards/repeat_soft/std': 0.014015505090355873, 'rewards/judge_quality/mean': 0.5874999761581421, 'rewards/judge_quality/std': 0.2043631225824356, 'rewards/total_composite/mean': 0.5512148141860962, 'rewards/total_composite/std': 0.1018567830324173, 'reward': 0.5512148141860962, 'reward_std': 0.1018567830324173, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.134989932179451, 'sampling/sampling_logp_difference/max': 1.8583168983459473, 'sampling/importance_sampling_ratio/min': 0.15593485534191132, 'sampling/importance_sampling_ratio/mean': 1.0240408182144165, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8467823192477226, 'clip_ratio/low_mean': 0.04137105867266655, 'clip_ratio/low_min': 0.04137105867266655, 'clip_ratio/high_mean': 0.08283164259046316, 'clip_ratio/high_max': 0.08283164259046316, 'clip_ratio/region_mean': 0.12420270126312971, 'reward_total_mean': 0.5512148141860962, 'reward_meter_mean': 0.6020994186401367, 'reward_meter_std': 0.3678845763206482, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9884793758392334, 'reward_repeat_soft_std': 0.014015505090355873, 'reward_judge_quality_mean': 0.5874999761581421, 'reward_judge_quality_std': 0.2043631225824356, 'reward_total_composite_mean': 0.5512148141860962, 'reward_total_composite_std': 0.1018567830324173, 'epoch': 0.16}

 49%|████▉     | 1632/3300 [4:00:11<3:04:47,  6.65s/it]INFO 04-13 11:21:09 [block_pool.py:378] Successfully reset prefix cache

 49%|████▉     | 1633/3300 [4:00:18<3:01:07,  6.52s/it]2026-04-13 11:21:15,954 | INFO | train_grpo_train | metrics_logged mode=train step=1633

                                                       
{'loss': 0.0113, 'grad_norm': 10.988968849182129, 'learning_rate': 5.054545454545455e-06, 'num_tokens': 2895417.0, 'completions/mean_length': 53.625, 'completions/min_length': 40.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.625, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9202272891998291, 'rewards/meter/std': 0.08080322295427322, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9281179904937744, 'rewards/repeat_soft/std': 0.07783205807209015, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.5684842467308044, 'rewards/total_composite/std': 0.06606169790029526, 'reward': 0.5684842467308044, 'reward_std': 0.06606169790029526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11435886472463608, 'sampling/sampling_logp_difference/max': 1.7910586595535278, 'sampling/importance_sampling_ratio/min': 0.16678351163864136, 'sampling/importance_sampling_ratio/mean': 1.0170116424560547, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6273412480950356, 'clip_ratio/low_mean': 0.019968220964074135, 'clip_ratio/low_min': 0.019968220964074135, 'clip_ratio/high_mean': 0.07719047367572784, 'clip_ratio/high_max': 0.07719047367572784, 'clip_ratio/region_mean': 0.09715869463980198, 'reward_total_mean': 0.5684842467308044, 'reward_meter_mean': 0.9202272891998291, 'reward_meter_std': 0.08080322295427322, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9281179904937744, 'reward_repeat_soft_std': 0.07783205807209015, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.5684842467308044, 'reward_total_composite_std': 0.06606169790029526, 'epoch': 0.16}

 49%|████▉     | 1633/3300 [4:00:18<3:01:07,  6.52s/it]INFO 04-13 11:21:16 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1634/3300 [4:00:30<3:51:24,  8.33s/it]2026-04-13 11:21:28,524 | INFO | train_grpo_train | metrics_logged mode=train step=1634

                                                       
{'loss': -0.1002, 'grad_norm': 3.135519504547119, 'learning_rate': 5.051515151515151e-06, 'num_tokens': 2896975.0, 'completions/mean_length': 112.75, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 55.71428680419922, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.7123268842697144, 'rewards/meter/std': 0.4074714779853821, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9069777727127075, 'rewards/repeat_soft/std': 0.09436126053333282, 'rewards/judge_quality/mean': 0.5612499713897705, 'rewards/judge_quality/std': 0.3223324716091156, 'rewards/total_composite/mean': 0.598398745059967, 'rewards/total_composite/std': 0.3048110604286194, 'reward': 0.598398745059967, 'reward_std': 0.3048110604286194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.142530158162117, 'sampling/sampling_logp_difference/max': 2.1631574630737305, 'sampling/importance_sampling_ratio/min': 0.11496156454086304, 'sampling/importance_sampling_ratio/mean': 1.003870964050293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7669676654040813, 'clip_ratio/low_mean': 0.03367326036095619, 'clip_ratio/low_min': 0.03367326036095619, 'clip_ratio/high_mean': 0.06216681748628616, 'clip_ratio/high_max': 0.06216681748628616, 'clip_ratio/region_mean': 0.09584007784724236, 'reward_total_mean': 0.598398745059967, 'reward_meter_mean': 0.7123268842697144, 'reward_meter_std': 0.4074714779853821, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9069777727127075, 'reward_repeat_soft_std': 0.09436126053333282, 'reward_judge_quality_mean': 0.5612499713897705, 'reward_judge_quality_std': 0.3223324716091156, 'reward_total_composite_mean': 0.598398745059967, 'reward_total_composite_std': 0.3048110604286194, 'epoch': 0.16}

 50%|████▉     | 1634/3300 [4:00:30<3:51:24,  8.33s/it]INFO 04-13 11:21:28 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1635/3300 [4:00:40<4:01:02,  8.69s/it]2026-04-13 11:21:38,031 | INFO | train_grpo_train | metrics_logged mode=train step=1635

                                                       
{'loss': 0.0696, 'grad_norm': 13.71398639678955, 'learning_rate': 5.048484848484849e-06, 'num_tokens': 2898864.0, 'completions/mean_length': 58.125, 'completions/min_length': 50.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7918387055397034, 'rewards/meter/std': 0.37128695845603943, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9755090475082397, 'rewards/repeat_soft/std': 0.019028011709451675, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.7015790939331055, 'rewards/total_composite/std': 0.21124167740345, 'reward': 0.7015790939331055, 'reward_std': 0.21124167740345, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1161414384841919, 'sampling/sampling_logp_difference/max': 3.696103096008301, 'sampling/importance_sampling_ratio/min': 0.024820059537887573, 'sampling/importance_sampling_ratio/mean': 1.0041608810424805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6581574194133282, 'clip_ratio/low_mean': 0.07275632116943598, 'clip_ratio/low_min': 0.07275632116943598, 'clip_ratio/high_mean': 0.03863636404275894, 'clip_ratio/high_max': 0.03863636404275894, 'clip_ratio/region_mean': 0.11139268521219492, 'reward_total_mean': 0.7015790939331055, 'reward_meter_mean': 0.7918387055397034, 'reward_meter_std': 0.37128695845603943, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9755090475082397, 'reward_repeat_soft_std': 0.019028011709451675, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.7015790939331055, 'reward_total_composite_std': 0.21124167740345, 'epoch': 0.16}

 50%|████▉     | 1635/3300 [4:00:40<4:01:02,  8.69s/it]INFO 04-13 11:21:38 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1636/3300 [4:00:46<3:41:44,  8.00s/it]2026-04-13 11:21:44,413 | INFO | train_grpo_train | metrics_logged mode=train step=1636

                                                       
{'loss': -0.0376, 'grad_norm': 17.088523864746094, 'learning_rate': 5.045454545454546e-06, 'num_tokens': 2900494.0, 'completions/mean_length': 51.75, 'completions/min_length': 42.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6968623399734497, 'rewards/meter/std': 0.33356785774230957, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8890321254730225, 'rewards/repeat_soft/std': 0.058889240026474, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5805325508117676, 'rewards/total_composite/std': 0.16673095524311066, 'reward': 0.5805325508117676, 'reward_std': 0.16673094034194946, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1255987584590912, 'sampling/sampling_logp_difference/max': 1.909731149673462, 'sampling/importance_sampling_ratio/min': 0.1481201946735382, 'sampling/importance_sampling_ratio/mean': 0.9984428286552429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8163897320628166, 'clip_ratio/low_mean': 0.046758043579757214, 'clip_ratio/low_min': 0.046758043579757214, 'clip_ratio/high_mean': 0.055573392659425735, 'clip_ratio/high_max': 0.055573392659425735, 'clip_ratio/region_mean': 0.10233143623918295, 'reward_total_mean': 0.5805325508117676, 'reward_meter_mean': 0.6968623399734497, 'reward_meter_std': 0.33356785774230957, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8890321254730225, 'reward_repeat_soft_std': 0.058889240026474, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5805325508117676, 'reward_total_composite_std': 0.16673095524311066, 'epoch': 0.16}

 50%|████▉     | 1636/3300 [4:00:46<3:41:44,  8.00s/it]INFO 04-13 11:21:44 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1637/3300 [4:00:55<3:48:00,  8.23s/it]2026-04-13 11:21:53,180 | INFO | train_grpo_train | metrics_logged mode=train step=1637

                                                       
{'loss': 0.668, 'grad_norm': 8.759028434753418, 'learning_rate': 5.042424242424243e-06, 'num_tokens': 2902433.0, 'completions/mean_length': 90.375, 'completions/min_length': 53.0, 'completions/max_length': 284.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.375, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 284.0, 'rewards/meter/mean': 0.6503992080688477, 'rewards/meter/std': 0.4289567172527313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9526631832122803, 'rewards/repeat_soft/std': 0.038023222237825394, 'rewards/judge_quality/mean': 0.4449999928474426, 'rewards/judge_quality/std': 0.21876277029514313, 'rewards/total_composite/mean': 0.5055826306343079, 'rewards/total_composite/std': 0.24735192954540253, 'reward': 0.5055826306343079, 'reward_std': 0.24735191464424133, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12658533453941345, 'sampling/sampling_logp_difference/max': 1.3045964241027832, 'sampling/importance_sampling_ratio/min': 0.27128201723098755, 'sampling/importance_sampling_ratio/mean': 1.0155060291290283, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.804808035492897, 'clip_ratio/low_mean': 0.02805097633972764, 'clip_ratio/low_min': 0.02805097633972764, 'clip_ratio/high_mean': 0.07687525823712349, 'clip_ratio/high_max': 0.07687525823712349, 'clip_ratio/region_mean': 0.10492623457685113, 'reward_total_mean': 0.5055826306343079, 'reward_meter_mean': 0.6503992080688477, 'reward_meter_std': 0.4289567172527313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9526631832122803, 'reward_repeat_soft_std': 0.038023222237825394, 'reward_judge_quality_mean': 0.4449999928474426, 'reward_judge_quality_std': 0.21876277029514313, 'reward_total_composite_mean': 0.5055826306343079, 'reward_total_composite_std': 0.24735192954540253, 'epoch': 0.16}

 50%|████▉     | 1637/3300 [4:00:55<3:48:00,  8.23s/it]INFO 04-13 11:21:53 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1638/3300 [4:01:06<4:13:43,  9.16s/it]2026-04-13 11:22:04,519 | INFO | train_grpo_train | metrics_logged mode=train step=1638

                                                       
{'loss': -0.2145, 'grad_norm': 2.6967246532440186, 'learning_rate': 5.0393939393939395e-06, 'num_tokens': 2904947.0, 'completions/mean_length': 174.25, 'completions/min_length': 109.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 126.00000762939453, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.9586606621742249, 'rewards/meter/std': 0.060762591660022736, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.18516401946544647, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7616796493530273, 'rewards/repeat_soft/std': 0.16934245824813843, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.14961259067058563, 'rewards/total_composite/mean': 0.44481372833251953, 'rewards/total_composite/std': 0.2030736356973648, 'reward': 0.44481372833251953, 'reward_std': 0.2030736207962036, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10499298572540283, 'sampling/sampling_logp_difference/max': 2.0107712745666504, 'sampling/importance_sampling_ratio/min': 0.13388536870479584, 'sampling/importance_sampling_ratio/mean': 1.0194429159164429, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7150713615119457, 'clip_ratio/low_mean': 0.017359061632305384, 'clip_ratio/low_min': 0.017359061632305384, 'clip_ratio/high_mean': 0.07144996337592602, 'clip_ratio/high_max': 0.07144996337592602, 'clip_ratio/region_mean': 0.0888090250082314, 'reward_total_mean': 0.44481372833251953, 'reward_meter_mean': 0.9586606621742249, 'reward_meter_std': 0.060762591660022736, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.18516401946544647, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7616796493530273, 'reward_repeat_soft_std': 0.16934245824813843, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.14961259067058563, 'reward_total_composite_mean': 0.44481372833251953, 'reward_total_composite_std': 0.2030736356973648, 'epoch': 0.16}

 50%|████▉     | 1638/3300 [4:01:06<4:13:43,  9.16s/it]INFO 04-13 11:22:04 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1639/3300 [4:01:17<4:29:25,  9.73s/it]2026-04-13 11:22:15,588 | INFO | train_grpo_train | metrics_logged mode=train step=1639

                                                       
{'loss': -0.1139, 'grad_norm': 2.492732048034668, 'learning_rate': 5.036363636363637e-06, 'num_tokens': 2906677.0, 'completions/mean_length': 118.25, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 62.000003814697266, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8640415072441101, 'rewards/meter/std': 0.32087600231170654, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9889659285545349, 'rewards/repeat_soft/std': 0.020190119743347168, 'rewards/judge_quality/mean': 0.6862499713897705, 'rewards/judge_quality/std': 0.34221702814102173, 'rewards/total_composite/mean': 0.7352422475814819, 'rewards/total_composite/std': 0.329959511756897, 'reward': 0.7352422475814819, 'reward_std': 0.329959511756897, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14623674750328064, 'sampling/sampling_logp_difference/max': 1.821061611175537, 'sampling/importance_sampling_ratio/min': 0.1618538349866867, 'sampling/importance_sampling_ratio/mean': 1.0083389282226562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7862942218780518, 'clip_ratio/low_mean': 0.02758458722382784, 'clip_ratio/low_min': 0.02758458722382784, 'clip_ratio/high_mean': 0.09788673743605614, 'clip_ratio/high_max': 0.09788673743605614, 'clip_ratio/region_mean': 0.12547132465988398, 'reward_total_mean': 0.7352422475814819, 'reward_meter_mean': 0.8640415072441101, 'reward_meter_std': 0.32087600231170654, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9889659285545349, 'reward_repeat_soft_std': 0.020190119743347168, 'reward_judge_quality_mean': 0.6862499713897705, 'reward_judge_quality_std': 0.34221702814102173, 'reward_total_composite_mean': 0.7352422475814819, 'reward_total_composite_std': 0.329959511756897, 'epoch': 0.16}

 50%|████▉     | 1639/3300 [4:01:17<4:29:25,  9.73s/it]INFO 04-13 11:22:15 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1640/3300 [4:01:23<3:59:42,  8.66s/it]2026-04-13 11:22:21,759 | INFO | train_grpo_train | metrics_logged mode=train step=1640

                                                       
{'loss': 0.0875, 'grad_norm': 10.618043899536133, 'learning_rate': 5.033333333333333e-06, 'num_tokens': 2908302.0, 'completions/mean_length': 36.125, 'completions/min_length': 30.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.7051137089729309, 'rewards/meter/std': 0.36694246530532837, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9104615449905396, 'rewards/repeat_soft/std': 0.08185256272554398, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.5875173807144165, 'rewards/total_composite/std': 0.16259489953517914, 'reward': 0.5875173807144165, 'reward_std': 0.16259488463401794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08619284629821777, 'sampling/sampling_logp_difference/max': 1.7844700813293457, 'sampling/importance_sampling_ratio/min': 0.16788600385189056, 'sampling/importance_sampling_ratio/mean': 1.0153688192367554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.47819822654128075, 'clip_ratio/low_mean': 0.023364828201010823, 'clip_ratio/low_min': 0.023364828201010823, 'clip_ratio/high_mean': 0.0380457378923893, 'clip_ratio/high_max': 0.0380457378923893, 'clip_ratio/region_mean': 0.06141056609340012, 'reward_total_mean': 0.5875173807144165, 'reward_meter_mean': 0.7051137089729309, 'reward_meter_std': 0.36694246530532837, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9104615449905396, 'reward_repeat_soft_std': 0.08185256272554398, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.5875173807144165, 'reward_total_composite_std': 0.16259489953517914, 'epoch': 0.16}

 50%|████▉     | 1640/3300 [4:01:23<3:59:42,  8.66s/it]INFO 04-13 11:22:21 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1641/3300 [4:01:31<3:51:09,  8.36s/it]2026-04-13 11:22:29,412 | INFO | train_grpo_train | metrics_logged mode=train step=1641

                                                       
{'loss': -0.0252, 'grad_norm': 8.12309741973877, 'learning_rate': 5.030303030303031e-06, 'num_tokens': 2910322.0, 'completions/mean_length': 83.5, 'completions/min_length': 60.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.5, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9841760396957397, 'rewards/meter/std': 0.008196291513741016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8510419130325317, 'rewards/repeat_soft/std': 0.07041285187005997, 'rewards/judge_quality/mean': 0.5237500071525574, 'rewards/judge_quality/std': 0.19078317284584045, 'rewards/total_composite/mean': 0.6627941131591797, 'rewards/total_composite/std': 0.1269497275352478, 'reward': 0.6627941131591797, 'reward_std': 0.1269497275352478, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10205010324716568, 'sampling/sampling_logp_difference/max': 1.4837443828582764, 'sampling/importance_sampling_ratio/min': 0.22678692638874054, 'sampling/importance_sampling_ratio/mean': 1.0135223865509033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.634688638150692, 'clip_ratio/low_mean': 0.05080424412153661, 'clip_ratio/low_min': 0.05080424412153661, 'clip_ratio/high_mean': 0.025846945121884346, 'clip_ratio/high_max': 0.025846945121884346, 'clip_ratio/region_mean': 0.07665118924342096, 'reward_total_mean': 0.6627941131591797, 'reward_meter_mean': 0.9841760396957397, 'reward_meter_std': 0.008196291513741016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8510419130325317, 'reward_repeat_soft_std': 0.07041285187005997, 'reward_judge_quality_mean': 0.5237500071525574, 'reward_judge_quality_std': 0.19078317284584045, 'reward_total_composite_mean': 0.6627941131591797, 'reward_total_composite_std': 0.1269497275352478, 'epoch': 0.16}

 50%|████▉     | 1641/3300 [4:01:31<3:51:09,  8.36s/it]INFO 04-13 11:22:29 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1642/3300 [4:01:43<4:20:40,  9.43s/it]2026-04-13 11:22:41,355 | INFO | train_grpo_train | metrics_logged mode=train step=1642

                                                       
{'loss': -0.202, 'grad_norm': 2.724552869796753, 'learning_rate': 5.027272727272728e-06, 'num_tokens': 2912716.0, 'completions/mean_length': 162.25, 'completions/min_length': 100.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.28572082519531, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.6643840074539185, 'rewards/meter/std': 0.3807678520679474, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9060827493667603, 'rewards/repeat_soft/std': 0.06389499455690384, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.46449315547943115, 'rewards/total_composite/std': 0.20578594505786896, 'reward': 0.46449315547943115, 'reward_std': 0.20578591525554657, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10945215076208115, 'sampling/sampling_logp_difference/max': 2.316504955291748, 'sampling/importance_sampling_ratio/min': 0.09861765801906586, 'sampling/importance_sampling_ratio/mean': 1.0101243257522583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6034212112426758, 'clip_ratio/low_mean': 0.03504893695935607, 'clip_ratio/low_min': 0.03504893695935607, 'clip_ratio/high_mean': 0.049963342025876045, 'clip_ratio/high_max': 0.049963342025876045, 'clip_ratio/region_mean': 0.08501227898523211, 'reward_total_mean': 0.46449315547943115, 'reward_meter_mean': 0.6643840074539185, 'reward_meter_std': 0.3807678520679474, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9060827493667603, 'reward_repeat_soft_std': 0.06389499455690384, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.46449315547943115, 'reward_total_composite_std': 0.20578594505786896, 'epoch': 0.16}

 50%|████▉     | 1642/3300 [4:01:43<4:20:40,  9.43s/it]INFO 04-13 11:22:41 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1643/3300 [4:01:55<4:41:02, 10.18s/it]2026-04-13 11:22:53,249 | INFO | train_grpo_train | metrics_logged mode=train step=1643

                                                       
{'loss': -0.0798, 'grad_norm': 1.5759207010269165, 'learning_rate': 5.024242424242425e-06, 'num_tokens': 2914122.0, 'completions/mean_length': 216.75, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 39.60000228881836, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8150025606155396, 'rewards/meter/std': 0.33757075667381287, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9947206974029541, 'rewards/repeat_soft/std': 0.00519803911447525, 'rewards/judge_quality/mean': 0.29750001430511475, 'rewards/judge_quality/std': 0.24188250303268433, 'rewards/total_composite/mean': 0.3926314413547516, 'rewards/total_composite/std': 0.3347933888435364, 'reward': 0.3926314413547516, 'reward_std': 0.334793359041214, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18363401293754578, 'sampling/sampling_logp_difference/max': 1.8588404655456543, 'sampling/importance_sampling_ratio/min': 0.1558532416820526, 'sampling/importance_sampling_ratio/mean': 1.0017668008804321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6053504720330238, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1028136657550931, 'clip_ratio/high_max': 0.1028136657550931, 'clip_ratio/region_mean': 0.1028136657550931, 'reward_total_mean': 0.3926314413547516, 'reward_meter_mean': 0.8150025606155396, 'reward_meter_std': 0.33757075667381287, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9947206974029541, 'reward_repeat_soft_std': 0.00519803911447525, 'reward_judge_quality_mean': 0.29750001430511475, 'reward_judge_quality_std': 0.24188250303268433, 'reward_total_composite_mean': 0.3926314413547516, 'reward_total_composite_std': 0.3347933888435364, 'epoch': 0.17}

 50%|████▉     | 1643/3300 [4:01:55<4:41:02, 10.18s/it]INFO 04-13 11:22:53 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1644/3300 [4:02:02<4:11:29,  9.11s/it]2026-04-13 11:22:59,946 | INFO | train_grpo_train | metrics_logged mode=train step=1644

                                                       
{'loss': 0.0057, 'grad_norm': 10.083914756774902, 'learning_rate': 5.021212121212121e-06, 'num_tokens': 2915782.0, 'completions/mean_length': 47.5, 'completions/min_length': 41.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7238696813583374, 'rewards/meter/std': 0.3476002812385559, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9268432855606079, 'rewards/repeat_soft/std': 0.035069454461336136, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.22984081506729126, 'rewards/total_composite/mean': 0.6061728596687317, 'rewards/total_composite/std': 0.14002163708209991, 'reward': 0.6061728596687317, 'reward_std': 0.14002162218093872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09343069046735764, 'sampling/sampling_logp_difference/max': 1.1515750885009766, 'sampling/importance_sampling_ratio/min': 0.31613844633102417, 'sampling/importance_sampling_ratio/mean': 1.0022430419921875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5562189929187298, 'clip_ratio/low_mean': 0.03513571806252003, 'clip_ratio/low_min': 0.03513571806252003, 'clip_ratio/high_mean': 0.04867568798363209, 'clip_ratio/high_max': 0.04867568798363209, 'clip_ratio/region_mean': 0.08381140604615211, 'reward_total_mean': 0.6061728596687317, 'reward_meter_mean': 0.7238696813583374, 'reward_meter_std': 0.3476002812385559, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9268432855606079, 'reward_repeat_soft_std': 0.035069454461336136, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.22984081506729126, 'reward_total_composite_mean': 0.6061728596687317, 'reward_total_composite_std': 0.14002163708209991, 'epoch': 0.17}

 50%|████▉     | 1644/3300 [4:02:02<4:11:29,  9.11s/it]INFO 04-13 11:23:00 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1645/3300 [4:02:09<3:55:14,  8.53s/it]2026-04-13 11:23:07,055 | INFO | train_grpo_train | metrics_logged mode=train step=1645

                                                       
{'loss': -0.0361, 'grad_norm': 17.102134704589844, 'learning_rate': 5.0181818181818186e-06, 'num_tokens': 2917231.0, 'completions/mean_length': 27.125, 'completions/min_length': 23.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7270609140396118, 'rewards/meter/std': 0.428354412317276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9311273097991943, 'rewards/repeat_soft/std': 0.07535543292760849, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.5775827169418335, 'rewards/total_composite/std': 0.17093253135681152, 'reward': 0.5775827169418335, 'reward_std': 0.17093253135681152, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1762944906949997, 'sampling/sampling_logp_difference/max': 1.8742389678955078, 'sampling/importance_sampling_ratio/min': 0.15347172319889069, 'sampling/importance_sampling_ratio/mean': 1.0031790733337402, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.177955947816372, 'clip_ratio/low_mean': 0.015416666865348816, 'clip_ratio/low_min': 0.015416666865348816, 'clip_ratio/high_mean': 0.129683512263, 'clip_ratio/high_max': 0.129683512263, 'clip_ratio/region_mean': 0.14510017912834883, 'reward_total_mean': 0.5775827169418335, 'reward_meter_mean': 0.7270609140396118, 'reward_meter_std': 0.428354412317276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9311273097991943, 'reward_repeat_soft_std': 0.07535543292760849, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.5775827169418335, 'reward_total_composite_std': 0.17093253135681152, 'epoch': 0.17}

 50%|████▉     | 1645/3300 [4:02:09<3:55:14,  8.53s/it]INFO 04-13 11:23:07 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1646/3300 [4:02:15<3:36:42,  7.86s/it]2026-04-13 11:23:13,366 | INFO | train_grpo_train | metrics_logged mode=train step=1646

                                                       
{'loss': 0.0579, 'grad_norm': 13.652124404907227, 'learning_rate': 5.015151515151515e-06, 'num_tokens': 2918812.0, 'completions/mean_length': 40.625, 'completions/min_length': 34.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7669195532798767, 'rewards/meter/std': 0.32380133867263794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8907057046890259, 'rewards/repeat_soft/std': 0.1519497185945511, 'rewards/judge_quality/mean': 0.3812499940395355, 'rewards/judge_quality/std': 0.08166787773370743, 'rewards/total_composite/mean': 0.5211167931556702, 'rewards/total_composite/std': 0.09286314249038696, 'reward': 0.5211167931556702, 'reward_std': 0.09286312758922577, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11913299560546875, 'sampling/sampling_logp_difference/max': 1.107813835144043, 'sampling/importance_sampling_ratio/min': 0.33028021454811096, 'sampling/importance_sampling_ratio/mean': 1.0003385543823242, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.713285006582737, 'clip_ratio/low_mean': 0.04548611235804856, 'clip_ratio/low_min': 0.04548611235804856, 'clip_ratio/high_mean': 0.07731873355805874, 'clip_ratio/high_max': 0.07731873355805874, 'clip_ratio/region_mean': 0.1228048459161073, 'reward_total_mean': 0.5211167931556702, 'reward_meter_mean': 0.7669195532798767, 'reward_meter_std': 0.32380133867263794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8907057046890259, 'reward_repeat_soft_std': 0.1519497185945511, 'reward_judge_quality_mean': 0.3812499940395355, 'reward_judge_quality_std': 0.08166787773370743, 'reward_total_composite_mean': 0.5211167931556702, 'reward_total_composite_std': 0.09286314249038696, 'epoch': 0.17}

 50%|████▉     | 1646/3300 [4:02:15<3:36:42,  7.86s/it]INFO 04-13 11:23:13 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1647/3300 [4:02:22<3:27:45,  7.54s/it]2026-04-13 11:23:20,159 | INFO | train_grpo_train | metrics_logged mode=train step=1647

                                                       
{'loss': 0.0052, 'grad_norm': 11.083194732666016, 'learning_rate': 5.012121212121212e-06, 'num_tokens': 2920479.0, 'completions/mean_length': 60.375, 'completions/min_length': 53.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.375, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6427329778671265, 'rewards/meter/std': 0.25448077917099, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9895933866500854, 'rewards/repeat_soft/std': 0.014044932089745998, 'rewards/judge_quality/mean': 0.5900000333786011, 'rewards/judge_quality/std': 0.27994897961616516, 'rewards/total_composite/mean': 0.5815964937210083, 'rewards/total_composite/std': 0.1397792249917984, 'reward': 0.5815964937210083, 'reward_std': 0.1397792249917984, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1341121792793274, 'sampling/sampling_logp_difference/max': 1.6723310947418213, 'sampling/importance_sampling_ratio/min': 0.18780875205993652, 'sampling/importance_sampling_ratio/mean': 1.0191738605499268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8402115479111671, 'clip_ratio/low_mean': 0.05557193420827389, 'clip_ratio/low_min': 0.05557193420827389, 'clip_ratio/high_mean': 0.07705580070614815, 'clip_ratio/high_max': 0.07705580070614815, 'clip_ratio/region_mean': 0.13262773491442204, 'reward_total_mean': 0.5815964937210083, 'reward_meter_mean': 0.6427329778671265, 'reward_meter_std': 0.25448077917099, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9895933866500854, 'reward_repeat_soft_std': 0.014044932089745998, 'reward_judge_quality_mean': 0.5900000333786011, 'reward_judge_quality_std': 0.27994897961616516, 'reward_total_composite_mean': 0.5815964937210083, 'reward_total_composite_std': 0.1397792249917984, 'epoch': 0.17}

 50%|████▉     | 1647/3300 [4:02:22<3:27:45,  7.54s/it]INFO 04-13 11:23:20 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1648/3300 [4:02:29<3:26:13,  7.49s/it]2026-04-13 11:23:27,523 | INFO | train_grpo_train | metrics_logged mode=train step=1648

                                                       
{'loss': -0.0047, 'grad_norm': 6.380144119262695, 'learning_rate': 5.009090909090909e-06, 'num_tokens': 2922243.0, 'completions/mean_length': 53.5, 'completions/min_length': 46.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.5, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9764459133148193, 'rewards/meter/std': 0.019814912229776382, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8258414268493652, 'rewards/repeat_soft/std': 0.12010334432125092, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.19334924221038818, 'rewards/total_composite/mean': 0.6210887432098389, 'rewards/total_composite/std': 0.1282896101474762, 'reward': 0.6210887432098389, 'reward_std': 0.1282896101474762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11061302572488785, 'sampling/sampling_logp_difference/max': 1.0346717834472656, 'sampling/importance_sampling_ratio/min': 0.3553430140018463, 'sampling/importance_sampling_ratio/mean': 1.0079593658447266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7249104119837284, 'clip_ratio/low_mean': 0.061141025042161345, 'clip_ratio/low_min': 0.061141025042161345, 'clip_ratio/high_mean': 0.040674603544175625, 'clip_ratio/high_max': 0.040674603544175625, 'clip_ratio/region_mean': 0.10181562858633697, 'reward_total_mean': 0.6210887432098389, 'reward_meter_mean': 0.9764459133148193, 'reward_meter_std': 0.019814912229776382, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8258414268493652, 'reward_repeat_soft_std': 0.12010334432125092, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.19334924221038818, 'reward_total_composite_mean': 0.6210887432098389, 'reward_total_composite_std': 0.1282896101474762, 'epoch': 0.17}

 50%|████▉     | 1648/3300 [4:02:29<3:26:13,  7.49s/it]INFO 04-13 11:23:27 [block_pool.py:378] Successfully reset prefix cache

 50%|████▉     | 1649/3300 [4:02:36<3:17:30,  7.18s/it]2026-04-13 11:23:33,969 | INFO | train_grpo_train | metrics_logged mode=train step=1649

                                                       
{'loss': -0.001, 'grad_norm': 8.52895450592041, 'learning_rate': 5.006060606060607e-06, 'num_tokens': 2923968.0, 'completions/mean_length': 61.625, 'completions/min_length': 54.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9864411354064941, 'rewards/meter/std': 0.01547372154891491, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9763901829719543, 'rewards/repeat_soft/std': 0.020429003983736038, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.11667262762784958, 'rewards/total_composite/mean': 0.6421948671340942, 'rewards/total_composite/std': 0.07390771061182022, 'reward': 0.6421948671340942, 'reward_std': 0.07390771806240082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1188473328948021, 'sampling/sampling_logp_difference/max': 1.487046718597412, 'sampling/importance_sampling_ratio/min': 0.22603923082351685, 'sampling/importance_sampling_ratio/mean': 1.012412428855896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9211339130997658, 'clip_ratio/low_mean': 0.08232017001137137, 'clip_ratio/low_min': 0.08232017001137137, 'clip_ratio/high_mean': 0.015625, 'clip_ratio/high_max': 0.015625, 'clip_ratio/region_mean': 0.09794517001137137, 'reward_total_mean': 0.6421948671340942, 'reward_meter_mean': 0.9864411354064941, 'reward_meter_std': 0.01547372154891491, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9763901829719543, 'reward_repeat_soft_std': 0.020429003983736038, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.11667262762784958, 'reward_total_composite_mean': 0.6421948671340942, 'reward_total_composite_std': 0.07390771061182022, 'epoch': 0.17}

 50%|████▉     | 1649/3300 [4:02:36<3:17:30,  7.18s/it]INFO 04-13 11:23:34 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1650/3300 [4:02:42<3:11:33,  6.97s/it]2026-04-13 11:23:40,434 | INFO | train_grpo_train | metrics_logged mode=train step=1650

                                                       
{'loss': -0.0375, 'grad_norm': 9.283758163452148, 'learning_rate': 5.003030303030303e-06, 'num_tokens': 2925857.0, 'completions/mean_length': 54.125, 'completions/min_length': 51.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9835355877876282, 'rewards/meter/std': 0.01232390571385622, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8640047907829285, 'rewards/repeat_soft/std': 0.033385492861270905, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6005144715309143, 'rewards/total_composite/std': 0.010814270935952663, 'reward': 0.6005144715309143, 'reward_std': 0.010814269073307514, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10571616888046265, 'sampling/sampling_logp_difference/max': 1.3834004402160645, 'sampling/importance_sampling_ratio/min': 0.25072452425956726, 'sampling/importance_sampling_ratio/mean': 1.0140377283096313, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5891980826854706, 'clip_ratio/low_mean': 0.056812843307852745, 'clip_ratio/low_min': 0.056812843307852745, 'clip_ratio/high_mean': 0.04984511248767376, 'clip_ratio/high_max': 0.04984511248767376, 'clip_ratio/region_mean': 0.1066579557955265, 'reward_total_mean': 0.6005144715309143, 'reward_meter_mean': 0.9835355877876282, 'reward_meter_std': 0.01232390571385622, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8640047907829285, 'reward_repeat_soft_std': 0.033385492861270905, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6005144715309143, 'reward_total_composite_std': 0.010814270935952663, 'epoch': 0.17}

 50%|█████     | 1650/3300 [4:02:42<3:11:33,  6.97s/it]INFO 04-13 11:23:40 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:16,  2.05s/it]

 30%|███       | 3/10 [00:08<00:22,  3.20s/it]

 40%|████      | 4/10 [00:17<00:31,  5.30s/it]

 50%|█████     | 5/10 [00:23<00:26,  5.35s/it]

 60%|██████    | 6/10 [00:28<00:21,  5.30s/it]

 70%|███████   | 7/10 [00:37<00:19,  6.45s/it]

 80%|████████  | 8/10 [00:46<00:14,  7.20s/it]

 90%|█████████ | 9/10 [00:55<00:07,  7.92s/it]

100%|██████████| 10/10 [01:04<00:00,  8.31s/it]2026-04-13 11:24:50,195 | INFO | train_grpo_train | metrics_logged mode=eval step=1650

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 69.6645, 'eval_samples_per_second': 1.148, 'eval_steps_per_second': 0.144, 'eval_num_tokens': 2925857.0, 'eval_completions/mean_length': 115.75, 'eval_completions/min_length': 37.9, 'eval_completions/max_length': 347.1, 'eval_completions/clipped_ratio': 0.0625, 'eval_completions/mean_terminated_length': 88.40357284545898, 'eval_completions/min_terminated_length': 37.9, 'eval_completions/max_terminated_length': 160.2, 'eval_rewards/meter/mean': 0.7606006145477295, 'eval_rewards/meter/std': 0.30303533375263214, 'eval_rewards/count_adherence/mean': 0.9537499904632568, 'eval_rewards/count_adherence/std': 0.08891427256166935, 'eval_rewards/hard_gate/mean': 0.925, 'eval_rewards/hard_gate/std': 0.18771235942840575, 'eval_rewards/repeat_soft/mean': 0.8713025391101837, 'eval_rewards/repeat_soft/std': 0.10102997198700905, 'eval_rewards/judge_quality/mean': 0.4048749953508377, 'eval_rewards/judge_quality/std': 0.1862373486161232, 'eval_rewards/total_composite/mean': 0.49997623860836027, 'eval_rewards/total_composite/std': 0.18340689837932586, 'eval_reward': 0.49997623860836027, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.05427191890776158, 'eval_sampling/sampling_logp_difference/max': 0.9719480037689209, 'eval_sampling/importance_sampling_ratio/min': 0.38455626368522644, 'eval_sampling/importance_sampling_ratio/mean': 1.0114875197410584, 'eval_sampling/importance_sampling_ratio/max': 1.4432477235794068, 'eval_entropy': 0.612224942445755, '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': 0.49997623860836027, 'eval_reward_meter_mean': 0.7606006145477295, 'eval_reward_meter_std': 0.30303533375263214, 'eval_reward_count_adherence_mean': 0.9537499904632568, 'eval_reward_count_adherence_std': 0.08891427256166935, 'eval_reward_hard_gate_mean': 0.925, 'eval_reward_hard_gate_std': 0.18771235942840575, 'eval_reward_repeat_soft_mean': 0.8713025391101837, 'eval_reward_repeat_soft_std': 0.10102997198700905, 'eval_reward_judge_quality_mean': 0.4048749953508377, 'eval_reward_judge_quality_std': 0.1862373486161232, 'eval_reward_total_composite_mean': 0.49997623860836027, 'eval_reward_total_composite_std': 0.18340689837932586, 'epoch': 0.17}

 50%|█████     | 1650/3300 [4:03:52<3:11:33,  6.97s/it]

100%|██████████| 10/10 [01:04<00:00,  8.31s/it]

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

 50%|█████     | 1651/3300 [4:04:02<13:11:15, 28.79s/it]2026-04-13 11:25:00,158 | INFO | train_grpo_train | metrics_logged mode=train step=1651

                                                        
{'loss': 0.0804, 'grad_norm': 12.425683975219727, 'learning_rate': 5e-06, 'num_tokens': 2927346.0, 'completions/mean_length': 44.125, 'completions/min_length': 34.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9057676792144775, 'rewards/meter/std': 0.121575728058815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8857040405273438, 'rewards/repeat_soft/std': 0.06436088681221008, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5860784649848938, 'rewards/total_composite/std': 0.027623461559414864, 'reward': 0.5860784649848938, 'reward_std': 0.027623457834124565, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09438827633857727, 'sampling/sampling_logp_difference/max': 1.3431239128112793, 'sampling/importance_sampling_ratio/min': 0.2610289752483368, 'sampling/importance_sampling_ratio/mean': 1.0196468830108643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6710145398974419, 'clip_ratio/low_mean': 0.05590006709098816, 'clip_ratio/low_min': 0.05590006709098816, 'clip_ratio/high_mean': 0.05511593818664551, 'clip_ratio/high_max': 0.05511593818664551, 'clip_ratio/region_mean': 0.11101600527763367, 'reward_total_mean': 0.5860784649848938, 'reward_meter_mean': 0.9057676792144775, 'reward_meter_std': 0.121575728058815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8857040405273438, 'reward_repeat_soft_std': 0.06436088681221008, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5860784649848938, 'reward_total_composite_std': 0.027623461559414864, 'epoch': 0.17}

 50%|█████     | 1651/3300 [4:04:02<13:11:15, 28.79s/it]INFO 04-13 11:25:00 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1652/3300 [4:04:09<10:13:14, 22.33s/it]2026-04-13 11:25:07,398 | INFO | train_grpo_train | metrics_logged mode=train step=1652

                                                        
{'loss': 0.0635, 'grad_norm': 7.243699550628662, 'learning_rate': 4.996969696969698e-06, 'num_tokens': 2929413.0, 'completions/mean_length': 87.375, 'completions/min_length': 75.0, 'completions/max_length': 116.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 87.375, 'completions/min_terminated_length': 75.0, 'completions/max_terminated_length': 116.0, 'rewards/meter/mean': 0.9921964406967163, 'rewards/meter/std': 0.0031764977611601353, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8434470891952515, 'rewards/repeat_soft/std': 0.09848575294017792, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5778021216392517, 'rewards/total_composite/std': 0.04142162203788757, 'reward': 0.5778021216392517, 'reward_std': 0.041421618312597275, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07206101715564728, 'sampling/sampling_logp_difference/max': 1.4006543159484863, 'sampling/importance_sampling_ratio/min': 0.24643567204475403, 'sampling/importance_sampling_ratio/mean': 1.0028376579284668, 'sampling/importance_sampling_ratio/max': 1.7866092920303345, 'entropy': 0.546353530138731, 'clip_ratio/low_mean': 0.024673645850270987, 'clip_ratio/low_min': 0.024673645850270987, 'clip_ratio/high_mean': 0.06221241131424904, 'clip_ratio/high_max': 0.06221241131424904, 'clip_ratio/region_mean': 0.08688605716452003, 'reward_total_mean': 0.5778021216392517, 'reward_meter_mean': 0.9921964406967163, 'reward_meter_std': 0.0031764977611601353, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8434470891952515, 'reward_repeat_soft_std': 0.09848575294017792, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5778021216392517, 'reward_total_composite_std': 0.04142162203788757, 'epoch': 0.17}

 50%|█████     | 1652/3300 [4:04:09<10:13:14, 22.33s/it]INFO 04-13 11:25:07 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1653/3300 [4:04:21<8:47:22, 19.21s/it] 2026-04-13 11:25:19,406 | INFO | train_grpo_train | metrics_logged mode=train step=1653

                                                       
{'loss': -0.1949, 'grad_norm': 1.912166953086853, 'learning_rate': 4.993939393939394e-06, 'num_tokens': 2931539.0, 'completions/mean_length': 159.75, 'completions/min_length': 90.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.42857360839844, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.9761732816696167, 'rewards/meter/std': 0.022720949724316597, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8665366768836975, 'rewards/repeat_soft/std': 0.044016994535923004, 'rewards/judge_quality/mean': 0.4112499952316284, 'rewards/judge_quality/std': 0.1797965168952942, 'rewards/total_composite/mean': 0.5456495881080627, 'rewards/total_composite/std': 0.23139263689517975, 'reward': 0.5456495881080627, 'reward_std': 0.23139263689517975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12424559146165848, 'sampling/sampling_logp_difference/max': 2.206782341003418, 'sampling/importance_sampling_ratio/min': 0.11005418747663498, 'sampling/importance_sampling_ratio/mean': 1.0153663158416748, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.656967006623745, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09762474708259106, 'clip_ratio/high_max': 0.09762474708259106, 'clip_ratio/region_mean': 0.09762474708259106, 'reward_total_mean': 0.5456495881080627, 'reward_meter_mean': 0.9761732816696167, 'reward_meter_std': 0.022720949724316597, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8665366768836975, 'reward_repeat_soft_std': 0.044016994535923004, 'reward_judge_quality_mean': 0.4112499952316284, 'reward_judge_quality_std': 0.1797965168952942, 'reward_total_composite_mean': 0.5456495881080627, 'reward_total_composite_std': 0.23139263689517975, 'epoch': 0.17}

 50%|█████     | 1653/3300 [4:04:21<8:47:22, 19.21s/it]INFO 04-13 11:25:19 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1654/3300 [4:04:33<7:50:11, 17.14s/it]2026-04-13 11:25:31,652 | INFO | train_grpo_train | metrics_logged mode=train step=1654

                                                       
{'loss': -0.1346, 'grad_norm': 1.924111247062683, 'learning_rate': 4.990909090909091e-06, 'num_tokens': 2933194.0, 'completions/mean_length': 110.875, 'completions/min_length': 45.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 53.57143020629883, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8242701292037964, 'rewards/meter/std': 0.3262185752391815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9393479824066162, 'rewards/repeat_soft/std': 0.049659643322229385, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.5063811540603638, 'rewards/total_composite/std': 0.20991861820220947, 'reward': 0.5063811540603638, 'reward_std': 0.20991861820220947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12239545583724976, 'sampling/sampling_logp_difference/max': 3.181450843811035, 'sampling/importance_sampling_ratio/min': 0.04152536392211914, 'sampling/importance_sampling_ratio/mean': 0.9930113554000854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5652026124298573, 'clip_ratio/low_mean': 0.007936508394777775, 'clip_ratio/low_min': 0.007936508394777775, 'clip_ratio/high_mean': 0.09462250303477049, 'clip_ratio/high_max': 0.09462250303477049, 'clip_ratio/region_mean': 0.10255901142954826, 'reward_total_mean': 0.5063811540603638, 'reward_meter_mean': 0.8242701292037964, 'reward_meter_std': 0.3262185752391815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9393479824066162, 'reward_repeat_soft_std': 0.049659643322229385, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.5063811540603638, 'reward_total_composite_std': 0.20991861820220947, 'epoch': 0.17}

 50%|█████     | 1654/3300 [4:04:33<7:50:11, 17.14s/it]INFO 04-13 11:25:31 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1655/3300 [4:04:40<6:21:10, 13.90s/it]2026-04-13 11:25:38,004 | INFO | train_grpo_train | metrics_logged mode=train step=1655

                                                       
{'loss': 0.0539, 'grad_norm': 17.1294002532959, 'learning_rate': 4.987878787878789e-06, 'num_tokens': 2934846.0, 'completions/mean_length': 40.5, 'completions/min_length': 35.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7644696831703186, 'rewards/meter/std': 0.2678174376487732, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9411360621452332, 'rewards/repeat_soft/std': 0.0657808780670166, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.620587944984436, 'rewards/total_composite/std': 0.16680742800235748, 'reward': 0.620587944984436, 'reward_std': 0.16680742800235748, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11558982729911804, 'sampling/sampling_logp_difference/max': 2.733036518096924, 'sampling/importance_sampling_ratio/min': 0.06502155214548111, 'sampling/importance_sampling_ratio/mean': 1.0181759595870972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4405171722173691, 'clip_ratio/low_mean': 0.08305060723796487, 'clip_ratio/low_min': 0.08305060723796487, 'clip_ratio/high_mean': 0.02430124208331108, 'clip_ratio/high_max': 0.02430124208331108, 'clip_ratio/region_mean': 0.10735184932127595, 'reward_total_mean': 0.620587944984436, 'reward_meter_mean': 0.7644696831703186, 'reward_meter_std': 0.2678174376487732, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9411360621452332, 'reward_repeat_soft_std': 0.0657808780670166, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.620587944984436, 'reward_total_composite_std': 0.16680742800235748, 'epoch': 0.17}

 50%|█████     | 1655/3300 [4:04:40<6:21:10, 13.90s/it]INFO 04-13 11:25:38 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1656/3300 [4:04:52<6:06:39, 13.38s/it]2026-04-13 11:25:50,165 | INFO | train_grpo_train | metrics_logged mode=train step=1656

                                                       
{'loss': -0.072, 'grad_norm': 1.673622965812683, 'learning_rate': 4.984848484848485e-06, 'num_tokens': 2936323.0, 'completions/mean_length': 80.625, 'completions/min_length': 17.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 19.0, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.857839822769165, 'rewards/meter/std': 0.32516783475875854, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9595891237258911, 'rewards/repeat_soft/std': 0.006366936955600977, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.5196996331214905, 'rewards/total_composite/std': 0.21318161487579346, 'reward': 0.5196996331214905, 'reward_std': 0.21318161487579346, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10437224060297012, 'sampling/sampling_logp_difference/max': 1.280099868774414, 'sampling/importance_sampling_ratio/min': 0.2780095338821411, 'sampling/importance_sampling_ratio/mean': 0.9901865720748901, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49581820517778397, 'clip_ratio/low_mean': 0.007352941203862429, 'clip_ratio/low_min': 0.007352941203862429, 'clip_ratio/high_mean': 0.10395781276747584, 'clip_ratio/high_max': 0.10395781276747584, 'clip_ratio/region_mean': 0.11131075397133827, 'reward_total_mean': 0.5196996331214905, 'reward_meter_mean': 0.857839822769165, 'reward_meter_std': 0.32516783475875854, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9595891237258911, 'reward_repeat_soft_std': 0.006366936955600977, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.5196996331214905, 'reward_total_composite_std': 0.21318161487579346, 'epoch': 0.17}

 50%|█████     | 1656/3300 [4:04:52<6:06:39, 13.38s/it]INFO 04-13 11:25:50 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1657/3300 [4:04:59<5:14:11, 11.47s/it]2026-04-13 11:25:57,188 | INFO | train_grpo_train | metrics_logged mode=train step=1657

                                                       
{'loss': 0.042, 'grad_norm': 8.607499122619629, 'learning_rate': 4.981818181818182e-06, 'num_tokens': 2938306.0, 'completions/mean_length': 80.875, 'completions/min_length': 74.0, 'completions/max_length': 94.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.875, 'completions/min_terminated_length': 74.0, 'completions/max_terminated_length': 94.0, 'rewards/meter/mean': 0.9017000794410706, 'rewards/meter/std': 0.16966751217842102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8983750343322754, 'rewards/repeat_soft/std': 0.03722051531076431, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.619802713394165, 'rewards/total_composite/std': 0.12518617510795593, 'reward': 0.619802713394165, 'reward_std': 0.12518619000911713, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12372054904699326, 'sampling/sampling_logp_difference/max': 2.9575610160827637, 'sampling/importance_sampling_ratio/min': 0.05194545537233353, 'sampling/importance_sampling_ratio/mean': 1.0086854696273804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5657336711883545, 'clip_ratio/low_mean': 0.10132492426782846, 'clip_ratio/low_min': 0.10132492426782846, 'clip_ratio/high_mean': 0.01461038924753666, 'clip_ratio/high_max': 0.01461038924753666, 'clip_ratio/region_mean': 0.11593531351536512, 'reward_total_mean': 0.619802713394165, 'reward_meter_mean': 0.9017000794410706, 'reward_meter_std': 0.16966751217842102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8983750343322754, 'reward_repeat_soft_std': 0.03722051531076431, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.619802713394165, 'reward_total_composite_std': 0.12518617510795593, 'epoch': 0.17}

 50%|█████     | 1657/3300 [4:04:59<5:14:11, 11.47s/it]INFO 04-13 11:25:57 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1658/3300 [4:05:10<5:14:23, 11.49s/it]2026-04-13 11:26:08,717 | INFO | train_grpo_train | metrics_logged mode=train step=1658

                                                       
{'loss': -0.2136, 'grad_norm': 1.5572788715362549, 'learning_rate': 4.978787878787879e-06, 'num_tokens': 2940588.0, 'completions/mean_length': 168.25, 'completions/min_length': 115.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 119.14286041259766, 'completions/min_terminated_length': 115.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.8625526428222656, 'rewards/meter/std': 0.3485822081565857, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8076668977737427, 'rewards/repeat_soft/std': 0.11250165104866028, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.5128768682479858, 'rewards/total_composite/std': 0.2076120376586914, 'reward': 0.5128768682479858, 'reward_std': 0.2076120227575302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08549162745475769, 'sampling/sampling_logp_difference/max': 5.35800313949585, 'sampling/importance_sampling_ratio/min': 0.004710302222520113, 'sampling/importance_sampling_ratio/mean': 1.001693606376648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3563239946961403, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06801584642380476, 'clip_ratio/high_max': 0.06801584642380476, 'clip_ratio/region_mean': 0.06801584642380476, 'reward_total_mean': 0.5128768682479858, 'reward_meter_mean': 0.8625526428222656, 'reward_meter_std': 0.3485822081565857, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8076668977737427, 'reward_repeat_soft_std': 0.11250165104866028, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.5128768682479858, 'reward_total_composite_std': 0.2076120376586914, 'epoch': 0.17}

 50%|█████     | 1658/3300 [4:05:10<5:14:23, 11.49s/it]INFO 04-13 11:26:08 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1659/3300 [4:05:18<4:39:25, 10.22s/it]2026-04-13 11:26:15,966 | INFO | train_grpo_train | metrics_logged mode=train step=1659

                                                       
{'loss': -0.0036, 'grad_norm': 6.4640045166015625, 'learning_rate': 4.975757575757576e-06, 'num_tokens': 2943088.0, 'completions/mean_length': 118.5, 'completions/min_length': 104.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.5, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.890613317489624, 'rewards/meter/std': 0.27192914485931396, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9004124999046326, 'rewards/repeat_soft/std': 0.04215488210320473, 'rewards/judge_quality/mean': 0.47749999165534973, 'rewards/judge_quality/std': 0.2303258776664734, 'rewards/total_composite/mean': 0.5966652035713196, 'rewards/total_composite/std': 0.15086136758327484, 'reward': 0.5966652035713196, 'reward_std': 0.15086136758327484, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12144719064235687, 'sampling/sampling_logp_difference/max': 3.8907980918884277, 'sampling/importance_sampling_ratio/min': 0.020429035648703575, 'sampling/importance_sampling_ratio/mean': 0.996963381767273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5733271576464176, 'clip_ratio/low_mean': 0.043103738222271204, 'clip_ratio/low_min': 0.043103738222271204, 'clip_ratio/high_mean': 0.06401834078133106, 'clip_ratio/high_max': 0.06401834078133106, 'clip_ratio/region_mean': 0.10712207900360227, 'reward_total_mean': 0.5966652035713196, 'reward_meter_mean': 0.890613317489624, 'reward_meter_std': 0.27192914485931396, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9004124999046326, 'reward_repeat_soft_std': 0.04215488210320473, 'reward_judge_quality_mean': 0.47749999165534973, 'reward_judge_quality_std': 0.2303258776664734, 'reward_total_composite_mean': 0.5966652035713196, 'reward_total_composite_std': 0.15086136758327484, 'epoch': 0.17}

 50%|█████     | 1659/3300 [4:05:18<4:39:25, 10.22s/it]INFO 04-13 11:26:16 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1660/3300 [4:05:24<4:06:18,  9.01s/it]2026-04-13 11:26:22,159 | INFO | train_grpo_train | metrics_logged mode=train step=1660

                                                       
{'loss': 0.0085, 'grad_norm': 8.742388725280762, 'learning_rate': 4.972727272727273e-06, 'num_tokens': 2944760.0, 'completions/mean_length': 44.0, 'completions/min_length': 35.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8351925611495972, 'rewards/meter/std': 0.27675724029541016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8899495005607605, 'rewards/repeat_soft/std': 0.02758779190480709, 'rewards/judge_quality/mean': 0.41749998927116394, 'rewards/judge_quality/std': 0.06902380287647247, 'rewards/total_composite/mean': 0.5577464699745178, 'rewards/total_composite/std': 0.08479692786931992, 'reward': 0.5577464699745178, 'reward_std': 0.08479694277048111, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11134511977434158, 'sampling/sampling_logp_difference/max': 1.9499166011810303, 'sampling/importance_sampling_ratio/min': 0.14228594303131104, 'sampling/importance_sampling_ratio/mean': 0.9990329742431641, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6182470545172691, 'clip_ratio/low_mean': 0.027862208895385265, 'clip_ratio/low_min': 0.027862208895385265, 'clip_ratio/high_mean': 0.10116859432309866, 'clip_ratio/high_max': 0.10116859432309866, 'clip_ratio/region_mean': 0.12903080321848392, 'reward_total_mean': 0.5577464699745178, 'reward_meter_mean': 0.8351925611495972, 'reward_meter_std': 0.27675724029541016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8899495005607605, 'reward_repeat_soft_std': 0.02758779190480709, 'reward_judge_quality_mean': 0.41749998927116394, 'reward_judge_quality_std': 0.06902380287647247, 'reward_total_composite_mean': 0.5577464699745178, 'reward_total_composite_std': 0.08479692786931992, 'epoch': 0.17}

 50%|█████     | 1660/3300 [4:05:24<4:06:18,  9.01s/it]INFO 04-13 11:26:22 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1661/3300 [4:05:30<3:46:21,  8.29s/it]2026-04-13 11:26:28,753 | INFO | train_grpo_train | metrics_logged mode=train step=1661

                                                       
{'loss': -0.0193, 'grad_norm': 8.281079292297363, 'learning_rate': 4.9696969696969696e-06, 'num_tokens': 2946558.0, 'completions/mean_length': 59.75, 'completions/min_length': 49.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.9679084420204163, 'rewards/meter/std': 0.031183168292045593, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9141935110092163, 'rewards/repeat_soft/std': 0.06871075928211212, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.7038965821266174, 'rewards/total_composite/std': 0.15150462090969086, 'reward': 0.7038965821266174, 'reward_std': 0.15150462090969086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09651017189025879, 'sampling/sampling_logp_difference/max': 2.2903401851654053, 'sampling/importance_sampling_ratio/min': 0.10123201459646225, 'sampling/importance_sampling_ratio/mean': 1.0104434490203857, 'sampling/importance_sampling_ratio/max': 1.9806965589523315, 'entropy': 0.5305213257670403, 'clip_ratio/low_mean': 0.038811630103737116, 'clip_ratio/low_min': 0.038811630103737116, 'clip_ratio/high_mean': 0.0311390720307827, 'clip_ratio/high_max': 0.0311390720307827, 'clip_ratio/region_mean': 0.06995070213451982, 'reward_total_mean': 0.7038965821266174, 'reward_meter_mean': 0.9679084420204163, 'reward_meter_std': 0.031183168292045593, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9141935110092163, 'reward_repeat_soft_std': 0.06871075928211212, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.7038965821266174, 'reward_total_composite_std': 0.15150462090969086, 'epoch': 0.17}

 50%|█████     | 1661/3300 [4:05:30<3:46:21,  8.29s/it]INFO 04-13 11:26:28 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1662/3300 [4:05:39<3:44:48,  8.23s/it]2026-04-13 11:26:36,865 | INFO | train_grpo_train | metrics_logged mode=train step=1662

                                                       
{'loss': -0.0129, 'grad_norm': 5.485751628875732, 'learning_rate': 4.966666666666667e-06, 'num_tokens': 2948814.0, 'completions/mean_length': 117.0, 'completions/min_length': 95.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.0, 'completions/min_terminated_length': 95.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.8648841381072998, 'rewards/meter/std': 0.16966873407363892, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6611753106117249, 'rewards/repeat_soft/std': 0.055279362946748734, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.522525429725647, 'rewards/total_composite/std': 0.06771218776702881, 'reward': 0.522525429725647, 'reward_std': 0.0677121952176094, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06389590352773666, 'sampling/sampling_logp_difference/max': 1.5982273817062378, 'sampling/importance_sampling_ratio/min': 0.20225471258163452, 'sampling/importance_sampling_ratio/mean': 1.0021018981933594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2970382459461689, 'clip_ratio/low_mean': 0.03604708984494209, 'clip_ratio/low_min': 0.03604708984494209, 'clip_ratio/high_mean': 0.024507308844476938, 'clip_ratio/high_max': 0.024507308844476938, 'clip_ratio/region_mean': 0.06055439868941903, 'reward_total_mean': 0.522525429725647, 'reward_meter_mean': 0.8648841381072998, 'reward_meter_std': 0.16966873407363892, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6611753106117249, 'reward_repeat_soft_std': 0.055279362946748734, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.522525429725647, 'reward_total_composite_std': 0.06771218776702881, 'epoch': 0.17}

 50%|█████     | 1662/3300 [4:05:39<3:44:48,  8.23s/it]INFO 04-13 11:26:37 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1663/3300 [4:05:50<4:11:18,  9.21s/it]2026-04-13 11:26:48,361 | INFO | train_grpo_train | metrics_logged mode=train step=1663

                                                       
{'loss': -0.1132, 'grad_norm': 2.1583328247070312, 'learning_rate': 4.963636363636364e-06, 'num_tokens': 2950458.0, 'completions/mean_length': 101.5, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 42.85714340209961, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.872140645980835, 'rewards/meter/std': 0.2246885895729065, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9181938767433167, 'rewards/repeat_soft/std': 0.0227973610162735, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.23439893126487732, 'rewards/total_composite/mean': 0.564523458480835, 'rewards/total_composite/std': 0.2551872432231903, 'reward': 0.564523458480835, 'reward_std': 0.2551872432231903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10454196482896805, 'sampling/sampling_logp_difference/max': 2.0216989517211914, 'sampling/importance_sampling_ratio/min': 0.1324302852153778, 'sampling/importance_sampling_ratio/mean': 1.0126889944076538, 'sampling/importance_sampling_ratio/max': 1.9438802003860474, 'entropy': 0.6456299051642418, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11845581140369177, 'clip_ratio/high_max': 0.11845581140369177, 'clip_ratio/region_mean': 0.11845581140369177, 'reward_total_mean': 0.564523458480835, 'reward_meter_mean': 0.872140645980835, 'reward_meter_std': 0.2246885895729065, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9181938767433167, 'reward_repeat_soft_std': 0.0227973610162735, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.23439893126487732, 'reward_total_composite_mean': 0.564523458480835, 'reward_total_composite_std': 0.2551872432231903, 'epoch': 0.17}

 50%|█████     | 1663/3300 [4:05:50<4:11:18,  9.21s/it]INFO 04-13 11:26:48 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1664/3300 [4:05:57<3:52:21,  8.52s/it]2026-04-13 11:26:55,271 | INFO | train_grpo_train | metrics_logged mode=train step=1664

                                                       
{'loss': 0.0608, 'grad_norm': 18.309833526611328, 'learning_rate': 4.9606060606060605e-06, 'num_tokens': 2951779.0, 'completions/mean_length': 27.125, 'completions/min_length': 19.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.125, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8375441431999207, 'rewards/meter/std': 0.310146689414978, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9536747932434082, 'rewards/repeat_soft/std': 0.012406822293996811, 'rewards/judge_quality/mean': 0.39625000953674316, 'rewards/judge_quality/std': 0.09085899591445923, 'rewards/total_composite/mean': 0.5535157918930054, 'rewards/total_composite/std': 0.08959093689918518, 'reward': 0.5535157918930054, 'reward_std': 0.08959092944860458, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12178554385900497, 'sampling/sampling_logp_difference/max': 0.7920396327972412, 'sampling/importance_sampling_ratio/min': 0.4529200494289398, 'sampling/importance_sampling_ratio/mean': 1.027703046798706, 'sampling/importance_sampling_ratio/max': 1.977829933166504, 'entropy': 1.03718501329422, 'clip_ratio/low_mean': 0.04992155823856592, 'clip_ratio/low_min': 0.04992155823856592, 'clip_ratio/high_mean': 0.07084077317267656, 'clip_ratio/high_max': 0.07084077317267656, 'clip_ratio/region_mean': 0.12076233141124249, 'reward_total_mean': 0.5535157918930054, 'reward_meter_mean': 0.8375441431999207, 'reward_meter_std': 0.310146689414978, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9536747932434082, 'reward_repeat_soft_std': 0.012406822293996811, 'reward_judge_quality_mean': 0.39625000953674316, 'reward_judge_quality_std': 0.09085899591445923, 'reward_total_composite_mean': 0.5535157918930054, 'reward_total_composite_std': 0.08959093689918518, 'epoch': 0.17}

 50%|█████     | 1664/3300 [4:05:57<3:52:21,  8.52s/it]INFO 04-13 11:26:55 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1665/3300 [4:06:04<3:37:42,  7.99s/it]2026-04-13 11:27:02,012 | INFO | train_grpo_train | metrics_logged mode=train step=1665

                                                       
{'loss': 0.0436, 'grad_norm': 9.189441680908203, 'learning_rate': 4.957575757575758e-06, 'num_tokens': 2953601.0, 'completions/mean_length': 56.75, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.5839331150054932, 'rewards/meter/std': 0.33002543449401855, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8933404088020325, 'rewards/repeat_soft/std': 0.05125471204519272, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.4816441833972931, 'rewards/total_composite/std': 0.08596338331699371, 'reward': 0.4816441833972931, 'reward_std': 0.08596337586641312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09927050769329071, 'sampling/sampling_logp_difference/max': 2.0201175212860107, 'sampling/importance_sampling_ratio/min': 0.13263988494873047, 'sampling/importance_sampling_ratio/mean': 1.0076899528503418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44833118095993996, 'clip_ratio/low_mean': 0.06354818027466536, 'clip_ratio/low_min': 0.06354818027466536, 'clip_ratio/high_mean': 0.03069634223356843, 'clip_ratio/high_max': 0.03069634223356843, 'clip_ratio/region_mean': 0.09424452250823379, 'reward_total_mean': 0.4816441833972931, 'reward_meter_mean': 0.5839331150054932, 'reward_meter_std': 0.33002543449401855, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8933404088020325, 'reward_repeat_soft_std': 0.05125471204519272, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.4816441833972931, 'reward_total_composite_std': 0.08596338331699371, 'epoch': 0.17}

 50%|█████     | 1665/3300 [4:06:04<3:37:42,  7.99s/it]INFO 04-13 11:27:02 [block_pool.py:378] Successfully reset prefix cache

 50%|█████     | 1666/3300 [4:06:16<4:14:47,  9.36s/it]2026-04-13 11:27:14,563 | INFO | train_grpo_train | metrics_logged mode=train step=1666

                                                       
{'loss': -0.0541, 'grad_norm': 4.4340715408325195, 'learning_rate': 4.954545454545455e-06, 'num_tokens': 2954941.0, 'completions/mean_length': 83.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 22.285715103149414, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.668758749961853, 'rewards/meter/std': 0.44776076078414917, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9480493068695068, 'rewards/repeat_soft/std': 0.03862285614013672, 'rewards/judge_quality/mean': 0.26249998807907104, 'rewards/judge_quality/std': 0.13562026619911194, 'rewards/total_composite/mean': 0.3588586449623108, 'rewards/total_composite/std': 0.23403485119342804, 'reward': 0.3588586449623108, 'reward_std': 0.23403485119342804, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15646137297153473, 'sampling/sampling_logp_difference/max': 1.7519347667694092, 'sampling/importance_sampling_ratio/min': 0.17343805730342865, 'sampling/importance_sampling_ratio/mean': 1.018505573272705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1319462731480598, 'clip_ratio/low_mean': 0.024801588151603937, 'clip_ratio/low_min': 0.024801588151603937, 'clip_ratio/high_mean': 0.0837974464520812, 'clip_ratio/high_max': 0.0837974464520812, 'clip_ratio/region_mean': 0.10859903460368514, 'reward_total_mean': 0.3588586449623108, 'reward_meter_mean': 0.668758749961853, 'reward_meter_std': 0.44776076078414917, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9480493068695068, 'reward_repeat_soft_std': 0.03862285614013672, 'reward_judge_quality_mean': 0.26249998807907104, 'reward_judge_quality_std': 0.13562026619911194, 'reward_total_composite_mean': 0.3588586449623108, 'reward_total_composite_std': 0.23403485119342804, 'epoch': 0.17}

 50%|█████     | 1666/3300 [4:06:16<4:14:47,  9.36s/it]INFO 04-13 11:27:14 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1667/3300 [4:06:29<4:39:17, 10.26s/it]2026-04-13 11:27:27,001 | INFO | train_grpo_train | metrics_logged mode=train step=1667

                                                       
{'loss': -0.2307, 'grad_norm': 1.59761381149292, 'learning_rate': 4.951515151515152e-06, 'num_tokens': 2957507.0, 'completions/mean_length': 206.75, 'completions/min_length': 127.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 163.1428680419922, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 185.0, 'rewards/meter/mean': 0.7981754541397095, 'rewards/meter/std': 0.328860342502594, 'rewards/count_adherence/mean': 0.8333333134651184, 'rewards/count_adherence/std': 0.28171807527542114, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7329895496368408, 'rewards/repeat_soft/std': 0.15630879998207092, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.14961259067058563, 'rewards/total_composite/mean': 0.4375540316104889, 'rewards/total_composite/std': 0.18709851801395416, 'reward': 0.4375540316104889, 'reward_std': 0.18709851801395416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06901933997869492, 'sampling/sampling_logp_difference/max': 1.4699207544326782, 'sampling/importance_sampling_ratio/min': 0.22994370758533478, 'sampling/importance_sampling_ratio/mean': 1.0060914754867554, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3053620681166649, 'clip_ratio/low_mean': 0.003937007859349251, 'clip_ratio/low_min': 0.003937007859349251, 'clip_ratio/high_mean': 0.0556541346013546, 'clip_ratio/high_max': 0.0556541346013546, 'clip_ratio/region_mean': 0.05959114246070385, 'reward_total_mean': 0.4375540316104889, 'reward_meter_mean': 0.7981754541397095, 'reward_meter_std': 0.328860342502594, 'reward_count_adherence_mean': 0.8333333134651184, 'reward_count_adherence_std': 0.28171807527542114, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7329895496368408, 'reward_repeat_soft_std': 0.15630879998207092, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.14961259067058563, 'reward_total_composite_mean': 0.4375540316104889, 'reward_total_composite_std': 0.18709851801395416, 'epoch': 0.17}

 51%|█████     | 1667/3300 [4:06:29<4:39:17, 10.26s/it]INFO 04-13 11:27:27 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1668/3300 [4:06:40<4:52:23, 10.75s/it]2026-04-13 11:27:38,827 | INFO | train_grpo_train | metrics_logged mode=train step=1668

                                                       
{'loss': -0.2336, 'grad_norm': 2.001739740371704, 'learning_rate': 4.9484848484848495e-06, 'num_tokens': 2959795.0, 'completions/mean_length': 285.0, 'completions/min_length': 127.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 148.8000030517578, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.6233569979667664, 'rewards/meter/std': 0.3930578827857971, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9499600529670715, 'rewards/repeat_soft/std': 0.0399932935833931, 'rewards/judge_quality/mean': 0.20749999582767487, 'rewards/judge_quality/std': 0.18312759697437286, 'rewards/total_composite/mean': 0.31414613127708435, 'rewards/total_composite/std': 0.2726595997810364, 'reward': 0.31414613127708435, 'reward_std': 0.2726595997810364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10766993463039398, 'sampling/sampling_logp_difference/max': 1.7286872863769531, 'sampling/importance_sampling_ratio/min': 0.17751729488372803, 'sampling/importance_sampling_ratio/mean': 1.0270744562149048, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6006448864936829, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06657080911099911, 'clip_ratio/high_max': 0.06657080911099911, 'clip_ratio/region_mean': 0.06657080911099911, 'reward_total_mean': 0.31414613127708435, 'reward_meter_mean': 0.6233569979667664, 'reward_meter_std': 0.3930578827857971, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9499600529670715, 'reward_repeat_soft_std': 0.0399932935833931, 'reward_judge_quality_mean': 0.20749999582767487, 'reward_judge_quality_std': 0.18312759697437286, 'reward_total_composite_mean': 0.31414613127708435, 'reward_total_composite_std': 0.2726595997810364, 'epoch': 0.17}

 51%|█████     | 1668/3300 [4:06:41<4:52:23, 10.75s/it]INFO 04-13 11:27:39 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1669/3300 [4:06:52<5:01:05, 11.08s/it]2026-04-13 11:27:50,666 | INFO | train_grpo_train | metrics_logged mode=train step=1669

                                                       
{'loss': -0.08, 'grad_norm': 2.2405269145965576, 'learning_rate': 4.945454545454546e-06, 'num_tokens': 2961143.0, 'completions/mean_length': 85.5, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 24.571430206298828, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.831732988357544, 'rewards/meter/std': 0.3342703580856323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9585226774215698, 'rewards/repeat_soft/std': 0.011249415576457977, 'rewards/judge_quality/mean': 0.45125001668930054, 'rewards/judge_quality/std': 0.23381540179252625, 'rewards/total_composite/mean': 0.5760042071342468, 'rewards/total_composite/std': 0.25966617465019226, 'reward': 0.5760042071342468, 'reward_std': 0.2596661448478699, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11087220162153244, 'sampling/sampling_logp_difference/max': 1.0156540870666504, 'sampling/importance_sampling_ratio/min': 0.3621654808521271, 'sampling/importance_sampling_ratio/mean': 1.0283763408660889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5842178240418434, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.07428115187212825, 'clip_ratio/high_max': 0.07428115187212825, 'clip_ratio/region_mean': 0.08990615187212825, 'reward_total_mean': 0.5760042071342468, 'reward_meter_mean': 0.831732988357544, 'reward_meter_std': 0.3342703580856323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9585226774215698, 'reward_repeat_soft_std': 0.011249415576457977, 'reward_judge_quality_mean': 0.45125001668930054, 'reward_judge_quality_std': 0.23381540179252625, 'reward_total_composite_mean': 0.5760042071342468, 'reward_total_composite_std': 0.25966617465019226, 'epoch': 0.17}

 51%|█████     | 1669/3300 [4:06:52<5:01:05, 11.08s/it]INFO 04-13 11:27:50 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1670/3300 [4:06:59<4:21:13,  9.62s/it]2026-04-13 11:27:56,875 | INFO | train_grpo_train | metrics_logged mode=train step=1670

                                                       
{'loss': 0.0438, 'grad_norm': 16.120502471923828, 'learning_rate': 4.942424242424243e-06, 'num_tokens': 2962746.0, 'completions/mean_length': 38.375, 'completions/min_length': 29.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.846521258354187, 'rewards/meter/std': 0.26176145672798157, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9566450119018555, 'rewards/repeat_soft/std': 0.06320735812187195, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.12906257808208466, 'rewards/total_composite/mean': 0.5743720531463623, 'rewards/total_composite/std': 0.06774458289146423, 'reward': 0.5743720531463623, 'reward_std': 0.06774458289146423, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1517239809036255, 'sampling/sampling_logp_difference/max': 2.793018341064453, 'sampling/importance_sampling_ratio/min': 0.061236102133989334, 'sampling/importance_sampling_ratio/mean': 0.9983158111572266, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6926508843898773, 'clip_ratio/low_mean': 0.025297619868069887, 'clip_ratio/low_min': 0.025297619868069887, 'clip_ratio/high_mean': 0.09588289353996515, 'clip_ratio/high_max': 0.09588289353996515, 'clip_ratio/region_mean': 0.12118051340803504, 'reward_total_mean': 0.5743720531463623, 'reward_meter_mean': 0.846521258354187, 'reward_meter_std': 0.26176145672798157, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9566450119018555, 'reward_repeat_soft_std': 0.06320735812187195, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.12906257808208466, 'reward_total_composite_mean': 0.5743720531463623, 'reward_total_composite_std': 0.06774458289146423, 'epoch': 0.17}

 51%|█████     | 1670/3300 [4:06:59<4:21:13,  9.62s/it]INFO 04-13 11:27:57 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1671/3300 [4:07:07<4:12:13,  9.29s/it]2026-04-13 11:28:05,403 | INFO | train_grpo_train | metrics_logged mode=train step=1671

                                                       
{'loss': 0.0323, 'grad_norm': 5.545763969421387, 'learning_rate': 4.93939393939394e-06, 'num_tokens': 2965486.0, 'completions/mean_length': 159.5, 'completions/min_length': 129.0, 'completions/max_length': 197.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.5, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 197.0, 'rewards/meter/mean': 0.6789463758468628, 'rewards/meter/std': 0.3376133143901825, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8708106875419617, 'rewards/repeat_soft/std': 0.12946419417858124, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.4800385534763336, 'rewards/total_composite/std': 0.07363229244947433, 'reward': 0.4800385534763336, 'reward_std': 0.07363227754831314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0955670103430748, 'sampling/sampling_logp_difference/max': 1.5123744010925293, 'sampling/importance_sampling_ratio/min': 0.2203860729932785, 'sampling/importance_sampling_ratio/mean': 1.004928469657898, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.565850742161274, 'clip_ratio/low_mean': 0.04794883681461215, 'clip_ratio/low_min': 0.04794883681461215, 'clip_ratio/high_mean': 0.035761090926826, 'clip_ratio/high_max': 0.035761090926826, 'clip_ratio/region_mean': 0.08370992774143815, 'reward_total_mean': 0.4800385534763336, 'reward_meter_mean': 0.6789463758468628, 'reward_meter_std': 0.3376133143901825, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8708106875419617, 'reward_repeat_soft_std': 0.12946419417858124, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.4800385534763336, 'reward_total_composite_std': 0.07363229244947433, 'epoch': 0.17}

 51%|█████     | 1671/3300 [4:07:07<4:12:13,  9.29s/it]INFO 04-13 11:28:05 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1672/3300 [4:07:13<3:46:20,  8.34s/it]2026-04-13 11:28:11,531 | INFO | train_grpo_train | metrics_logged mode=train step=1672

                                                       
{'loss': 0.2415, 'grad_norm': 18.02611541748047, 'learning_rate': 4.936363636363637e-06, 'num_tokens': 2966841.0, 'completions/mean_length': 26.375, 'completions/min_length': 20.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.375, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8848724961280823, 'rewards/meter/std': 0.294619083404541, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.713463306427002, 'rewards/repeat_soft/std': 0.34030842781066895, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.19255799055099487, 'rewards/total_composite/mean': 0.5555944442749023, 'rewards/total_composite/std': 0.20079876482486725, 'reward': 0.5555944442749023, 'reward_std': 0.20079874992370605, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11104923486709595, 'sampling/sampling_logp_difference/max': 1.8065016269683838, 'sampling/importance_sampling_ratio/min': 0.1642276644706726, 'sampling/importance_sampling_ratio/mean': 1.0184675455093384, 'sampling/importance_sampling_ratio/max': 1.7557411193847656, 'entropy': 0.7442312054336071, 'clip_ratio/low_mean': 0.029363354202359915, 'clip_ratio/low_min': 0.029363354202359915, 'clip_ratio/high_mean': 0.05451839976012707, 'clip_ratio/high_max': 0.05451839976012707, 'clip_ratio/region_mean': 0.08388175396248698, 'reward_total_mean': 0.5555944442749023, 'reward_meter_mean': 0.8848724961280823, 'reward_meter_std': 0.294619083404541, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.713463306427002, 'reward_repeat_soft_std': 0.34030842781066895, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.19255799055099487, 'reward_total_composite_mean': 0.5555944442749023, 'reward_total_composite_std': 0.20079876482486725, 'epoch': 0.17}

 51%|█████     | 1672/3300 [4:07:13<3:46:20,  8.34s/it]INFO 04-13 11:28:11 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1673/3300 [4:07:20<3:36:20,  7.98s/it]2026-04-13 11:28:18,661 | INFO | train_grpo_train | metrics_logged mode=train step=1673

                                                       
{'loss': -0.0, 'grad_norm': 7.407658100128174, 'learning_rate': 4.933333333333334e-06, 'num_tokens': 2968831.0, 'completions/mean_length': 80.75, 'completions/min_length': 64.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.75, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.6783466339111328, 'rewards/meter/std': 0.36895039677619934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.720346212387085, 'rewards/repeat_soft/std': 0.13622376322746277, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.49563726782798767, 'rewards/total_composite/std': 0.09533317387104034, 'reward': 0.49563726782798767, 'reward_std': 0.09533317387104034, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0666607990860939, 'sampling/sampling_logp_difference/max': 1.689176321029663, 'sampling/importance_sampling_ratio/min': 0.1846715807914734, 'sampling/importance_sampling_ratio/mean': 1.0037403106689453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3736836090683937, 'clip_ratio/low_mean': 0.023216268629767, 'clip_ratio/low_min': 0.023216268629767, 'clip_ratio/high_mean': 0.036169617902487516, 'clip_ratio/high_max': 0.036169617902487516, 'clip_ratio/region_mean': 0.05938588653225452, 'reward_total_mean': 0.49563726782798767, 'reward_meter_mean': 0.6783466339111328, 'reward_meter_std': 0.36895039677619934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.720346212387085, 'reward_repeat_soft_std': 0.13622376322746277, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.49563726782798767, 'reward_total_composite_std': 0.09533317387104034, 'epoch': 0.17}

 51%|█████     | 1673/3300 [4:07:20<3:36:20,  7.98s/it]INFO 04-13 11:28:18 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1674/3300 [4:07:28<3:30:42,  7.78s/it]2026-04-13 11:28:25,963 | INFO | train_grpo_train | metrics_logged mode=train step=1674

                                                       
{'loss': 0.021, 'grad_norm': 8.999605178833008, 'learning_rate': 4.9303030303030305e-06, 'num_tokens': 2970778.0, 'completions/mean_length': 84.375, 'completions/min_length': 62.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.375, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.32025760412216187, 'rewards/meter/std': 0.1794765591621399, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8883340358734131, 'rewards/repeat_soft/std': 0.07298793643712997, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.4041918218135834, 'rewards/total_composite/std': 0.1147385984659195, 'reward': 0.4041918218135834, 'reward_std': 0.11473860591650009, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11148685216903687, 'sampling/sampling_logp_difference/max': 2.3669281005859375, 'sampling/importance_sampling_ratio/min': 0.09376832842826843, 'sampling/importance_sampling_ratio/mean': 0.9885243773460388, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39864677004516125, 'clip_ratio/low_mean': 0.0698322867974639, 'clip_ratio/low_min': 0.0698322867974639, 'clip_ratio/high_mean': 0.03228022065013647, 'clip_ratio/high_max': 0.03228022065013647, 'clip_ratio/region_mean': 0.10211250744760036, 'reward_total_mean': 0.4041918218135834, 'reward_meter_mean': 0.32025760412216187, 'reward_meter_std': 0.1794765591621399, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8883340358734131, 'reward_repeat_soft_std': 0.07298793643712997, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.4041918218135834, 'reward_total_composite_std': 0.1147385984659195, 'epoch': 0.17}

 51%|█████     | 1674/3300 [4:07:28<3:30:42,  7.78s/it]INFO 04-13 11:28:26 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1675/3300 [4:07:35<3:29:38,  7.74s/it]2026-04-13 11:28:33,625 | INFO | train_grpo_train | metrics_logged mode=train step=1675

                                                       
{'loss': -0.0038, 'grad_norm': 4.344687461853027, 'learning_rate': 4.927272727272728e-06, 'num_tokens': 2973342.0, 'completions/mean_length': 126.5, 'completions/min_length': 114.0, 'completions/max_length': 141.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.5, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 141.0, 'rewards/meter/mean': 0.8118463754653931, 'rewards/meter/std': 0.30178484320640564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6909312605857849, 'rewards/repeat_soft/std': 0.048823073506355286, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5161367654800415, 'rewards/total_composite/std': 0.09046050906181335, 'reward': 0.5161367654800415, 'reward_std': 0.09046050906181335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05651190131902695, 'sampling/sampling_logp_difference/max': 1.9631762504577637, 'sampling/importance_sampling_ratio/min': 0.14041173458099365, 'sampling/importance_sampling_ratio/mean': 1.006211280822754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.301896084100008, 'clip_ratio/low_mean': 0.013461179565638304, 'clip_ratio/low_min': 0.013461179565638304, 'clip_ratio/high_mean': 0.0389620391651988, 'clip_ratio/high_max': 0.0389620391651988, 'clip_ratio/region_mean': 0.05242321873083711, 'reward_total_mean': 0.5161367654800415, 'reward_meter_mean': 0.8118463754653931, 'reward_meter_std': 0.30178484320640564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6909312605857849, 'reward_repeat_soft_std': 0.048823073506355286, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5161367654800415, 'reward_total_composite_std': 0.09046050906181335, 'epoch': 0.17}

 51%|█████     | 1675/3300 [4:07:35<3:29:38,  7.74s/it]INFO 04-13 11:28:33 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1676/3300 [4:07:44<3:38:18,  8.07s/it]2026-04-13 11:28:42,445 | INFO | train_grpo_train | metrics_logged mode=train step=1676

                                                       
{'loss': 0.0526, 'grad_norm': 4.972901821136475, 'learning_rate': 4.924242424242425e-06, 'num_tokens': 2976235.0, 'completions/mean_length': 161.625, 'completions/min_length': 131.0, 'completions/max_length': 212.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 161.625, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.8618706464767456, 'rewards/meter/std': 0.2439996600151062, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7606981992721558, 'rewards/repeat_soft/std': 0.13278447091579437, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.503989040851593, 'rewards/total_composite/std': 0.06762686371803284, 'reward': 0.503989040851593, 'reward_std': 0.06762687116861343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07873500883579254, 'sampling/sampling_logp_difference/max': 1.9043989181518555, 'sampling/importance_sampling_ratio/min': 0.1527395099401474, 'sampling/importance_sampling_ratio/mean': 1.0007163286209106, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40678513795137405, 'clip_ratio/low_mean': 0.022617980604991317, 'clip_ratio/low_min': 0.022617980604991317, 'clip_ratio/high_mean': 0.045249105896800756, 'clip_ratio/high_max': 0.045249105896800756, 'clip_ratio/region_mean': 0.06786708650179207, 'reward_total_mean': 0.503989040851593, 'reward_meter_mean': 0.8618706464767456, 'reward_meter_std': 0.2439996600151062, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7606981992721558, 'reward_repeat_soft_std': 0.13278447091579437, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.503989040851593, 'reward_total_composite_std': 0.06762686371803284, 'epoch': 0.17}

 51%|█████     | 1676/3300 [4:07:44<3:38:18,  8.07s/it]INFO 04-13 11:28:42 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1677/3300 [4:07:50<3:24:01,  7.54s/it]2026-04-13 11:28:48,762 | INFO | train_grpo_train | metrics_logged mode=train step=1677

                                                       
{'loss': -0.0091, 'grad_norm': 9.968160629272461, 'learning_rate': 4.9212121212121214e-06, 'num_tokens': 2977843.0, 'completions/mean_length': 40.0, 'completions/min_length': 35.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.5374641418457031, 'rewards/meter/std': 0.4309203624725342, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8733203411102295, 'rewards/repeat_soft/std': 0.058413829654455185, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.4737306237220764, 'rewards/total_composite/std': 0.26079532504081726, 'reward': 0.4737306237220764, 'reward_std': 0.2607952952384949, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08989720791578293, 'sampling/sampling_logp_difference/max': 2.8600518703460693, 'sampling/importance_sampling_ratio/min': 0.057265788316726685, 'sampling/importance_sampling_ratio/mean': 1.0158300399780273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49360642209649086, 'clip_ratio/low_mean': 0.026151658268645406, 'clip_ratio/low_min': 0.026151658268645406, 'clip_ratio/high_mean': 0.03813495417125523, 'clip_ratio/high_max': 0.03813495417125523, 'clip_ratio/region_mean': 0.06428661243990064, 'reward_total_mean': 0.4737306237220764, 'reward_meter_mean': 0.5374641418457031, 'reward_meter_std': 0.4309203624725342, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8733203411102295, 'reward_repeat_soft_std': 0.058413829654455185, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.4737306237220764, 'reward_total_composite_std': 0.26079532504081726, 'epoch': 0.17}

 51%|█████     | 1677/3300 [4:07:50<3:24:01,  7.54s/it]INFO 04-13 11:28:49 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1678/3300 [4:07:59<3:32:38,  7.87s/it]2026-04-13 11:28:57,391 | INFO | train_grpo_train | metrics_logged mode=train step=1678

                                                       
{'loss': -0.0214, 'grad_norm': 2.942268133163452, 'learning_rate': 4.918181818181819e-06, 'num_tokens': 2980783.0, 'completions/mean_length': 186.5, 'completions/min_length': 154.0, 'completions/max_length': 226.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 186.5, 'completions/min_terminated_length': 154.0, 'completions/max_terminated_length': 226.0, 'rewards/meter/mean': 0.8959339261054993, 'rewards/meter/std': 0.21870744228363037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.5695617198944092, 'rewards/repeat_soft/std': 0.0780857726931572, 'rewards/judge_quality/mean': 0.22499999403953552, 'rewards/judge_quality/std': 0.04629100486636162, 'rewards/total_composite/mean': 0.4149327278137207, 'rewards/total_composite/std': 0.04342221841216087, 'reward': 0.4149327278137207, 'reward_std': 0.04342221841216087, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04488222673535347, 'sampling/sampling_logp_difference/max': 2.6392836570739746, 'sampling/importance_sampling_ratio/min': 0.07141240686178207, 'sampling/importance_sampling_ratio/mean': 1.0022094249725342, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.274970768019557, 'clip_ratio/low_mean': 0.011812139535322785, 'clip_ratio/low_min': 0.011812139535322785, 'clip_ratio/high_mean': 0.03270714543759823, 'clip_ratio/high_max': 0.03270714543759823, 'clip_ratio/region_mean': 0.044519284972921014, 'reward_total_mean': 0.4149327278137207, 'reward_meter_mean': 0.8959339261054993, 'reward_meter_std': 0.21870744228363037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.5695617198944092, 'reward_repeat_soft_std': 0.0780857726931572, 'reward_judge_quality_mean': 0.22499999403953552, 'reward_judge_quality_std': 0.04629100486636162, 'reward_total_composite_mean': 0.4149327278137207, 'reward_total_composite_std': 0.04342221841216087, 'epoch': 0.17}

 51%|█████     | 1678/3300 [4:07:59<3:32:38,  7.87s/it]INFO 04-13 11:28:57 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1679/3300 [4:08:11<4:02:33,  8.98s/it]2026-04-13 11:29:08,962 | INFO | train_grpo_train | metrics_logged mode=train step=1679

                                                       
{'loss': -0.2272, 'grad_norm': 1.6492377519607544, 'learning_rate': 4.915151515151516e-06, 'num_tokens': 2983300.0, 'completions/mean_length': 199.625, 'completions/min_length': 99.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 155.0, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.769597589969635, 'rewards/meter/std': 0.20830576121807098, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.21380899846553802, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.6938011646270752, 'rewards/repeat_soft/std': 0.13908973336219788, 'rewards/judge_quality/mean': 0.2887499928474426, 'rewards/judge_quality/std': 0.15797263383865356, 'rewards/total_composite/mean': 0.39985230565071106, 'rewards/total_composite/std': 0.17132367193698883, 'reward': 0.39985230565071106, 'reward_std': 0.17132367193698883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0714179128408432, 'sampling/sampling_logp_difference/max': 3.210151195526123, 'sampling/importance_sampling_ratio/min': 0.04035051539540291, 'sampling/importance_sampling_ratio/mean': 1.013409972190857, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30607396736741066, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.045120069524273276, 'clip_ratio/high_max': 0.045120069524273276, 'clip_ratio/region_mean': 0.048498447984457016, 'reward_total_mean': 0.39985230565071106, 'reward_meter_mean': 0.769597589969635, 'reward_meter_std': 0.20830576121807098, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.21380899846553802, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.6938011646270752, 'reward_repeat_soft_std': 0.13908973336219788, 'reward_judge_quality_mean': 0.2887499928474426, 'reward_judge_quality_std': 0.15797263383865356, 'reward_total_composite_mean': 0.39985230565071106, 'reward_total_composite_std': 0.17132367193698883, 'epoch': 0.17}

 51%|█████     | 1679/3300 [4:08:11<4:02:33,  8.98s/it]INFO 04-13 11:29:09 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1680/3300 [4:08:18<3:45:53,  8.37s/it]2026-04-13 11:29:15,904 | INFO | train_grpo_train | metrics_logged mode=train step=1680

                                                       
{'loss': 0.077, 'grad_norm': 9.330824851989746, 'learning_rate': 4.912121212121212e-06, 'num_tokens': 2985351.0, 'completions/mean_length': 94.375, 'completions/min_length': 82.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.375, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.8564003109931946, 'rewards/meter/std': 0.1464369297027588, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7046080827713013, 'rewards/repeat_soft/std': 0.07144229859113693, 'rewards/judge_quality/mean': 0.38624998927116394, 'rewards/judge_quality/std': 0.09545940905809402, 'rewards/total_composite/mean': 0.470175176858902, 'rewards/total_composite/std': 0.06205389276146889, 'reward': 0.470175176858902, 'reward_std': 0.062053896486759186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08212559670209885, 'sampling/sampling_logp_difference/max': 2.9680633544921875, 'sampling/importance_sampling_ratio/min': 0.05140276253223419, 'sampling/importance_sampling_ratio/mean': 0.9970459342002869, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.302042568102479, 'clip_ratio/low_mean': 0.011065413244068623, 'clip_ratio/low_min': 0.011065413244068623, 'clip_ratio/high_mean': 0.05521555768791586, 'clip_ratio/high_max': 0.05521555768791586, 'clip_ratio/region_mean': 0.06628097093198448, 'reward_total_mean': 0.470175176858902, 'reward_meter_mean': 0.8564003109931946, 'reward_meter_std': 0.1464369297027588, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7046080827713013, 'reward_repeat_soft_std': 0.07144229859113693, 'reward_judge_quality_mean': 0.38624998927116394, 'reward_judge_quality_std': 0.09545940905809402, 'reward_total_composite_mean': 0.470175176858902, 'reward_total_composite_std': 0.06205389276146889, 'epoch': 0.17}

 51%|█████     | 1680/3300 [4:08:18<3:45:53,  8.37s/it]INFO 04-13 11:29:16 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1681/3300 [4:08:24<3:28:49,  7.74s/it]2026-04-13 11:29:22,175 | INFO | train_grpo_train | metrics_logged mode=train step=1681

                                                       
{'loss': 0.0655, 'grad_norm': 9.672286987304688, 'learning_rate': 4.90909090909091e-06, 'num_tokens': 2987043.0, 'completions/mean_length': 56.5, 'completions/min_length': 45.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.5, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.9728363156318665, 'rewards/meter/std': 0.018391400575637817, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8520519137382507, 'rewards/repeat_soft/std': 0.04583599418401718, 'rewards/judge_quality/mean': 0.32624998688697815, 'rewards/judge_quality/std': 0.11350739002227783, 'rewards/total_composite/mean': 0.5336079597473145, 'rewards/total_composite/std': 0.07198585569858551, 'reward': 0.5336079597473145, 'reward_std': 0.07198584824800491, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0933646410703659, 'sampling/sampling_logp_difference/max': 1.170951008796692, 'sampling/importance_sampling_ratio/min': 0.3100719451904297, 'sampling/importance_sampling_ratio/mean': 1.0184705257415771, 'sampling/importance_sampling_ratio/max': 1.6749534606933594, 'entropy': 0.626958291977644, 'clip_ratio/low_mean': 0.032328391214832664, 'clip_ratio/low_min': 0.032328391214832664, 'clip_ratio/high_mean': 0.0491899773478508, 'clip_ratio/high_max': 0.0491899773478508, 'clip_ratio/region_mean': 0.08151836856268346, 'reward_total_mean': 0.5336079597473145, 'reward_meter_mean': 0.9728363156318665, 'reward_meter_std': 0.018391400575637817, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8520519137382507, 'reward_repeat_soft_std': 0.04583599418401718, 'reward_judge_quality_mean': 0.32624998688697815, 'reward_judge_quality_std': 0.11350739002227783, 'reward_total_composite_mean': 0.5336079597473145, 'reward_total_composite_std': 0.07198585569858551, 'epoch': 0.17}

 51%|█████     | 1681/3300 [4:08:24<3:28:49,  7.74s/it]INFO 04-13 11:29:22 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1682/3300 [4:08:31<3:21:26,  7.47s/it]2026-04-13 11:29:29,024 | INFO | train_grpo_train | metrics_logged mode=train step=1682

                                                       
{'loss': 0.041, 'grad_norm': 12.904158592224121, 'learning_rate': 4.906060606060606e-06, 'num_tokens': 2988684.0, 'completions/mean_length': 54.125, 'completions/min_length': 49.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9888840317726135, 'rewards/meter/std': 0.013273422606289387, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8677555322647095, 'rewards/repeat_soft/std': 0.07466918230056763, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5888320207595825, 'rewards/total_composite/std': 0.04263556748628616, 'reward': 0.5888320207595825, 'reward_std': 0.04263556748628616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09531761705875397, 'sampling/sampling_logp_difference/max': 1.3383307456970215, 'sampling/importance_sampling_ratio/min': 0.2622831165790558, 'sampling/importance_sampling_ratio/mean': 1.0110734701156616, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5609508864581585, 'clip_ratio/low_mean': 0.028822056017816067, 'clip_ratio/low_min': 0.028822056017816067, 'clip_ratio/high_mean': 0.06253244495019317, 'clip_ratio/high_max': 0.06253244495019317, 'clip_ratio/region_mean': 0.09135450096800923, 'reward_total_mean': 0.5888320207595825, 'reward_meter_mean': 0.9888840317726135, 'reward_meter_std': 0.013273422606289387, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8677555322647095, 'reward_repeat_soft_std': 0.07466918230056763, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5888320207595825, 'reward_total_composite_std': 0.04263556748628616, 'epoch': 0.17}

 51%|█████     | 1682/3300 [4:08:31<3:21:26,  7.47s/it]INFO 04-13 11:29:29 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1683/3300 [4:08:37<3:11:35,  7.11s/it]2026-04-13 11:29:35,279 | INFO | train_grpo_train | metrics_logged mode=train step=1683

                                                       
{'loss': 0.0611, 'grad_norm': 12.600188255310059, 'learning_rate': 4.903030303030303e-06, 'num_tokens': 2990363.0, 'completions/mean_length': 51.875, 'completions/min_length': 42.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.49862271547317505, 'rewards/meter/std': 0.32598933577537537, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9934655427932739, 'rewards/repeat_soft/std': 0.011311059817671776, 'rewards/judge_quality/mean': 0.6349999904632568, 'rewards/judge_quality/std': 0.2235429286956787, 'rewards/total_composite/mean': 0.5552878379821777, 'rewards/total_composite/std': 0.13822627067565918, 'reward': 0.5552878379821777, 'reward_std': 0.13822627067565918, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12635992467403412, 'sampling/sampling_logp_difference/max': 1.9483773708343506, 'sampling/importance_sampling_ratio/min': 0.14250512421131134, 'sampling/importance_sampling_ratio/mean': 1.0187221765518188, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.514126792550087, 'clip_ratio/low_mean': 0.05596538633108139, 'clip_ratio/low_min': 0.05596538633108139, 'clip_ratio/high_mean': 0.04068396287038922, 'clip_ratio/high_max': 0.04068396287038922, 'clip_ratio/region_mean': 0.09664934920147061, 'reward_total_mean': 0.5552878379821777, 'reward_meter_mean': 0.49862271547317505, 'reward_meter_std': 0.32598933577537537, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9934655427932739, 'reward_repeat_soft_std': 0.011311059817671776, 'reward_judge_quality_mean': 0.6349999904632568, 'reward_judge_quality_std': 0.2235429286956787, 'reward_total_composite_mean': 0.5552878379821777, 'reward_total_composite_std': 0.13822627067565918, 'epoch': 0.17}

 51%|█████     | 1683/3300 [4:08:37<3:11:35,  7.11s/it]INFO 04-13 11:29:35 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1684/3300 [4:08:43<3:07:04,  6.95s/it]2026-04-13 11:29:41,907 | INFO | train_grpo_train | metrics_logged mode=train step=1684

                                                       
{'loss': 0.1425, 'grad_norm': 13.452690124511719, 'learning_rate': 4.9000000000000005e-06, 'num_tokens': 2991738.0, 'completions/mean_length': 32.875, 'completions/min_length': 20.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.875, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6996747255325317, 'rewards/meter/std': 0.4190917909145355, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.4629100561141968, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7916808128356934, 'rewards/repeat_soft/std': 0.1934211254119873, 'rewards/judge_quality/mean': 0.3349999785423279, 'rewards/judge_quality/std': 0.12972497940063477, 'rewards/total_composite/mean': 0.4318060874938965, 'rewards/total_composite/std': 0.12287203222513199, 'reward': 0.4318060874938965, 'reward_std': 0.12287202477455139, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12454545497894287, 'sampling/sampling_logp_difference/max': 2.507312774658203, 'sampling/importance_sampling_ratio/min': 0.08148691803216934, 'sampling/importance_sampling_ratio/mean': 0.9963033199310303, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7385776005685329, 'clip_ratio/low_mean': 0.04498840123414993, 'clip_ratio/low_min': 0.04498840123414993, 'clip_ratio/high_mean': 0.053281772416085005, 'clip_ratio/high_max': 0.053281772416085005, 'clip_ratio/region_mean': 0.09827017365023494, 'reward_total_mean': 0.4318060874938965, 'reward_meter_mean': 0.6996747255325317, 'reward_meter_std': 0.4190917909145355, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.4629100561141968, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7916808128356934, 'reward_repeat_soft_std': 0.1934211254119873, 'reward_judge_quality_mean': 0.3349999785423279, 'reward_judge_quality_std': 0.12972497940063477, 'reward_total_composite_mean': 0.4318060874938965, 'reward_total_composite_std': 0.12287203222513199, 'epoch': 0.17}

 51%|█████     | 1684/3300 [4:08:44<3:07:04,  6.95s/it]INFO 04-13 11:29:42 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1685/3300 [4:08:51<3:14:32,  7.23s/it]2026-04-13 11:29:49,745 | INFO | train_grpo_train | metrics_logged mode=train step=1685

                                                       
{'loss': 0.0517, 'grad_norm': 6.260843276977539, 'learning_rate': 4.896969696969697e-06, 'num_tokens': 2994191.0, 'completions/mean_length': 130.625, 'completions/min_length': 109.0, 'completions/max_length': 151.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.625, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 151.0, 'rewards/meter/mean': 0.8883123397827148, 'rewards/meter/std': 0.19294770061969757, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9430127143859863, 'rewards/repeat_soft/std': 0.026210656389594078, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.5498702526092529, 'rewards/total_composite/std': 0.08809173852205276, 'reward': 0.5498702526092529, 'reward_std': 0.08809174597263336, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1032494604587555, 'sampling/sampling_logp_difference/max': 1.8032281398773193, 'sampling/importance_sampling_ratio/min': 0.1647661328315735, 'sampling/importance_sampling_ratio/mean': 1.0128101110458374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6937706656754017, 'clip_ratio/low_mean': 0.02845947165042162, 'clip_ratio/low_min': 0.02845947165042162, 'clip_ratio/high_mean': 0.06688423454761505, 'clip_ratio/high_max': 0.06688423454761505, 'clip_ratio/region_mean': 0.09534370619803667, 'reward_total_mean': 0.5498702526092529, 'reward_meter_mean': 0.8883123397827148, 'reward_meter_std': 0.19294770061969757, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9430127143859863, 'reward_repeat_soft_std': 0.026210656389594078, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.5498702526092529, 'reward_total_composite_std': 0.08809173852205276, 'epoch': 0.17}

 51%|█████     | 1685/3300 [4:08:51<3:14:32,  7.23s/it]INFO 04-13 11:29:49 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1686/3300 [4:08:59<3:16:43,  7.31s/it]2026-04-13 11:29:57,253 | INFO | train_grpo_train | metrics_logged mode=train step=1686

                                                       
{'loss': -0.0353, 'grad_norm': 5.953235626220703, 'learning_rate': 4.893939393939394e-06, 'num_tokens': 2996821.0, 'completions/mean_length': 132.75, 'completions/min_length': 111.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 132.75, 'completions/min_terminated_length': 111.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.8653354644775391, 'rewards/meter/std': 0.222174271941185, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8975305557250977, 'rewards/repeat_soft/std': 0.09665177762508392, 'rewards/judge_quality/mean': 0.30124998092651367, 'rewards/judge_quality/std': 0.10398317128419876, 'rewards/total_composite/mean': 0.507197916507721, 'rewards/total_composite/std': 0.08576513826847076, 'reward': 0.507197916507721, 'reward_std': 0.08576513826847076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0919460579752922, 'sampling/sampling_logp_difference/max': 3.7899551391601562, 'sampling/importance_sampling_ratio/min': 0.02259661629796028, 'sampling/importance_sampling_ratio/mean': 1.0128374099731445, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5677018612623215, 'clip_ratio/low_mean': 0.05248046666383743, 'clip_ratio/low_min': 0.05248046666383743, 'clip_ratio/high_mean': 0.03497450239956379, 'clip_ratio/high_max': 0.03497450239956379, 'clip_ratio/region_mean': 0.08745496906340122, 'reward_total_mean': 0.507197916507721, 'reward_meter_mean': 0.8653354644775391, 'reward_meter_std': 0.222174271941185, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8975305557250977, 'reward_repeat_soft_std': 0.09665177762508392, 'reward_judge_quality_mean': 0.30124998092651367, 'reward_judge_quality_std': 0.10398317128419876, 'reward_total_composite_mean': 0.507197916507721, 'reward_total_composite_std': 0.08576513826847076, 'epoch': 0.17}

 51%|█████     | 1686/3300 [4:08:59<3:16:43,  7.31s/it]INFO 04-13 11:29:57 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1687/3300 [4:09:06<3:11:24,  7.12s/it]2026-04-13 11:30:03,910 | INFO | train_grpo_train | metrics_logged mode=train step=1687

                                                       
{'loss': 0.0425, 'grad_norm': 9.852559089660645, 'learning_rate': 4.8909090909090914e-06, 'num_tokens': 2998704.0, 'completions/mean_length': 64.375, 'completions/min_length': 51.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.375, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.9665820598602295, 'rewards/meter/std': 0.06644105166196823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9778945446014404, 'rewards/repeat_soft/std': 0.03151845559477806, 'rewards/judge_quality/mean': 0.4137499928474426, 'rewards/judge_quality/std': 0.06781013309955597, 'rewards/total_composite/mean': 0.6065137386322021, 'rewards/total_composite/std': 0.046656109392642975, 'reward': 0.6065137386322021, 'reward_std': 0.04665609821677208, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11938251554965973, 'sampling/sampling_logp_difference/max': 2.54801869392395, 'sampling/importance_sampling_ratio/min': 0.07823652774095535, 'sampling/importance_sampling_ratio/mean': 1.0205943584442139, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8639738485217094, 'clip_ratio/low_mean': 0.05129922367632389, 'clip_ratio/low_min': 0.05129922367632389, 'clip_ratio/high_mean': 0.06819419329985976, 'clip_ratio/high_max': 0.06819419329985976, 'clip_ratio/region_mean': 0.11949341697618365, 'reward_total_mean': 0.6065137386322021, 'reward_meter_mean': 0.9665820598602295, 'reward_meter_std': 0.06644105166196823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9778945446014404, 'reward_repeat_soft_std': 0.03151845559477806, 'reward_judge_quality_mean': 0.4137499928474426, 'reward_judge_quality_std': 0.06781013309955597, 'reward_total_composite_mean': 0.6065137386322021, 'reward_total_composite_std': 0.046656109392642975, 'epoch': 0.17}

 51%|█████     | 1687/3300 [4:09:06<3:11:24,  7.12s/it]INFO 04-13 11:30:04 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1688/3300 [4:09:14<3:18:59,  7.41s/it]2026-04-13 11:30:11,993 | INFO | train_grpo_train | metrics_logged mode=train step=1688

                                                       
{'loss': -0.0767, 'grad_norm': 3.8066842555999756, 'learning_rate': 4.887878787878788e-06, 'num_tokens': 3001705.0, 'completions/mean_length': 177.125, 'completions/min_length': 130.0, 'completions/max_length': 207.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 177.125, 'completions/min_terminated_length': 130.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.9835997223854065, 'rewards/meter/std': 0.01157785952091217, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6306673884391785, 'rewards/repeat_soft/std': 0.13541024923324585, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.5140981674194336, 'rewards/total_composite/std': 0.07761696726083755, 'reward': 0.5140981674194336, 'reward_std': 0.07761697471141815, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0640011578798294, 'sampling/sampling_logp_difference/max': 2.27779483795166, 'sampling/importance_sampling_ratio/min': 0.1025100126862526, 'sampling/importance_sampling_ratio/mean': 1.0001037120819092, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3475971445441246, 'clip_ratio/low_mean': 0.016462384024634957, 'clip_ratio/low_min': 0.016462384024634957, 'clip_ratio/high_mean': 0.040552848717197776, 'clip_ratio/high_max': 0.040552848717197776, 'clip_ratio/region_mean': 0.05701523274183273, 'reward_total_mean': 0.5140981674194336, 'reward_meter_mean': 0.9835997223854065, 'reward_meter_std': 0.01157785952091217, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6306673884391785, 'reward_repeat_soft_std': 0.13541024923324585, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.5140981674194336, 'reward_total_composite_std': 0.07761696726083755, 'epoch': 0.17}

 51%|█████     | 1688/3300 [4:09:14<3:18:59,  7.41s/it]INFO 04-13 11:30:12 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1689/3300 [4:09:20<3:12:13,  7.16s/it]2026-04-13 11:30:18,571 | INFO | train_grpo_train | metrics_logged mode=train step=1689

                                                       
{'loss': 0.0366, 'grad_norm': 10.181629180908203, 'learning_rate': 4.884848484848485e-06, 'num_tokens': 3003449.0, 'completions/mean_length': 59.0, 'completions/min_length': 53.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8670849204063416, 'rewards/meter/std': 0.30219969153404236, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9106879234313965, 'rewards/repeat_soft/std': 0.030124902725219727, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.5828097462654114, 'rewards/total_composite/std': 0.09014017134904861, 'reward': 0.5828097462654114, 'reward_std': 0.09014015644788742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11158782988786697, 'sampling/sampling_logp_difference/max': 1.7126024961471558, 'sampling/importance_sampling_ratio/min': 0.18039570748806, 'sampling/importance_sampling_ratio/mean': 1.0069172382354736, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5254568979144096, 'clip_ratio/low_mean': 0.01411290280520916, 'clip_ratio/low_min': 0.01411290280520916, 'clip_ratio/high_mean': 0.08197389869019389, 'clip_ratio/high_max': 0.08197389869019389, 'clip_ratio/region_mean': 0.09608680149540305, 'reward_total_mean': 0.5828097462654114, 'reward_meter_mean': 0.8670849204063416, 'reward_meter_std': 0.30219969153404236, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9106879234313965, 'reward_repeat_soft_std': 0.030124902725219727, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.5828097462654114, 'reward_total_composite_std': 0.09014017134904861, 'epoch': 0.17}

 51%|█████     | 1689/3300 [4:09:20<3:12:13,  7.16s/it]INFO 04-13 11:30:18 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1690/3300 [4:09:27<3:05:10,  6.90s/it]2026-04-13 11:30:24,873 | INFO | train_grpo_train | metrics_logged mode=train step=1690

                                                       
{'loss': 0.0718, 'grad_norm': 17.32733917236328, 'learning_rate': 4.881818181818182e-06, 'num_tokens': 3004870.0, 'completions/mean_length': 19.625, 'completions/min_length': 17.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.625, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9793528318405151, 'rewards/meter/std': 0.013812100514769554, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8885176777839661, 'rewards/repeat_soft/std': 0.06772414594888687, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.6445207595825195, 'rewards/total_composite/std': 0.09985730051994324, 'reward': 0.6445207595825195, 'reward_std': 0.09985730797052383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12900254130363464, 'sampling/sampling_logp_difference/max': 1.0521681308746338, 'sampling/importance_sampling_ratio/min': 0.34917983412742615, 'sampling/importance_sampling_ratio/mean': 1.0159752368927002, 'sampling/importance_sampling_ratio/max': 1.860042691230774, 'entropy': 0.7945980131626129, 'clip_ratio/low_mean': 0.12428737711161375, 'clip_ratio/low_min': 0.12428737711161375, 'clip_ratio/high_mean': 0.022058824077248573, 'clip_ratio/high_max': 0.022058824077248573, 'clip_ratio/region_mean': 0.14634620118886232, 'reward_total_mean': 0.6445207595825195, 'reward_meter_mean': 0.9793528318405151, 'reward_meter_std': 0.013812100514769554, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8885176777839661, 'reward_repeat_soft_std': 0.06772414594888687, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.6445207595825195, 'reward_total_composite_std': 0.09985730051994324, 'epoch': 0.17}

 51%|█████     | 1690/3300 [4:09:27<3:05:10,  6.90s/it]INFO 04-13 11:30:25 [block_pool.py:378] Successfully reset prefix cache

 51%|█████     | 1691/3300 [4:09:34<3:08:08,  7.02s/it]2026-04-13 11:30:32,155 | INFO | train_grpo_train | metrics_logged mode=train step=1691

                                                       
{'loss': 0.0555, 'grad_norm': 12.87840747833252, 'learning_rate': 4.878787878787879e-06, 'num_tokens': 3006625.0, 'completions/mean_length': 61.375, 'completions/min_length': 39.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.8104844093322754, 'rewards/meter/std': 0.19354458153247833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9004038572311401, 'rewards/repeat_soft/std': 0.0796278715133667, 'rewards/judge_quality/mean': 0.4987500011920929, 'rewards/judge_quality/std': 0.21357084810733795, 'rewards/total_composite/mean': 0.5965396165847778, 'rewards/total_composite/std': 0.11422868072986603, 'reward': 0.5965396165847778, 'reward_std': 0.11422866582870483, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13484059274196625, 'sampling/sampling_logp_difference/max': 2.514012098312378, 'sampling/importance_sampling_ratio/min': 0.08094283193349838, 'sampling/importance_sampling_ratio/mean': 1.0066410303115845, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6124727204442024, 'clip_ratio/low_mean': 0.09606262110173702, 'clip_ratio/low_min': 0.09606262110173702, 'clip_ratio/high_mean': 0.042786071076989174, 'clip_ratio/high_max': 0.042786071076989174, 'clip_ratio/region_mean': 0.1388486921787262, 'reward_total_mean': 0.5965396165847778, 'reward_meter_mean': 0.8104844093322754, 'reward_meter_std': 0.19354458153247833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9004038572311401, 'reward_repeat_soft_std': 0.0796278715133667, 'reward_judge_quality_mean': 0.4987500011920929, 'reward_judge_quality_std': 0.21357084810733795, 'reward_total_composite_mean': 0.5965396165847778, 'reward_total_composite_std': 0.11422868072986603, 'epoch': 0.17}

 51%|█████     | 1691/3300 [4:09:34<3:08:08,  7.02s/it]INFO 04-13 11:30:32 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1692/3300 [4:09:40<3:03:44,  6.86s/it]2026-04-13 11:30:38,648 | INFO | train_grpo_train | metrics_logged mode=train step=1692

                                                       
{'loss': 0.0305, 'grad_norm': 6.780965805053711, 'learning_rate': 4.875757575757576e-06, 'num_tokens': 3008567.0, 'completions/mean_length': 73.75, 'completions/min_length': 69.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 73.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9741885662078857, 'rewards/meter/std': 0.01744021102786064, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6959117650985718, 'rewards/repeat_soft/std': 0.1492699682712555, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.5219532251358032, 'rewards/total_composite/std': 0.08748581260442734, 'reward': 0.5219532251358032, 'reward_std': 0.08748582005500793, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06695970892906189, 'sampling/sampling_logp_difference/max': 1.565828800201416, 'sampling/importance_sampling_ratio/min': 0.20891478657722473, 'sampling/importance_sampling_ratio/mean': 0.9970290064811707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3333785180002451, 'clip_ratio/low_mean': 0.004982864367775619, 'clip_ratio/low_min': 0.004982864367775619, 'clip_ratio/high_mean': 0.03747419686987996, 'clip_ratio/high_max': 0.03747419686987996, 'clip_ratio/region_mean': 0.04245706123765558, 'reward_total_mean': 0.5219532251358032, 'reward_meter_mean': 0.9741885662078857, 'reward_meter_std': 0.01744021102786064, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6959117650985718, 'reward_repeat_soft_std': 0.1492699682712555, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.5219532251358032, 'reward_total_composite_std': 0.08748581260442734, 'epoch': 0.17}

 51%|█████▏    | 1692/3300 [4:09:40<3:03:44,  6.86s/it]INFO 04-13 11:30:38 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1693/3300 [4:09:52<3:40:21,  8.23s/it]2026-04-13 11:30:50,084 | INFO | train_grpo_train | metrics_logged mode=train step=1693

                                                       
{'loss': -0.1995, 'grad_norm': 2.3482632637023926, 'learning_rate': 4.872727272727273e-06, 'num_tokens': 3011066.0, 'completions/mean_length': 174.375, 'completions/min_length': 110.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 126.14286041259766, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.9050018787384033, 'rewards/meter/std': 0.20063656568527222, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7749311923980713, 'rewards/repeat_soft/std': 0.09003739804029465, 'rewards/judge_quality/mean': 0.3187499940395355, 'rewards/judge_quality/std': 0.14961259067058563, 'rewards/total_composite/mean': 0.43696027994155884, 'rewards/total_composite/std': 0.19049564003944397, 'reward': 0.43696027994155884, 'reward_std': 0.19049562513828278, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08963243663311005, 'sampling/sampling_logp_difference/max': 1.7019519805908203, 'sampling/importance_sampling_ratio/min': 0.1823272705078125, 'sampling/importance_sampling_ratio/mean': 1.0067193508148193, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.430629450827837, 'clip_ratio/low_mean': 0.010019906796514988, 'clip_ratio/low_min': 0.010019906796514988, 'clip_ratio/high_mean': 0.04426609817892313, 'clip_ratio/high_max': 0.04426609817892313, 'clip_ratio/region_mean': 0.05428600497543812, 'reward_total_mean': 0.43696027994155884, 'reward_meter_mean': 0.9050018787384033, 'reward_meter_std': 0.20063656568527222, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7749311923980713, 'reward_repeat_soft_std': 0.09003739804029465, 'reward_judge_quality_mean': 0.3187499940395355, 'reward_judge_quality_std': 0.14961259067058563, 'reward_total_composite_mean': 0.43696027994155884, 'reward_total_composite_std': 0.19049564003944397, 'epoch': 0.17}

 51%|█████▏    | 1693/3300 [4:09:52<3:40:21,  8.23s/it]INFO 04-13 11:30:50 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1694/3300 [4:09:58<3:23:01,  7.58s/it]2026-04-13 11:30:56,178 | INFO | train_grpo_train | metrics_logged mode=train step=1694

                                                       
{'loss': 0.0164, 'grad_norm': 9.841991424560547, 'learning_rate': 4.8696969696969705e-06, 'num_tokens': 3012833.0, 'completions/mean_length': 37.875, 'completions/min_length': 36.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.941387414932251, 'rewards/meter/std': 0.041819099336862564, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7620973587036133, 'rewards/repeat_soft/std': 0.06565750390291214, 'rewards/judge_quality/mean': 0.4024999737739563, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5602039098739624, 'rewards/total_composite/std': 0.040808357298374176, 'reward': 0.5602039098739624, 'reward_std': 0.040808361023664474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04996867850422859, 'sampling/sampling_logp_difference/max': 1.131651759147644, 'sampling/importance_sampling_ratio/min': 0.3225001394748688, 'sampling/importance_sampling_ratio/mean': 1.0246801376342773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.24913730844855309, 'clip_ratio/low_mean': 0.019243421033024788, 'clip_ratio/low_min': 0.019243421033024788, 'clip_ratio/high_mean': 0.026587442494928837, 'clip_ratio/high_max': 0.026587442494928837, 'clip_ratio/region_mean': 0.045830863527953625, 'reward_total_mean': 0.5602039098739624, 'reward_meter_mean': 0.941387414932251, 'reward_meter_std': 0.041819099336862564, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7620973587036133, 'reward_repeat_soft_std': 0.06565750390291214, 'reward_judge_quality_mean': 0.4024999737739563, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5602039098739624, 'reward_total_composite_std': 0.040808357298374176, 'epoch': 0.17}

 51%|█████▏    | 1694/3300 [4:09:58<3:23:01,  7.58s/it]INFO 04-13 11:30:56 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1695/3300 [4:10:05<3:18:18,  7.41s/it]2026-04-13 11:31:03,180 | INFO | train_grpo_train | metrics_logged mode=train step=1695

                                                       
{'loss': 0.0639, 'grad_norm': 13.700176239013672, 'learning_rate': 4.866666666666667e-06, 'num_tokens': 3014377.0, 'completions/mean_length': 34.0, 'completions/min_length': 32.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8843082785606384, 'rewards/meter/std': 0.2113535851240158, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7810590267181396, 'rewards/repeat_soft/std': 0.09941042214632034, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5585750341415405, 'rewards/total_composite/std': 0.04502662271261215, 'reward': 0.5585750341415405, 'reward_std': 0.04502662271261215, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08787927776575089, 'sampling/sampling_logp_difference/max': 1.5108418464660645, 'sampling/importance_sampling_ratio/min': 0.27825117111206055, 'sampling/importance_sampling_ratio/mean': 1.0241531133651733, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35129158198833466, 'clip_ratio/low_mean': 0.02302631549537182, 'clip_ratio/low_min': 0.02302631549537182, 'clip_ratio/high_mean': 0.048711445881053805, 'clip_ratio/high_max': 0.048711445881053805, 'clip_ratio/region_mean': 0.07173776137642562, 'reward_total_mean': 0.5585750341415405, 'reward_meter_mean': 0.8843082785606384, 'reward_meter_std': 0.2113535851240158, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7810590267181396, 'reward_repeat_soft_std': 0.09941042214632034, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5585750341415405, 'reward_total_composite_std': 0.04502662271261215, 'epoch': 0.17}

 51%|█████▏    | 1695/3300 [4:10:05<3:18:18,  7.41s/it]INFO 04-13 11:31:03 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1696/3300 [4:10:12<3:18:18,  7.42s/it]2026-04-13 11:31:10,598 | INFO | train_grpo_train | metrics_logged mode=train step=1696

                                                       
{'loss': 0.1091, 'grad_norm': 7.173166275024414, 'learning_rate': 4.863636363636364e-06, 'num_tokens': 3016793.0, 'completions/mean_length': 119.0, 'completions/min_length': 98.0, 'completions/max_length': 152.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.0, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 152.0, 'rewards/meter/mean': 0.9621785283088684, 'rewards/meter/std': 0.07766371965408325, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8270488977432251, 'rewards/repeat_soft/std': 0.09540286660194397, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.586732029914856, 'rewards/total_composite/std': 0.018162593245506287, 'reward': 0.586732029914856, 'reward_std': 0.018162589520215988, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09017331898212433, 'sampling/sampling_logp_difference/max': 2.4064292907714844, 'sampling/importance_sampling_ratio/min': 0.13300110399723053, 'sampling/importance_sampling_ratio/mean': 1.0062044858932495, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44541386142373085, 'clip_ratio/low_mean': 0.015144678764045238, 'clip_ratio/low_min': 0.015144678764045238, 'clip_ratio/high_mean': 0.06057947035878897, 'clip_ratio/high_max': 0.06057947035878897, 'clip_ratio/region_mean': 0.0757241491228342, 'reward_total_mean': 0.586732029914856, 'reward_meter_mean': 0.9621785283088684, 'reward_meter_std': 0.07766371965408325, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8270488977432251, 'reward_repeat_soft_std': 0.09540286660194397, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.586732029914856, 'reward_total_composite_std': 0.018162593245506287, 'epoch': 0.17}

 51%|█████▏    | 1696/3300 [4:10:12<3:18:18,  7.42s/it]INFO 04-13 11:31:10 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1697/3300 [4:10:19<3:14:20,  7.27s/it]2026-04-13 11:31:17,548 | INFO | train_grpo_train | metrics_logged mode=train step=1697

                                                       
{'loss': 0.002, 'grad_norm': 3.858222246170044, 'learning_rate': 4.8606060606060615e-06, 'num_tokens': 3018934.0, 'completions/mean_length': 104.625, 'completions/min_length': 94.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.625, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.9845653772354126, 'rewards/meter/std': 0.011520404368638992, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.66322922706604, 'rewards/repeat_soft/std': 0.10307466238737106, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6074063777923584, 'rewards/total_composite/std': 0.09574397653341293, 'reward': 0.6074063777923584, 'reward_std': 0.09574397653341293, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.059490155428647995, 'sampling/sampling_logp_difference/max': 3.0207901000976562, 'sampling/importance_sampling_ratio/min': 0.04876267537474632, 'sampling/importance_sampling_ratio/mean': 1.0033087730407715, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29337758757174015, 'clip_ratio/low_mean': 0.04706494836136699, 'clip_ratio/low_min': 0.04706494836136699, 'clip_ratio/high_mean': 0.004807692486792803, 'clip_ratio/high_max': 0.004807692486792803, 'clip_ratio/region_mean': 0.05187264084815979, 'reward_total_mean': 0.6074063777923584, 'reward_meter_mean': 0.9845653772354126, 'reward_meter_std': 0.011520404368638992, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.66322922706604, 'reward_repeat_soft_std': 0.10307466238737106, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6074063777923584, 'reward_total_composite_std': 0.09574397653341293, 'epoch': 0.17}

 51%|█████▏    | 1697/3300 [4:10:19<3:14:20,  7.27s/it]INFO 04-13 11:31:17 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1698/3300 [4:10:25<3:04:53,  6.92s/it]2026-04-13 11:31:23,651 | INFO | train_grpo_train | metrics_logged mode=train step=1698

                                                       
{'loss': 0.0165, 'grad_norm': 10.221336364746094, 'learning_rate': 4.857575757575758e-06, 'num_tokens': 3020568.0, 'completions/mean_length': 35.25, 'completions/min_length': 32.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9896067976951599, 'rewards/meter/std': 0.0032773904968053102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9455275535583496, 'rewards/repeat_soft/std': 0.021588625386357307, 'rewards/judge_quality/mean': 0.7112500071525574, 'rewards/judge_quality/std': 0.24793073534965515, 'rewards/total_composite/mean': 0.7994704246520996, 'rewards/total_composite/std': 0.1603584885597229, 'reward': 0.7994704246520996, 'reward_std': 0.1603584885597229, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08277014642953873, 'sampling/sampling_logp_difference/max': 1.3099803924560547, 'sampling/importance_sampling_ratio/min': 0.26982536911964417, 'sampling/importance_sampling_ratio/mean': 1.0136631727218628, 'sampling/importance_sampling_ratio/max': 1.829031229019165, 'entropy': 0.36017119884490967, 'clip_ratio/low_mean': 0.03167505795136094, 'clip_ratio/low_min': 0.03167505795136094, 'clip_ratio/high_mean': 0.04956510406918824, 'clip_ratio/high_max': 0.04956510406918824, 'clip_ratio/region_mean': 0.08124016202054918, 'reward_total_mean': 0.7994704246520996, 'reward_meter_mean': 0.9896067976951599, 'reward_meter_std': 0.0032773904968053102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9455275535583496, 'reward_repeat_soft_std': 0.021588625386357307, 'reward_judge_quality_mean': 0.7112500071525574, 'reward_judge_quality_std': 0.24793073534965515, 'reward_total_composite_mean': 0.7994704246520996, 'reward_total_composite_std': 0.1603584885597229, 'epoch': 0.17}

 51%|█████▏    | 1698/3300 [4:10:25<3:04:53,  6.92s/it]INFO 04-13 11:31:23 [block_pool.py:378] Successfully reset prefix cache

 51%|█████▏    | 1699/3300 [4:10:37<3:39:42,  8.23s/it]2026-04-13 11:31:34,930 | INFO | train_grpo_train | metrics_logged mode=train step=1699

                                                       
{'loss': -0.1534, 'grad_norm': 1.69248366355896, 'learning_rate': 4.854545454545455e-06, 'num_tokens': 3022221.0, 'completions/mean_length': 114.625, 'completions/min_length': 42.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.857147216796875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.8519529104232788, 'rewards/meter/std': 0.3445809781551361, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9030239582061768, 'rewards/repeat_soft/std': 0.06341450661420822, 'rewards/judge_quality/mean': 0.3812500238418579, 'rewards/judge_quality/std': 0.13452960550785065, 'rewards/total_composite/mean': 0.5289825201034546, 'rewards/total_composite/std': 0.21403877437114716, 'reward': 0.5289825201034546, 'reward_std': 0.21403877437114716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11025475710630417, 'sampling/sampling_logp_difference/max': 1.3818187713623047, 'sampling/importance_sampling_ratio/min': 0.2511214017868042, 'sampling/importance_sampling_ratio/mean': 1.0042784214019775, 'sampling/importance_sampling_ratio/max': 1.9752295017242432, 'entropy': 0.5506503693759441, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1055633183568716, 'clip_ratio/high_max': 0.1055633183568716, 'clip_ratio/region_mean': 0.1055633183568716, 'reward_total_mean': 0.5289825201034546, 'reward_meter_mean': 0.8519529104232788, 'reward_meter_std': 0.3445809781551361, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9030239582061768, 'reward_repeat_soft_std': 0.06341450661420822, 'reward_judge_quality_mean': 0.3812500238418579, 'reward_judge_quality_std': 0.13452960550785065, 'reward_total_composite_mean': 0.5289825201034546, 'reward_total_composite_std': 0.21403877437114716, 'epoch': 0.17}

 51%|█████▏    | 1699/3300 [4:10:37<3:39:42,  8.23s/it]INFO 04-13 11:31:35 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1700/3300 [4:10:44<3:32:13,  7.96s/it]2026-04-13 11:31:42,249 | INFO | train_grpo_train | metrics_logged mode=train step=1700

                                                       
{'loss': 0.0074, 'grad_norm': 5.452632904052734, 'learning_rate': 4.851515151515152e-06, 'num_tokens': 3024522.0, 'completions/mean_length': 120.625, 'completions/min_length': 106.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.625, 'completions/min_terminated_length': 106.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.830030620098114, 'rewards/meter/std': 0.2582976520061493, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7376050353050232, 'rewards/repeat_soft/std': 0.09343899041414261, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.537239134311676, 'rewards/total_composite/std': 0.06643049418926239, 'reward': 0.537239134311676, 'reward_std': 0.06643049418926239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08235330134630203, 'sampling/sampling_logp_difference/max': 1.233177661895752, 'sampling/importance_sampling_ratio/min': 0.29136523604393005, 'sampling/importance_sampling_ratio/mean': 1.004504919052124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43990712240338326, 'clip_ratio/low_mean': 0.01977576781064272, 'clip_ratio/low_min': 0.01977576781064272, 'clip_ratio/high_mean': 0.06406635232269764, 'clip_ratio/high_max': 0.06406635232269764, 'clip_ratio/region_mean': 0.08384212013334036, 'reward_total_mean': 0.537239134311676, 'reward_meter_mean': 0.830030620098114, 'reward_meter_std': 0.2582976520061493, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7376050353050232, 'reward_repeat_soft_std': 0.09343899041414261, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.537239134311676, 'reward_total_composite_std': 0.06643049418926239, 'epoch': 0.17}

 52%|█████▏    | 1700/3300 [4:10:44<3:32:13,  7.96s/it]INFO 04-13 11:31:42 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.87s/it]

 30%|███       | 3/10 [00:09<00:23,  3.33s/it]

 40%|████      | 4/10 [00:13<00:21,  3.56s/it]

 50%|█████     | 5/10 [00:17<00:19,  3.91s/it]

 60%|██████    | 6/10 [00:26<00:22,  5.54s/it]

 70%|███████   | 7/10 [00:30<00:15,  5.20s/it]

 80%|████████  | 8/10 [00:34<00:09,  4.80s/it]

 90%|█████████ | 9/10 [00:43<00:06,  6.06s/it]

100%|██████████| 10/10 [00:47<00:00,  5.22s/it]2026-04-13 11:32:34,116 | INFO | train_grpo_train | metrics_logged mode=eval step=1700

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 51.766, 'eval_samples_per_second': 1.545, 'eval_steps_per_second': 0.193, 'eval_num_tokens': 3024522.0, 'eval_completions/mean_length': 96.95, 'eval_completions/min_length': 34.7, 'eval_completions/max_length': 210.6, 'eval_completions/clipped_ratio': 0.0375, 'eval_completions/mean_terminated_length': 81.07321472167969, 'eval_completions/min_terminated_length': 34.7, 'eval_completions/max_terminated_length': 145.1, 'eval_rewards/meter/mean': 0.7788607656955719, 'eval_rewards/meter/std': 0.2661908954381943, 'eval_rewards/count_adherence/mean': 0.951874989271164, 'eval_rewards/count_adherence/std': 0.08645724020898342, 'eval_rewards/hard_gate/mean': 0.9375, 'eval_rewards/hard_gate/std': 0.09804592728614807, 'eval_rewards/repeat_soft/mean': 0.8310980260372162, 'eval_rewards/repeat_soft/std': 0.11172637566924096, 'eval_rewards/judge_quality/mean': 0.42000000476837157, 'eval_rewards/judge_quality/std': 0.1566985785961151, 'eval_rewards/total_composite/mean': 0.5148498684167861, 'eval_rewards/total_composite/std': 0.1532291118055582, 'eval_reward': 0.5148498684167861, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.04819921813905239, 'eval_sampling/sampling_logp_difference/max': 1.0375221252441407, 'eval_sampling/importance_sampling_ratio/min': 0.36269562840461733, 'eval_sampling/importance_sampling_ratio/mean': 1.0100383281707763, 'eval_sampling/importance_sampling_ratio/max': 1.4731812477111816, 'eval_entropy': 0.49511538743972777, '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': 0.5148498684167861, 'eval_reward_meter_mean': 0.7788607656955719, 'eval_reward_meter_std': 0.2661908954381943, 'eval_reward_count_adherence_mean': 0.951874989271164, 'eval_reward_count_adherence_std': 0.08645724020898342, 'eval_reward_hard_gate_mean': 0.9375, 'eval_reward_hard_gate_std': 0.09804592728614807, 'eval_reward_repeat_soft_mean': 0.8310980260372162, 'eval_reward_repeat_soft_std': 0.11172637566924096, 'eval_reward_judge_quality_mean': 0.42000000476837157, 'eval_reward_judge_quality_std': 0.1566985785961151, 'eval_reward_total_composite_mean': 0.5148498684167861, 'eval_reward_total_composite_std': 0.1532291118055582, 'epoch': 0.17}

 52%|█████▏    | 1700/3300 [4:11:36<3:32:13,  7.96s/it]

100%|██████████| 10/10 [00:47<00:00,  5.22s/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(

 52%|█████▏    | 1701/3300 [4:11:46<10:43:17, 24.14s/it]2026-04-13 11:32:44,133 | INFO | train_grpo_train | metrics_logged mode=train step=1701

                                                        
{'loss': -0.0625, 'grad_norm': 6.982199668884277, 'learning_rate': 4.848484848484849e-06, 'num_tokens': 3026667.0, 'completions/mean_length': 100.125, 'completions/min_length': 73.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.125, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.9745658040046692, 'rewards/meter/std': 0.030551763251423836, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7237980365753174, 'rewards/repeat_soft/std': 0.06728465110063553, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.56345534324646, 'rewards/total_composite/std': 0.04501413553953171, 'reward': 0.56345534324646, 'reward_std': 0.04501413553953171, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07348770648241043, 'sampling/sampling_logp_difference/max': 1.2052745819091797, 'sampling/importance_sampling_ratio/min': 0.2996097207069397, 'sampling/importance_sampling_ratio/mean': 1.001634955406189, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.450840562582016, 'clip_ratio/low_mean': 0.012467292603105307, 'clip_ratio/low_min': 0.012467292603105307, 'clip_ratio/high_mean': 0.05064171692356467, 'clip_ratio/high_max': 0.05064171692356467, 'clip_ratio/region_mean': 0.06310900952666998, 'reward_total_mean': 0.56345534324646, 'reward_meter_mean': 0.9745658040046692, 'reward_meter_std': 0.030551763251423836, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7237980365753174, 'reward_repeat_soft_std': 0.06728465110063553, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.56345534324646, 'reward_total_composite_std': 0.04501413553953171, 'epoch': 0.17}

 52%|█████▏    | 1701/3300 [4:11:46<10:43:17, 24.14s/it]INFO 04-13 11:32:44 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1702/3300 [4:11:53<8:24:45, 18.95s/it] 2026-04-13 11:32:51,002 | INFO | train_grpo_train | metrics_logged mode=train step=1702

                                                       
{'loss': 0.0219, 'grad_norm': 5.328232288360596, 'learning_rate': 4.845454545454546e-06, 'num_tokens': 3028905.0, 'completions/mean_length': 110.75, 'completions/min_length': 93.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.75, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9815160632133484, 'rewards/meter/std': 0.010495602153241634, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7253273725509644, 'rewards/repeat_soft/std': 0.08580384403467178, 'rewards/judge_quality/mean': 0.36500000953674316, 'rewards/judge_quality/std': 0.10528871417045593, 'rewards/total_composite/mean': 0.5415375232696533, 'rewards/total_composite/std': 0.07670989632606506, 'reward': 0.5415375232696533, 'reward_std': 0.07670990377664566, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07422178238630295, 'sampling/sampling_logp_difference/max': 1.5198631286621094, 'sampling/importance_sampling_ratio/min': 0.28725555539131165, 'sampling/importance_sampling_ratio/mean': 1.0084192752838135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4771435223519802, 'clip_ratio/low_mean': 0.009179437882266939, 'clip_ratio/low_min': 0.009179437882266939, 'clip_ratio/high_mean': 0.06365141971036792, 'clip_ratio/high_max': 0.06365141971036792, 'clip_ratio/region_mean': 0.07283085759263486, 'reward_total_mean': 0.5415375232696533, 'reward_meter_mean': 0.9815160632133484, 'reward_meter_std': 0.010495602153241634, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7253273725509644, 'reward_repeat_soft_std': 0.08580384403467178, 'reward_judge_quality_mean': 0.36500000953674316, 'reward_judge_quality_std': 0.10528871417045593, 'reward_total_composite_mean': 0.5415375232696533, 'reward_total_composite_std': 0.07670989632606506, 'epoch': 0.17}

 52%|█████▏    | 1702/3300 [4:11:53<8:24:45, 18.95s/it]INFO 04-13 11:32:51 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1703/3300 [4:12:00<6:54:46, 15.58s/it]2026-04-13 11:32:58,716 | INFO | train_grpo_train | metrics_logged mode=train step=1703

                                                       
{'loss': 0.0537, 'grad_norm': 7.986655235290527, 'learning_rate': 4.842424242424243e-06, 'num_tokens': 3031039.0, 'completions/mean_length': 91.75, 'completions/min_length': 77.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.8958242535591125, 'rewards/meter/std': 0.2371399998664856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8050409555435181, 'rewards/repeat_soft/std': 0.049173615872859955, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.517031192779541, 'rewards/total_composite/std': 0.0822354108095169, 'reward': 0.517031192779541, 'reward_std': 0.0822354108095169, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09671947360038757, 'sampling/sampling_logp_difference/max': 2.264767646789551, 'sampling/importance_sampling_ratio/min': 0.10385415703058243, 'sampling/importance_sampling_ratio/mean': 1.0112695693969727, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5040853545069695, 'clip_ratio/low_mean': 0.043156355153769255, 'clip_ratio/low_min': 0.043156355153769255, 'clip_ratio/high_mean': 0.044979365542531013, 'clip_ratio/high_max': 0.044979365542531013, 'clip_ratio/region_mean': 0.08813572069630027, 'reward_total_mean': 0.517031192779541, 'reward_meter_mean': 0.8958242535591125, 'reward_meter_std': 0.2371399998664856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8050409555435181, 'reward_repeat_soft_std': 0.049173615872859955, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.517031192779541, 'reward_total_composite_std': 0.0822354108095169, 'epoch': 0.17}

 52%|█████▏    | 1703/3300 [4:12:00<6:54:46, 15.58s/it]INFO 04-13 11:32:58 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1704/3300 [4:12:07<5:40:11, 12.79s/it]2026-04-13 11:33:04,982 | INFO | train_grpo_train | metrics_logged mode=train step=1704

                                                       
{'loss': 0.0574, 'grad_norm': 12.21061897277832, 'learning_rate': 4.83939393939394e-06, 'num_tokens': 3032720.0, 'completions/mean_length': 41.125, 'completions/min_length': 31.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7094885110855103, 'rewards/meter/std': 0.27525606751441956, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9190400242805481, 'rewards/repeat_soft/std': 0.03655832260847092, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.1888640969991684, 'rewards/total_composite/mean': 0.5927785634994507, 'rewards/total_composite/std': 0.13393543660640717, 'reward': 0.5927785634994507, 'reward_std': 0.13393543660640717, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11570199579000473, 'sampling/sampling_logp_difference/max': 2.3757834434509277, 'sampling/importance_sampling_ratio/min': 0.09294164925813675, 'sampling/importance_sampling_ratio/mean': 1.0100421905517578, 'sampling/importance_sampling_ratio/max': 1.8284733295440674, 'entropy': 0.6329236328601837, 'clip_ratio/low_mean': 0.04719053162261844, 'clip_ratio/low_min': 0.04719053162261844, 'clip_ratio/high_mean': 0.04858295898884535, 'clip_ratio/high_max': 0.04858295898884535, 'clip_ratio/region_mean': 0.09577349061146379, 'reward_total_mean': 0.5927785634994507, 'reward_meter_mean': 0.7094885110855103, 'reward_meter_std': 0.27525606751441956, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9190400242805481, 'reward_repeat_soft_std': 0.03655832260847092, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.1888640969991684, 'reward_total_composite_mean': 0.5927785634994507, 'reward_total_composite_std': 0.13393543660640717, 'epoch': 0.17}

 52%|█████▏    | 1704/3300 [4:12:07<5:40:11, 12.79s/it]INFO 04-13 11:33:05 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1705/3300 [4:12:12<4:44:48, 10.71s/it]2026-04-13 11:33:10,857 | INFO | train_grpo_train | metrics_logged mode=train step=1705

                                                       
{'loss': -0.0042, 'grad_norm': 19.694028854370117, 'learning_rate': 4.836363636363637e-06, 'num_tokens': 3034208.0, 'completions/mean_length': 24.0, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.0, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.8004680871963501, 'rewards/meter/std': 0.2228754162788391, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9544928669929504, 'rewards/repeat_soft/std': 0.015905901789665222, 'rewards/judge_quality/mean': 0.3387500047683716, 'rewards/judge_quality/std': 0.13292719423770905, 'rewards/total_composite/mean': 0.5244325995445251, 'rewards/total_composite/std': 0.09833145141601562, 'reward': 0.5244325995445251, 'reward_std': 0.09833145141601562, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11064853519201279, 'sampling/sampling_logp_difference/max': 1.2437875270843506, 'sampling/importance_sampling_ratio/min': 0.28829023241996765, 'sampling/importance_sampling_ratio/mean': 0.9975305795669556, 'sampling/importance_sampling_ratio/max': 1.6610095500946045, 'entropy': 0.86296296864748, 'clip_ratio/low_mean': 0.041220239363610744, 'clip_ratio/low_min': 0.041220239363610744, 'clip_ratio/high_mean': 0.04556785896420479, 'clip_ratio/high_max': 0.04556785896420479, 'clip_ratio/region_mean': 0.08678809832781553, 'reward_total_mean': 0.5244325995445251, 'reward_meter_mean': 0.8004680871963501, 'reward_meter_std': 0.2228754162788391, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9544928669929504, 'reward_repeat_soft_std': 0.015905901789665222, 'reward_judge_quality_mean': 0.3387500047683716, 'reward_judge_quality_std': 0.13292719423770905, 'reward_total_composite_mean': 0.5244325995445251, 'reward_total_composite_std': 0.09833145141601562, 'epoch': 0.17}

 52%|█████▏    | 1705/3300 [4:12:13<4:44:48, 10.71s/it]INFO 04-13 11:33:11 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1706/3300 [4:12:19<4:13:35,  9.55s/it]2026-04-13 11:33:17,683 | INFO | train_grpo_train | metrics_logged mode=train step=1706

                                                       
{'loss': 0.0292, 'grad_norm': 12.84737777709961, 'learning_rate': 4.833333333333333e-06, 'num_tokens': 3035743.0, 'completions/mean_length': 43.875, 'completions/min_length': 38.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.7723362445831299, 'rewards/meter/std': 0.2725357413291931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9816555976867676, 'rewards/repeat_soft/std': 0.04615623131394386, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955310225486755, 'rewards/total_composite/mean': 0.6426393985748291, 'rewards/total_composite/std': 0.1737646460533142, 'reward': 0.6426393985748291, 'reward_std': 0.1737646460533142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1071280688047409, 'sampling/sampling_logp_difference/max': 1.0546283721923828, 'sampling/importance_sampling_ratio/min': 0.3483218550682068, 'sampling/importance_sampling_ratio/mean': 1.0121512413024902, 'sampling/importance_sampling_ratio/max': 1.9893280267715454, 'entropy': 0.6153622791171074, 'clip_ratio/low_mean': 0.100666300393641, 'clip_ratio/low_min': 0.100666300393641, 'clip_ratio/high_mean': 0.03223467618227005, 'clip_ratio/high_max': 0.03223467618227005, 'clip_ratio/region_mean': 0.13290097657591105, 'reward_total_mean': 0.6426393985748291, 'reward_meter_mean': 0.7723362445831299, 'reward_meter_std': 0.2725357413291931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9816555976867676, 'reward_repeat_soft_std': 0.04615623131394386, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955310225486755, 'reward_total_composite_mean': 0.6426393985748291, 'reward_total_composite_std': 0.1737646460533142, 'epoch': 0.17}

 52%|█████▏    | 1706/3300 [4:12:19<4:13:35,  9.55s/it]INFO 04-13 11:33:17 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1707/3300 [4:12:26<3:49:12,  8.63s/it]2026-04-13 11:33:24,186 | INFO | train_grpo_train | metrics_logged mode=train step=1707

                                                       
{'loss': -0.0433, 'grad_norm': 12.744470596313477, 'learning_rate': 4.830303030303031e-06, 'num_tokens': 3037470.0, 'completions/mean_length': 53.875, 'completions/min_length': 45.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9339768886566162, 'rewards/meter/std': 0.12653842568397522, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9758355021476746, 'rewards/repeat_soft/std': 0.037519022822380066, 'rewards/judge_quality/mean': 0.5274999737739563, 'rewards/judge_quality/std': 0.18873640894889832, 'rewards/total_composite/mean': 0.6698579788208008, 'rewards/total_composite/std': 0.13340696692466736, 'reward': 0.6698579788208008, 'reward_std': 0.13340695202350616, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12084158509969711, 'sampling/sampling_logp_difference/max': 1.859917163848877, 'sampling/importance_sampling_ratio/min': 0.15568551421165466, 'sampling/importance_sampling_ratio/mean': 1.0078500509262085, 'sampling/importance_sampling_ratio/max': 1.920824646949768, 'entropy': 0.8653598427772522, 'clip_ratio/low_mean': 0.07423791708424687, 'clip_ratio/low_min': 0.07423791708424687, 'clip_ratio/high_mean': 0.031413814052939415, 'clip_ratio/high_max': 0.031413814052939415, 'clip_ratio/region_mean': 0.10565173113718629, 'reward_total_mean': 0.6698579788208008, 'reward_meter_mean': 0.9339768886566162, 'reward_meter_std': 0.12653842568397522, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9758355021476746, 'reward_repeat_soft_std': 0.037519022822380066, 'reward_judge_quality_mean': 0.5274999737739563, 'reward_judge_quality_std': 0.18873640894889832, 'reward_total_composite_mean': 0.6698579788208008, 'reward_total_composite_std': 0.13340696692466736, 'epoch': 0.17}

 52%|█████▏    | 1707/3300 [4:12:26<3:49:12,  8.63s/it]INFO 04-13 11:33:24 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1708/3300 [4:12:33<3:33:42,  8.05s/it]2026-04-13 11:33:30,881 | INFO | train_grpo_train | metrics_logged mode=train step=1708

                                                       
{'loss': -0.0337, 'grad_norm': 8.233144760131836, 'learning_rate': 4.827272727272728e-06, 'num_tokens': 3039121.0, 'completions/mean_length': 37.375, 'completions/min_length': 31.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.375, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.45985865592956543, 'rewards/meter/std': 0.42175498604774475, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8944973945617676, 'rewards/repeat_soft/std': 0.04428502172231674, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.4863719940185547, 'rewards/total_composite/std': 0.15772169828414917, 'reward': 0.4863719940185547, 'reward_std': 0.15772169828414917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09548614919185638, 'sampling/sampling_logp_difference/max': 1.1301515102386475, 'sampling/importance_sampling_ratio/min': 0.3229843080043793, 'sampling/importance_sampling_ratio/mean': 1.0027332305908203, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.586298007518053, 'clip_ratio/low_mean': 0.05920380586758256, 'clip_ratio/low_min': 0.05920380586758256, 'clip_ratio/high_mean': 0.037892288994044065, 'clip_ratio/high_max': 0.037892288994044065, 'clip_ratio/region_mean': 0.09709609486162663, 'reward_total_mean': 0.4863719940185547, 'reward_meter_mean': 0.45985865592956543, 'reward_meter_std': 0.42175498604774475, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8944973945617676, 'reward_repeat_soft_std': 0.04428502172231674, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.4863719940185547, 'reward_total_composite_std': 0.15772169828414917, 'epoch': 0.17}

 52%|█████▏    | 1708/3300 [4:12:33<3:33:42,  8.05s/it]INFO 04-13 11:33:31 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1709/3300 [4:12:40<3:26:29,  7.79s/it]2026-04-13 11:33:38,045 | INFO | train_grpo_train | metrics_logged mode=train step=1709

                                                       
{'loss': 0.0363, 'grad_norm': 9.889695167541504, 'learning_rate': 4.824242424242424e-06, 'num_tokens': 3040825.0, 'completions/mean_length': 47.0, 'completions/min_length': 40.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9627807140350342, 'rewards/meter/std': 0.07612529397010803, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9161338806152344, 'rewards/repeat_soft/std': 0.044543150812387466, 'rewards/judge_quality/mean': 0.47999998927116394, 'rewards/judge_quality/std': 0.10993505269289017, 'rewards/total_composite/mean': 0.6330803632736206, 'rewards/total_composite/std': 0.033940430730581284, 'reward': 0.6330803632736206, 'reward_std': 0.03394043818116188, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10801807045936584, 'sampling/sampling_logp_difference/max': 1.71560537815094, 'sampling/importance_sampling_ratio/min': 0.17985482513904572, 'sampling/importance_sampling_ratio/mean': 0.9994314908981323, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5231927931308746, 'clip_ratio/low_mean': 0.05567022366449237, 'clip_ratio/low_min': 0.05567022366449237, 'clip_ratio/high_mean': 0.0384084889665246, 'clip_ratio/high_max': 0.0384084889665246, 'clip_ratio/region_mean': 0.09407871263101697, 'reward_total_mean': 0.6330803632736206, 'reward_meter_mean': 0.9627807140350342, 'reward_meter_std': 0.07612529397010803, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9161338806152344, 'reward_repeat_soft_std': 0.044543150812387466, 'reward_judge_quality_mean': 0.47999998927116394, 'reward_judge_quality_std': 0.10993505269289017, 'reward_total_composite_mean': 0.6330803632736206, 'reward_total_composite_std': 0.033940430730581284, 'epoch': 0.17}

 52%|█████▏    | 1709/3300 [4:12:40<3:26:29,  7.79s/it]INFO 04-13 11:33:38 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1710/3300 [4:12:47<3:22:20,  7.64s/it]2026-04-13 11:33:45,324 | INFO | train_grpo_train | metrics_logged mode=train step=1710

                                                       
{'loss': 0.1373, 'grad_norm': 7.542257785797119, 'learning_rate': 4.8212121212121215e-06, 'num_tokens': 3042836.0, 'completions/mean_length': 83.375, 'completions/min_length': 66.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.375, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.3569932281970978, 'rewards/meter/std': 0.2999321222305298, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8213039636611938, 'rewards/repeat_soft/std': 0.10220811516046524, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.41422730684280396, 'rewards/total_composite/std': 0.09932152181863785, 'reward': 0.41422730684280396, 'reward_std': 0.09932152181863785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09379032999277115, 'sampling/sampling_logp_difference/max': 1.533194899559021, 'sampling/importance_sampling_ratio/min': 0.21584497392177582, 'sampling/importance_sampling_ratio/mean': 1.0104491710662842, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5609025619924068, 'clip_ratio/low_mean': 0.04872157471254468, 'clip_ratio/low_min': 0.04872157471254468, 'clip_ratio/high_mean': 0.04706374555826187, 'clip_ratio/high_max': 0.04706374555826187, 'clip_ratio/region_mean': 0.09578532027080655, 'reward_total_mean': 0.41422730684280396, 'reward_meter_mean': 0.3569932281970978, 'reward_meter_std': 0.2999321222305298, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8213039636611938, 'reward_repeat_soft_std': 0.10220811516046524, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.41422730684280396, 'reward_total_composite_std': 0.09932152181863785, 'epoch': 0.17}

 52%|█████▏    | 1710/3300 [4:12:47<3:22:20,  7.64s/it]INFO 04-13 11:33:45 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1711/3300 [4:12:58<3:53:06,  8.80s/it]2026-04-13 11:33:56,855 | INFO | train_grpo_train | metrics_logged mode=train step=1711

                                                       
{'loss': -0.2413, 'grad_norm': 2.1527371406555176, 'learning_rate': 4.818181818181819e-06, 'num_tokens': 3045448.0, 'completions/mean_length': 205.5, 'completions/min_length': 128.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 161.71429443359375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 181.0, 'rewards/meter/mean': 0.8527488112449646, 'rewards/meter/std': 0.23914027214050293, 'rewards/count_adherence/mean': 0.8958333730697632, 'rewards/count_adherence/std': 0.23464766144752502, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.853449285030365, 'rewards/repeat_soft/std': 0.08920948952436447, 'rewards/judge_quality/mean': 0.24250000715255737, 'rewards/judge_quality/std': 0.12947696447372437, 'rewards/total_composite/mean': 0.4225049614906311, 'rewards/total_composite/std': 0.18557484447956085, 'reward': 0.4225049614906311, 'reward_std': 0.18557484447956085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08878102898597717, 'sampling/sampling_logp_difference/max': 3.4642715454101562, 'sampling/importance_sampling_ratio/min': 0.03129579499363899, 'sampling/importance_sampling_ratio/mean': 1.012242317199707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5966197475790977, 'clip_ratio/low_mean': 0.007867133244872093, 'clip_ratio/low_min': 0.007867133244872093, 'clip_ratio/high_mean': 0.07043249066919088, 'clip_ratio/high_max': 0.07043249066919088, 'clip_ratio/region_mean': 0.07829962391406298, 'reward_total_mean': 0.4225049614906311, 'reward_meter_mean': 0.8527488112449646, 'reward_meter_std': 0.23914027214050293, 'reward_count_adherence_mean': 0.8958333730697632, 'reward_count_adherence_std': 0.23464766144752502, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.853449285030365, 'reward_repeat_soft_std': 0.08920948952436447, 'reward_judge_quality_mean': 0.24250000715255737, 'reward_judge_quality_std': 0.12947696447372437, 'reward_total_composite_mean': 0.4225049614906311, 'reward_total_composite_std': 0.18557484447956085, 'epoch': 0.17}

 52%|█████▏    | 1711/3300 [4:12:59<3:53:06,  8.80s/it]INFO 04-13 11:33:57 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1712/3300 [4:13:10<4:15:09,  9.64s/it]2026-04-13 11:34:08,453 | INFO | train_grpo_train | metrics_logged mode=train step=1712

                                                       
{'loss': -0.1346, 'grad_norm': 3.0581750869750977, 'learning_rate': 4.815151515151515e-06, 'num_tokens': 3047159.0, 'completions/mean_length': 113.875, 'completions/min_length': 53.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 57.000003814697266, 'completions/min_terminated_length': 53.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.7386385202407837, 'rewards/meter/std': 0.4336848556995392, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.983648419380188, 'rewards/repeat_soft/std': 0.016031138598918915, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.5108710527420044, 'rewards/total_composite/std': 0.22471150755882263, 'reward': 0.5108710527420044, 'reward_std': 0.22471149265766144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11820080131292343, 'sampling/sampling_logp_difference/max': 2.4018521308898926, 'sampling/importance_sampling_ratio/min': 0.09055008739233017, 'sampling/importance_sampling_ratio/mean': 1.0040462017059326, 'sampling/importance_sampling_ratio/max': 1.9066431522369385, 'entropy': 0.8000173345208168, 'clip_ratio/low_mean': 0.004545454401522875, 'clip_ratio/low_min': 0.004545454401522875, 'clip_ratio/high_mean': 0.08814913779497147, 'clip_ratio/high_max': 0.08814913779497147, 'clip_ratio/region_mean': 0.09269459219649434, 'reward_total_mean': 0.5108710527420044, 'reward_meter_mean': 0.7386385202407837, 'reward_meter_std': 0.4336848556995392, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.983648419380188, 'reward_repeat_soft_std': 0.016031138598918915, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.5108710527420044, 'reward_total_composite_std': 0.22471150755882263, 'epoch': 0.17}

 52%|█████▏    | 1712/3300 [4:13:10<4:15:09,  9.64s/it]INFO 04-13 11:34:08 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1713/3300 [4:13:17<3:54:51,  8.88s/it]2026-04-13 11:34:15,554 | INFO | train_grpo_train | metrics_logged mode=train step=1713

                                                       
{'loss': 0.0401, 'grad_norm': 8.389473915100098, 'learning_rate': 4.8121212121212125e-06, 'num_tokens': 3049314.0, 'completions/mean_length': 110.375, 'completions/min_length': 82.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.375, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.802635669708252, 'rewards/meter/std': 0.19809673726558685, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9171019792556763, 'rewards/repeat_soft/std': 0.03333454951643944, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5566848516464233, 'rewards/total_composite/std': 0.05519948527216911, 'reward': 0.5566848516464233, 'reward_std': 0.05519949272274971, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10923045128583908, 'sampling/sampling_logp_difference/max': 2.0601401329040527, 'sampling/importance_sampling_ratio/min': 0.12743611633777618, 'sampling/importance_sampling_ratio/mean': 1.003833532333374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5601277053356171, 'clip_ratio/low_mean': 0.022562035359442234, 'clip_ratio/low_min': 0.022562035359442234, 'clip_ratio/high_mean': 0.05748207215219736, 'clip_ratio/high_max': 0.05748207215219736, 'clip_ratio/region_mean': 0.0800441075116396, 'reward_total_mean': 0.5566848516464233, 'reward_meter_mean': 0.802635669708252, 'reward_meter_std': 0.19809673726558685, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9171019792556763, 'reward_repeat_soft_std': 0.03333454951643944, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5566848516464233, 'reward_total_composite_std': 0.05519948527216911, 'epoch': 0.17}

 52%|█████▏    | 1713/3300 [4:13:17<3:54:51,  8.88s/it]INFO 04-13 11:34:15 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1714/3300 [4:13:23<3:34:06,  8.10s/it]2026-04-13 11:34:21,845 | INFO | train_grpo_train | metrics_logged mode=train step=1714

                                                       
{'loss': 0.0001, 'grad_norm': 9.25256633758545, 'learning_rate': 4.80909090909091e-06, 'num_tokens': 3050797.0, 'completions/mean_length': 36.375, 'completions/min_length': 34.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.4204741418361664, 'rewards/meter/std': 0.336512953042984, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8884314894676208, 'rewards/repeat_soft/std': 0.050489891320466995, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4530245363712311, 'rewards/total_composite/std': 0.09045282006263733, 'reward': 0.4530245363712311, 'reward_std': 0.09045282006263733, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0816180482506752, 'sampling/sampling_logp_difference/max': 2.278085231781006, 'sampling/importance_sampling_ratio/min': 0.10248024016618729, 'sampling/importance_sampling_ratio/mean': 0.9892809987068176, 'sampling/importance_sampling_ratio/max': 1.8046834468841553, 'entropy': 0.44643400982022285, 'clip_ratio/low_mean': 0.03467508847825229, 'clip_ratio/low_min': 0.03467508847825229, 'clip_ratio/high_mean': 0.0642518401145935, 'clip_ratio/high_max': 0.0642518401145935, 'clip_ratio/region_mean': 0.0989269285928458, 'reward_total_mean': 0.4530245363712311, 'reward_meter_mean': 0.4204741418361664, 'reward_meter_std': 0.336512953042984, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8884314894676208, 'reward_repeat_soft_std': 0.050489891320466995, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4530245363712311, 'reward_total_composite_std': 0.09045282006263733, 'epoch': 0.17}

 52%|█████▏    | 1714/3300 [4:13:24<3:34:06,  8.10s/it]INFO 04-13 11:34:22 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1715/3300 [4:13:30<3:19:02,  7.53s/it]2026-04-13 11:34:28,047 | INFO | train_grpo_train | metrics_logged mode=train step=1715

                                                       
{'loss': -0.0584, 'grad_norm': 9.323890686035156, 'learning_rate': 4.806060606060606e-06, 'num_tokens': 3052563.0, 'completions/mean_length': 50.75, 'completions/min_length': 38.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9446964263916016, 'rewards/meter/std': 0.09015833586454391, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8592045903205872, 'rewards/repeat_soft/std': 0.026658890768885612, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.7154438495635986, 'rewards/total_composite/std': 0.1673603504896164, 'reward': 0.7154438495635986, 'reward_std': 0.1673603504896164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07247844338417053, 'sampling/sampling_logp_difference/max': 0.961169958114624, 'sampling/importance_sampling_ratio/min': 0.38244518637657166, 'sampling/importance_sampling_ratio/mean': 1.01378333568573, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4564861096441746, 'clip_ratio/low_mean': 0.05797531921416521, 'clip_ratio/low_min': 0.05797531921416521, 'clip_ratio/high_mean': 0.04102809727191925, 'clip_ratio/high_max': 0.04102809727191925, 'clip_ratio/region_mean': 0.09900341648608446, 'reward_total_mean': 0.7154438495635986, 'reward_meter_mean': 0.9446964263916016, 'reward_meter_std': 0.09015833586454391, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8592045903205872, 'reward_repeat_soft_std': 0.026658890768885612, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.7154438495635986, 'reward_total_composite_std': 0.1673603504896164, 'epoch': 0.17}

 52%|█████▏    | 1715/3300 [4:13:30<3:19:02,  7.53s/it]INFO 04-13 11:34:28 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1716/3300 [4:13:36<3:11:11,  7.24s/it]2026-04-13 11:34:34,621 | INFO | train_grpo_train | metrics_logged mode=train step=1716

                                                       
{'loss': -0.0046, 'grad_norm': 13.035734176635742, 'learning_rate': 4.803030303030303e-06, 'num_tokens': 3054350.0, 'completions/mean_length': 52.375, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.375, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8809199929237366, 'rewards/meter/std': 0.20748481154441833, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9217410683631897, 'rewards/repeat_soft/std': 0.033956173807382584, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.25949129462242126, 'rewards/total_composite/mean': 0.7109391093254089, 'rewards/total_composite/std': 0.15203580260276794, 'reward': 0.7109391093254089, 'reward_std': 0.15203580260276794, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13106243312358856, 'sampling/sampling_logp_difference/max': 2.2116916179656982, 'sampling/importance_sampling_ratio/min': 0.1095152348279953, 'sampling/importance_sampling_ratio/mean': 1.009244680404663, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6339414864778519, 'clip_ratio/low_mean': 0.06317034922540188, 'clip_ratio/low_min': 0.06317034922540188, 'clip_ratio/high_mean': 0.04133064579218626, 'clip_ratio/high_max': 0.04133064579218626, 'clip_ratio/region_mean': 0.10450099501758814, 'reward_total_mean': 0.7109391093254089, 'reward_meter_mean': 0.8809199929237366, 'reward_meter_std': 0.20748481154441833, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9217410683631897, 'reward_repeat_soft_std': 0.033956173807382584, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.25949129462242126, 'reward_total_composite_mean': 0.7109391093254089, 'reward_total_composite_std': 0.15203580260276794, 'epoch': 0.17}

 52%|█████▏    | 1716/3300 [4:13:36<3:11:11,  7.24s/it]INFO 04-13 11:34:34 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1717/3300 [4:13:43<3:03:44,  6.96s/it]2026-04-13 11:34:40,940 | INFO | train_grpo_train | metrics_logged mode=train step=1717

                                                       
{'loss': 0.0767, 'grad_norm': 14.608574867248535, 'learning_rate': 4.800000000000001e-06, 'num_tokens': 3056031.0, 'completions/mean_length': 40.125, 'completions/min_length': 32.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.4885815680027008, 'rewards/meter/std': 0.3847001791000366, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.997380793094635, 'rewards/repeat_soft/std': 0.003361919429153204, 'rewards/judge_quality/mean': 0.7987500429153442, 'rewards/judge_quality/std': 0.22465451061725616, 'rewards/total_composite/mean': 0.6047747731208801, 'rewards/total_composite/std': 0.22018471360206604, 'reward': 0.6047747731208801, 'reward_std': 0.22018472850322723, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12028294056653976, 'sampling/sampling_logp_difference/max': 1.6750097274780273, 'sampling/importance_sampling_ratio/min': 0.18730635941028595, 'sampling/importance_sampling_ratio/mean': 0.9817869663238525, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4833027049899101, 'clip_ratio/low_mean': 0.06514029623940587, 'clip_ratio/low_min': 0.06514029623940587, 'clip_ratio/high_mean': 0.037118902895599604, 'clip_ratio/high_max': 0.037118902895599604, 'clip_ratio/region_mean': 0.10225919913500547, 'reward_total_mean': 0.6047747731208801, 'reward_meter_mean': 0.4885815680027008, 'reward_meter_std': 0.3847001791000366, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.997380793094635, 'reward_repeat_soft_std': 0.003361919429153204, 'reward_judge_quality_mean': 0.7987500429153442, 'reward_judge_quality_std': 0.22465451061725616, 'reward_total_composite_mean': 0.6047747731208801, 'reward_total_composite_std': 0.22018471360206604, 'epoch': 0.17}

 52%|█████▏    | 1717/3300 [4:13:43<3:03:44,  6.96s/it]INFO 04-13 11:34:41 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1718/3300 [4:13:49<2:56:17,  6.69s/it]2026-04-13 11:34:46,966 | INFO | train_grpo_train | metrics_logged mode=train step=1718

                                                       
{'loss': 0.0068, 'grad_norm': 14.311399459838867, 'learning_rate': 4.796969696969697e-06, 'num_tokens': 3057577.0, 'completions/mean_length': 21.25, 'completions/min_length': 20.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9603111147880554, 'rewards/meter/std': 0.06461824476718903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6089694499969482, 'rewards/total_composite/std': 0.0203370600938797, 'reward': 0.6089694499969482, 'reward_std': 0.020337076857686043, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07003674656152725, 'sampling/sampling_logp_difference/max': 0.9432845115661621, 'sampling/importance_sampling_ratio/min': 0.3893469274044037, 'sampling/importance_sampling_ratio/mean': 0.9982831478118896, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36864886805415154, 'clip_ratio/low_mean': 0.030113637447357178, 'clip_ratio/low_min': 0.030113637447357178, 'clip_ratio/high_mean': 0.0525162355042994, 'clip_ratio/high_max': 0.0525162355042994, 'clip_ratio/region_mean': 0.08262987295165658, 'reward_total_mean': 0.6089694499969482, 'reward_meter_mean': 0.9603111147880554, 'reward_meter_std': 0.06461824476718903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6089694499969482, 'reward_total_composite_std': 0.0203370600938797, 'epoch': 0.17}

 52%|█████▏    | 1718/3300 [4:13:49<2:56:17,  6.69s/it]INFO 04-13 11:34:47 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1719/3300 [4:13:56<2:59:37,  6.82s/it]2026-04-13 11:34:54,079 | INFO | train_grpo_train | metrics_logged mode=train step=1719

                                                       
{'loss': -0.0779, 'grad_norm': 15.311482429504395, 'learning_rate': 4.793939393939394e-06, 'num_tokens': 3059250.0, 'completions/mean_length': 46.125, 'completions/min_length': 38.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.44882452487945557, 'rewards/meter/std': 0.4111330211162567, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8626612424850464, 'rewards/repeat_soft/std': 0.13374517858028412, 'rewards/judge_quality/mean': 0.7325000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5474143624305725, 'rewards/total_composite/std': 0.22817520797252655, 'reward': 0.5474143624305725, 'reward_std': 0.22817519307136536, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11525116860866547, 'sampling/sampling_logp_difference/max': 2.725635290145874, 'sampling/importance_sampling_ratio/min': 0.06550458073616028, 'sampling/importance_sampling_ratio/mean': 1.0321239233016968, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5838736332952976, 'clip_ratio/low_mean': 0.06504121143370867, 'clip_ratio/low_min': 0.06504121143370867, 'clip_ratio/high_mean': 0.020101881120353937, 'clip_ratio/high_max': 0.020101881120353937, 'clip_ratio/region_mean': 0.0851430925540626, 'reward_total_mean': 0.5474143624305725, 'reward_meter_mean': 0.44882452487945557, 'reward_meter_std': 0.4111330211162567, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8626612424850464, 'reward_repeat_soft_std': 0.13374517858028412, 'reward_judge_quality_mean': 0.7325000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5474143624305725, 'reward_total_composite_std': 0.22817520797252655, 'epoch': 0.17}

 52%|█████▏    | 1719/3300 [4:13:56<2:59:37,  6.82s/it]INFO 04-13 11:34:54 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1720/3300 [4:14:03<3:06:59,  7.10s/it]2026-04-13 11:35:01,906 | INFO | train_grpo_train | metrics_logged mode=train step=1720

                                                       
{'loss': 0.0429, 'grad_norm': 6.603409290313721, 'learning_rate': 4.790909090909091e-06, 'num_tokens': 3061572.0, 'completions/mean_length': 110.25, 'completions/min_length': 99.0, 'completions/max_length': 125.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 110.25, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 125.0, 'rewards/meter/mean': 0.9676520228385925, 'rewards/meter/std': 0.032065149396657944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7556529641151428, 'rewards/repeat_soft/std': 0.08880481123924255, 'rewards/judge_quality/mean': 0.5325000286102295, 'rewards/judge_quality/std': 0.15526477992534637, 'rewards/total_composite/mean': 0.6472546458244324, 'rewards/total_composite/std': 0.10062187165021896, 'reward': 0.6472546458244324, 'reward_std': 0.10062186419963837, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09675262123346329, 'sampling/sampling_logp_difference/max': 1.920860767364502, 'sampling/importance_sampling_ratio/min': 0.14648081362247467, 'sampling/importance_sampling_ratio/mean': 1.0061516761779785, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5555191040039062, 'clip_ratio/low_mean': 0.053047848865389824, 'clip_ratio/low_min': 0.053047848865389824, 'clip_ratio/high_mean': 0.03863520547747612, 'clip_ratio/high_max': 0.03863520547747612, 'clip_ratio/region_mean': 0.09168305434286594, 'reward_total_mean': 0.6472546458244324, 'reward_meter_mean': 0.9676520228385925, 'reward_meter_std': 0.032065149396657944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7556529641151428, 'reward_repeat_soft_std': 0.08880481123924255, 'reward_judge_quality_mean': 0.5325000286102295, 'reward_judge_quality_std': 0.15526477992534637, 'reward_total_composite_mean': 0.6472546458244324, 'reward_total_composite_std': 0.10062187165021896, 'epoch': 0.17}

 52%|█████▏    | 1720/3300 [4:14:04<3:06:59,  7.10s/it]INFO 04-13 11:35:02 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1721/3300 [4:14:11<3:07:12,  7.11s/it]2026-04-13 11:35:09,053 | INFO | train_grpo_train | metrics_logged mode=train step=1721

                                                       
{'loss': 0.0166, 'grad_norm': 12.26235294342041, 'learning_rate': 4.787878787878788e-06, 'num_tokens': 3063239.0, 'completions/mean_length': 45.375, 'completions/min_length': 38.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.375, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8829458951950073, 'rewards/meter/std': 0.08990895003080368, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9137102961540222, 'rewards/repeat_soft/std': 0.03713247925043106, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5805026292800903, 'rewards/total_composite/std': 0.030275192111730576, 'reward': 0.5805026292800903, 'reward_std': 0.03027520515024662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11579596996307373, 'sampling/sampling_logp_difference/max': 1.4301106929779053, 'sampling/importance_sampling_ratio/min': 0.23928244411945343, 'sampling/importance_sampling_ratio/mean': 0.9996203780174255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7600813508033752, 'clip_ratio/low_mean': 0.04526778403669596, 'clip_ratio/low_min': 0.04526778403669596, 'clip_ratio/high_mean': 0.05815478228032589, 'clip_ratio/high_max': 0.05815478228032589, 'clip_ratio/region_mean': 0.10342256631702185, 'reward_total_mean': 0.5805026292800903, 'reward_meter_mean': 0.8829458951950073, 'reward_meter_std': 0.08990895003080368, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9137102961540222, 'reward_repeat_soft_std': 0.03713247925043106, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5805026292800903, 'reward_total_composite_std': 0.030275192111730576, 'epoch': 0.17}

 52%|█████▏    | 1721/3300 [4:14:11<3:07:12,  7.11s/it]INFO 04-13 11:35:09 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1722/3300 [4:14:17<3:05:00,  7.03s/it]2026-04-13 11:35:15,894 | INFO | train_grpo_train | metrics_logged mode=train step=1722

                                                       
{'loss': 0.1096, 'grad_norm': 15.155242919921875, 'learning_rate': 4.784848484848485e-06, 'num_tokens': 3064764.0, 'completions/mean_length': 32.625, 'completions/min_length': 26.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.625, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.6672650575637817, 'rewards/meter/std': 0.4079885482788086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9153906106948853, 'rewards/repeat_soft/std': 0.05205296352505684, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.5362443923950195, 'rewards/total_composite/std': 0.10083714127540588, 'reward': 0.5362443923950195, 'reward_std': 0.10083714127540588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12003348022699356, 'sampling/sampling_logp_difference/max': 2.377509593963623, 'sampling/importance_sampling_ratio/min': 0.09278135746717453, 'sampling/importance_sampling_ratio/mean': 0.9632998704910278, 'sampling/importance_sampling_ratio/max': 1.7206178903579712, 'entropy': 0.4664684534072876, 'clip_ratio/low_mean': 0.01881533139385283, 'clip_ratio/low_min': 0.01881533139385283, 'clip_ratio/high_mean': 0.10188171919435263, 'clip_ratio/high_max': 0.10188171919435263, 'clip_ratio/region_mean': 0.12069705058820546, 'reward_total_mean': 0.5362443923950195, 'reward_meter_mean': 0.6672650575637817, 'reward_meter_std': 0.4079885482788086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9153906106948853, 'reward_repeat_soft_std': 0.05205296352505684, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.5362443923950195, 'reward_total_composite_std': 0.10083714127540588, 'epoch': 0.17}

 52%|█████▏    | 1722/3300 [4:14:18<3:05:00,  7.03s/it]INFO 04-13 11:35:16 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1723/3300 [4:14:24<3:03:36,  6.99s/it]2026-04-13 11:35:22,773 | INFO | train_grpo_train | metrics_logged mode=train step=1723

                                                       
{'loss': 0.0051, 'grad_norm': 13.15866756439209, 'learning_rate': 4.7818181818181825e-06, 'num_tokens': 3066345.0, 'completions/mean_length': 43.625, 'completions/min_length': 37.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.7076498866081238, 'rewards/meter/std': 0.28890377283096313, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9060574769973755, 'rewards/repeat_soft/std': 0.05195293948054314, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5321946740150452, 'rewards/total_composite/std': 0.0805114358663559, 'reward': 0.5321946740150452, 'reward_std': 0.0805114284157753, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10051216930150986, 'sampling/sampling_logp_difference/max': 2.5790698528289795, 'sampling/importance_sampling_ratio/min': 0.07584451884031296, 'sampling/importance_sampling_ratio/mean': 1.0028880834579468, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5359084531664848, 'clip_ratio/low_mean': 0.030447634868323803, 'clip_ratio/low_min': 0.030447634868323803, 'clip_ratio/high_mean': 0.05426136334426701, 'clip_ratio/high_max': 0.05426136334426701, 'clip_ratio/region_mean': 0.08470899821259081, 'reward_total_mean': 0.5321946740150452, 'reward_meter_mean': 0.7076498866081238, 'reward_meter_std': 0.28890377283096313, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9060574769973755, 'reward_repeat_soft_std': 0.05195293948054314, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5321946740150452, 'reward_total_composite_std': 0.0805114358663559, 'epoch': 0.17}

 52%|█████▏    | 1723/3300 [4:14:25<3:03:36,  6.99s/it]INFO 04-13 11:35:23 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1724/3300 [4:14:31<3:03:45,  7.00s/it]2026-04-13 11:35:29,792 | INFO | train_grpo_train | metrics_logged mode=train step=1724

                                                       
{'loss': 0.0253, 'grad_norm': 14.858102798461914, 'learning_rate': 4.77878787878788e-06, 'num_tokens': 3067837.0, 'completions/mean_length': 40.5, 'completions/min_length': 36.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.7685610055923462, 'rewards/meter/std': 0.25966876745224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9639634490013123, 'rewards/repeat_soft/std': 0.032831985503435135, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5945678353309631, 'rewards/total_composite/std': 0.15389123558998108, 'reward': 0.5945678353309631, 'reward_std': 0.15389125049114227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12579135596752167, 'sampling/sampling_logp_difference/max': 1.070995807647705, 'sampling/importance_sampling_ratio/min': 0.3426671028137207, 'sampling/importance_sampling_ratio/mean': 1.0026395320892334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7431009709835052, 'clip_ratio/low_mean': 0.08824184723198414, 'clip_ratio/low_min': 0.08824184723198414, 'clip_ratio/high_mean': 0.0321815712377429, 'clip_ratio/high_max': 0.0321815712377429, 'clip_ratio/region_mean': 0.12042341846972704, 'reward_total_mean': 0.5945678353309631, 'reward_meter_mean': 0.7685610055923462, 'reward_meter_std': 0.25966876745224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9639634490013123, 'reward_repeat_soft_std': 0.032831985503435135, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5945678353309631, 'reward_total_composite_std': 0.15389123558998108, 'epoch': 0.17}

 52%|█████▏    | 1724/3300 [4:14:32<3:03:45,  7.00s/it]INFO 04-13 11:35:30 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1725/3300 [4:14:38<3:02:26,  6.95s/it]2026-04-13 11:35:36,644 | INFO | train_grpo_train | metrics_logged mode=train step=1725

                                                       
{'loss': 0.0081, 'grad_norm': 15.199424743652344, 'learning_rate': 4.775757575757576e-06, 'num_tokens': 3069466.0, 'completions/mean_length': 35.625, 'completions/min_length': 34.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.7659406661987305, 'rewards/meter/std': 0.12717434763908386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8684053421020508, 'rewards/repeat_soft/std': 0.021665874868631363, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5410131216049194, 'rewards/total_composite/std': 0.03264806792140007, 'reward': 0.5410131216049194, 'reward_std': 0.03264806047081947, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07706625759601593, 'sampling/sampling_logp_difference/max': 1.8642072677612305, 'sampling/importance_sampling_ratio/min': 0.15501905977725983, 'sampling/importance_sampling_ratio/mean': 0.9940878748893738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.21618294902145863, 'clip_ratio/low_mean': 0.03824948891997337, 'clip_ratio/low_min': 0.03824948891997337, 'clip_ratio/high_mean': 0.010416666744276881, 'clip_ratio/high_max': 0.010416666744276881, 'clip_ratio/region_mean': 0.048666155664250255, 'reward_total_mean': 0.5410131216049194, 'reward_meter_mean': 0.7659406661987305, 'reward_meter_std': 0.12717434763908386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8684053421020508, 'reward_repeat_soft_std': 0.021665874868631363, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5410131216049194, 'reward_total_composite_std': 0.03264806792140007, 'epoch': 0.17}

 52%|█████▏    | 1725/3300 [4:14:38<3:02:26,  6.95s/it]INFO 04-13 11:35:37 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1726/3300 [4:14:46<3:12:22,  7.33s/it]2026-04-13 11:35:44,864 | INFO | train_grpo_train | metrics_logged mode=train step=1726

                                                       
{'loss': 0.0364, 'grad_norm': 9.00948429107666, 'learning_rate': 4.772727272727273e-06, 'num_tokens': 3071996.0, 'completions/mean_length': 111.25, 'completions/min_length': 99.0, 'completions/max_length': 126.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 111.25, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 126.0, 'rewards/meter/mean': 0.790528416633606, 'rewards/meter/std': 0.3705079257488251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.818536102771759, 'rewards/repeat_soft/std': 0.13772985339164734, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5385947227478027, 'rewards/total_composite/std': 0.09109405428171158, 'reward': 0.5385947227478027, 'reward_std': 0.09109406173229218, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08634468913078308, 'sampling/sampling_logp_difference/max': 1.504404067993164, 'sampling/importance_sampling_ratio/min': 0.22214964032173157, 'sampling/importance_sampling_ratio/mean': 1.007115364074707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49379777163267136, 'clip_ratio/low_mean': 0.01975108217447996, 'clip_ratio/low_min': 0.01975108217447996, 'clip_ratio/high_mean': 0.06179565005004406, 'clip_ratio/high_max': 0.06179565005004406, 'clip_ratio/region_mean': 0.08154673222452402, 'reward_total_mean': 0.5385947227478027, 'reward_meter_mean': 0.790528416633606, 'reward_meter_std': 0.3705079257488251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.818536102771759, 'reward_repeat_soft_std': 0.13772985339164734, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5385947227478027, 'reward_total_composite_std': 0.09109405428171158, 'epoch': 0.17}

 52%|█████▏    | 1726/3300 [4:14:47<3:12:22,  7.33s/it]INFO 04-13 11:35:45 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1727/3300 [4:14:54<3:17:56,  7.55s/it]2026-04-13 11:35:52,865 | INFO | train_grpo_train | metrics_logged mode=train step=1727

                                                       
{'loss': 0.0281, 'grad_norm': 5.4726033210754395, 'learning_rate': 4.769696969696971e-06, 'num_tokens': 3074126.0, 'completions/mean_length': 108.25, 'completions/min_length': 86.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.25, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9783064723014832, 'rewards/meter/std': 0.007912681438028812, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7211712598800659, 'rewards/repeat_soft/std': 0.06373512744903564, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.6012626886367798, 'rewards/total_composite/std': 0.06286019831895828, 'reward': 0.6012626886367798, 'reward_std': 0.06286019831895828, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08809080719947815, 'sampling/sampling_logp_difference/max': 1.8019466400146484, 'sampling/importance_sampling_ratio/min': 0.25487732887268066, 'sampling/importance_sampling_ratio/mean': 1.01508367061615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5446457751095295, 'clip_ratio/low_mean': 0.06396431918255985, 'clip_ratio/low_min': 0.06396431918255985, 'clip_ratio/high_mean': 0.012019230984151363, 'clip_ratio/high_max': 0.012019230984151363, 'clip_ratio/region_mean': 0.07598355016671121, 'reward_total_mean': 0.6012626886367798, 'reward_meter_mean': 0.9783064723014832, 'reward_meter_std': 0.007912681438028812, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7211712598800659, 'reward_repeat_soft_std': 0.06373512744903564, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.6012626886367798, 'reward_total_composite_std': 0.06286019831895828, 'epoch': 0.17}

 52%|█████▏    | 1727/3300 [4:14:55<3:17:56,  7.55s/it]INFO 04-13 11:35:53 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1728/3300 [4:15:02<3:14:42,  7.43s/it]2026-04-13 11:36:00,073 | INFO | train_grpo_train | metrics_logged mode=train step=1728

                                                       
{'loss': -0.0284, 'grad_norm': 10.919466018676758, 'learning_rate': 4.766666666666667e-06, 'num_tokens': 3075944.0, 'completions/mean_length': 51.25, 'completions/min_length': 42.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8725104928016663, 'rewards/meter/std': 0.17084768414497375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8668285608291626, 'rewards/repeat_soft/std': 0.0349387563765049, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5713319778442383, 'rewards/total_composite/std': 0.043708302080631256, 'reward': 0.5713319778442383, 'reward_std': 0.043708302080631256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09125392138957977, 'sampling/sampling_logp_difference/max': 1.3393590450286865, 'sampling/importance_sampling_ratio/min': 0.2620135545730591, 'sampling/importance_sampling_ratio/mean': 1.034496784210205, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6422306261956692, 'clip_ratio/low_mean': 0.03758142422884703, 'clip_ratio/low_min': 0.03758142422884703, 'clip_ratio/high_mean': 0.0488913943991065, 'clip_ratio/high_max': 0.0488913943991065, 'clip_ratio/region_mean': 0.08647281862795353, 'reward_total_mean': 0.5713319778442383, 'reward_meter_mean': 0.8725104928016663, 'reward_meter_std': 0.17084768414497375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8668285608291626, 'reward_repeat_soft_std': 0.0349387563765049, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5713319778442383, 'reward_total_composite_std': 0.043708302080631256, 'epoch': 0.17}

 52%|█████▏    | 1728/3300 [4:15:02<3:14:42,  7.43s/it]INFO 04-13 11:36:00 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1729/3300 [4:15:08<3:04:54,  7.06s/it]2026-04-13 11:36:06,214 | INFO | train_grpo_train | metrics_logged mode=train step=1729

                                                       
{'loss': -0.0457, 'grad_norm': 15.969704627990723, 'learning_rate': 4.763636363636364e-06, 'num_tokens': 3077334.0, 'completions/mean_length': 22.75, 'completions/min_length': 16.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8908631801605225, 'rewards/meter/std': 0.20336231589317322, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9284578561782837, 'rewards/repeat_soft/std': 0.034619759768247604, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5687318444252014, 'rewards/total_composite/std': 0.05835019424557686, 'reward': 0.5687318444252014, 'reward_std': 0.05835019052028656, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08931823819875717, 'sampling/sampling_logp_difference/max': 0.7789598703384399, 'sampling/importance_sampling_ratio/min': 0.47100359201431274, 'sampling/importance_sampling_ratio/mean': 1.0275490283966064, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7520938664674759, 'clip_ratio/low_mean': 0.011363636702299118, 'clip_ratio/low_min': 0.011363636702299118, 'clip_ratio/high_mean': 0.05802327301353216, 'clip_ratio/high_max': 0.05802327301353216, 'clip_ratio/region_mean': 0.06938690971583128, 'reward_total_mean': 0.5687318444252014, 'reward_meter_mean': 0.8908631801605225, 'reward_meter_std': 0.20336231589317322, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9284578561782837, 'reward_repeat_soft_std': 0.034619759768247604, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5687318444252014, 'reward_total_composite_std': 0.05835019424557686, 'epoch': 0.17}

 52%|█████▏    | 1729/3300 [4:15:08<3:04:54,  7.06s/it]INFO 04-13 11:36:06 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1730/3300 [4:15:14<2:56:59,  6.76s/it]2026-04-13 11:36:12,275 | INFO | train_grpo_train | metrics_logged mode=train step=1730

                                                       
{'loss': 0.1095, 'grad_norm': 12.536210060119629, 'learning_rate': 4.760606060606061e-06, 'num_tokens': 3078969.0, 'completions/mean_length': 38.375, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.375, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.49436667561531067, 'rewards/meter/std': 0.4479481875896454, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8966197371482849, 'rewards/repeat_soft/std': 0.059394724667072296, 'rewards/judge_quality/mean': 0.4387499988079071, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4764440953731537, 'rewards/total_composite/std': 0.13087791204452515, 'reward': 0.4764440953731537, 'reward_std': 0.13087791204452515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11847681552171707, 'sampling/sampling_logp_difference/max': 1.4087927341461182, 'sampling/importance_sampling_ratio/min': 0.24443821609020233, 'sampling/importance_sampling_ratio/mean': 1.0156583786010742, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.568436685949564, 'clip_ratio/low_mean': 0.04864378087222576, 'clip_ratio/low_min': 0.04864378087222576, 'clip_ratio/high_mean': 0.048414388904348016, 'clip_ratio/high_max': 0.048414388904348016, 'clip_ratio/region_mean': 0.09705816977657378, 'reward_total_mean': 0.4764440953731537, 'reward_meter_mean': 0.49436667561531067, 'reward_meter_std': 0.4479481875896454, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8966197371482849, 'reward_repeat_soft_std': 0.059394724667072296, 'reward_judge_quality_mean': 0.4387499988079071, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4764440953731537, 'reward_total_composite_std': 0.13087791204452515, 'epoch': 0.17}

 52%|█████▏    | 1730/3300 [4:15:14<2:56:59,  6.76s/it]INFO 04-13 11:36:12 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1731/3300 [4:15:23<3:11:32,  7.32s/it]2026-04-13 11:36:20,974 | INFO | train_grpo_train | metrics_logged mode=train step=1731

                                                       
{'loss': 0.0197, 'grad_norm': 4.172364711761475, 'learning_rate': 4.757575757575758e-06, 'num_tokens': 3081969.0, 'completions/mean_length': 166.0, 'completions/min_length': 120.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.0, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.9715299606323242, 'rewards/meter/std': 0.010508740320801735, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.660871684551239, 'rewards/repeat_soft/std': 0.10618186742067337, 'rewards/judge_quality/mean': 0.4025000035762787, 'rewards/judge_quality/std': 0.06250713765621185, 'rewards/total_composite/mean': 0.5408867597579956, 'rewards/total_composite/std': 0.04914843291044235, 'reward': 0.5408867597579956, 'reward_std': 0.049148425459861755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07430362701416016, 'sampling/sampling_logp_difference/max': 9.649591445922852, 'sampling/importance_sampling_ratio/min': 6.445188773795962e-05, 'sampling/importance_sampling_ratio/mean': 1.0056936740875244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34446943178772926, 'clip_ratio/low_mean': 0.011553122196346521, 'clip_ratio/low_min': 0.011553122196346521, 'clip_ratio/high_mean': 0.04789688950404525, 'clip_ratio/high_max': 0.04789688950404525, 'clip_ratio/region_mean': 0.05945001170039177, 'reward_total_mean': 0.5408867597579956, 'reward_meter_mean': 0.9715299606323242, 'reward_meter_std': 0.010508740320801735, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.660871684551239, 'reward_repeat_soft_std': 0.10618186742067337, 'reward_judge_quality_mean': 0.4025000035762787, 'reward_judge_quality_std': 0.06250713765621185, 'reward_total_composite_mean': 0.5408867597579956, 'reward_total_composite_std': 0.04914843291044235, 'epoch': 0.17}

 52%|█████▏    | 1731/3300 [4:15:23<3:11:32,  7.32s/it]INFO 04-13 11:36:21 [block_pool.py:378] Successfully reset prefix cache

 52%|█████▏    | 1732/3300 [4:15:30<3:09:13,  7.24s/it]2026-04-13 11:36:28,009 | INFO | train_grpo_train | metrics_logged mode=train step=1732

                                                       
{'loss': 0.045, 'grad_norm': 8.757774353027344, 'learning_rate': 4.754545454545455e-06, 'num_tokens': 3083735.0, 'completions/mean_length': 58.75, 'completions/min_length': 54.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.75, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9289580583572388, 'rewards/meter/std': 0.052348531782627106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8491862416267395, 'rewards/repeat_soft/std': 0.05479966104030609, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5675961971282959, 'rewards/total_composite/std': 0.04271506518125534, 'reward': 0.5675961971282959, 'reward_std': 0.04271508380770683, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07935873419046402, 'sampling/sampling_logp_difference/max': 1.4487967491149902, 'sampling/importance_sampling_ratio/min': 0.23485270142555237, 'sampling/importance_sampling_ratio/mean': 1.002606987953186, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33651405945420265, 'clip_ratio/low_mean': 0.023050089366734028, 'clip_ratio/low_min': 0.023050089366734028, 'clip_ratio/high_mean': 0.045291122049093246, 'clip_ratio/high_max': 0.045291122049093246, 'clip_ratio/region_mean': 0.06834121141582727, 'reward_total_mean': 0.5675961971282959, 'reward_meter_mean': 0.9289580583572388, 'reward_meter_std': 0.052348531782627106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8491862416267395, 'reward_repeat_soft_std': 0.05479966104030609, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5675961971282959, 'reward_total_composite_std': 0.04271506518125534, 'epoch': 0.17}

 52%|█████▏    | 1732/3300 [4:15:30<3:09:13,  7.24s/it]INFO 04-13 11:36:28 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1733/3300 [4:15:37<3:08:09,  7.20s/it]2026-04-13 11:36:35,141 | INFO | train_grpo_train | metrics_logged mode=train step=1733

                                                       
{'loss': -0.0296, 'grad_norm': 8.731679916381836, 'learning_rate': 4.751515151515152e-06, 'num_tokens': 3085653.0, 'completions/mean_length': 61.75, 'completions/min_length': 55.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7169671058654785, 'rewards/meter/std': 0.36305248737335205, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8858823180198669, 'rewards/repeat_soft/std': 0.0730840340256691, 'rewards/judge_quality/mean': 0.6075000166893005, 'rewards/judge_quality/std': 0.25877460837364197, 'rewards/total_composite/mean': 0.5934597253799438, 'rewards/total_composite/std': 0.16547562181949615, 'reward': 0.5934597253799438, 'reward_std': 0.16547560691833496, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07510776072740555, 'sampling/sampling_logp_difference/max': 1.4382802248001099, 'sampling/importance_sampling_ratio/min': 0.23733556270599365, 'sampling/importance_sampling_ratio/mean': 1.0047154426574707, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.305077251046896, 'clip_ratio/low_mean': 0.048390478594228625, 'clip_ratio/low_min': 0.048390478594228625, 'clip_ratio/high_mean': 0.027283618226647377, 'clip_ratio/high_max': 0.027283618226647377, 'clip_ratio/region_mean': 0.075674096820876, 'reward_total_mean': 0.5934597253799438, 'reward_meter_mean': 0.7169671058654785, 'reward_meter_std': 0.36305248737335205, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8858823180198669, 'reward_repeat_soft_std': 0.0730840340256691, 'reward_judge_quality_mean': 0.6075000166893005, 'reward_judge_quality_std': 0.25877460837364197, 'reward_total_composite_mean': 0.5934597253799438, 'reward_total_composite_std': 0.16547562181949615, 'epoch': 0.17}

 53%|█████▎    | 1733/3300 [4:15:37<3:08:09,  7.20s/it]INFO 04-13 11:36:35 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1734/3300 [4:15:44<3:11:17,  7.33s/it]2026-04-13 11:36:42,752 | INFO | train_grpo_train | metrics_logged mode=train step=1734

                                                       
{'loss': 0.0396, 'grad_norm': 9.175833702087402, 'learning_rate': 4.748484848484849e-06, 'num_tokens': 3087311.0, 'completions/mean_length': 57.25, 'completions/min_length': 50.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9507708549499512, 'rewards/meter/std': 0.06551821529865265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.923712968826294, 'rewards/repeat_soft/std': 0.03284953162074089, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.6728672385215759, 'rewards/total_composite/std': 0.1251244693994522, 'reward': 0.6728672385215759, 'reward_std': 0.1251244693994522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09673381596803665, 'sampling/sampling_logp_difference/max': 0.9068347215652466, 'sampling/importance_sampling_ratio/min': 0.4449917674064636, 'sampling/importance_sampling_ratio/mean': 1.0263419151306152, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.643095389008522, 'clip_ratio/low_mean': 0.08720999862998724, 'clip_ratio/low_min': 0.08720999862998724, 'clip_ratio/high_mean': 0.02271186374127865, 'clip_ratio/high_max': 0.02271186374127865, 'clip_ratio/region_mean': 0.10992186237126589, 'reward_total_mean': 0.6728672385215759, 'reward_meter_mean': 0.9507708549499512, 'reward_meter_std': 0.06551821529865265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.923712968826294, 'reward_repeat_soft_std': 0.03284953162074089, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.6728672385215759, 'reward_total_composite_std': 0.1251244693994522, 'epoch': 0.17}

 53%|█████▎    | 1734/3300 [4:15:44<3:11:17,  7.33s/it]INFO 04-13 11:36:43 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1735/3300 [4:15:51<3:03:53,  7.05s/it]2026-04-13 11:36:49,099 | INFO | train_grpo_train | metrics_logged mode=train step=1735

                                                       
{'loss': -0.0233, 'grad_norm': 11.077441215515137, 'learning_rate': 4.745454545454546e-06, 'num_tokens': 3088892.0, 'completions/mean_length': 47.625, 'completions/min_length': 41.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9870283007621765, 'rewards/meter/std': 0.009919442236423492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9679203629493713, 'rewards/repeat_soft/std': 0.039826054126024246, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.6194697618484497, 'rewards/total_composite/std': 0.01266756933182478, 'reward': 0.6194697618484497, 'reward_std': 0.012667574919760227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09732978045940399, 'sampling/sampling_logp_difference/max': 1.9471259117126465, 'sampling/importance_sampling_ratio/min': 0.14268356561660767, 'sampling/importance_sampling_ratio/mean': 1.0145297050476074, 'sampling/importance_sampling_ratio/max': 1.7899833917617798, 'entropy': 0.7135571166872978, 'clip_ratio/low_mean': 0.04798766737803817, 'clip_ratio/low_min': 0.04798766737803817, 'clip_ratio/high_mean': 0.04465630929917097, 'clip_ratio/high_max': 0.04465630929917097, 'clip_ratio/region_mean': 0.09264397667720914, 'reward_total_mean': 0.6194697618484497, 'reward_meter_mean': 0.9870283007621765, 'reward_meter_std': 0.009919442236423492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9679203629493713, 'reward_repeat_soft_std': 0.039826054126024246, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.6194697618484497, 'reward_total_composite_std': 0.01266756933182478, 'epoch': 0.17}

 53%|█████▎    | 1735/3300 [4:15:51<3:03:53,  7.05s/it]INFO 04-13 11:36:49 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1736/3300 [4:15:58<3:04:21,  7.07s/it]2026-04-13 11:36:56,231 | INFO | train_grpo_train | metrics_logged mode=train step=1736

                                                       
{'loss': 0.0929, 'grad_norm': 7.1619768142700195, 'learning_rate': 4.7424242424242426e-06, 'num_tokens': 3091085.0, 'completions/mean_length': 88.125, 'completions/min_length': 66.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.125, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.9182576537132263, 'rewards/meter/std': 0.07804084569215775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7472658753395081, 'rewards/repeat_soft/std': 0.1533329039812088, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.557198703289032, 'rewards/total_composite/std': 0.08052248507738113, 'reward': 0.557198703289032, 'reward_std': 0.08052248507738113, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08153990656137466, 'sampling/sampling_logp_difference/max': 1.4716401100158691, 'sampling/importance_sampling_ratio/min': 0.22954869270324707, 'sampling/importance_sampling_ratio/mean': 1.0046273469924927, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5766340754926205, 'clip_ratio/low_mean': 0.04738907259888947, 'clip_ratio/low_min': 0.04738907259888947, 'clip_ratio/high_mean': 0.029976489953696728, 'clip_ratio/high_max': 0.029976489953696728, 'clip_ratio/region_mean': 0.0773655625525862, 'reward_total_mean': 0.557198703289032, 'reward_meter_mean': 0.9182576537132263, 'reward_meter_std': 0.07804084569215775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7472658753395081, 'reward_repeat_soft_std': 0.1533329039812088, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.557198703289032, 'reward_total_composite_std': 0.08052248507738113, 'epoch': 0.17}

 53%|█████▎    | 1736/3300 [4:15:58<3:04:21,  7.07s/it]INFO 04-13 11:36:56 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1737/3300 [4:16:05<3:06:33,  7.16s/it]2026-04-13 11:37:03,654 | INFO | train_grpo_train | metrics_logged mode=train step=1737

                                                       
{'loss': 0.122, 'grad_norm': 15.391199111938477, 'learning_rate': 4.73939393939394e-06, 'num_tokens': 3092727.0, 'completions/mean_length': 51.25, 'completions/min_length': 38.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.6474987268447876, 'rewards/meter/std': 0.4506898522377014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9824091196060181, 'rewards/repeat_soft/std': 0.030868539586663246, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5271519422531128, 'rewards/total_composite/std': 0.12375441193580627, 'reward': 0.5271519422531128, 'reward_std': 0.12375440448522568, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14935629069805145, 'sampling/sampling_logp_difference/max': 1.8774235248565674, 'sampling/importance_sampling_ratio/min': 0.15298376977443695, 'sampling/importance_sampling_ratio/mean': 1.0210734605789185, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.114188238978386, 'clip_ratio/low_mean': 0.044731407426297665, 'clip_ratio/low_min': 0.044731407426297665, 'clip_ratio/high_mean': 0.07409316394478083, 'clip_ratio/high_max': 0.07409316394478083, 'clip_ratio/region_mean': 0.11882457137107849, 'reward_total_mean': 0.5271519422531128, 'reward_meter_mean': 0.6474987268447876, 'reward_meter_std': 0.4506898522377014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9824091196060181, 'reward_repeat_soft_std': 0.030868539586663246, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5271519422531128, 'reward_total_composite_std': 0.12375441193580627, 'epoch': 0.17}

 53%|█████▎    | 1737/3300 [4:16:05<3:06:33,  7.16s/it]INFO 04-13 11:37:04 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1738/3300 [4:16:13<3:14:50,  7.48s/it]2026-04-13 11:37:11,830 | INFO | train_grpo_train | metrics_logged mode=train step=1738

                                                       
{'loss': 0.0749, 'grad_norm': 9.294628143310547, 'learning_rate': 4.736363636363637e-06, 'num_tokens': 3095324.0, 'completions/mean_length': 113.625, 'completions/min_length': 102.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 113.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9556190967559814, 'rewards/meter/std': 0.10261744260787964, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7098491191864014, 'rewards/repeat_soft/std': 0.07084794342517853, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5585969686508179, 'rewards/total_composite/std': 0.05751483142375946, 'reward': 0.5585969686508179, 'reward_std': 0.05751482769846916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08325659483671188, 'sampling/sampling_logp_difference/max': 1.6753778457641602, 'sampling/importance_sampling_ratio/min': 0.18723741173744202, 'sampling/importance_sampling_ratio/mean': 0.9986563920974731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3381666596978903, 'clip_ratio/low_mean': 0.01459030108526349, 'clip_ratio/low_min': 0.01459030108526349, 'clip_ratio/high_mean': 0.06721780961379409, 'clip_ratio/high_max': 0.06721780961379409, 'clip_ratio/region_mean': 0.08180811069905758, 'reward_total_mean': 0.5585969686508179, 'reward_meter_mean': 0.9556190967559814, 'reward_meter_std': 0.10261744260787964, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7098491191864014, 'reward_repeat_soft_std': 0.07084794342517853, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5585969686508179, 'reward_total_composite_std': 0.05751483142375946, 'epoch': 0.17}

 53%|█████▎    | 1738/3300 [4:16:14<3:14:50,  7.48s/it]INFO 04-13 11:37:12 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1739/3300 [4:16:20<3:08:17,  7.24s/it]2026-04-13 11:37:18,493 | INFO | train_grpo_train | metrics_logged mode=train step=1739

                                                       
{'loss': 0.033, 'grad_norm': 15.374832153320312, 'learning_rate': 4.7333333333333335e-06, 'num_tokens': 3096913.0, 'completions/mean_length': 33.625, 'completions/min_length': 30.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8024061918258667, 'rewards/meter/std': 0.32132992148399353, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8930982947349548, 'rewards/repeat_soft/std': 0.02109317108988762, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.267261266708374, 'rewards/total_composite/mean': 0.6610607504844666, 'rewards/total_composite/std': 0.18794965744018555, 'reward': 0.6610607504844666, 'reward_std': 0.18794964253902435, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.076873779296875, 'sampling/sampling_logp_difference/max': 1.4227275848388672, 'sampling/importance_sampling_ratio/min': 0.24105563759803772, 'sampling/importance_sampling_ratio/mean': 1.011893391609192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4238227345049381, 'clip_ratio/low_mean': 0.05576387792825699, 'clip_ratio/low_min': 0.05576387792825699, 'clip_ratio/high_mean': 0.022368420846760273, 'clip_ratio/high_max': 0.022368420846760273, 'clip_ratio/region_mean': 0.07813229877501726, 'reward_total_mean': 0.6610607504844666, 'reward_meter_mean': 0.8024061918258667, 'reward_meter_std': 0.32132992148399353, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8930982947349548, 'reward_repeat_soft_std': 0.02109317108988762, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.267261266708374, 'reward_total_composite_mean': 0.6610607504844666, 'reward_total_composite_std': 0.18794965744018555, 'epoch': 0.17}

 53%|█████▎    | 1739/3300 [4:16:20<3:08:17,  7.24s/it]INFO 04-13 11:37:18 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1740/3300 [4:16:28<3:14:39,  7.49s/it]2026-04-13 11:37:26,623 | INFO | train_grpo_train | metrics_logged mode=train step=1740

                                                       
{'loss': -0.0102, 'grad_norm': 5.100631237030029, 'learning_rate': 4.730303030303031e-06, 'num_tokens': 3099420.0, 'completions/mean_length': 126.375, 'completions/min_length': 107.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.375, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.9850823879241943, 'rewards/meter/std': 0.007084795273840427, 'rewards/count_adherence/mean': 0.8500000238418579, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7304711937904358, 'rewards/repeat_soft/std': 0.09397655725479126, 'rewards/judge_quality/mean': 0.32249999046325684, 'rewards/judge_quality/std': 0.10925068706274033, 'rewards/total_composite/mean': 0.4860914349555969, 'rewards/total_composite/std': 0.07439238578081131, 'reward': 0.4860914349555969, 'reward_std': 0.07439238578081131, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061496373265981674, 'sampling/sampling_logp_difference/max': 1.6928315162658691, 'sampling/importance_sampling_ratio/min': 0.18399780988693237, 'sampling/importance_sampling_ratio/mean': 1.0103132724761963, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3876885324716568, 'clip_ratio/low_mean': 0.022239373065531254, 'clip_ratio/low_min': 0.022239373065531254, 'clip_ratio/high_mean': 0.028220489853993058, 'clip_ratio/high_max': 0.028220489853993058, 'clip_ratio/region_mean': 0.05045986291952431, 'reward_total_mean': 0.4860914349555969, 'reward_meter_mean': 0.9850823879241943, 'reward_meter_std': 0.007084795273840427, 'reward_count_adherence_mean': 0.8500000238418579, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7304711937904358, 'reward_repeat_soft_std': 0.09397655725479126, 'reward_judge_quality_mean': 0.32249999046325684, 'reward_judge_quality_std': 0.10925068706274033, 'reward_total_composite_mean': 0.4860914349555969, 'reward_total_composite_std': 0.07439238578081131, 'epoch': 0.17}

 53%|█████▎    | 1740/3300 [4:16:28<3:14:39,  7.49s/it]INFO 04-13 11:37:26 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1741/3300 [4:16:35<3:12:10,  7.40s/it]2026-04-13 11:37:33,800 | INFO | train_grpo_train | metrics_logged mode=train step=1741

                                                       
{'loss': -0.0792, 'grad_norm': 12.408956527709961, 'learning_rate': 4.727272727272728e-06, 'num_tokens': 3101128.0, 'completions/mean_length': 50.5, 'completions/min_length': 42.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.4509848952293396, 'rewards/meter/std': 0.30584847927093506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.991193413734436, 'rewards/repeat_soft/std': 0.009842580184340477, 'rewards/judge_quality/mean': 0.59375, 'rewards/judge_quality/std': 0.2775370180606842, 'rewards/total_composite/mean': 0.5134477019309998, 'rewards/total_composite/std': 0.13883522152900696, 'reward': 0.5134477019309998, 'reward_std': 0.13883522152900696, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10810422897338867, 'sampling/sampling_logp_difference/max': 1.3345818519592285, 'sampling/importance_sampling_ratio/min': 0.26326823234558105, 'sampling/importance_sampling_ratio/mean': 1.00209641456604, 'sampling/importance_sampling_ratio/max': 1.9134043455123901, 'entropy': 0.44106733053922653, 'clip_ratio/low_mean': 0.05428210739046335, 'clip_ratio/low_min': 0.05428210739046335, 'clip_ratio/high_mean': 0.0437433160841465, 'clip_ratio/high_max': 0.0437433160841465, 'clip_ratio/region_mean': 0.09802542347460985, 'reward_total_mean': 0.5134477019309998, 'reward_meter_mean': 0.4509848952293396, 'reward_meter_std': 0.30584847927093506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.991193413734436, 'reward_repeat_soft_std': 0.009842580184340477, 'reward_judge_quality_mean': 0.59375, 'reward_judge_quality_std': 0.2775370180606842, 'reward_total_composite_mean': 0.5134477019309998, 'reward_total_composite_std': 0.13883522152900696, 'epoch': 0.17}

 53%|█████▎    | 1741/3300 [4:16:36<3:12:10,  7.40s/it]INFO 04-13 11:37:34 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1742/3300 [4:16:44<3:21:56,  7.78s/it]2026-04-13 11:37:42,463 | INFO | train_grpo_train | metrics_logged mode=train step=1742

                                                       
{'loss': 0.0969, 'grad_norm': 5.192895889282227, 'learning_rate': 4.724242424242424e-06, 'num_tokens': 3103687.0, 'completions/mean_length': 131.875, 'completions/min_length': 98.0, 'completions/max_length': 187.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.875, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 187.0, 'rewards/meter/mean': 0.8247295022010803, 'rewards/meter/std': 0.2682240605354309, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6811105012893677, 'rewards/repeat_soft/std': 0.08826851844787598, 'rewards/judge_quality/mean': 0.3037499785423279, 'rewards/judge_quality/std': 0.13362392783164978, 'rewards/total_composite/mean': 0.4580414593219757, 'rewards/total_composite/std': 0.09204375743865967, 'reward': 0.4580414593219757, 'reward_std': 0.09204376488924026, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06757155805826187, 'sampling/sampling_logp_difference/max': 1.2468342781066895, 'sampling/importance_sampling_ratio/min': 0.28741323947906494, 'sampling/importance_sampling_ratio/mean': 1.0102301836013794, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5221900083124638, 'clip_ratio/low_mean': 0.022508766502141953, 'clip_ratio/low_min': 0.022508766502141953, 'clip_ratio/high_mean': 0.027129985624924302, 'clip_ratio/high_max': 0.027129985624924302, 'clip_ratio/region_mean': 0.049638752127066255, 'reward_total_mean': 0.4580414593219757, 'reward_meter_mean': 0.8247295022010803, 'reward_meter_std': 0.2682240605354309, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6811105012893677, 'reward_repeat_soft_std': 0.08826851844787598, 'reward_judge_quality_mean': 0.3037499785423279, 'reward_judge_quality_std': 0.13362392783164978, 'reward_total_composite_mean': 0.4580414593219757, 'reward_total_composite_std': 0.09204375743865967, 'epoch': 0.17}

 53%|█████▎    | 1742/3300 [4:16:44<3:21:56,  7.78s/it]INFO 04-13 11:37:42 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1743/3300 [4:16:56<3:55:13,  9.06s/it]2026-04-13 11:37:54,485 | INFO | train_grpo_train | metrics_logged mode=train step=1743

                                                       
{'loss': -0.1299, 'grad_norm': 1.7194968461990356, 'learning_rate': 4.721212121212122e-06, 'num_tokens': 3105323.0, 'completions/mean_length': 169.5, 'completions/min_length': 47.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 55.333335876464844, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.8567893505096436, 'rewards/meter/std': 0.3435137867927551, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.934806764125824, 'rewards/repeat_soft/std': 0.044237542897462845, 'rewards/judge_quality/mean': 0.32749998569488525, 'rewards/judge_quality/std': 0.17127670347690582, 'rewards/total_composite/mean': 0.45591509342193604, 'rewards/total_composite/std': 0.28141120076179504, 'reward': 0.45591509342193604, 'reward_std': 0.28141117095947266, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.134735569357872, 'sampling/sampling_logp_difference/max': 1.4726686477661133, 'sampling/importance_sampling_ratio/min': 0.2293127328157425, 'sampling/importance_sampling_ratio/mean': 1.016987919807434, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5698898136615753, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09488785662688315, 'clip_ratio/high_max': 0.09488785662688315, 'clip_ratio/region_mean': 0.09488785662688315, 'reward_total_mean': 0.45591509342193604, 'reward_meter_mean': 0.8567893505096436, 'reward_meter_std': 0.3435137867927551, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.934806764125824, 'reward_repeat_soft_std': 0.044237542897462845, 'reward_judge_quality_mean': 0.32749998569488525, 'reward_judge_quality_std': 0.17127670347690582, 'reward_total_composite_mean': 0.45591509342193604, 'reward_total_composite_std': 0.28141120076179504, 'epoch': 0.18}

 53%|█████▎    | 1743/3300 [4:16:56<3:55:13,  9.06s/it]INFO 04-13 11:37:54 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1744/3300 [4:17:02<3:32:34,  8.20s/it]2026-04-13 11:38:00,662 | INFO | train_grpo_train | metrics_logged mode=train step=1744

                                                       
{'loss': 0.0997, 'grad_norm': 18.683486938476562, 'learning_rate': 4.718181818181818e-06, 'num_tokens': 3106680.0, 'completions/mean_length': 25.625, 'completions/min_length': 18.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.9062577486038208, 'rewards/meter/std': 0.11527059972286224, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9534684419631958, 'rewards/repeat_soft/std': 0.021279437467455864, 'rewards/judge_quality/mean': 0.44624999165534973, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.6056767702102661, 'rewards/total_composite/std': 0.03251388669013977, 'reward': 0.6056767702102661, 'reward_std': 0.032513879239559174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14203035831451416, 'sampling/sampling_logp_difference/max': 1.0328145027160645, 'sampling/importance_sampling_ratio/min': 0.3560035824775696, 'sampling/importance_sampling_ratio/mean': 1.0251483917236328, 'sampling/importance_sampling_ratio/max': 1.7394042015075684, 'entropy': 1.2420547008514404, 'clip_ratio/low_mean': 0.023888888768851757, 'clip_ratio/low_min': 0.023888888768851757, 'clip_ratio/high_mean': 0.09716971544548869, 'clip_ratio/high_max': 0.09716971544548869, 'clip_ratio/region_mean': 0.12105860421434045, 'reward_total_mean': 0.6056767702102661, 'reward_meter_mean': 0.9062577486038208, 'reward_meter_std': 0.11527059972286224, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9534684419631958, 'reward_repeat_soft_std': 0.021279437467455864, 'reward_judge_quality_mean': 0.44624999165534973, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.6056767702102661, 'reward_total_composite_std': 0.03251388669013977, 'epoch': 0.18}

 53%|█████▎    | 1744/3300 [4:17:02<3:32:34,  8.20s/it]INFO 04-13 11:38:00 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1745/3300 [4:17:08<3:13:57,  7.48s/it]2026-04-13 11:38:06,476 | INFO | train_grpo_train | metrics_logged mode=train step=1745

                                                       
{'loss': 0.0503, 'grad_norm': 14.6497802734375, 'learning_rate': 4.715151515151515e-06, 'num_tokens': 3108114.0, 'completions/mean_length': 24.25, 'completions/min_length': 19.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.9924908876419067, 'rewards/meter/std': 0.003887118538841605, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9176196455955505, 'rewards/repeat_soft/std': 0.052994225174188614, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.6182574033737183, 'rewards/total_composite/std': 0.012917676940560341, 'reward': 0.6182574033737183, 'reward_std': 0.012917687185108662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10061626881361008, 'sampling/sampling_logp_difference/max': 1.249585509300232, 'sampling/importance_sampling_ratio/min': 0.28662359714508057, 'sampling/importance_sampling_ratio/mean': 1.0225293636322021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9781959131360054, 'clip_ratio/low_mean': 0.06482456158846617, 'clip_ratio/low_min': 0.06482456158846617, 'clip_ratio/high_mean': 0.030974460765719414, 'clip_ratio/high_max': 0.030974460765719414, 'clip_ratio/region_mean': 0.09579902235418558, 'reward_total_mean': 0.6182574033737183, 'reward_meter_mean': 0.9924908876419067, 'reward_meter_std': 0.003887118538841605, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9176196455955505, 'reward_repeat_soft_std': 0.052994225174188614, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.6182574033737183, 'reward_total_composite_std': 0.012917676940560341, 'epoch': 0.18}

 53%|█████▎    | 1745/3300 [4:17:08<3:13:57,  7.48s/it]INFO 04-13 11:38:06 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1746/3300 [4:17:15<3:11:22,  7.39s/it]2026-04-13 11:38:13,642 | INFO | train_grpo_train | metrics_logged mode=train step=1746

                                                       
{'loss': 0.0244, 'grad_norm': 6.043193340301514, 'learning_rate': 4.7121212121212126e-06, 'num_tokens': 3110838.0, 'completions/mean_length': 126.5, 'completions/min_length': 107.0, 'completions/max_length': 148.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.5, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 148.0, 'rewards/meter/mean': 0.9032549262046814, 'rewards/meter/std': 0.1314523071050644, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7391740083694458, 'rewards/repeat_soft/std': 0.1549014300107956, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.1249857097864151, 'rewards/total_composite/mean': 0.47975093126296997, 'rewards/total_composite/std': 0.0910043939948082, 'reward': 0.47975093126296997, 'reward_std': 0.0910043865442276, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09540028870105743, 'sampling/sampling_logp_difference/max': 1.770749568939209, 'sampling/importance_sampling_ratio/min': 0.17020536959171295, 'sampling/importance_sampling_ratio/mean': 1.003059983253479, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5396188572049141, 'clip_ratio/low_mean': 0.0195392930181697, 'clip_ratio/low_min': 0.0195392930181697, 'clip_ratio/high_mean': 0.05596997123211622, 'clip_ratio/high_max': 0.05596997123211622, 'clip_ratio/region_mean': 0.07550926425028592, 'reward_total_mean': 0.47975093126296997, 'reward_meter_mean': 0.9032549262046814, 'reward_meter_std': 0.1314523071050644, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7391740083694458, 'reward_repeat_soft_std': 0.1549014300107956, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.1249857097864151, 'reward_total_composite_mean': 0.47975093126296997, 'reward_total_composite_std': 0.0910043939948082, 'epoch': 0.18}

 53%|█████▎    | 1746/3300 [4:17:15<3:11:22,  7.39s/it]INFO 04-13 11:38:13 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1747/3300 [4:17:27<3:47:08,  8.78s/it]2026-04-13 11:38:25,653 | INFO | train_grpo_train | metrics_logged mode=train step=1747

                                                       
{'loss': -0.1828, 'grad_norm': 2.840977191925049, 'learning_rate': 4.709090909090909e-06, 'num_tokens': 3113230.0, 'completions/mean_length': 172.0, 'completions/min_length': 103.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 123.42857360839844, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9895341396331787, 'rewards/meter/std': 0.007784141227602959, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.971744179725647, 'rewards/repeat_soft/std': 0.01658444106578827, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.2308950126171112, 'rewards/total_composite/mean': 0.5592749118804932, 'rewards/total_composite/std': 0.2532850205898285, 'reward': 0.5592749118804932, 'reward_std': 0.2532850205898285, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12874051928520203, 'sampling/sampling_logp_difference/max': 2.107656478881836, 'sampling/importance_sampling_ratio/min': 0.12152242660522461, 'sampling/importance_sampling_ratio/mean': 1.0061155557632446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6671543717384338, 'clip_ratio/low_mean': 0.018292502965778112, 'clip_ratio/low_min': 0.018292502965778112, 'clip_ratio/high_mean': 0.0698564164340496, 'clip_ratio/high_max': 0.0698564164340496, 'clip_ratio/region_mean': 0.08814891939982772, 'reward_total_mean': 0.5592749118804932, 'reward_meter_mean': 0.9895341396331787, 'reward_meter_std': 0.007784141227602959, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.971744179725647, 'reward_repeat_soft_std': 0.01658444106578827, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.2308950126171112, 'reward_total_composite_mean': 0.5592749118804932, 'reward_total_composite_std': 0.2532850205898285, 'epoch': 0.18}

 53%|█████▎    | 1747/3300 [4:17:27<3:47:08,  8.78s/it]INFO 04-13 11:38:25 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1748/3300 [4:17:34<3:31:42,  8.18s/it]2026-04-13 11:38:32,457 | INFO | train_grpo_train | metrics_logged mode=train step=1748

                                                       
{'loss': 0.0005, 'grad_norm': 11.052653312683105, 'learning_rate': 4.706060606060606e-06, 'num_tokens': 3114991.0, 'completions/mean_length': 53.125, 'completions/min_length': 50.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.5705060958862305, 'rewards/meter/std': 0.3067757189273834, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8631938099861145, 'rewards/repeat_soft/std': 0.0716160461306572, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4884698987007141, 'rewards/total_composite/std': 0.08246802538633347, 'reward': 0.4884698987007141, 'reward_std': 0.08246801048517227, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1203719899058342, 'sampling/sampling_logp_difference/max': 1.6194086074829102, 'sampling/importance_sampling_ratio/min': 0.19801577925682068, 'sampling/importance_sampling_ratio/mean': 1.0158250331878662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7180077284574509, 'clip_ratio/low_mean': 0.0695974649861455, 'clip_ratio/low_min': 0.0695974649861455, 'clip_ratio/high_mean': 0.03645964711904526, 'clip_ratio/high_max': 0.03645964711904526, 'clip_ratio/region_mean': 0.10605711210519075, 'reward_total_mean': 0.4884698987007141, 'reward_meter_mean': 0.5705060958862305, 'reward_meter_std': 0.3067757189273834, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8631938099861145, 'reward_repeat_soft_std': 0.0716160461306572, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4884698987007141, 'reward_total_composite_std': 0.08246802538633347, 'epoch': 0.18}

 53%|█████▎    | 1748/3300 [4:17:34<3:31:42,  8.18s/it]INFO 04-13 11:38:32 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1749/3300 [4:17:42<3:29:07,  8.09s/it]2026-04-13 11:38:40,330 | INFO | train_grpo_train | metrics_logged mode=train step=1749

                                                       
{'loss': 0.0629, 'grad_norm': 13.392763137817383, 'learning_rate': 4.7030303030303035e-06, 'num_tokens': 3116363.0, 'completions/mean_length': 24.5, 'completions/min_length': 21.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.9797687530517578, 'rewards/meter/std': 0.02011459320783615, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9197984933853149, 'rewards/repeat_soft/std': 0.05981236323714256, 'rewards/judge_quality/mean': 0.39249998331069946, 'rewards/judge_quality/std': 0.08892211318016052, 'rewards/total_composite/mean': 0.5875974297523499, 'rewards/total_composite/std': 0.05380617454648018, 'reward': 0.5875974297523499, 'reward_std': 0.053806155920028687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13312344253063202, 'sampling/sampling_logp_difference/max': 1.723677635192871, 'sampling/importance_sampling_ratio/min': 0.17840883135795593, 'sampling/importance_sampling_ratio/mean': 1.022372841835022, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9063016846776009, 'clip_ratio/low_mean': 0.06807692348957062, 'clip_ratio/low_min': 0.06807692348957062, 'clip_ratio/high_mean': 0.09576756693422794, 'clip_ratio/high_max': 0.09576756693422794, 'clip_ratio/region_mean': 0.16384449042379856, 'reward_total_mean': 0.5875974297523499, 'reward_meter_mean': 0.9797687530517578, 'reward_meter_std': 0.02011459320783615, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9197984933853149, 'reward_repeat_soft_std': 0.05981236323714256, 'reward_judge_quality_mean': 0.39249998331069946, 'reward_judge_quality_std': 0.08892211318016052, 'reward_total_composite_mean': 0.5875974297523499, 'reward_total_composite_std': 0.05380617454648018, 'epoch': 0.18}

 53%|█████▎    | 1749/3300 [4:17:42<3:29:07,  8.09s/it]INFO 04-13 11:38:40 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1750/3300 [4:17:49<3:22:05,  7.82s/it]2026-04-13 11:38:47,529 | INFO | train_grpo_train | metrics_logged mode=train step=1750

                                                       
{'loss': 0.0417, 'grad_norm': 7.338882923126221, 'learning_rate': 4.7e-06, 'num_tokens': 3118599.0, 'completions/mean_length': 98.5, 'completions/min_length': 88.0, 'completions/max_length': 106.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.5, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 106.0, 'rewards/meter/mean': 0.6081505417823792, 'rewards/meter/std': 0.31105008721351624, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8411466479301453, 'rewards/repeat_soft/std': 0.053297799080610275, 'rewards/judge_quality/mean': 0.5575000047683716, 'rewards/judge_quality/std': 0.19955308735370636, 'rewards/total_composite/mean': 0.5617530941963196, 'rewards/total_composite/std': 0.1653565913438797, 'reward': 0.5617530941963196, 'reward_std': 0.1653565764427185, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09892749786376953, 'sampling/sampling_logp_difference/max': 1.4269309043884277, 'sampling/importance_sampling_ratio/min': 0.2400445193052292, 'sampling/importance_sampling_ratio/mean': 1.0135704278945923, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6103395111858845, 'clip_ratio/low_mean': 0.05503624491393566, 'clip_ratio/low_min': 0.05503624491393566, 'clip_ratio/high_mean': 0.04551269207149744, 'clip_ratio/high_max': 0.04551269207149744, 'clip_ratio/region_mean': 0.1005489369854331, 'reward_total_mean': 0.5617530941963196, 'reward_meter_mean': 0.6081505417823792, 'reward_meter_std': 0.31105008721351624, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8411466479301453, 'reward_repeat_soft_std': 0.053297799080610275, 'reward_judge_quality_mean': 0.5575000047683716, 'reward_judge_quality_std': 0.19955308735370636, 'reward_total_composite_mean': 0.5617530941963196, 'reward_total_composite_std': 0.1653565913438797, 'epoch': 0.18}

 53%|█████▎    | 1750/3300 [4:17:49<3:22:05,  7.82s/it]INFO 04-13 11:38:47 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:08<00:34,  4.36s/it]

 30%|███       | 3/10 [00:13<00:32,  4.59s/it]

 40%|████      | 4/10 [00:22<00:36,  6.08s/it]

 50%|█████     | 5/10 [00:26<00:27,  5.59s/it]

 60%|██████    | 6/10 [00:31<00:21,  5.33s/it]

 70%|███████   | 7/10 [00:36<00:15,  5.13s/it]

 80%|████████  | 8/10 [00:40<00:09,  4.74s/it]

 90%|█████████ | 9/10 [00:46<00:05,  5.11s/it]

100%|██████████| 10/10 [00:55<00:00,  6.33s/it]2026-04-13 11:39:47,401 | INFO | train_grpo_train | metrics_logged mode=eval step=1750

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 59.7674, 'eval_samples_per_second': 1.339, 'eval_steps_per_second': 0.167, 'eval_num_tokens': 3118599.0, 'eval_completions/mean_length': 93.05, 'eval_completions/min_length': 36.2, 'eval_completions/max_length': 221.7, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 81.93571472167969, 'eval_completions/min_terminated_length': 36.2, 'eval_completions/max_terminated_length': 143.2, 'eval_rewards/meter/mean': 0.7724583685398102, 'eval_rewards/meter/std': 0.28240539878606796, 'eval_rewards/count_adherence/mean': 0.9633333265781403, 'eval_rewards/count_adherence/std': 0.07533912770450116, 'eval_rewards/hard_gate/mean': 0.9625, 'eval_rewards/hard_gate/std': 0.10606601536273956, 'eval_rewards/repeat_soft/mean': 0.862103009223938, 'eval_rewards/repeat_soft/std': 0.09512052573263645, 'eval_rewards/judge_quality/mean': 0.4583750069141388, 'eval_rewards/judge_quality/std': 0.1511122086085379, 'eval_rewards/total_composite/mean': 0.5376160174608231, 'eval_rewards/total_composite/std': 0.15217913947999478, 'eval_reward': 0.5376160174608231, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.049692163988947866, 'eval_sampling/sampling_logp_difference/max': 0.954959774017334, 'eval_sampling/importance_sampling_ratio/min': 0.3967170387506485, 'eval_sampling/importance_sampling_ratio/mean': 1.0106980443000793, 'eval_sampling/importance_sampling_ratio/max': 1.4370436906814574, 'eval_entropy': 0.5266555488109589, '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': 0.5376160174608231, 'eval_reward_meter_mean': 0.7724583685398102, 'eval_reward_meter_std': 0.28240539878606796, 'eval_reward_count_adherence_mean': 0.9633333265781403, 'eval_reward_count_adherence_std': 0.07533912770450116, 'eval_reward_hard_gate_mean': 0.9625, 'eval_reward_hard_gate_std': 0.10606601536273956, 'eval_reward_repeat_soft_mean': 0.862103009223938, 'eval_reward_repeat_soft_std': 0.09512052573263645, 'eval_reward_judge_quality_mean': 0.4583750069141388, 'eval_reward_judge_quality_std': 0.1511122086085379, 'eval_reward_total_composite_mean': 0.5376160174608231, 'eval_reward_total_composite_std': 0.15217913947999478, 'epoch': 0.18}

 53%|█████▎    | 1750/3300 [4:18:49<3:22:05,  7.82s/it]

100%|██████████| 10/10 [00:55<00:00,  6.33s/it]

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

 53%|█████▎    | 1751/3300 [4:19:01<11:38:08, 27.04s/it]2026-04-13 11:39:59,413 | INFO | train_grpo_train | metrics_logged mode=train step=1751

                                                        
{'loss': 0.095, 'grad_norm': 6.663627624511719, 'learning_rate': 4.696969696969698e-06, 'num_tokens': 3120994.0, 'completions/mean_length': 123.375, 'completions/min_length': 104.0, 'completions/max_length': 169.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 123.375, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 169.0, 'rewards/meter/mean': 0.9890388250350952, 'rewards/meter/std': 0.008543361909687519, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6901615858078003, 'rewards/repeat_soft/std': 0.04421984776854515, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.5125491619110107, 'rewards/total_composite/std': 0.055993881076574326, 'reward': 0.5125491619110107, 'reward_std': 0.055993881076574326, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0809365063905716, 'sampling/sampling_logp_difference/max': 4.661493301391602, 'sampling/importance_sampling_ratio/min': 0.009452336467802525, 'sampling/importance_sampling_ratio/mean': 1.0108307600021362, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41733942180871964, 'clip_ratio/low_mean': 0.019566303584724665, 'clip_ratio/low_min': 0.019566303584724665, 'clip_ratio/high_mean': 0.047922331374138594, 'clip_ratio/high_max': 0.047922331374138594, 'clip_ratio/region_mean': 0.06748863495886326, 'reward_total_mean': 0.5125491619110107, 'reward_meter_mean': 0.9890388250350952, 'reward_meter_std': 0.008543361909687519, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6901615858078003, 'reward_repeat_soft_std': 0.04421984776854515, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.5125491619110107, 'reward_total_composite_std': 0.055993881076574326, 'epoch': 0.18}

 53%|█████▎    | 1751/3300 [4:19:01<11:38:08, 27.04s/it]INFO 04-13 11:39:59 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1752/3300 [4:19:07<8:55:57, 20.77s/it] 2026-04-13 11:40:05,565 | INFO | train_grpo_train | metrics_logged mode=train step=1752

                                                       
{'loss': -0.0441, 'grad_norm': 14.835586547851562, 'learning_rate': 4.693939393939394e-06, 'num_tokens': 3122487.0, 'completions/mean_length': 39.625, 'completions/min_length': 33.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.7192169427871704, 'rewards/meter/std': 0.40678057074546814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9895240068435669, 'rewards/repeat_soft/std': 0.014799713157117367, 'rewards/judge_quality/mean': 0.48625001311302185, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5864670276641846, 'rewards/total_composite/std': 0.17704661190509796, 'reward': 0.5864670276641846, 'reward_std': 0.17704661190509796, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11205822974443436, 'sampling/sampling_logp_difference/max': 1.7428555488586426, 'sampling/importance_sampling_ratio/min': 0.17501990497112274, 'sampling/importance_sampling_ratio/mean': 1.0069360733032227, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5704237520694733, 'clip_ratio/low_mean': 0.036381255835294724, 'clip_ratio/low_min': 0.036381255835294724, 'clip_ratio/high_mean': 0.08526288438588381, 'clip_ratio/high_max': 0.08526288438588381, 'clip_ratio/region_mean': 0.12164414022117853, 'reward_total_mean': 0.5864670276641846, 'reward_meter_mean': 0.7192169427871704, 'reward_meter_std': 0.40678057074546814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9895240068435669, 'reward_repeat_soft_std': 0.014799713157117367, 'reward_judge_quality_mean': 0.48625001311302185, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5864670276641846, 'reward_total_composite_std': 0.17704661190509796, 'epoch': 0.18}

 53%|█████▎    | 1752/3300 [4:19:07<8:55:57, 20.77s/it]INFO 04-13 11:40:05 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1753/3300 [4:19:13<7:00:55, 16.33s/it]2026-04-13 11:40:11,511 | INFO | train_grpo_train | metrics_logged mode=train step=1753

                                                       
{'loss': 0.0608, 'grad_norm': 15.959066390991211, 'learning_rate': 4.690909090909092e-06, 'num_tokens': 3123841.0, 'completions/mean_length': 21.25, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8230654001235962, 'rewards/meter/std': 0.3459331691265106, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9470235109329224, 'rewards/repeat_soft/std': 0.03704577311873436, 'rewards/judge_quality/mean': 0.40625, 'rewards/judge_quality/std': 0.0645727664232254, 'rewards/total_composite/mean': 0.5581273436546326, 'rewards/total_composite/std': 0.0982942134141922, 'reward': 0.5581273436546326, 'reward_std': 0.098294198513031, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14953167736530304, 'sampling/sampling_logp_difference/max': 1.1728765964508057, 'sampling/importance_sampling_ratio/min': 0.3094754219055176, 'sampling/importance_sampling_ratio/mean': 0.9900863170623779, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7450743019580841, 'clip_ratio/low_mean': 0.02890037652105093, 'clip_ratio/low_min': 0.02890037652105093, 'clip_ratio/high_mean': 0.06461187684908509, 'clip_ratio/high_max': 0.06461187684908509, 'clip_ratio/region_mean': 0.09351225337013602, 'reward_total_mean': 0.5581273436546326, 'reward_meter_mean': 0.8230654001235962, 'reward_meter_std': 0.3459331691265106, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9470235109329224, 'reward_repeat_soft_std': 0.03704577311873436, 'reward_judge_quality_mean': 0.40625, 'reward_judge_quality_std': 0.0645727664232254, 'reward_total_composite_mean': 0.5581273436546326, 'reward_total_composite_std': 0.0982942134141922, 'epoch': 0.18}

 53%|█████▎    | 1753/3300 [4:19:13<7:00:55, 16.33s/it]INFO 04-13 11:40:11 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1754/3300 [4:19:25<6:27:20, 15.03s/it]2026-04-13 11:40:23,521 | INFO | train_grpo_train | metrics_logged mode=train step=1754

                                                       
{'loss': -0.1499, 'grad_norm': 2.3400802612304688, 'learning_rate': 4.687878787878788e-06, 'num_tokens': 3125545.0, 'completions/mean_length': 123.0, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 67.42857360839844, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 97.0, 'rewards/meter/mean': 0.32548457384109497, 'rewards/meter/std': 0.15839283168315887, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.15430334210395813, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9319171905517578, 'rewards/repeat_soft/std': 0.0698607787489891, 'rewards/judge_quality/mean': 0.45625001192092896, 'rewards/judge_quality/std': 0.20982562005519867, 'rewards/total_composite/mean': 0.333653062582016, 'rewards/total_composite/std': 0.14164356887340546, 'reward': 0.333653062582016, 'reward_std': 0.14164355397224426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14890842139720917, 'sampling/sampling_logp_difference/max': 1.6867527961730957, 'sampling/importance_sampling_ratio/min': 0.1851196587085724, 'sampling/importance_sampling_ratio/mean': 1.0020707845687866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8648200929164886, 'clip_ratio/low_mean': 0.019736841320991516, 'clip_ratio/low_min': 0.019736841320991516, 'clip_ratio/high_mean': 0.102735442109406, 'clip_ratio/high_max': 0.102735442109406, 'clip_ratio/region_mean': 0.12247228343039751, 'reward_total_mean': 0.333653062582016, 'reward_meter_mean': 0.32548457384109497, 'reward_meter_std': 0.15839283168315887, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.15430334210395813, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9319171905517578, 'reward_repeat_soft_std': 0.0698607787489891, 'reward_judge_quality_mean': 0.45625001192092896, 'reward_judge_quality_std': 0.20982562005519867, 'reward_total_composite_mean': 0.333653062582016, 'reward_total_composite_std': 0.14164356887340546, 'epoch': 0.18}

 53%|█████▎    | 1754/3300 [4:19:25<6:27:20, 15.03s/it]INFO 04-13 11:40:23 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1755/3300 [4:19:37<6:00:24, 14.00s/it]2026-04-13 11:40:35,107 | INFO | train_grpo_train | metrics_logged mode=train step=1755

                                                       
{'loss': -0.18, 'grad_norm': 3.3729248046875, 'learning_rate': 4.684848484848485e-06, 'num_tokens': 3127940.0, 'completions/mean_length': 172.375, 'completions/min_length': 102.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 123.85714721679688, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 146.0, 'rewards/meter/mean': 0.8067696690559387, 'rewards/meter/std': 0.24337105453014374, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7807559370994568, 'rewards/repeat_soft/std': 0.09562687575817108, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.5023723840713501, 'rewards/total_composite/std': 0.25182026624679565, 'reward': 0.5023723840713501, 'reward_std': 0.25182023644447327, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08725365996360779, 'sampling/sampling_logp_difference/max': 1.9357357025146484, 'sampling/importance_sampling_ratio/min': 0.14431805908679962, 'sampling/importance_sampling_ratio/mean': 1.0076463222503662, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48959536850452423, 'clip_ratio/low_mean': 0.02950372127816081, 'clip_ratio/low_min': 0.02950372127816081, 'clip_ratio/high_mean': 0.03864203952252865, 'clip_ratio/high_max': 0.03864203952252865, 'clip_ratio/region_mean': 0.06814576080068946, 'reward_total_mean': 0.5023723840713501, 'reward_meter_mean': 0.8067696690559387, 'reward_meter_std': 0.24337105453014374, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7807559370994568, 'reward_repeat_soft_std': 0.09562687575817108, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.5023723840713501, 'reward_total_composite_std': 0.25182026624679565, 'epoch': 0.18}

 53%|█████▎    | 1755/3300 [4:19:37<6:00:24, 14.00s/it]INFO 04-13 11:40:35 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1756/3300 [4:19:44<5:06:11, 11.90s/it]2026-04-13 11:40:42,104 | INFO | train_grpo_train | metrics_logged mode=train step=1756

                                                       
{'loss': -0.0363, 'grad_norm': 9.370224952697754, 'learning_rate': 4.681818181818183e-06, 'num_tokens': 3129970.0, 'completions/mean_length': 86.75, 'completions/min_length': 71.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.75, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.36734187602996826, 'rewards/meter/std': 0.3477254807949066, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8463608622550964, 'rewards/repeat_soft/std': 0.059881750494241714, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.21954500675201416, 'rewards/total_composite/mean': 0.4815005958080292, 'rewards/total_composite/std': 0.17383673787117004, 'reward': 0.4815005958080292, 'reward_std': 0.17383673787117004, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08453039824962616, 'sampling/sampling_logp_difference/max': 2.1572823524475098, 'sampling/importance_sampling_ratio/min': 0.11563895642757416, 'sampling/importance_sampling_ratio/mean': 1.0083303451538086, 'sampling/importance_sampling_ratio/max': 1.9933003187179565, 'entropy': 0.41204744577407837, 'clip_ratio/low_mean': 0.05621515866369009, 'clip_ratio/low_min': 0.05621515866369009, 'clip_ratio/high_mean': 0.03240876737982035, 'clip_ratio/high_max': 0.03240876737982035, 'clip_ratio/region_mean': 0.08862392604351044, 'reward_total_mean': 0.4815005958080292, 'reward_meter_mean': 0.36734187602996826, 'reward_meter_std': 0.3477254807949066, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8463608622550964, 'reward_repeat_soft_std': 0.059881750494241714, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.21954500675201416, 'reward_total_composite_mean': 0.4815005958080292, 'reward_total_composite_std': 0.17383673787117004, 'epoch': 0.18}

 53%|█████▎    | 1756/3300 [4:19:44<5:06:11, 11.90s/it]INFO 04-13 11:40:42 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1757/3300 [4:19:50<4:19:38, 10.10s/it]2026-04-13 11:40:48,001 | INFO | train_grpo_train | metrics_logged mode=train step=1757

                                                       
{'loss': 0.0843, 'grad_norm': 16.338285446166992, 'learning_rate': 4.678787878787879e-06, 'num_tokens': 3131366.0, 'completions/mean_length': 23.5, 'completions/min_length': 16.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8964908719062805, 'rewards/meter/std': 0.13775919377803802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9565219283103943, 'rewards/repeat_soft/std': 0.016908472403883934, 'rewards/judge_quality/mean': 0.6737500429153442, 'rewards/judge_quality/std': 0.263435423374176, 'rewards/total_composite/mean': 0.7228672504425049, 'rewards/total_composite/std': 0.13182733952999115, 'reward': 0.7228672504425049, 'reward_std': 0.13182730972766876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10597249120473862, 'sampling/sampling_logp_difference/max': 0.9990348815917969, 'sampling/importance_sampling_ratio/min': 0.36823469400405884, 'sampling/importance_sampling_ratio/mean': 1.0141311883926392, 'sampling/importance_sampling_ratio/max': 1.60276460647583, 'entropy': 0.8347270488739014, 'clip_ratio/low_mean': 0.07963878381997347, 'clip_ratio/low_min': 0.07963878381997347, 'clip_ratio/high_mean': 0.03393665188923478, 'clip_ratio/high_max': 0.03393665188923478, 'clip_ratio/region_mean': 0.11357543570920825, 'reward_total_mean': 0.7228672504425049, 'reward_meter_mean': 0.8964908719062805, 'reward_meter_std': 0.13775919377803802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9565219283103943, 'reward_repeat_soft_std': 0.016908472403883934, 'reward_judge_quality_mean': 0.6737500429153442, 'reward_judge_quality_std': 0.263435423374176, 'reward_total_composite_mean': 0.7228672504425049, 'reward_total_composite_std': 0.13182733952999115, 'epoch': 0.18}

 53%|█████▎    | 1757/3300 [4:19:50<4:19:38, 10.10s/it]INFO 04-13 11:40:48 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1758/3300 [4:19:56<3:50:07,  8.95s/it]2026-04-13 11:40:54,287 | INFO | train_grpo_train | metrics_logged mode=train step=1758

                                                       
{'loss': 0.0337, 'grad_norm': 16.53756332397461, 'learning_rate': 4.675757575757576e-06, 'num_tokens': 3132882.0, 'completions/mean_length': 35.5, 'completions/min_length': 28.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.8374530076980591, 'rewards/meter/std': 0.16281481087207794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8819432258605957, 'rewards/repeat_soft/std': 0.10562364757061005, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6346421241760254, 'rewards/total_composite/std': 0.1582137793302536, 'reward': 0.6346421241760254, 'reward_std': 0.1582137942314148, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11359891295433044, 'sampling/sampling_logp_difference/max': 1.0051372051239014, 'sampling/importance_sampling_ratio/min': 0.3659944236278534, 'sampling/importance_sampling_ratio/mean': 1.0092700719833374, 'sampling/importance_sampling_ratio/max': 1.958479642868042, 'entropy': 0.6273365765810013, 'clip_ratio/low_mean': 0.05642253835685551, 'clip_ratio/low_min': 0.05642253835685551, 'clip_ratio/high_mean': 0.020551801659166813, 'clip_ratio/high_max': 0.020551801659166813, 'clip_ratio/region_mean': 0.07697434001602232, 'reward_total_mean': 0.6346421241760254, 'reward_meter_mean': 0.8374530076980591, 'reward_meter_std': 0.16281481087207794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8819432258605957, 'reward_repeat_soft_std': 0.10562364757061005, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6346421241760254, 'reward_total_composite_std': 0.1582137793302536, 'epoch': 0.18}

 53%|█████▎    | 1758/3300 [4:19:56<3:50:07,  8.95s/it]INFO 04-13 11:40:54 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1759/3300 [4:20:04<3:42:09,  8.65s/it]2026-04-13 11:41:02,288 | INFO | train_grpo_train | metrics_logged mode=train step=1759

                                                       
{'loss': 0.0614, 'grad_norm': 5.83065938949585, 'learning_rate': 4.6727272727272735e-06, 'num_tokens': 3135179.0, 'completions/mean_length': 116.125, 'completions/min_length': 100.0, 'completions/max_length': 164.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 116.125, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 164.0, 'rewards/meter/mean': 0.8520817160606384, 'rewards/meter/std': 0.22800517082214355, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7942557334899902, 'rewards/repeat_soft/std': 0.10239209234714508, 'rewards/judge_quality/mean': 0.6024999618530273, 'rewards/judge_quality/std': 0.25217628479003906, 'rewards/total_composite/mean': 0.5905674695968628, 'rewards/total_composite/std': 0.15837174654006958, 'reward': 0.5905674695968628, 'reward_std': 0.15837174654006958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0862930417060852, 'sampling/sampling_logp_difference/max': 4.108579635620117, 'sampling/importance_sampling_ratio/min': 0.01643109694123268, 'sampling/importance_sampling_ratio/mean': 1.0071531534194946, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5726190023124218, 'clip_ratio/low_mean': 0.055389934219419956, 'clip_ratio/low_min': 0.055389934219419956, 'clip_ratio/high_mean': 0.019607843831181526, 'clip_ratio/high_max': 0.019607843831181526, 'clip_ratio/region_mean': 0.07499777805060148, 'reward_total_mean': 0.5905674695968628, 'reward_meter_mean': 0.8520817160606384, 'reward_meter_std': 0.22800517082214355, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7942557334899902, 'reward_repeat_soft_std': 0.10239209234714508, 'reward_judge_quality_mean': 0.6024999618530273, 'reward_judge_quality_std': 0.25217628479003906, 'reward_total_composite_mean': 0.5905674695968628, 'reward_total_composite_std': 0.15837174654006958, 'epoch': 0.18}

 53%|█████▎    | 1759/3300 [4:20:04<3:42:09,  8.65s/it]INFO 04-13 11:41:02 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1760/3300 [4:20:10<3:23:33,  7.93s/it]2026-04-13 11:41:08,485 | INFO | train_grpo_train | metrics_logged mode=train step=1760

                                                       
{'loss': 0.0066, 'grad_norm': 9.593058586120605, 'learning_rate': 4.66969696969697e-06, 'num_tokens': 3136704.0, 'completions/mean_length': 33.625, 'completions/min_length': 31.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9631706476211548, 'rewards/meter/std': 0.012758648954331875, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8936579823493958, 'rewards/repeat_soft/std': 0.032854970544576645, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6759107112884521, 'rewards/total_composite/std': 0.14938969910144806, 'reward': 0.6759107112884521, 'reward_std': 0.14938968420028687, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07251586765050888, 'sampling/sampling_logp_difference/max': 1.6966581344604492, 'sampling/importance_sampling_ratio/min': 0.18329504132270813, 'sampling/importance_sampling_ratio/mean': 0.9984449744224548, 'sampling/importance_sampling_ratio/max': 1.776847243309021, 'entropy': 0.3891554996371269, 'clip_ratio/low_mean': 0.04817867837846279, 'clip_ratio/low_min': 0.04817867837846279, 'clip_ratio/high_mean': 0.040088385343551636, 'clip_ratio/high_max': 0.040088385343551636, 'clip_ratio/region_mean': 0.08826706372201443, 'reward_total_mean': 0.6759107112884521, 'reward_meter_mean': 0.9631706476211548, 'reward_meter_std': 0.012758648954331875, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8936579823493958, 'reward_repeat_soft_std': 0.032854970544576645, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6759107112884521, 'reward_total_composite_std': 0.14938969910144806, 'epoch': 0.18}

 53%|█████▎    | 1760/3300 [4:20:10<3:23:33,  7.93s/it]INFO 04-13 11:41:08 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1761/3300 [4:20:18<3:22:45,  7.90s/it]2026-04-13 11:41:16,332 | INFO | train_grpo_train | metrics_logged mode=train step=1761

                                                       
{'loss': 0.0463, 'grad_norm': 5.957404136657715, 'learning_rate': 4.666666666666667e-06, 'num_tokens': 3139106.0, 'completions/mean_length': 105.25, 'completions/min_length': 96.0, 'completions/max_length': 120.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 105.25, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 120.0, 'rewards/meter/mean': 0.7072615027427673, 'rewards/meter/std': 0.19589649140834808, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7279385328292847, 'rewards/repeat_soft/std': 0.06776217371225357, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6447337865829468, 'rewards/total_composite/std': 0.17704874277114868, 'reward': 0.6447337865829468, 'reward_std': 0.17704874277114868, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0715804472565651, 'sampling/sampling_logp_difference/max': 2.039142370223999, 'sampling/importance_sampling_ratio/min': 0.1301402747631073, 'sampling/importance_sampling_ratio/mean': 1.0107367038726807, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38171785324811935, 'clip_ratio/low_mean': 0.03635486401617527, 'clip_ratio/low_min': 0.03635486401617527, 'clip_ratio/high_mean': 0.040356777142733335, 'clip_ratio/high_max': 0.040356777142733335, 'clip_ratio/region_mean': 0.0767116411589086, 'reward_total_mean': 0.6447337865829468, 'reward_meter_mean': 0.7072615027427673, 'reward_meter_std': 0.19589649140834808, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7279385328292847, 'reward_repeat_soft_std': 0.06776217371225357, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6447337865829468, 'reward_total_composite_std': 0.17704874277114868, 'epoch': 0.18}

 53%|█████▎    | 1761/3300 [4:20:18<3:22:45,  7.90s/it]INFO 04-13 11:41:16 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1762/3300 [4:20:30<3:55:55,  9.20s/it]2026-04-13 11:41:28,569 | INFO | train_grpo_train | metrics_logged mode=train step=1762

                                                       
{'loss': -0.11, 'grad_norm': 2.159605026245117, 'learning_rate': 4.663636363636364e-06, 'num_tokens': 3140556.0, 'completions/mean_length': 95.25, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.71428680419922, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.6219979524612427, 'rewards/meter/std': 0.4170207977294922, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.933821439743042, 'rewards/repeat_soft/std': 0.043045803904533386, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.4652155339717865, 'rewards/total_composite/std': 0.20921353995800018, 'reward': 0.4652155339717865, 'reward_std': 0.209213525056839, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0978228896856308, 'sampling/sampling_logp_difference/max': 1.2196571826934814, 'sampling/importance_sampling_ratio/min': 0.2953313887119293, 'sampling/importance_sampling_ratio/mean': 1.0135092735290527, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5758384056389332, 'clip_ratio/low_mean': 0.016607505502179265, 'clip_ratio/low_min': 0.016607505502179265, 'clip_ratio/high_mean': 0.08218754455447197, 'clip_ratio/high_max': 0.08218754455447197, 'clip_ratio/region_mean': 0.09879505005665123, 'reward_total_mean': 0.4652155339717865, 'reward_meter_mean': 0.6219979524612427, 'reward_meter_std': 0.4170207977294922, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.933821439743042, 'reward_repeat_soft_std': 0.043045803904533386, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.4652155339717865, 'reward_total_composite_std': 0.20921353995800018, 'epoch': 0.18}

 53%|█████▎    | 1762/3300 [4:20:30<3:55:55,  9.20s/it]INFO 04-13 11:41:28 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1763/3300 [4:20:38<3:43:14,  8.71s/it]2026-04-13 11:41:36,140 | INFO | train_grpo_train | metrics_logged mode=train step=1763

                                                       
{'loss': -0.0068, 'grad_norm': 5.532903671264648, 'learning_rate': 4.660606060606061e-06, 'num_tokens': 3142956.0, 'completions/mean_length': 119.0, 'completions/min_length': 103.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.0, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.5499364733695984, 'rewards/meter/std': 0.3466138243675232, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7667504549026489, 'rewards/repeat_soft/std': 0.045128677040338516, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4173145890235901, 'rewards/total_composite/std': 0.09702639281749725, 'reward': 0.4173145890235901, 'reward_std': 0.09702640026807785, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07352930307388306, 'sampling/sampling_logp_difference/max': 1.6493425369262695, 'sampling/importance_sampling_ratio/min': 0.21601814031600952, 'sampling/importance_sampling_ratio/mean': 1.007866382598877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46876227855682373, 'clip_ratio/low_mean': 0.026812857948243618, 'clip_ratio/low_min': 0.026812857948243618, 'clip_ratio/high_mean': 0.05130196502432227, 'clip_ratio/high_max': 0.05130196502432227, 'clip_ratio/region_mean': 0.07811482297256589, 'reward_total_mean': 0.4173145890235901, 'reward_meter_mean': 0.5499364733695984, 'reward_meter_std': 0.3466138243675232, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7667504549026489, 'reward_repeat_soft_std': 0.045128677040338516, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4173145890235901, 'reward_total_composite_std': 0.09702639281749725, 'epoch': 0.18}

 53%|█████▎    | 1763/3300 [4:20:38<3:43:14,  8.71s/it]INFO 04-13 11:41:36 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1764/3300 [4:20:44<3:27:05,  8.09s/it]2026-04-13 11:41:42,768 | INFO | train_grpo_train | metrics_logged mode=train step=1764

                                                       
{'loss': 0.0115, 'grad_norm': 12.079869270324707, 'learning_rate': 4.657575757575758e-06, 'num_tokens': 3144733.0, 'completions/mean_length': 54.125, 'completions/min_length': 47.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.125, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8532311916351318, 'rewards/meter/std': 0.2813609540462494, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9801090955734253, 'rewards/repeat_soft/std': 0.027251332998275757, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.10507651418447495, 'rewards/total_composite/mean': 0.6062156558036804, 'rewards/total_composite/std': 0.11235784739255905, 'reward': 0.6062156558036804, 'reward_std': 0.11235783994197845, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13048695027828217, 'sampling/sampling_logp_difference/max': 1.5955047607421875, 'sampling/importance_sampling_ratio/min': 0.20280614495277405, 'sampling/importance_sampling_ratio/mean': 1.0123000144958496, 'sampling/importance_sampling_ratio/max': 1.858381748199463, 'entropy': 0.8582986369729042, 'clip_ratio/low_mean': 0.048888481222093105, 'clip_ratio/low_min': 0.048888481222093105, 'clip_ratio/high_mean': 0.09009198471903801, 'clip_ratio/high_max': 0.09009198471903801, 'clip_ratio/region_mean': 0.13898046594113111, 'reward_total_mean': 0.6062156558036804, 'reward_meter_mean': 0.8532311916351318, 'reward_meter_std': 0.2813609540462494, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9801090955734253, 'reward_repeat_soft_std': 0.027251332998275757, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.10507651418447495, 'reward_total_composite_mean': 0.6062156558036804, 'reward_total_composite_std': 0.11235784739255905, 'epoch': 0.18}

 53%|█████▎    | 1764/3300 [4:20:44<3:27:05,  8.09s/it]INFO 04-13 11:41:43 [block_pool.py:378] Successfully reset prefix cache

 53%|█████▎    | 1765/3300 [4:20:57<3:57:58,  9.30s/it]2026-04-13 11:41:54,896 | INFO | train_grpo_train | metrics_logged mode=train step=1765

                                                       
{'loss': -0.1219, 'grad_norm': 1.6977033615112305, 'learning_rate': 4.654545454545455e-06, 'num_tokens': 3146297.0, 'completions/mean_length': 234.5, 'completions/min_length': 57.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 68.0, 'completions/min_terminated_length': 57.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.6177030801773071, 'rewards/meter/std': 0.39913487434387207, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.17817415297031403, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.8289181590080261, 'rewards/repeat_soft/std': 0.1014774739742279, 'rewards/judge_quality/mean': 0.25999999046325684, 'rewards/judge_quality/std': 0.18314708769321442, 'rewards/total_composite/mean': 0.2794647812843323, 'rewards/total_composite/std': 0.2390124350786209, 'reward': 0.2794647812843323, 'reward_std': 0.2390124350786209, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09115440398454666, 'sampling/sampling_logp_difference/max': 1.0756086111068726, 'sampling/importance_sampling_ratio/min': 0.3410901129245758, 'sampling/importance_sampling_ratio/mean': 1.0201588869094849, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42579662427306175, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05242840852588415, 'clip_ratio/high_max': 0.05242840852588415, 'clip_ratio/region_mean': 0.05242840852588415, 'reward_total_mean': 0.2794647812843323, 'reward_meter_mean': 0.6177030801773071, 'reward_meter_std': 0.39913487434387207, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.17817415297031403, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.8289181590080261, 'reward_repeat_soft_std': 0.1014774739742279, 'reward_judge_quality_mean': 0.25999999046325684, 'reward_judge_quality_std': 0.18314708769321442, 'reward_total_composite_mean': 0.2794647812843323, 'reward_total_composite_std': 0.2390124350786209, 'epoch': 0.18}

 53%|█████▎    | 1765/3300 [4:20:57<3:57:58,  9.30s/it]INFO 04-13 11:41:55 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1766/3300 [4:21:03<3:36:10,  8.46s/it]2026-04-13 11:42:01,380 | INFO | train_grpo_train | metrics_logged mode=train step=1766

                                                       
{'loss': -0.0363, 'grad_norm': 12.41996955871582, 'learning_rate': 4.651515151515152e-06, 'num_tokens': 3147997.0, 'completions/mean_length': 64.5, 'completions/min_length': 40.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.7062945365905762, 'rewards/meter/std': 0.3823467791080475, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9083278179168701, 'rewards/repeat_soft/std': 0.05187005177140236, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.5370651483535767, 'rewards/total_composite/std': 0.14727263152599335, 'reward': 0.5370651483535767, 'reward_std': 0.14727263152599335, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14444056153297424, 'sampling/sampling_logp_difference/max': 2.5635123252868652, 'sampling/importance_sampling_ratio/min': 0.07703369855880737, 'sampling/importance_sampling_ratio/mean': 0.9915181398391724, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5885419137775898, 'clip_ratio/low_mean': 0.05729159014299512, 'clip_ratio/low_min': 0.05729159014299512, 'clip_ratio/high_mean': 0.05323988106101751, 'clip_ratio/high_max': 0.05323988106101751, 'clip_ratio/region_mean': 0.11053147120401263, 'reward_total_mean': 0.5370651483535767, 'reward_meter_mean': 0.7062945365905762, 'reward_meter_std': 0.3823467791080475, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9083278179168701, 'reward_repeat_soft_std': 0.05187005177140236, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.5370651483535767, 'reward_total_composite_std': 0.14727263152599335, 'epoch': 0.18}

 54%|█████▎    | 1766/3300 [4:21:03<3:36:10,  8.46s/it]INFO 04-13 11:42:01 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1767/3300 [4:21:09<3:17:29,  7.73s/it]2026-04-13 11:42:07,416 | INFO | train_grpo_train | metrics_logged mode=train step=1767

                                                       
{'loss': -0.0102, 'grad_norm': 13.42194938659668, 'learning_rate': 4.648484848484849e-06, 'num_tokens': 3149559.0, 'completions/mean_length': 32.25, 'completions/min_length': 31.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.958257794380188, 'rewards/meter/std': 0.005651597864925861, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7578088045120239, 'rewards/repeat_soft/std': 0.05588805302977562, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.8086973428726196, 'rewards/total_composite/std': 0.1457444578409195, 'reward': 0.8086973428726196, 'reward_std': 0.1457444280385971, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.039556313306093216, 'sampling/sampling_logp_difference/max': 1.269411563873291, 'sampling/importance_sampling_ratio/min': 0.2809969186782837, 'sampling/importance_sampling_ratio/mean': 0.9951732754707336, 'sampling/importance_sampling_ratio/max': 1.454006552696228, 'entropy': 0.1566576324403286, 'clip_ratio/low_mean': 0.008064515888690948, 'clip_ratio/low_min': 0.008064515888690948, 'clip_ratio/high_mean': 0.022741199005395174, 'clip_ratio/high_max': 0.022741199005395174, 'clip_ratio/region_mean': 0.030805714894086123, 'reward_total_mean': 0.8086973428726196, 'reward_meter_mean': 0.958257794380188, 'reward_meter_std': 0.005651597864925861, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7578088045120239, 'reward_repeat_soft_std': 0.05588805302977562, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.8086973428726196, 'reward_total_composite_std': 0.1457444578409195, 'epoch': 0.18}

 54%|█████▎    | 1767/3300 [4:21:09<3:17:29,  7.73s/it]INFO 04-13 11:42:07 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1768/3300 [4:21:15<3:05:34,  7.27s/it]2026-04-13 11:42:13,595 | INFO | train_grpo_train | metrics_logged mode=train step=1768

                                                       
{'loss': 0.0193, 'grad_norm': 12.585282325744629, 'learning_rate': 4.645454545454545e-06, 'num_tokens': 3151470.0, 'completions/mean_length': 48.875, 'completions/min_length': 46.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.875, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8644855618476868, 'rewards/meter/std': 0.3265610635280609, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9649581909179688, 'rewards/repeat_soft/std': 0.03800182044506073, 'rewards/judge_quality/mean': 0.4424999952316284, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5927733778953552, 'rewards/total_composite/std': 0.09315738081932068, 'reward': 0.5927733778953552, 'reward_std': 0.09315737336874008, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12805289030075073, 'sampling/sampling_logp_difference/max': 1.685804843902588, 'sampling/importance_sampling_ratio/min': 0.1852952390909195, 'sampling/importance_sampling_ratio/mean': 1.0057452917099, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7312658205628395, 'clip_ratio/low_mean': 0.0076530613005161285, 'clip_ratio/low_min': 0.0076530613005161285, 'clip_ratio/high_mean': 0.11520470632240176, 'clip_ratio/high_max': 0.11520470632240176, 'clip_ratio/region_mean': 0.12285776762291789, 'reward_total_mean': 0.5927733778953552, 'reward_meter_mean': 0.8644855618476868, 'reward_meter_std': 0.3265610635280609, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9649581909179688, 'reward_repeat_soft_std': 0.03800182044506073, 'reward_judge_quality_mean': 0.4424999952316284, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5927733778953552, 'reward_total_composite_std': 0.09315738081932068, 'epoch': 0.18}

 54%|█████▎    | 1768/3300 [4:21:15<3:05:34,  7.27s/it]INFO 04-13 11:42:13 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1769/3300 [4:21:24<3:13:52,  7.60s/it]2026-04-13 11:42:21,970 | INFO | train_grpo_train | metrics_logged mode=train step=1769

                                                       
{'loss': 0.0357, 'grad_norm': 4.545206069946289, 'learning_rate': 4.642424242424243e-06, 'num_tokens': 3154059.0, 'completions/mean_length': 143.625, 'completions/min_length': 119.0, 'completions/max_length': 167.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.625, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.8274730443954468, 'rewards/meter/std': 0.3358956575393677, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7043778896331787, 'rewards/repeat_soft/std': 0.10080349445343018, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.2834985554218292, 'rewards/total_composite/mean': 0.5214362144470215, 'rewards/total_composite/std': 0.200816810131073, 'reward': 0.5214362144470215, 'reward_std': 0.200816810131073, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07771973311901093, 'sampling/sampling_logp_difference/max': 2.0672922134399414, 'sampling/importance_sampling_ratio/min': 0.12652792036533356, 'sampling/importance_sampling_ratio/mean': 1.002564787864685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.42923061549663544, 'clip_ratio/low_mean': 0.029658550396561623, 'clip_ratio/low_min': 0.029658550396561623, 'clip_ratio/high_mean': 0.0478785103186965, 'clip_ratio/high_max': 0.0478785103186965, 'clip_ratio/region_mean': 0.07753706071525812, 'reward_total_mean': 0.5214362144470215, 'reward_meter_mean': 0.8274730443954468, 'reward_meter_std': 0.3358956575393677, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7043778896331787, 'reward_repeat_soft_std': 0.10080349445343018, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.2834985554218292, 'reward_total_composite_mean': 0.5214362144470215, 'reward_total_composite_std': 0.200816810131073, 'epoch': 0.18}

 54%|█████▎    | 1769/3300 [4:21:24<3:13:52,  7.60s/it]INFO 04-13 11:42:22 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1770/3300 [4:21:31<3:09:42,  7.44s/it]2026-04-13 11:42:29,047 | INFO | train_grpo_train | metrics_logged mode=train step=1770

                                                       
{'loss': 0.086, 'grad_norm': 10.989616394042969, 'learning_rate': 4.63939393939394e-06, 'num_tokens': 3156199.0, 'completions/mean_length': 85.5, 'completions/min_length': 68.0, 'completions/max_length': 99.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.5, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 99.0, 'rewards/meter/mean': 0.9170221090316772, 'rewards/meter/std': 0.15822051465511322, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7349417209625244, 'rewards/repeat_soft/std': 0.14873848855495453, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.49723243713378906, 'rewards/total_composite/std': 0.04901236668229103, 'reward': 0.49723243713378906, 'reward_std': 0.04901236668229103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09723883122205734, 'sampling/sampling_logp_difference/max': 1.4950122833251953, 'sampling/importance_sampling_ratio/min': 0.2242458611726761, 'sampling/importance_sampling_ratio/mean': 1.004080891609192, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6100311912596226, 'clip_ratio/low_mean': 0.029673343524336815, 'clip_ratio/low_min': 0.029673343524336815, 'clip_ratio/high_mean': 0.06719653774052858, 'clip_ratio/high_max': 0.06719653774052858, 'clip_ratio/region_mean': 0.0968698812648654, 'reward_total_mean': 0.49723243713378906, 'reward_meter_mean': 0.9170221090316772, 'reward_meter_std': 0.15822051465511322, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7349417209625244, 'reward_repeat_soft_std': 0.14873848855495453, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.49723243713378906, 'reward_total_composite_std': 0.04901236668229103, 'epoch': 0.18}

 54%|█████▎    | 1770/3300 [4:21:31<3:09:42,  7.44s/it]INFO 04-13 11:42:29 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1771/3300 [4:21:39<3:16:17,  7.70s/it]2026-04-13 11:42:37,354 | INFO | train_grpo_train | metrics_logged mode=train step=1771

                                                       
{'loss': -0.0555, 'grad_norm': 8.31665325164795, 'learning_rate': 4.636363636363636e-06, 'num_tokens': 3159009.0, 'completions/mean_length': 165.25, 'completions/min_length': 124.0, 'completions/max_length': 195.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 165.25, 'completions/min_terminated_length': 124.0, 'completions/max_terminated_length': 195.0, 'rewards/meter/mean': 0.9113659858703613, 'rewards/meter/std': 0.203917995095253, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8120887875556946, 'rewards/repeat_soft/std': 0.09801527857780457, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5306162238121033, 'rewards/total_composite/std': 0.0546850860118866, 'reward': 0.5306162238121033, 'reward_std': 0.054685078561306, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0938008725643158, 'sampling/sampling_logp_difference/max': 2.30631160736084, 'sampling/importance_sampling_ratio/min': 0.0996280387043953, 'sampling/importance_sampling_ratio/mean': 1.0084292888641357, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.559680700302124, 'clip_ratio/low_mean': 0.019789081532508135, 'clip_ratio/low_min': 0.019789081532508135, 'clip_ratio/high_mean': 0.06842813175171614, 'clip_ratio/high_max': 0.06842813175171614, 'clip_ratio/region_mean': 0.08821721328422427, 'reward_total_mean': 0.5306162238121033, 'reward_meter_mean': 0.9113659858703613, 'reward_meter_std': 0.203917995095253, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8120887875556946, 'reward_repeat_soft_std': 0.09801527857780457, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5306162238121033, 'reward_total_composite_std': 0.0546850860118866, 'epoch': 0.18}

 54%|█████▎    | 1771/3300 [4:21:39<3:16:17,  7.70s/it]INFO 04-13 11:42:37 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1772/3300 [4:21:46<3:14:21,  7.63s/it]2026-04-13 11:42:44,827 | INFO | train_grpo_train | metrics_logged mode=train step=1772

                                                       
{'loss': 0.0047, 'grad_norm': 7.811924457550049, 'learning_rate': 4.633333333333334e-06, 'num_tokens': 3161518.0, 'completions/mean_length': 121.625, 'completions/min_length': 102.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 121.625, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.4056808650493622, 'rewards/meter/std': 0.32659587264060974, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.819111704826355, 'rewards/repeat_soft/std': 0.07080240547657013, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4286176264286041, 'rewards/total_composite/std': 0.08320928364992142, 'reward': 0.4286176264286041, 'reward_std': 0.08320929110050201, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1050177738070488, 'sampling/sampling_logp_difference/max': 3.019057273864746, 'sampling/importance_sampling_ratio/min': 0.04884724318981171, 'sampling/importance_sampling_ratio/mean': 1.0052545070648193, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6332890838384628, 'clip_ratio/low_mean': 0.04819859517738223, 'clip_ratio/low_min': 0.04819859517738223, 'clip_ratio/high_mean': 0.03790410188958049, 'clip_ratio/high_max': 0.03790410188958049, 'clip_ratio/region_mean': 0.08610269706696272, 'reward_total_mean': 0.4286176264286041, 'reward_meter_mean': 0.4056808650493622, 'reward_meter_std': 0.32659587264060974, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.819111704826355, 'reward_repeat_soft_std': 0.07080240547657013, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4286176264286041, 'reward_total_composite_std': 0.08320928364992142, 'epoch': 0.18}

 54%|█████▎    | 1772/3300 [4:21:47<3:14:21,  7.63s/it]INFO 04-13 11:42:45 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▎    | 1773/3300 [4:21:54<3:13:10,  7.59s/it]2026-04-13 11:42:52,320 | INFO | train_grpo_train | metrics_logged mode=train step=1773

                                                       
{'loss': 0.0723, 'grad_norm': 6.276249408721924, 'learning_rate': 4.630303030303031e-06, 'num_tokens': 3163854.0, 'completions/mean_length': 120.0, 'completions/min_length': 104.0, 'completions/max_length': 135.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.0, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 135.0, 'rewards/meter/mean': 0.4395008683204651, 'rewards/meter/std': 0.27812936902046204, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7932449579238892, 'rewards/repeat_soft/std': 0.0670030266046524, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.343326210975647, 'rewards/total_composite/std': 0.15900444984436035, 'reward': 0.343326210975647, 'reward_std': 0.15900444984436035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09043148159980774, 'sampling/sampling_logp_difference/max': 1.6271238327026367, 'sampling/importance_sampling_ratio/min': 0.19649390876293182, 'sampling/importance_sampling_ratio/mean': 1.0063484907150269, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6767326816916466, 'clip_ratio/low_mean': 0.045293652918189764, 'clip_ratio/low_min': 0.045293652918189764, 'clip_ratio/high_mean': 0.04355377238243818, 'clip_ratio/high_max': 0.04355377238243818, 'clip_ratio/region_mean': 0.08884742530062795, 'reward_total_mean': 0.343326210975647, 'reward_meter_mean': 0.4395008683204651, 'reward_meter_std': 0.27812936902046204, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7932449579238892, 'reward_repeat_soft_std': 0.0670030266046524, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.343326210975647, 'reward_total_composite_std': 0.15900444984436035, 'epoch': 0.18}

 54%|█████▎    | 1773/3300 [4:21:54<3:13:10,  7.59s/it]INFO 04-13 11:42:52 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1774/3300 [4:22:00<3:03:28,  7.21s/it]2026-04-13 11:42:58,655 | INFO | train_grpo_train | metrics_logged mode=train step=1774

                                                       
{'loss': 0.2022, 'grad_norm': 21.477895736694336, 'learning_rate': 4.627272727272727e-06, 'num_tokens': 3165287.0, 'completions/mean_length': 20.125, 'completions/min_length': 12.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.125, 'completions/min_terminated_length': 12.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.3956363797187805, 'rewards/meter/std': 0.359622985124588, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9515777826309204, 'rewards/repeat_soft/std': 0.02880319207906723, 'rewards/judge_quality/mean': 0.3799999952316284, 'rewards/judge_quality/std': 0.11501552164554596, 'rewards/total_composite/mean': 0.4245898127555847, 'rewards/total_composite/std': 0.13764996826648712, 'reward': 0.4245898127555847, 'reward_std': 0.13764995336532593, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15270955860614777, 'sampling/sampling_logp_difference/max': 1.2393665313720703, 'sampling/importance_sampling_ratio/min': 0.28956758975982666, 'sampling/importance_sampling_ratio/mean': 1.0259755849838257, 'sampling/importance_sampling_ratio/max': 1.914949655532837, 'entropy': 1.2450411096215248, 'clip_ratio/low_mean': 0.09484980208799243, 'clip_ratio/low_min': 0.09484980208799243, 'clip_ratio/high_mean': 0.056746033020317554, 'clip_ratio/high_max': 0.056746033020317554, 'clip_ratio/region_mean': 0.15159583510830998, 'reward_total_mean': 0.4245898127555847, 'reward_meter_mean': 0.3956363797187805, 'reward_meter_std': 0.359622985124588, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9515777826309204, 'reward_repeat_soft_std': 0.02880319207906723, 'reward_judge_quality_mean': 0.3799999952316284, 'reward_judge_quality_std': 0.11501552164554596, 'reward_total_composite_mean': 0.4245898127555847, 'reward_total_composite_std': 0.13764996826648712, 'epoch': 0.18}

 54%|█████▍    | 1774/3300 [4:22:00<3:03:28,  7.21s/it]INFO 04-13 11:42:58 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1775/3300 [4:22:08<3:04:11,  7.25s/it]2026-04-13 11:43:05,982 | INFO | train_grpo_train | metrics_logged mode=train step=1775

                                                       
{'loss': 0.0943, 'grad_norm': 7.26712703704834, 'learning_rate': 4.6242424242424245e-06, 'num_tokens': 3167474.0, 'completions/mean_length': 101.375, 'completions/min_length': 82.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.375, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.5810152888298035, 'rewards/meter/std': 0.2275850921869278, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8646507263183594, 'rewards/repeat_soft/std': 0.06669933348894119, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.42982417345046997, 'rewards/total_composite/std': 0.05832752585411072, 'reward': 0.42982417345046997, 'reward_std': 0.05832751840353012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09182518720626831, 'sampling/sampling_logp_difference/max': 2.051914930343628, 'sampling/importance_sampling_ratio/min': 0.12848861515522003, 'sampling/importance_sampling_ratio/mean': 1.00731360912323, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5825636647641659, 'clip_ratio/low_mean': 0.05231765937060118, 'clip_ratio/low_min': 0.05231765937060118, 'clip_ratio/high_mean': 0.04964388348162174, 'clip_ratio/high_max': 0.04964388348162174, 'clip_ratio/region_mean': 0.10196154285222292, 'reward_total_mean': 0.42982417345046997, 'reward_meter_mean': 0.5810152888298035, 'reward_meter_std': 0.2275850921869278, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8646507263183594, 'reward_repeat_soft_std': 0.06669933348894119, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.42982417345046997, 'reward_total_composite_std': 0.05832752585411072, 'epoch': 0.18}

 54%|█████▍    | 1775/3300 [4:22:08<3:04:11,  7.25s/it]INFO 04-13 11:43:06 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1776/3300 [4:22:14<2:55:12,  6.90s/it]2026-04-13 11:43:12,054 | INFO | train_grpo_train | metrics_logged mode=train step=1776

                                                       
{'loss': 0.0004, 'grad_norm': 19.19764518737793, 'learning_rate': 4.621212121212122e-06, 'num_tokens': 3169086.0, 'completions/mean_length': 22.5, 'completions/min_length': 19.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.5, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.7644567489624023, 'rewards/meter/std': 0.37200382351875305, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.95320725440979, 'rewards/repeat_soft/std': 0.022005343809723854, 'rewards/judge_quality/mean': 0.38874998688697815, 'rewards/judge_quality/std': 0.08675704896450043, 'rewards/total_composite/mean': 0.5397839546203613, 'rewards/total_composite/std': 0.11333271116018295, 'reward': 0.5397839546203613, 'reward_std': 0.11333271116018295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14350047707557678, 'sampling/sampling_logp_difference/max': 1.3546104431152344, 'sampling/importance_sampling_ratio/min': 0.25804781913757324, 'sampling/importance_sampling_ratio/mean': 1.0243251323699951, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0798679664731026, 'clip_ratio/low_mean': 0.02336956560611725, 'clip_ratio/low_min': 0.02336956560611725, 'clip_ratio/high_mean': 0.10425302013754845, 'clip_ratio/high_max': 0.10425302013754845, 'clip_ratio/region_mean': 0.1276225857436657, 'reward_total_mean': 0.5397839546203613, 'reward_meter_mean': 0.7644567489624023, 'reward_meter_std': 0.37200382351875305, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.95320725440979, 'reward_repeat_soft_std': 0.022005343809723854, 'reward_judge_quality_mean': 0.38874998688697815, 'reward_judge_quality_std': 0.08675704896450043, 'reward_total_composite_mean': 0.5397839546203613, 'reward_total_composite_std': 0.11333271116018295, 'epoch': 0.18}

 54%|█████▍    | 1776/3300 [4:22:14<2:55:12,  6.90s/it]INFO 04-13 11:43:12 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1777/3300 [4:22:20<2:48:39,  6.64s/it]2026-04-13 11:43:18,128 | INFO | train_grpo_train | metrics_logged mode=train step=1777

                                                       
{'loss': 0.0508, 'grad_norm': 14.726597785949707, 'learning_rate': 4.618181818181818e-06, 'num_tokens': 3170659.0, 'completions/mean_length': 34.625, 'completions/min_length': 32.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8540171980857849, 'rewards/meter/std': 0.294059693813324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8245512843132019, 'rewards/repeat_soft/std': 0.07131902873516083, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5434920787811279, 'rewards/total_composite/std': 0.0831054225564003, 'reward': 0.5434920787811279, 'reward_std': 0.0831054225564003, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07951752841472626, 'sampling/sampling_logp_difference/max': 1.0437231063842773, 'sampling/importance_sampling_ratio/min': 0.3521411716938019, 'sampling/importance_sampling_ratio/mean': 0.9965921640396118, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4809534139931202, 'clip_ratio/low_mean': 0.023838141933083534, 'clip_ratio/low_min': 0.023838141933083534, 'clip_ratio/high_mean': 0.05456614587455988, 'clip_ratio/high_max': 0.05456614587455988, 'clip_ratio/region_mean': 0.07840428780764341, 'reward_total_mean': 0.5434920787811279, 'reward_meter_mean': 0.8540171980857849, 'reward_meter_std': 0.294059693813324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8245512843132019, 'reward_repeat_soft_std': 0.07131902873516083, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5434920787811279, 'reward_total_composite_std': 0.0831054225564003, 'epoch': 0.18}

 54%|█████▍    | 1777/3300 [4:22:20<2:48:39,  6.64s/it]INFO 04-13 11:43:18 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1778/3300 [4:22:27<2:51:05,  6.74s/it]2026-04-13 11:43:25,110 | INFO | train_grpo_train | metrics_logged mode=train step=1778

                                                       
{'loss': 0.0254, 'grad_norm': 11.550619125366211, 'learning_rate': 4.615151515151515e-06, 'num_tokens': 3172276.0, 'completions/mean_length': 35.125, 'completions/min_length': 33.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9001180529594421, 'rewards/meter/std': 0.1616150140762329, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.76820969581604, 'rewards/repeat_soft/std': 0.07499588280916214, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5609637498855591, 'rewards/total_composite/std': 0.045429080724716187, 'reward': 0.5609637498855591, 'reward_std': 0.04542907327413559, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04459671676158905, 'sampling/sampling_logp_difference/max': 0.7482872009277344, 'sampling/importance_sampling_ratio/min': 0.4731763005256653, 'sampling/importance_sampling_ratio/mean': 0.9990738034248352, 'sampling/importance_sampling_ratio/max': 1.7169886827468872, 'entropy': 0.27748790569603443, 'clip_ratio/low_mean': 0.0033783784601837397, 'clip_ratio/low_min': 0.0033783784601837397, 'clip_ratio/high_mean': 0.043360243551433086, 'clip_ratio/high_max': 0.043360243551433086, 'clip_ratio/region_mean': 0.046738622011616826, 'reward_total_mean': 0.5609637498855591, 'reward_meter_mean': 0.9001180529594421, 'reward_meter_std': 0.1616150140762329, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.76820969581604, 'reward_repeat_soft_std': 0.07499588280916214, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5609637498855591, 'reward_total_composite_std': 0.045429080724716187, 'epoch': 0.18}

 54%|█████▍    | 1778/3300 [4:22:27<2:51:05,  6.74s/it]INFO 04-13 11:43:25 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1779/3300 [4:22:36<3:08:26,  7.43s/it]2026-04-13 11:43:34,142 | INFO | train_grpo_train | metrics_logged mode=train step=1779

                                                       
{'loss': -0.0005, 'grad_norm': 3.2185769081115723, 'learning_rate': 4.612121212121212e-06, 'num_tokens': 3175499.0, 'completions/mean_length': 209.875, 'completions/min_length': 169.0, 'completions/max_length': 240.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 209.875, 'completions/min_terminated_length': 169.0, 'completions/max_terminated_length': 240.0, 'rewards/meter/mean': 0.8809875845909119, 'rewards/meter/std': 0.14933769404888153, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.07715165615081787, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6478923559188843, 'rewards/repeat_soft/std': 0.08954145014286041, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.45875000953674316, 'rewards/total_composite/std': 0.06916475296020508, 'reward': 0.45875000953674316, 'reward_std': 0.06916474550962448, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05661586672067642, 'sampling/sampling_logp_difference/max': 1.8036327362060547, 'sampling/importance_sampling_ratio/min': 0.1646994948387146, 'sampling/importance_sampling_ratio/mean': 1.0127066373825073, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3152757529169321, 'clip_ratio/low_mean': 0.02137986244633794, 'clip_ratio/low_min': 0.02137986244633794, 'clip_ratio/high_mean': 0.025456495000980794, 'clip_ratio/high_max': 0.025456495000980794, 'clip_ratio/region_mean': 0.04683635744731873, 'reward_total_mean': 0.45875000953674316, 'reward_meter_mean': 0.8809875845909119, 'reward_meter_std': 0.14933769404888153, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.07715165615081787, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6478923559188843, 'reward_repeat_soft_std': 0.08954145014286041, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.45875000953674316, 'reward_total_composite_std': 0.06916475296020508, 'epoch': 0.18}

 54%|█████▍    | 1779/3300 [4:22:36<3:08:26,  7.43s/it]INFO 04-13 11:43:34 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1780/3300 [4:22:42<2:59:49,  7.10s/it]2026-04-13 11:43:40,470 | INFO | train_grpo_train | metrics_logged mode=train step=1780

                                                       
{'loss': 0.0483, 'grad_norm': 8.797980308532715, 'learning_rate': 4.60909090909091e-06, 'num_tokens': 3177122.0, 'completions/mean_length': 52.875, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8468008637428284, 'rewards/meter/std': 0.22864817082881927, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9050402045249939, 'rewards/repeat_soft/std': 0.03166838362812996, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.6148871183395386, 'rewards/total_composite/std': 0.12613748013973236, 'reward': 0.6148871183395386, 'reward_std': 0.12613748013973236, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12379765510559082, 'sampling/sampling_logp_difference/max': 1.970367431640625, 'sampling/importance_sampling_ratio/min': 0.13940562307834625, 'sampling/importance_sampling_ratio/mean': 1.0071300268173218, 'sampling/importance_sampling_ratio/max': 1.9179096221923828, 'entropy': 0.7289240285754204, 'clip_ratio/low_mean': 0.08008087985217571, 'clip_ratio/low_min': 0.08008087985217571, 'clip_ratio/high_mean': 0.026988636702299118, 'clip_ratio/high_max': 0.026988636702299118, 'clip_ratio/region_mean': 0.10706951655447483, 'reward_total_mean': 0.6148871183395386, 'reward_meter_mean': 0.8468008637428284, 'reward_meter_std': 0.22864817082881927, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9050402045249939, 'reward_repeat_soft_std': 0.03166838362812996, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.6148871183395386, 'reward_total_composite_std': 0.12613748013973236, 'epoch': 0.18}

 54%|█████▍    | 1780/3300 [4:22:42<2:59:49,  7.10s/it]INFO 04-13 11:43:40 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1781/3300 [4:22:50<3:03:11,  7.24s/it]2026-04-13 11:43:48,027 | INFO | train_grpo_train | metrics_logged mode=train step=1781

                                                       
{'loss': 0.0441, 'grad_norm': 7.625504493713379, 'learning_rate': 4.606060606060606e-06, 'num_tokens': 3179254.0, 'completions/mean_length': 95.5, 'completions/min_length': 78.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.5, 'completions/min_terminated_length': 78.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.6000702381134033, 'rewards/meter/std': 0.3523312211036682, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.1178511381149292, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7699798941612244, 'rewards/repeat_soft/std': 0.07029587030410767, 'rewards/judge_quality/mean': 0.5487500429153442, 'rewards/judge_quality/std': 0.22937415540218353, 'rewards/total_composite/mean': 0.43533459305763245, 'rewards/total_composite/std': 0.11670365929603577, 'reward': 0.43533459305763245, 'reward_std': 0.11670365929603577, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09412666410207748, 'sampling/sampling_logp_difference/max': 1.4652800559997559, 'sampling/importance_sampling_ratio/min': 0.23101328313350677, 'sampling/importance_sampling_ratio/mean': 0.9994523525238037, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5815066248178482, 'clip_ratio/low_mean': 0.028647522209212184, 'clip_ratio/low_min': 0.028647522209212184, 'clip_ratio/high_mean': 0.05417865049093962, 'clip_ratio/high_max': 0.05417865049093962, 'clip_ratio/region_mean': 0.0828261727001518, 'reward_total_mean': 0.43533459305763245, 'reward_meter_mean': 0.6000702381134033, 'reward_meter_std': 0.3523312211036682, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.1178511381149292, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7699798941612244, 'reward_repeat_soft_std': 0.07029587030410767, 'reward_judge_quality_mean': 0.5487500429153442, 'reward_judge_quality_std': 0.22937415540218353, 'reward_total_composite_mean': 0.43533459305763245, 'reward_total_composite_std': 0.11670365929603577, 'epoch': 0.18}

 54%|█████▍    | 1781/3300 [4:22:50<3:03:11,  7.24s/it]INFO 04-13 11:43:48 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1782/3300 [4:22:57<3:00:46,  7.15s/it]2026-04-13 11:43:54,958 | INFO | train_grpo_train | metrics_logged mode=train step=1782

                                                       
{'loss': 0.0737, 'grad_norm': 16.86492347717285, 'learning_rate': 4.603030303030304e-06, 'num_tokens': 3180719.0, 'completions/mean_length': 23.125, 'completions/min_length': 17.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6610732674598694, 'rewards/meter/std': 0.45391368865966797, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9396507740020752, 'rewards/repeat_soft/std': 0.04000112786889076, 'rewards/judge_quality/mean': 0.26749998331069946, 'rewards/judge_quality/std': 0.10375107079744339, 'rewards/total_composite/mean': 0.46557289361953735, 'rewards/total_composite/std': 0.10590554028749466, 'reward': 0.46557289361953735, 'reward_std': 0.10590554773807526, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19363300502300262, 'sampling/sampling_logp_difference/max': 1.2690849304199219, 'sampling/importance_sampling_ratio/min': 0.2810887396335602, 'sampling/importance_sampling_ratio/mean': 1.028009295463562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.6260251477360725, 'clip_ratio/low_mean': 0.11799331195652485, 'clip_ratio/low_min': 0.11799331195652485, 'clip_ratio/high_mean': 0.0705706044100225, 'clip_ratio/high_max': 0.0705706044100225, 'clip_ratio/region_mean': 0.18856391636654735, 'reward_total_mean': 0.46557289361953735, 'reward_meter_mean': 0.6610732674598694, 'reward_meter_std': 0.45391368865966797, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9396507740020752, 'reward_repeat_soft_std': 0.04000112786889076, 'reward_judge_quality_mean': 0.26749998331069946, 'reward_judge_quality_std': 0.10375107079744339, 'reward_total_composite_mean': 0.46557289361953735, 'reward_total_composite_std': 0.10590554028749466, 'epoch': 0.18}

 54%|█████▍    | 1782/3300 [4:22:57<3:00:46,  7.15s/it]INFO 04-13 11:43:55 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1783/3300 [4:23:03<2:57:59,  7.04s/it]2026-04-13 11:44:01,760 | INFO | train_grpo_train | metrics_logged mode=train step=1783

                                                       
{'loss': 0.0934, 'grad_norm': 7.992717742919922, 'learning_rate': 4.600000000000001e-06, 'num_tokens': 3182501.0, 'completions/mean_length': 57.75, 'completions/min_length': 39.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.75, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.2404651790857315, 'rewards/meter/std': 0.21960659325122833, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8838211297988892, 'rewards/repeat_soft/std': 0.044775135815143585, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.3238157331943512, 'rewards/total_composite/std': 0.06840462237596512, 'reward': 0.3238157331943512, 'reward_std': 0.06840461492538452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10930705070495605, 'sampling/sampling_logp_difference/max': 3.122370958328247, 'sampling/importance_sampling_ratio/min': 0.044052597135305405, 'sampling/importance_sampling_ratio/mean': 1.0260294675827026, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7410463839769363, 'clip_ratio/low_mean': 0.02844047429971397, 'clip_ratio/low_min': 0.02844047429971397, 'clip_ratio/high_mean': 0.05527160316705704, 'clip_ratio/high_max': 0.05527160316705704, 'clip_ratio/region_mean': 0.083712077466771, 'reward_total_mean': 0.3238157331943512, 'reward_meter_mean': 0.2404651790857315, 'reward_meter_std': 0.21960659325122833, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8838211297988892, 'reward_repeat_soft_std': 0.044775135815143585, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.3238157331943512, 'reward_total_composite_std': 0.06840462237596512, 'epoch': 0.18}

 54%|█████▍    | 1783/3300 [4:23:03<2:57:59,  7.04s/it]INFO 04-13 11:44:02 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1784/3300 [4:23:10<2:55:41,  6.95s/it]2026-04-13 11:44:08,494 | INFO | train_grpo_train | metrics_logged mode=train step=1784

                                                       
{'loss': 0.0604, 'grad_norm': 11.344841957092285, 'learning_rate': 4.596969696969697e-06, 'num_tokens': 3184251.0, 'completions/mean_length': 53.75, 'completions/min_length': 41.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.75, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.9381598234176636, 'rewards/meter/std': 0.13511887192726135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9147299528121948, 'rewards/repeat_soft/std': 0.05088740214705467, 'rewards/judge_quality/mean': 0.4724999964237213, 'rewards/judge_quality/std': 0.1011011004447937, 'rewards/total_composite/mean': 0.6177034378051758, 'rewards/total_composite/std': 0.012627576477825642, 'reward': 0.6177034378051758, 'reward_std': 0.012627576477825642, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11097406595945358, 'sampling/sampling_logp_difference/max': 1.2670319080352783, 'sampling/importance_sampling_ratio/min': 0.2816663980484009, 'sampling/importance_sampling_ratio/mean': 1.0102325677871704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.744365319609642, 'clip_ratio/low_mean': 0.04959837393835187, 'clip_ratio/low_min': 0.04959837393835187, 'clip_ratio/high_mean': 0.05398035328835249, 'clip_ratio/high_max': 0.05398035328835249, 'clip_ratio/region_mean': 0.10357872722670436, 'reward_total_mean': 0.6177034378051758, 'reward_meter_mean': 0.9381598234176636, 'reward_meter_std': 0.13511887192726135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9147299528121948, 'reward_repeat_soft_std': 0.05088740214705467, 'reward_judge_quality_mean': 0.4724999964237213, 'reward_judge_quality_std': 0.1011011004447937, 'reward_total_composite_mean': 0.6177034378051758, 'reward_total_composite_std': 0.012627576477825642, 'epoch': 0.18}

 54%|█████▍    | 1784/3300 [4:23:10<2:55:41,  6.95s/it]INFO 04-13 11:44:08 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1785/3300 [4:23:18<3:04:22,  7.30s/it]2026-04-13 11:44:16,625 | INFO | train_grpo_train | metrics_logged mode=train step=1785

                                                       
{'loss': 0.0597, 'grad_norm': 8.982309341430664, 'learning_rate': 4.5939393939393945e-06, 'num_tokens': 3187298.0, 'completions/mean_length': 188.875, 'completions/min_length': 170.0, 'completions/max_length': 206.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 188.875, 'completions/min_terminated_length': 170.0, 'completions/max_terminated_length': 206.0, 'rewards/meter/mean': 0.9501059055328369, 'rewards/meter/std': 0.0778055265545845, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7613427639007568, 'rewards/repeat_soft/std': 0.031513430178165436, 'rewards/judge_quality/mean': 0.2887499928474426, 'rewards/judge_quality/std': 0.11630470305681229, 'rewards/total_composite/mean': 0.45387911796569824, 'rewards/total_composite/std': 0.07856108993291855, 'reward': 0.45387911796569824, 'reward_std': 0.07856109738349915, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07520117610692978, 'sampling/sampling_logp_difference/max': 1.9786715507507324, 'sampling/importance_sampling_ratio/min': 0.13825276494026184, 'sampling/importance_sampling_ratio/mean': 1.0020201206207275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49315447732806206, 'clip_ratio/low_mean': 0.045178765431046486, 'clip_ratio/low_min': 0.045178765431046486, 'clip_ratio/high_mean': 0.024684758856892586, 'clip_ratio/high_max': 0.024684758856892586, 'clip_ratio/region_mean': 0.06986352428793907, 'reward_total_mean': 0.45387911796569824, 'reward_meter_mean': 0.9501059055328369, 'reward_meter_std': 0.0778055265545845, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7613427639007568, 'reward_repeat_soft_std': 0.031513430178165436, 'reward_judge_quality_mean': 0.2887499928474426, 'reward_judge_quality_std': 0.11630470305681229, 'reward_total_composite_mean': 0.45387911796569824, 'reward_total_composite_std': 0.07856108993291855, 'epoch': 0.18}

 54%|█████▍    | 1785/3300 [4:23:18<3:04:22,  7.30s/it]INFO 04-13 11:44:16 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1786/3300 [4:23:26<3:04:16,  7.30s/it]2026-04-13 11:44:23,915 | INFO | train_grpo_train | metrics_logged mode=train step=1786

                                                       
{'loss': 0.2066, 'grad_norm': 12.4069242477417, 'learning_rate': 4.590909090909092e-06, 'num_tokens': 3189002.0, 'completions/mean_length': 58.0, 'completions/min_length': 44.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.0, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.8881990909576416, 'rewards/meter/std': 0.18764106929302216, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9975974559783936, 'rewards/repeat_soft/std': 0.002893527038395405, 'rewards/judge_quality/mean': 0.6775000095367432, 'rewards/judge_quality/std': 0.2271406203508377, 'rewards/total_composite/mean': 0.7222679853439331, 'rewards/total_composite/std': 0.16378124058246613, 'reward': 0.7222679853439331, 'reward_std': 0.16378121078014374, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15097936987876892, 'sampling/sampling_logp_difference/max': 1.3733134269714355, 'sampling/importance_sampling_ratio/min': 0.2532663643360138, 'sampling/importance_sampling_ratio/mean': 0.993457019329071, 'sampling/importance_sampling_ratio/max': 1.794889211654663, 'entropy': 1.1973204538226128, 'clip_ratio/low_mean': 0.049005559645593166, 'clip_ratio/low_min': 0.049005559645593166, 'clip_ratio/high_mean': 0.09182675927877426, 'clip_ratio/high_max': 0.09182675927877426, 'clip_ratio/region_mean': 0.14083231892436743, 'reward_total_mean': 0.7222679853439331, 'reward_meter_mean': 0.8881990909576416, 'reward_meter_std': 0.18764106929302216, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9975974559783936, 'reward_repeat_soft_std': 0.002893527038395405, 'reward_judge_quality_mean': 0.6775000095367432, 'reward_judge_quality_std': 0.2271406203508377, 'reward_total_composite_mean': 0.7222679853439331, 'reward_total_composite_std': 0.16378124058246613, 'epoch': 0.18}

 54%|█████▍    | 1786/3300 [4:23:26<3:04:16,  7.30s/it]INFO 04-13 11:44:24 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1787/3300 [4:23:33<3:06:11,  7.38s/it]2026-04-13 11:44:31,488 | INFO | train_grpo_train | metrics_logged mode=train step=1787

                                                       
{'loss': 0.0232, 'grad_norm': 5.080633163452148, 'learning_rate': 4.587878787878788e-06, 'num_tokens': 3191384.0, 'completions/mean_length': 117.75, 'completions/min_length': 113.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.75, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.8378202319145203, 'rewards/meter/std': 0.16670092940330505, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6724264621734619, 'rewards/repeat_soft/std': 0.07522182166576385, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.4547361731529236, 'rewards/total_composite/std': 0.06651253998279572, 'reward': 0.4547361731529236, 'reward_std': 0.06651252508163452, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.061867937445640564, 'sampling/sampling_logp_difference/max': 2.2654361724853516, 'sampling/importance_sampling_ratio/min': 0.10378475487232208, 'sampling/importance_sampling_ratio/mean': 0.9977706670761108, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29621799662709236, 'clip_ratio/low_mean': 0.0259244404733181, 'clip_ratio/low_min': 0.0259244404733181, 'clip_ratio/high_mean': 0.02461230894550681, 'clip_ratio/high_max': 0.02461230894550681, 'clip_ratio/region_mean': 0.05053674941882491, 'reward_total_mean': 0.4547361731529236, 'reward_meter_mean': 0.8378202319145203, 'reward_meter_std': 0.16670092940330505, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6724264621734619, 'reward_repeat_soft_std': 0.07522182166576385, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.4547361731529236, 'reward_total_composite_std': 0.06651253998279572, 'epoch': 0.18}

 54%|█████▍    | 1787/3300 [4:23:33<3:06:11,  7.38s/it]INFO 04-13 11:44:31 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1788/3300 [4:23:42<3:16:05,  7.78s/it]2026-04-13 11:44:40,192 | INFO | train_grpo_train | metrics_logged mode=train step=1788

                                                       
{'loss': 0.0396, 'grad_norm': 5.669865608215332, 'learning_rate': 4.5848484848484854e-06, 'num_tokens': 3194393.0, 'completions/mean_length': 180.125, 'completions/min_length': 149.0, 'completions/max_length': 213.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 180.125, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 213.0, 'rewards/meter/mean': 0.9179707765579224, 'rewards/meter/std': 0.1364118903875351, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7013819813728333, 'rewards/repeat_soft/std': 0.09097636491060257, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.10966669768095016, 'rewards/total_composite/mean': 0.42079004645347595, 'rewards/total_composite/std': 0.07772660255432129, 'reward': 0.42079004645347595, 'reward_std': 0.07772661000490189, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0762377679347992, 'sampling/sampling_logp_difference/max': 2.0151617527008057, 'sampling/importance_sampling_ratio/min': 0.1332988440990448, 'sampling/importance_sampling_ratio/mean': 1.004361629486084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45144612342119217, 'clip_ratio/low_mean': 0.0337425353936851, 'clip_ratio/low_min': 0.0337425353936851, 'clip_ratio/high_mean': 0.037409089505672455, 'clip_ratio/high_max': 0.037409089505672455, 'clip_ratio/region_mean': 0.07115162489935756, 'reward_total_mean': 0.42079004645347595, 'reward_meter_mean': 0.9179707765579224, 'reward_meter_std': 0.1364118903875351, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7013819813728333, 'reward_repeat_soft_std': 0.09097636491060257, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.10966669768095016, 'reward_total_composite_mean': 0.42079004645347595, 'reward_total_composite_std': 0.07772660255432129, 'epoch': 0.18}

 54%|█████▍    | 1788/3300 [4:23:42<3:16:05,  7.78s/it]INFO 04-13 11:44:40 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1789/3300 [4:23:49<3:13:55,  7.70s/it]2026-04-13 11:44:47,702 | INFO | train_grpo_train | metrics_logged mode=train step=1789

                                                       
{'loss': -0.0047, 'grad_norm': 4.517831325531006, 'learning_rate': 4.581818181818183e-06, 'num_tokens': 3196695.0, 'completions/mean_length': 112.75, 'completions/min_length': 105.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.75, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.7732725143432617, 'rewards/meter/std': 0.24383077025413513, 'rewards/count_adherence/mean': 0.53125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.628362238407135, 'rewards/repeat_soft/std': 0.0992504432797432, 'rewards/judge_quality/mean': 0.35249999165534973, 'rewards/judge_quality/std': 0.1249857097864151, 'rewards/total_composite/mean': 0.3709080219268799, 'rewards/total_composite/std': 0.0777900367975235, 'reward': 0.3709080219268799, 'reward_std': 0.0777900367975235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06778769195079803, 'sampling/sampling_logp_difference/max': 2.028944492340088, 'sampling/importance_sampling_ratio/min': 0.13147422671318054, 'sampling/importance_sampling_ratio/mean': 0.9983298182487488, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34576996602118015, 'clip_ratio/low_mean': 0.028252380900084972, 'clip_ratio/low_min': 0.028252380900084972, 'clip_ratio/high_mean': 0.041362200397998095, 'clip_ratio/high_max': 0.041362200397998095, 'clip_ratio/region_mean': 0.06961458129808307, 'reward_total_mean': 0.3709080219268799, 'reward_meter_mean': 0.7732725143432617, 'reward_meter_std': 0.24383077025413513, 'reward_count_adherence_mean': 0.53125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.628362238407135, 'reward_repeat_soft_std': 0.0992504432797432, 'reward_judge_quality_mean': 0.35249999165534973, 'reward_judge_quality_std': 0.1249857097864151, 'reward_total_composite_mean': 0.3709080219268799, 'reward_total_composite_std': 0.0777900367975235, 'epoch': 0.18}

 54%|█████▍    | 1789/3300 [4:23:49<3:13:55,  7.70s/it]INFO 04-13 11:44:47 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1790/3300 [4:23:55<3:00:22,  7.17s/it]2026-04-13 11:44:53,627 | INFO | train_grpo_train | metrics_logged mode=train step=1790

                                                       
{'loss': 0.0722, 'grad_norm': 11.943293571472168, 'learning_rate': 4.578787878787879e-06, 'num_tokens': 3198182.0, 'completions/mean_length': 26.875, 'completions/min_length': 24.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.875, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.590266227722168, 'rewards/meter/std': 0.448323518037796, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9335227012634277, 'rewards/repeat_soft/std': 0.04519963264465332, 'rewards/judge_quality/mean': 0.6225000023841858, 'rewards/judge_quality/std': 0.24656209349632263, 'rewards/total_composite/mean': 0.5792888402938843, 'rewards/total_composite/std': 0.19423989951610565, 'reward': 0.5792888402938843, 'reward_std': 0.19423988461494446, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13793045282363892, 'sampling/sampling_logp_difference/max': 1.7279443740844727, 'sampling/importance_sampling_ratio/min': 0.17764921486377716, 'sampling/importance_sampling_ratio/mean': 1.0088661909103394, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9380409270524979, 'clip_ratio/low_mean': 0.04325311351567507, 'clip_ratio/low_min': 0.04325311351567507, 'clip_ratio/high_mean': 0.07876068446785212, 'clip_ratio/high_max': 0.07876068446785212, 'clip_ratio/region_mean': 0.12201379798352718, 'reward_total_mean': 0.5792888402938843, 'reward_meter_mean': 0.590266227722168, 'reward_meter_std': 0.448323518037796, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9335227012634277, 'reward_repeat_soft_std': 0.04519963264465332, 'reward_judge_quality_mean': 0.6225000023841858, 'reward_judge_quality_std': 0.24656209349632263, 'reward_total_composite_mean': 0.5792888402938843, 'reward_total_composite_std': 0.19423989951610565, 'epoch': 0.18}

 54%|█████▍    | 1790/3300 [4:23:55<3:00:22,  7.17s/it]INFO 04-13 11:44:53 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1791/3300 [4:24:04<3:15:39,  7.78s/it]2026-04-13 11:45:02,837 | INFO | train_grpo_train | metrics_logged mode=train step=1791

                                                       
{'loss': -0.0217, 'grad_norm': 4.378878593444824, 'learning_rate': 4.575757575757576e-06, 'num_tokens': 3201323.0, 'completions/mean_length': 183.625, 'completions/min_length': 142.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 183.625, 'completions/min_terminated_length': 142.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.9574471712112427, 'rewards/meter/std': 0.07274672389030457, 'rewards/count_adherence/mean': 0.6000000238418579, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7691037654876709, 'rewards/repeat_soft/std': 0.11240530014038086, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.469510018825531, 'rewards/total_composite/std': 0.05962883681058884, 'reward': 0.469510018825531, 'reward_std': 0.05962882936000824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08641237020492554, 'sampling/sampling_logp_difference/max': 1.7538394927978516, 'sampling/importance_sampling_ratio/min': 0.25902318954467773, 'sampling/importance_sampling_ratio/mean': 1.0056226253509521, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5196529664099216, 'clip_ratio/low_mean': 0.024049400817602873, 'clip_ratio/low_min': 0.024049400817602873, 'clip_ratio/high_mean': 0.06004335265606642, 'clip_ratio/high_max': 0.06004335265606642, 'clip_ratio/region_mean': 0.08409275347366929, 'reward_total_mean': 0.469510018825531, 'reward_meter_mean': 0.9574471712112427, 'reward_meter_std': 0.07274672389030457, 'reward_count_adherence_mean': 0.6000000238418579, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7691037654876709, 'reward_repeat_soft_std': 0.11240530014038086, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.469510018825531, 'reward_total_composite_std': 0.05962883681058884, 'epoch': 0.18}

 54%|█████▍    | 1791/3300 [4:24:05<3:15:39,  7.78s/it]INFO 04-13 11:45:03 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1792/3300 [4:24:12<3:10:26,  7.58s/it]2026-04-13 11:45:09,942 | INFO | train_grpo_train | metrics_logged mode=train step=1792

                                                       
{'loss': -0.09, 'grad_norm': 8.902054786682129, 'learning_rate': 4.572727272727273e-06, 'num_tokens': 3203121.0, 'completions/mean_length': 62.75, 'completions/min_length': 42.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 62.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7635539174079895, 'rewards/meter/std': 0.3895902633666992, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8927844166755676, 'rewards/repeat_soft/std': 0.032342758029699326, 'rewards/judge_quality/mean': 0.5275000333786011, 'rewards/judge_quality/std': 0.2499571591615677, 'rewards/total_composite/mean': 0.6104133129119873, 'rewards/total_composite/std': 0.21616396307945251, 'reward': 0.6104133129119873, 'reward_std': 0.21616393327713013, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11459516733884811, 'sampling/sampling_logp_difference/max': 1.3093681335449219, 'sampling/importance_sampling_ratio/min': 0.26999059319496155, 'sampling/importance_sampling_ratio/mean': 1.02434241771698, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0145499035716057, 'clip_ratio/low_mean': 0.0704440912231803, 'clip_ratio/low_min': 0.0704440912231803, 'clip_ratio/high_mean': 0.025892728939652443, 'clip_ratio/high_max': 0.025892728939652443, 'clip_ratio/region_mean': 0.09633682016283274, 'reward_total_mean': 0.6104133129119873, 'reward_meter_mean': 0.7635539174079895, 'reward_meter_std': 0.3895902633666992, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8927844166755676, 'reward_repeat_soft_std': 0.032342758029699326, 'reward_judge_quality_mean': 0.5275000333786011, 'reward_judge_quality_std': 0.2499571591615677, 'reward_total_composite_mean': 0.6104133129119873, 'reward_total_composite_std': 0.21616396307945251, 'epoch': 0.18}

 54%|█████▍    | 1792/3300 [4:24:12<3:10:26,  7.58s/it]INFO 04-13 11:45:10 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1793/3300 [4:24:18<3:04:43,  7.35s/it]2026-04-13 11:45:16,776 | INFO | train_grpo_train | metrics_logged mode=train step=1793

                                                       
{'loss': 0.0053, 'grad_norm': 12.37386703491211, 'learning_rate': 4.56969696969697e-06, 'num_tokens': 3204599.0, 'completions/mean_length': 34.75, 'completions/min_length': 32.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.964697539806366, 'rewards/meter/std': 0.003814697964116931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8172740936279297, 'rewards/repeat_soft/std': 0.10937722027301788, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6250307559967041, 'rewards/total_composite/std': 0.12092455476522446, 'reward': 0.6250307559967041, 'reward_std': 0.12092454731464386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06922697275876999, 'sampling/sampling_logp_difference/max': 1.4014520645141602, 'sampling/importance_sampling_ratio/min': 0.24623914062976837, 'sampling/importance_sampling_ratio/mean': 1.0026189088821411, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35333433747291565, 'clip_ratio/low_mean': 0.06949255149811506, 'clip_ratio/low_min': 0.06949255149811506, 'clip_ratio/high_mean': 0.013888888992369175, 'clip_ratio/high_max': 0.013888888992369175, 'clip_ratio/region_mean': 0.08338144049048424, 'reward_total_mean': 0.6250307559967041, 'reward_meter_mean': 0.964697539806366, 'reward_meter_std': 0.003814697964116931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8172740936279297, 'reward_repeat_soft_std': 0.10937722027301788, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6250307559967041, 'reward_total_composite_std': 0.12092455476522446, 'epoch': 0.18}

 54%|█████▍    | 1793/3300 [4:24:19<3:04:43,  7.35s/it]INFO 04-13 11:45:17 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1794/3300 [4:24:26<3:03:12,  7.30s/it]2026-04-13 11:45:23,947 | INFO | train_grpo_train | metrics_logged mode=train step=1794

                                                       
{'loss': 0.042, 'grad_norm': 8.380064010620117, 'learning_rate': 4.566666666666667e-06, 'num_tokens': 3206257.0, 'completions/mean_length': 47.25, 'completions/min_length': 40.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.9883179068565369, 'rewards/meter/std': 0.0074955374002456665, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8761883974075317, 'rewards/repeat_soft/std': 0.03360023722052574, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.6012390851974487, 'rewards/total_composite/std': 0.0041947257705032825, 'reward': 0.6012390851974487, 'reward_std': 0.004194718785583973, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10523198544979095, 'sampling/sampling_logp_difference/max': 1.8004121780395508, 'sampling/importance_sampling_ratio/min': 0.16523078083992004, 'sampling/importance_sampling_ratio/mean': 1.0199353694915771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7152277380228043, 'clip_ratio/low_mean': 0.04757572757080197, 'clip_ratio/low_min': 0.04757572757080197, 'clip_ratio/high_mean': 0.040652749594300985, 'clip_ratio/high_max': 0.040652749594300985, 'clip_ratio/region_mean': 0.08822847716510296, 'reward_total_mean': 0.6012390851974487, 'reward_meter_mean': 0.9883179068565369, 'reward_meter_std': 0.0074955374002456665, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8761883974075317, 'reward_repeat_soft_std': 0.03360023722052574, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.6012390851974487, 'reward_total_composite_std': 0.0041947257705032825, 'epoch': 0.18}

 54%|█████▍    | 1794/3300 [4:24:26<3:03:12,  7.30s/it]INFO 04-13 11:45:24 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1795/3300 [4:24:33<3:04:25,  7.35s/it]2026-04-13 11:45:31,422 | INFO | train_grpo_train | metrics_logged mode=train step=1795

                                                       
{'loss': 0.0115, 'grad_norm': 6.115444660186768, 'learning_rate': 4.563636363636364e-06, 'num_tokens': 3208312.0, 'completions/mean_length': 101.875, 'completions/min_length': 90.0, 'completions/max_length': 113.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.875, 'completions/min_terminated_length': 90.0, 'completions/max_terminated_length': 113.0, 'rewards/meter/mean': 0.8848729133605957, 'rewards/meter/std': 0.18064017593860626, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8048514127731323, 'rewards/repeat_soft/std': 0.07562015950679779, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5051406621932983, 'rewards/total_composite/std': 0.09334918111562729, 'reward': 0.5051406621932983, 'reward_std': 0.09334918111562729, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10673120617866516, 'sampling/sampling_logp_difference/max': 1.910182237625122, 'sampling/importance_sampling_ratio/min': 0.14805340766906738, 'sampling/importance_sampling_ratio/mean': 1.0092767477035522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6201447658240795, 'clip_ratio/low_mean': 0.04387442488223314, 'clip_ratio/low_min': 0.04387442488223314, 'clip_ratio/high_mean': 0.045673233456909657, 'clip_ratio/high_max': 0.045673233456909657, 'clip_ratio/region_mean': 0.0895476583391428, 'reward_total_mean': 0.5051406621932983, 'reward_meter_mean': 0.8848729133605957, 'reward_meter_std': 0.18064017593860626, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8048514127731323, 'reward_repeat_soft_std': 0.07562015950679779, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5051406621932983, 'reward_total_composite_std': 0.09334918111562729, 'epoch': 0.18}

 54%|█████▍    | 1795/3300 [4:24:33<3:04:25,  7.35s/it]INFO 04-13 11:45:31 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1796/3300 [4:24:41<3:09:41,  7.57s/it]2026-04-13 11:45:39,490 | INFO | train_grpo_train | metrics_logged mode=train step=1796

                                                       
{'loss': 0.065, 'grad_norm': 5.339907169342041, 'learning_rate': 4.560606060606061e-06, 'num_tokens': 3210589.0, 'completions/mean_length': 119.625, 'completions/min_length': 104.0, 'completions/max_length': 157.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.625, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 157.0, 'rewards/meter/mean': 0.9552260637283325, 'rewards/meter/std': 0.07213979214429855, 'rewards/count_adherence/mean': 0.625, 'rewards/count_adherence/std': 0.1178511381149292, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8133866786956787, 'rewards/repeat_soft/std': 0.061037372797727585, 'rewards/judge_quality/mean': 0.44874998927116394, 'rewards/judge_quality/std': 0.16137246787548065, 'rewards/total_composite/mean': 0.5194798707962036, 'rewards/total_composite/std': 0.07868607342243195, 'reward': 0.5194798707962036, 'reward_std': 0.07868607342243195, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08553514629602432, 'sampling/sampling_logp_difference/max': 3.378880500793457, 'sampling/importance_sampling_ratio/min': 0.03408559411764145, 'sampling/importance_sampling_ratio/mean': 1.0060369968414307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5528635047376156, 'clip_ratio/low_mean': 0.04500382672995329, 'clip_ratio/low_min': 0.04500382672995329, 'clip_ratio/high_mean': 0.03737408481538296, 'clip_ratio/high_max': 0.03737408481538296, 'clip_ratio/region_mean': 0.08237791154533625, 'reward_total_mean': 0.5194798707962036, 'reward_meter_mean': 0.9552260637283325, 'reward_meter_std': 0.07213979214429855, 'reward_count_adherence_mean': 0.625, 'reward_count_adherence_std': 0.1178511381149292, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8133866786956787, 'reward_repeat_soft_std': 0.061037372797727585, 'reward_judge_quality_mean': 0.44874998927116394, 'reward_judge_quality_std': 0.16137246787548065, 'reward_total_composite_mean': 0.5194798707962036, 'reward_total_composite_std': 0.07868607342243195, 'epoch': 0.18}

 54%|█████▍    | 1796/3300 [4:24:41<3:09:41,  7.57s/it]INFO 04-13 11:45:39 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1797/3300 [4:24:53<3:39:28,  8.76s/it]2026-04-13 11:45:51,039 | INFO | train_grpo_train | metrics_logged mode=train step=1797

                                                       
{'loss': -0.2245, 'grad_norm': 2.382840156555176, 'learning_rate': 4.557575757575758e-06, 'num_tokens': 3213457.0, 'completions/mean_length': 231.5, 'completions/min_length': 164.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 191.42857360839844, 'completions/min_terminated_length': 164.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.6678950190544128, 'rewards/meter/std': 0.4299687445163727, 'rewards/count_adherence/mean': 0.7250000238418579, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8232831358909607, 'rewards/repeat_soft/std': 0.07086049765348434, 'rewards/judge_quality/mean': 0.3399999737739563, 'rewards/judge_quality/std': 0.1505228877067566, 'rewards/total_composite/mean': 0.3902186155319214, 'rewards/total_composite/std': 0.18478281795978546, 'reward': 0.3902186155319214, 'reward_std': 0.18478280305862427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09489186853170395, 'sampling/sampling_logp_difference/max': 1.3858833312988281, 'sampling/importance_sampling_ratio/min': 0.25010278820991516, 'sampling/importance_sampling_ratio/mean': 1.0041186809539795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5956591069698334, 'clip_ratio/low_mean': 0.021478006150573492, 'clip_ratio/low_min': 0.021478006150573492, 'clip_ratio/high_mean': 0.05395158566534519, 'clip_ratio/high_max': 0.05395158566534519, 'clip_ratio/region_mean': 0.07542959181591868, 'reward_total_mean': 0.3902186155319214, 'reward_meter_mean': 0.6678950190544128, 'reward_meter_std': 0.4299687445163727, 'reward_count_adherence_mean': 0.7250000238418579, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8232831358909607, 'reward_repeat_soft_std': 0.07086049765348434, 'reward_judge_quality_mean': 0.3399999737739563, 'reward_judge_quality_std': 0.1505228877067566, 'reward_total_composite_mean': 0.3902186155319214, 'reward_total_composite_std': 0.18478281795978546, 'epoch': 0.18}

 54%|█████▍    | 1797/3300 [4:24:53<3:39:28,  8.76s/it]INFO 04-13 11:45:51 [block_pool.py:378] Successfully reset prefix cache

 54%|█████▍    | 1798/3300 [4:25:00<3:27:49,  8.30s/it]2026-04-13 11:45:58,268 | INFO | train_grpo_train | metrics_logged mode=train step=1798

                                                       
{'loss': 0.2943, 'grad_norm': 11.488846778869629, 'learning_rate': 4.554545454545455e-06, 'num_tokens': 3215041.0, 'completions/mean_length': 48.0, 'completions/min_length': 27.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.857647716999054, 'rewards/meter/std': 0.15262115001678467, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.37796446681022644, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9310461282730103, 'rewards/repeat_soft/std': 0.08830144256353378, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.5560691952705383, 'rewards/total_composite/std': 0.11330970376729965, 'reward': 0.5560691952705383, 'reward_std': 0.11330970376729965, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13383953273296356, 'sampling/sampling_logp_difference/max': 1.622748851776123, 'sampling/importance_sampling_ratio/min': 0.1973554491996765, 'sampling/importance_sampling_ratio/mean': 1.0150392055511475, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0053318589925766, 'clip_ratio/low_mean': 0.04046074906364083, 'clip_ratio/low_min': 0.04046074906364083, 'clip_ratio/high_mean': 0.08967884071171284, 'clip_ratio/high_max': 0.08967884071171284, 'clip_ratio/region_mean': 0.13013958977535367, 'reward_total_mean': 0.5560691952705383, 'reward_meter_mean': 0.857647716999054, 'reward_meter_std': 0.15262115001678467, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.37796446681022644, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9310461282730103, 'reward_repeat_soft_std': 0.08830144256353378, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.5560691952705383, 'reward_total_composite_std': 0.11330970376729965, 'epoch': 0.18}

 54%|█████▍    | 1798/3300 [4:25:00<3:27:49,  8.30s/it]INFO 04-13 11:45:58 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1799/3300 [4:25:08<3:25:49,  8.23s/it]2026-04-13 11:46:06,323 | INFO | train_grpo_train | metrics_logged mode=train step=1799

                                                       
{'loss': 0.0474, 'grad_norm': 7.578016757965088, 'learning_rate': 4.551515151515152e-06, 'num_tokens': 3217700.0, 'completions/mean_length': 167.375, 'completions/min_length': 136.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 167.375, 'completions/min_terminated_length': 136.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.9869016408920288, 'rewards/meter/std': 0.010118884965777397, 'rewards/count_adherence/mean': 0.53125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6848177909851074, 'rewards/repeat_soft/std': 0.0947936475276947, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4647965431213379, 'rewards/total_composite/std': 0.040095239877700806, 'reward': 0.4647965431213379, 'reward_std': 0.040095243602991104, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07609271258115768, 'sampling/sampling_logp_difference/max': 1.9342234134674072, 'sampling/importance_sampling_ratio/min': 0.14453648030757904, 'sampling/importance_sampling_ratio/mean': 1.0183758735656738, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5007193610072136, 'clip_ratio/low_mean': 0.021510645747184753, 'clip_ratio/low_min': 0.021510645747184753, 'clip_ratio/high_mean': 0.04436182277277112, 'clip_ratio/high_max': 0.04436182277277112, 'clip_ratio/region_mean': 0.06587246851995587, 'reward_total_mean': 0.4647965431213379, 'reward_meter_mean': 0.9869016408920288, 'reward_meter_std': 0.010118884965777397, 'reward_count_adherence_mean': 0.53125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6848177909851074, 'reward_repeat_soft_std': 0.0947936475276947, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4647965431213379, 'reward_total_composite_std': 0.040095239877700806, 'epoch': 0.18}

 55%|█████▍    | 1799/3300 [4:25:08<3:25:49,  8.23s/it]INFO 04-13 11:46:06 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1800/3300 [4:25:16<3:27:23,  8.30s/it]2026-04-13 11:46:14,770 | INFO | train_grpo_train | metrics_logged mode=train step=1800

                                                       
{'loss': 0.0076, 'grad_norm': 4.887317180633545, 'learning_rate': 4.548484848484849e-06, 'num_tokens': 3220775.0, 'completions/mean_length': 200.375, 'completions/min_length': 162.0, 'completions/max_length': 242.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 200.375, 'completions/min_terminated_length': 162.0, 'completions/max_terminated_length': 242.0, 'rewards/meter/mean': 0.8649101257324219, 'rewards/meter/std': 0.2364814579486847, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8354235887527466, 'rewards/repeat_soft/std': 0.04172105714678764, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5114339590072632, 'rewards/total_composite/std': 0.06003821641206741, 'reward': 0.5114339590072632, 'reward_std': 0.060038212686777115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10037865489721298, 'sampling/sampling_logp_difference/max': 1.4967641830444336, 'sampling/importance_sampling_ratio/min': 0.22385334968566895, 'sampling/importance_sampling_ratio/mean': 1.0184109210968018, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8853504583239555, 'clip_ratio/low_mean': 0.027874689549207687, 'clip_ratio/low_min': 0.027874689549207687, 'clip_ratio/high_mean': 0.0581750962883234, 'clip_ratio/high_max': 0.0581750962883234, 'clip_ratio/region_mean': 0.08604978583753109, 'reward_total_mean': 0.5114339590072632, 'reward_meter_mean': 0.8649101257324219, 'reward_meter_std': 0.2364814579486847, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8354235887527466, 'reward_repeat_soft_std': 0.04172105714678764, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5114339590072632, 'reward_total_composite_std': 0.06003821641206741, 'epoch': 0.18}

 55%|█████▍    | 1800/3300 [4:25:16<3:27:23,  8.30s/it]INFO 04-13 11:46:15 [block_pool.py:378] Successfully reset prefix cache
/root/workspace/Shaer/grpo/.venv/lib/python3.11/site-packages/trl/trainer/grpo_trainer.py:1450: UserWarning: std(): degrees of freedom is <= 0. Correction should be strictly less than the reduction factor (input numel divided by output numel). (Triggered internally at /pytorch/aten/src/ATen/native/ReduceOps.cpp:1839.)
  std_rewards = rewards.view(-1, self.num_generations).std(dim=1)


  0%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:18,  2.32s/it]

 30%|███       | 3/10 [00:09<00:23,  3.41s/it]

 40%|████      | 4/10 [00:13<00:22,  3.76s/it]

 50%|█████     | 5/10 [00:19<00:22,  4.56s/it]

 60%|██████    | 6/10 [00:25<00:19,  4.76s/it]

 70%|███████   | 7/10 [00:29<00:13,  4.64s/it]

 80%|████████  | 8/10 [00:33<00:09,  4.57s/it]

 90%|█████████ | 9/10 [00:38<00:04,  4.71s/it]

100%|██████████| 10/10 [00:43<00:00,  4.55s/it]2026-04-13 11:47:03,198 | INFO | train_grpo_train | metrics_logged mode=eval step=1800

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 48.3282, 'eval_samples_per_second': 1.655, 'eval_steps_per_second': 0.207, 'eval_num_tokens': 3220775.0, 'eval_completions/mean_length': 103.4125, 'eval_completions/min_length': 34.7, 'eval_completions/max_length': 180.3, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 103.4125, 'eval_completions/min_terminated_length': 34.7, 'eval_completions/max_terminated_length': 180.3, 'eval_rewards/meter/mean': 0.7815804183483124, 'eval_rewards/meter/std': 0.280592598952353, 'eval_rewards/count_adherence/mean': 0.7295833468437195, 'eval_rewards/count_adherence/std': 0.21119236797094346, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/repeat_soft/mean': 0.8449050545692444, 'eval_rewards/repeat_soft/std': 0.09524145871400833, 'eval_rewards/judge_quality/mean': 0.47274999916553495, 'eval_rewards/judge_quality/std': 0.15114980656653643, 'eval_rewards/total_composite/mean': 0.5103550374507904, 'eval_rewards/total_composite/std': 0.13814009353518486, 'eval_reward': 0.5103550374507904, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0535921260714531, 'eval_sampling/sampling_logp_difference/max': 0.9655992984771729, 'eval_sampling/importance_sampling_ratio/min': 0.38459205329418183, 'eval_sampling/importance_sampling_ratio/mean': 1.0134925723075867, 'eval_sampling/importance_sampling_ratio/max': 1.4427119255065919, 'eval_entropy': 0.592803618311882, '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': 0.5103550374507904, 'eval_reward_meter_mean': 0.7815804183483124, 'eval_reward_meter_std': 0.280592598952353, 'eval_reward_count_adherence_mean': 0.7295833468437195, 'eval_reward_count_adherence_std': 0.21119236797094346, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_repeat_soft_mean': 0.8449050545692444, 'eval_reward_repeat_soft_std': 0.09524145871400833, 'eval_reward_judge_quality_mean': 0.47274999916553495, 'eval_reward_judge_quality_std': 0.15114980656653643, 'eval_reward_total_composite_mean': 0.5103550374507904, 'eval_reward_total_composite_std': 0.13814009353518486, 'epoch': 0.18}

 55%|█████▍    | 1800/3300 [4:26:05<3:27:23,  8.30s/it]

100%|██████████| 10/10 [00:43<00:00,  4.55s/it]

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

 55%|█████▍    | 1801/3300 [4:26:15<9:40:51, 23.25s/it]2026-04-13 11:47:12,926 | INFO | train_grpo_train | metrics_logged mode=train step=1801

                                                       
{'loss': 0.0177, 'grad_norm': 9.770906448364258, 'learning_rate': 4.5454545454545455e-06, 'num_tokens': 3222243.0, 'completions/mean_length': 35.5, 'completions/min_length': 34.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.9592167735099792, 'rewards/meter/std': 0.02322445809841156, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7318859696388245, 'rewards/repeat_soft/std': 0.0112870829179883, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5716490745544434, 'rewards/total_composite/std': 0.0063031138852238655, 'reward': 0.5716490745544434, 'reward_std': 0.006303121335804462, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.04629416763782501, 'sampling/sampling_logp_difference/max': 1.4752837419509888, 'sampling/importance_sampling_ratio/min': 0.22871382534503937, 'sampling/importance_sampling_ratio/mean': 1.0094786882400513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2513156123459339, 'clip_ratio/low_mean': 0.010135134682059288, 'clip_ratio/low_min': 0.010135134682059288, 'clip_ratio/high_mean': 0.02460317499935627, 'clip_ratio/high_max': 0.02460317499935627, 'clip_ratio/region_mean': 0.03473830968141556, 'reward_total_mean': 0.5716490745544434, 'reward_meter_mean': 0.9592167735099792, 'reward_meter_std': 0.02322445809841156, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7318859696388245, 'reward_repeat_soft_std': 0.0112870829179883, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5716490745544434, 'reward_total_composite_std': 0.0063031138852238655, 'epoch': 0.18}

 55%|█████▍    | 1801/3300 [4:26:15<9:40:51, 23.25s/it]INFO 04-13 11:47:13 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1802/3300 [4:26:23<7:47:36, 18.73s/it]2026-04-13 11:47:21,099 | INFO | train_grpo_train | metrics_logged mode=train step=1802

                                                       
{'loss': 0.0007, 'grad_norm': 5.378952503204346, 'learning_rate': 4.542424242424243e-06, 'num_tokens': 3225235.0, 'completions/mean_length': 184.0, 'completions/min_length': 156.0, 'completions/max_length': 209.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 184.0, 'completions/min_terminated_length': 156.0, 'completions/max_terminated_length': 209.0, 'rewards/meter/mean': 0.7820342183113098, 'rewards/meter/std': 0.33812469244003296, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8114336729049683, 'rewards/repeat_soft/std': 0.049185801297426224, 'rewards/judge_quality/mean': 0.3687500059604645, 'rewards/judge_quality/std': 0.10802611708641052, 'rewards/total_composite/mean': 0.4131154417991638, 'rewards/total_composite/std': 0.09631349891424179, 'reward': 0.4131154417991638, 'reward_std': 0.09631349891424179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10045631974935532, 'sampling/sampling_logp_difference/max': 1.3578910827636719, 'sampling/importance_sampling_ratio/min': 0.2572026252746582, 'sampling/importance_sampling_ratio/mean': 1.0220563411712646, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9590811803936958, 'clip_ratio/low_mean': 0.0426053018309176, 'clip_ratio/low_min': 0.0426053018309176, 'clip_ratio/high_mean': 0.044604139402508736, 'clip_ratio/high_max': 0.044604139402508736, 'clip_ratio/region_mean': 0.08720944123342633, 'reward_total_mean': 0.4131154417991638, 'reward_meter_mean': 0.7820342183113098, 'reward_meter_std': 0.33812469244003296, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8114336729049683, 'reward_repeat_soft_std': 0.049185801297426224, 'reward_judge_quality_mean': 0.3687500059604645, 'reward_judge_quality_std': 0.10802611708641052, 'reward_total_composite_mean': 0.4131154417991638, 'reward_total_composite_std': 0.09631349891424179, 'epoch': 0.18}

 55%|█████▍    | 1802/3300 [4:26:23<7:47:36, 18.73s/it]INFO 04-13 11:47:21 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1803/3300 [4:26:31<6:26:11, 15.48s/it]2026-04-13 11:47:28,986 | INFO | train_grpo_train | metrics_logged mode=train step=1803

                                                       
{'loss': 0.0786, 'grad_norm': 11.218101501464844, 'learning_rate': 4.539393939393939e-06, 'num_tokens': 3227036.0, 'completions/mean_length': 65.125, 'completions/min_length': 52.0, 'completions/max_length': 74.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 65.125, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 74.0, 'rewards/meter/mean': 0.6598147749900818, 'rewards/meter/std': 0.3006100654602051, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8704584836959839, 'rewards/repeat_soft/std': 0.06427149474620819, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5993191599845886, 'rewards/total_composite/std': 0.1949187070131302, 'reward': 0.5993191599845886, 'reward_std': 0.1949187070131302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10370650142431259, 'sampling/sampling_logp_difference/max': 1.9163055419921875, 'sampling/importance_sampling_ratio/min': 0.14714960753917694, 'sampling/importance_sampling_ratio/mean': 0.9993343353271484, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4846961982548237, 'clip_ratio/low_mean': 0.06705652084201574, 'clip_ratio/low_min': 0.06705652084201574, 'clip_ratio/high_mean': 0.0369061091914773, 'clip_ratio/high_max': 0.0369061091914773, 'clip_ratio/region_mean': 0.10396263003349304, 'reward_total_mean': 0.5993191599845886, 'reward_meter_mean': 0.6598147749900818, 'reward_meter_std': 0.3006100654602051, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8704584836959839, 'reward_repeat_soft_std': 0.06427149474620819, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5993191599845886, 'reward_total_composite_std': 0.1949187070131302, 'epoch': 0.18}

 55%|█████▍    | 1803/3300 [4:26:31<6:26:11, 15.48s/it]INFO 04-13 11:47:29 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1804/3300 [4:26:38<5:23:54, 12.99s/it]2026-04-13 11:47:36,172 | INFO | train_grpo_train | metrics_logged mode=train step=1804

                                                       
{'loss': 0.0992, 'grad_norm': 10.028509140014648, 'learning_rate': 4.5363636363636364e-06, 'num_tokens': 3228749.0, 'completions/mean_length': 56.125, 'completions/min_length': 44.0, 'completions/max_length': 69.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.125, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9606437683105469, 'rewards/meter/std': 0.0385453999042511, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9285358190536499, 'rewards/repeat_soft/std': 0.04393455758690834, 'rewards/judge_quality/mean': 0.668749988079071, 'rewards/judge_quality/std': 0.25842589139938354, 'rewards/total_composite/mean': 0.7556654810905457, 'rewards/total_composite/std': 0.16067703068256378, 'reward': 0.7556654810905457, 'reward_std': 0.16067704558372498, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11988765001296997, 'sampling/sampling_logp_difference/max': 1.3343610763549805, 'sampling/importance_sampling_ratio/min': 0.26332637667655945, 'sampling/importance_sampling_ratio/mean': 0.998325526714325, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.775499127805233, 'clip_ratio/low_mean': 0.03212598082609475, 'clip_ratio/low_min': 0.03212598082609475, 'clip_ratio/high_mean': 0.06991446064785123, 'clip_ratio/high_max': 0.06991446064785123, 'clip_ratio/region_mean': 0.10204044147394598, 'reward_total_mean': 0.7556654810905457, 'reward_meter_mean': 0.9606437683105469, 'reward_meter_std': 0.0385453999042511, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9285358190536499, 'reward_repeat_soft_std': 0.04393455758690834, 'reward_judge_quality_mean': 0.668749988079071, 'reward_judge_quality_std': 0.25842589139938354, 'reward_total_composite_mean': 0.7556654810905457, 'reward_total_composite_std': 0.16067703068256378, 'epoch': 0.18}

 55%|█████▍    | 1804/3300 [4:26:38<5:23:54, 12.99s/it]INFO 04-13 11:47:36 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1805/3300 [4:26:45<4:40:28, 11.26s/it]2026-04-13 11:47:43,401 | INFO | train_grpo_train | metrics_logged mode=train step=1805

                                                       
{'loss': -0.0555, 'grad_norm': 10.76839542388916, 'learning_rate': 4.533333333333334e-06, 'num_tokens': 3230528.0, 'completions/mean_length': 57.375, 'completions/min_length': 43.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.6418136954307556, 'rewards/meter/std': 0.36736324429512024, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.25877460837364197, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9308949708938599, 'rewards/repeat_soft/std': 0.11350374668836594, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5372963547706604, 'rewards/total_composite/std': 0.14024575054645538, 'reward': 0.5372963547706604, 'reward_std': 0.14024576544761658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1133660301566124, 'sampling/sampling_logp_difference/max': 1.6175422668457031, 'sampling/importance_sampling_ratio/min': 0.19838567078113556, 'sampling/importance_sampling_ratio/mean': 1.0129854679107666, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8203057646751404, 'clip_ratio/low_mean': 0.06033054692670703, 'clip_ratio/low_min': 0.06033054692670703, 'clip_ratio/high_mean': 0.04288742411881685, 'clip_ratio/high_max': 0.04288742411881685, 'clip_ratio/region_mean': 0.10321797104552388, 'reward_total_mean': 0.5372963547706604, 'reward_meter_mean': 0.6418136954307556, 'reward_meter_std': 0.36736324429512024, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.25877460837364197, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9308949708938599, 'reward_repeat_soft_std': 0.11350374668836594, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5372963547706604, 'reward_total_composite_std': 0.14024575054645538, 'epoch': 0.18}

 55%|█████▍    | 1805/3300 [4:26:45<4:40:28, 11.26s/it]INFO 04-13 11:47:43 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1806/3300 [4:26:54<4:21:14, 10.49s/it]2026-04-13 11:47:52,110 | INFO | train_grpo_train | metrics_logged mode=train step=1806

                                                       
{'loss': 0.0055, 'grad_norm': 4.9276909828186035, 'learning_rate': 4.53030303030303e-06, 'num_tokens': 3233534.0, 'completions/mean_length': 181.75, 'completions/min_length': 146.0, 'completions/max_length': 223.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 181.75, 'completions/min_terminated_length': 146.0, 'completions/max_terminated_length': 223.0, 'rewards/meter/mean': 0.7319014072418213, 'rewards/meter/std': 0.27093157172203064, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.07715165615081787, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7889889478683472, 'rewards/repeat_soft/std': 0.03599509969353676, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4598240852355957, 'rewards/total_composite/std': 0.07499600946903229, 'reward': 0.4598240852355957, 'reward_std': 0.07499600946903229, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08227083832025528, 'sampling/sampling_logp_difference/max': 1.5979204177856445, 'sampling/importance_sampling_ratio/min': 0.20231682062149048, 'sampling/importance_sampling_ratio/mean': 1.0073206424713135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6394183374941349, 'clip_ratio/low_mean': 0.026449366007000208, 'clip_ratio/low_min': 0.026449366007000208, 'clip_ratio/high_mean': 0.052284734789282084, 'clip_ratio/high_max': 0.052284734789282084, 'clip_ratio/region_mean': 0.07873410079628229, 'reward_total_mean': 0.4598240852355957, 'reward_meter_mean': 0.7319014072418213, 'reward_meter_std': 0.27093157172203064, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.07715165615081787, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7889889478683472, 'reward_repeat_soft_std': 0.03599509969353676, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4598240852355957, 'reward_total_composite_std': 0.07499600946903229, 'epoch': 0.18}

 55%|█████▍    | 1806/3300 [4:26:54<4:21:14, 10.49s/it]INFO 04-13 11:47:52 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1807/3300 [4:27:00<3:47:02,  9.12s/it]2026-04-13 11:47:58,035 | INFO | train_grpo_train | metrics_logged mode=train step=1807

                                                       
{'loss': 0.0895, 'grad_norm': 16.027671813964844, 'learning_rate': 4.527272727272727e-06, 'num_tokens': 3234932.0, 'completions/mean_length': 24.75, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9225627779960632, 'rewards/meter/std': 0.09814183413982391, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9316083192825317, 'rewards/repeat_soft/std': 0.05721777305006981, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.5849047303199768, 'rewards/total_composite/std': 0.04630429670214653, 'reward': 0.5849047303199768, 'reward_std': 0.04630429670214653, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1425008922815323, 'sampling/sampling_logp_difference/max': 1.8376264572143555, 'sampling/importance_sampling_ratio/min': 0.15919482707977295, 'sampling/importance_sampling_ratio/mean': 1.0083662271499634, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0856371745467186, 'clip_ratio/low_mean': 0.06887955591082573, 'clip_ratio/low_min': 0.06887955591082573, 'clip_ratio/high_mean': 0.12037668284028769, 'clip_ratio/high_max': 0.12037668284028769, 'clip_ratio/region_mean': 0.18925623875111341, 'reward_total_mean': 0.5849047303199768, 'reward_meter_mean': 0.9225627779960632, 'reward_meter_std': 0.09814183413982391, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9316083192825317, 'reward_repeat_soft_std': 0.05721777305006981, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.5849047303199768, 'reward_total_composite_std': 0.04630429670214653, 'epoch': 0.18}

 55%|█████▍    | 1807/3300 [4:27:00<3:47:02,  9.12s/it]INFO 04-13 11:47:58 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1808/3300 [4:27:07<3:30:28,  8.46s/it]2026-04-13 11:48:04,962 | INFO | train_grpo_train | metrics_logged mode=train step=1808

                                                       
{'loss': 0.0453, 'grad_norm': 8.079506874084473, 'learning_rate': 4.524242424242425e-06, 'num_tokens': 3237030.0, 'completions/mean_length': 95.25, 'completions/min_length': 83.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.25, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.9181572794914246, 'rewards/meter/std': 0.12369729578495026, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8471939563751221, 'rewards/repeat_soft/std': 0.021875236183404922, 'rewards/judge_quality/mean': 0.8362500667572021, 'rewards/judge_quality/std': 0.23688077926635742, 'rewards/total_composite/mean': 0.7605818510055542, 'rewards/total_composite/std': 0.15978670120239258, 'reward': 0.7605818510055542, 'reward_std': 0.15978671610355377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09632844477891922, 'sampling/sampling_logp_difference/max': 1.2662220001220703, 'sampling/importance_sampling_ratio/min': 0.28189462423324585, 'sampling/importance_sampling_ratio/mean': 1.019246220588684, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7270065546035767, 'clip_ratio/low_mean': 0.019042464904487133, 'clip_ratio/low_min': 0.019042464904487133, 'clip_ratio/high_mean': 0.06389599200338125, 'clip_ratio/high_max': 0.06389599200338125, 'clip_ratio/region_mean': 0.08293845690786839, 'reward_total_mean': 0.7605818510055542, 'reward_meter_mean': 0.9181572794914246, 'reward_meter_std': 0.12369729578495026, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8471939563751221, 'reward_repeat_soft_std': 0.021875236183404922, 'reward_judge_quality_mean': 0.8362500667572021, 'reward_judge_quality_std': 0.23688077926635742, 'reward_total_composite_mean': 0.7605818510055542, 'reward_total_composite_std': 0.15978670120239258, 'epoch': 0.18}

 55%|█████▍    | 1808/3300 [4:27:07<3:30:28,  8.46s/it]INFO 04-13 11:48:05 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1809/3300 [4:27:14<3:21:36,  8.11s/it]2026-04-13 11:48:12,254 | INFO | train_grpo_train | metrics_logged mode=train step=1809

                                                       
{'loss': -0.0039, 'grad_norm': 11.031441688537598, 'learning_rate': 4.521212121212122e-06, 'num_tokens': 3239058.0, 'completions/mean_length': 90.5, 'completions/min_length': 76.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.5, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.9052593111991882, 'rewards/meter/std': 0.10923722386360168, 'rewards/count_adherence/mean': 0.4166666865348816, 'rewards/count_adherence/std': 0.15430335700511932, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8219844102859497, 'rewards/repeat_soft/std': 0.07050145417451859, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.4561029076576233, 'rewards/total_composite/std': 0.10142374038696289, 'reward': 0.4561029076576233, 'reward_std': 0.10142374038696289, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10839082300662994, 'sampling/sampling_logp_difference/max': 2.2924723625183105, 'sampling/importance_sampling_ratio/min': 0.10101640224456787, 'sampling/importance_sampling_ratio/mean': 1.0194932222366333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6266350857913494, 'clip_ratio/low_mean': 0.03545480594038963, 'clip_ratio/low_min': 0.03545480594038963, 'clip_ratio/high_mean': 0.05669600097462535, 'clip_ratio/high_max': 0.05669600097462535, 'clip_ratio/region_mean': 0.09215080691501498, 'reward_total_mean': 0.4561029076576233, 'reward_meter_mean': 0.9052593111991882, 'reward_meter_std': 0.10923722386360168, 'reward_count_adherence_mean': 0.4166666865348816, 'reward_count_adherence_std': 0.15430335700511932, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8219844102859497, 'reward_repeat_soft_std': 0.07050145417451859, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.4561029076576233, 'reward_total_composite_std': 0.10142374038696289, 'epoch': 0.18}

 55%|█████▍    | 1809/3300 [4:27:14<3:21:36,  8.11s/it]INFO 04-13 11:48:12 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1810/3300 [4:27:22<3:22:41,  8.16s/it]2026-04-13 11:48:20,523 | INFO | train_grpo_train | metrics_logged mode=train step=1810

                                                       
{'loss': -0.025, 'grad_norm': 5.897127151489258, 'learning_rate': 4.518181818181819e-06, 'num_tokens': 3241821.0, 'completions/mean_length': 170.375, 'completions/min_length': 148.0, 'completions/max_length': 212.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 170.375, 'completions/min_terminated_length': 148.0, 'completions/max_terminated_length': 212.0, 'rewards/meter/mean': 0.6426774263381958, 'rewards/meter/std': 0.20387646555900574, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9584013819694519, 'rewards/repeat_soft/std': 0.021950701251626015, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.4701296389102936, 'rewards/total_composite/std': 0.05986081436276436, 'reward': 0.4701296389102936, 'reward_std': 0.05986081436276436, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13232646882534027, 'sampling/sampling_logp_difference/max': 2.335085868835449, 'sampling/importance_sampling_ratio/min': 0.09680216759443283, 'sampling/importance_sampling_ratio/mean': 1.0158830881118774, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.075805090367794, 'clip_ratio/low_mean': 0.0722449030727148, 'clip_ratio/low_min': 0.0722449030727148, 'clip_ratio/high_mean': 0.04664424154907465, 'clip_ratio/high_max': 0.04664424154907465, 'clip_ratio/region_mean': 0.11888914462178946, 'reward_total_mean': 0.4701296389102936, 'reward_meter_mean': 0.6426774263381958, 'reward_meter_std': 0.20387646555900574, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9584013819694519, 'reward_repeat_soft_std': 0.021950701251626015, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.4701296389102936, 'reward_total_composite_std': 0.05986081436276436, 'epoch': 0.18}

 55%|█████▍    | 1810/3300 [4:27:22<3:22:41,  8.16s/it]INFO 04-13 11:48:20 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1811/3300 [4:27:29<3:13:56,  7.81s/it]2026-04-13 11:48:27,540 | INFO | train_grpo_train | metrics_logged mode=train step=1811

                                                       
{'loss': -0.0265, 'grad_norm': 8.900130271911621, 'learning_rate': 4.5151515151515155e-06, 'num_tokens': 3243646.0, 'completions/mean_length': 69.125, 'completions/min_length': 56.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.7467904686927795, 'rewards/meter/std': 0.29527267813682556, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8589749336242676, 'rewards/repeat_soft/std': 0.03796645998954773, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4660533666610718, 'rewards/total_composite/std': 0.0844200924038887, 'reward': 0.4660533666610718, 'reward_std': 0.0844200998544693, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11518388241529465, 'sampling/sampling_logp_difference/max': 1.473679542541504, 'sampling/importance_sampling_ratio/min': 0.3138110935688019, 'sampling/importance_sampling_ratio/mean': 1.0072176456451416, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7754762470722198, 'clip_ratio/low_mean': 0.04373500728979707, 'clip_ratio/low_min': 0.04373500728979707, 'clip_ratio/high_mean': 0.07083044201135635, 'clip_ratio/high_max': 0.07083044201135635, 'clip_ratio/region_mean': 0.11456544930115342, 'reward_total_mean': 0.4660533666610718, 'reward_meter_mean': 0.7467904686927795, 'reward_meter_std': 0.29527267813682556, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8589749336242676, 'reward_repeat_soft_std': 0.03796645998954773, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4660533666610718, 'reward_total_composite_std': 0.0844200924038887, 'epoch': 0.18}

 55%|█████▍    | 1811/3300 [4:27:29<3:13:56,  7.81s/it]INFO 04-13 11:48:27 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1812/3300 [4:27:35<3:00:32,  7.28s/it]2026-04-13 11:48:33,558 | INFO | train_grpo_train | metrics_logged mode=train step=1812

                                                       
{'loss': 0.0516, 'grad_norm': 11.70666790008545, 'learning_rate': 4.512121212121213e-06, 'num_tokens': 3244994.0, 'completions/mean_length': 25.5, 'completions/min_length': 21.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7291864156723022, 'rewards/meter/std': 0.4397353231906891, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9311109185218811, 'rewards/repeat_soft/std': 0.02573113888502121, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.62016761302948, 'rewards/total_composite/std': 0.2084752470254898, 'reward': 0.62016761302948, 'reward_std': 0.2084752470254898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11912678182125092, 'sampling/sampling_logp_difference/max': 0.9539971351623535, 'sampling/importance_sampling_ratio/min': 0.3851982653141022, 'sampling/importance_sampling_ratio/mean': 1.0331926345825195, 'sampling/importance_sampling_ratio/max': 1.9330251216888428, 'entropy': 0.7726024240255356, 'clip_ratio/low_mean': 0.04123463947325945, 'clip_ratio/low_min': 0.04123463947325945, 'clip_ratio/high_mean': 0.049452862702310085, 'clip_ratio/high_max': 0.049452862702310085, 'clip_ratio/region_mean': 0.09068750217556953, 'reward_total_mean': 0.62016761302948, 'reward_meter_mean': 0.7291864156723022, 'reward_meter_std': 0.4397353231906891, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9311109185218811, 'reward_repeat_soft_std': 0.02573113888502121, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.62016761302948, 'reward_total_composite_std': 0.2084752470254898, 'epoch': 0.18}

 55%|█████▍    | 1812/3300 [4:27:35<3:00:32,  7.28s/it]INFO 04-13 11:48:33 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1813/3300 [4:27:42<2:54:33,  7.04s/it]2026-04-13 11:48:40,058 | INFO | train_grpo_train | metrics_logged mode=train step=1813

                                                       
{'loss': 0.0103, 'grad_norm': 4.901209354400635, 'learning_rate': 4.50909090909091e-06, 'num_tokens': 3246798.0, 'completions/mean_length': 76.5, 'completions/min_length': 70.0, 'completions/max_length': 80.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.5, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 80.0, 'rewards/meter/mean': 0.9378914833068848, 'rewards/meter/std': 0.040167856961488724, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7148953676223755, 'rewards/repeat_soft/std': 0.04254239797592163, 'rewards/judge_quality/mean': 0.3687500059604645, 'rewards/judge_quality/std': 0.10802611708641052, 'rewards/total_composite/mean': 0.46721765398979187, 'rewards/total_composite/std': 0.06920699775218964, 'reward': 0.46721765398979187, 'reward_std': 0.06920700520277023, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06374968588352203, 'sampling/sampling_logp_difference/max': 0.9852927923202515, 'sampling/importance_sampling_ratio/min': 0.37332990765571594, 'sampling/importance_sampling_ratio/mean': 1.0055588483810425, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.36710847169160843, 'clip_ratio/low_mean': 0.01134446682408452, 'clip_ratio/low_min': 0.01134446682408452, 'clip_ratio/high_mean': 0.04556079301983118, 'clip_ratio/high_max': 0.04556079301983118, 'clip_ratio/region_mean': 0.0569052598439157, 'reward_total_mean': 0.46721765398979187, 'reward_meter_mean': 0.9378914833068848, 'reward_meter_std': 0.040167856961488724, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7148953676223755, 'reward_repeat_soft_std': 0.04254239797592163, 'reward_judge_quality_mean': 0.3687500059604645, 'reward_judge_quality_std': 0.10802611708641052, 'reward_total_composite_mean': 0.46721765398979187, 'reward_total_composite_std': 0.06920699775218964, 'epoch': 0.18}

 55%|█████▍    | 1813/3300 [4:27:42<2:54:33,  7.04s/it]INFO 04-13 11:48:40 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▍    | 1814/3300 [4:27:49<2:59:47,  7.26s/it]2026-04-13 11:48:47,833 | INFO | train_grpo_train | metrics_logged mode=train step=1814

                                                       
{'loss': 0.0555, 'grad_norm': 4.9915266036987305, 'learning_rate': 4.5060606060606065e-06, 'num_tokens': 3249228.0, 'completions/mean_length': 125.75, 'completions/min_length': 107.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 125.75, 'completions/min_terminated_length': 107.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.780795156955719, 'rewards/meter/std': 0.2588658928871155, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7797104120254517, 'rewards/repeat_soft/std': 0.04299363121390343, 'rewards/judge_quality/mean': 0.6450000405311584, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.5325644612312317, 'rewards/total_composite/std': 0.13640287518501282, 'reward': 0.5325644612312317, 'reward_std': 0.136402890086174, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07033336907625198, 'sampling/sampling_logp_difference/max': 1.2939739227294922, 'sampling/importance_sampling_ratio/min': 0.27417904138565063, 'sampling/importance_sampling_ratio/mean': 1.0076937675476074, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5893510244786739, 'clip_ratio/low_mean': 0.03880102909170091, 'clip_ratio/low_min': 0.03880102909170091, 'clip_ratio/high_mean': 0.01608090242370963, 'clip_ratio/high_max': 0.01608090242370963, 'clip_ratio/region_mean': 0.05488193151541054, 'reward_total_mean': 0.5325644612312317, 'reward_meter_mean': 0.780795156955719, 'reward_meter_std': 0.2588658928871155, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7797104120254517, 'reward_repeat_soft_std': 0.04299363121390343, 'reward_judge_quality_mean': 0.6450000405311584, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.5325644612312317, 'reward_total_composite_std': 0.13640287518501282, 'epoch': 0.18}

 55%|█████▍    | 1814/3300 [4:27:50<2:59:47,  7.26s/it]INFO 04-13 11:48:48 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1815/3300 [4:27:56<2:52:36,  6.97s/it]2026-04-13 11:48:54,130 | INFO | train_grpo_train | metrics_logged mode=train step=1815

                                                       
{'loss': 0.0526, 'grad_norm': 14.766071319580078, 'learning_rate': 4.503030303030304e-06, 'num_tokens': 3250851.0, 'completions/mean_length': 47.875, 'completions/min_length': 44.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8778351545333862, 'rewards/meter/std': 0.2011054903268814, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9835158586502075, 'rewards/repeat_soft/std': 0.01885959878563881, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6672040224075317, 'rewards/total_composite/std': 0.17567037045955658, 'reward': 0.6672040224075317, 'reward_std': 0.17567035555839539, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1272457093000412, 'sampling/sampling_logp_difference/max': 1.2632765769958496, 'sampling/importance_sampling_ratio/min': 0.28272613883018494, 'sampling/importance_sampling_ratio/mean': 1.005624771118164, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8602741062641144, 'clip_ratio/low_mean': 0.06548884371295571, 'clip_ratio/low_min': 0.06548884371295571, 'clip_ratio/high_mean': 0.02700046170502901, 'clip_ratio/high_max': 0.02700046170502901, 'clip_ratio/region_mean': 0.09248930541798472, 'reward_total_mean': 0.6672040224075317, 'reward_meter_mean': 0.8778351545333862, 'reward_meter_std': 0.2011054903268814, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9835158586502075, 'reward_repeat_soft_std': 0.01885959878563881, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6672040224075317, 'reward_total_composite_std': 0.17567037045955658, 'epoch': 0.18}

 55%|█████▌    | 1815/3300 [4:27:56<2:52:36,  6.97s/it]INFO 04-13 11:48:54 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1816/3300 [4:28:04<2:59:11,  7.25s/it]2026-04-13 11:49:02,000 | INFO | train_grpo_train | metrics_logged mode=train step=1816

                                                       
{'loss': -0.0518, 'grad_norm': 5.025565147399902, 'learning_rate': 4.5e-06, 'num_tokens': 3253580.0, 'completions/mean_length': 153.125, 'completions/min_length': 120.0, 'completions/max_length': 188.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 153.125, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 188.0, 'rewards/meter/mean': 0.8800153732299805, 'rewards/meter/std': 0.16372206807136536, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.788138747215271, 'rewards/repeat_soft/std': 0.09071363508701324, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.45846500992774963, 'rewards/total_composite/std': 0.04814787209033966, 'reward': 0.45846500992774963, 'reward_std': 0.04814785718917847, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09667257964611053, 'sampling/sampling_logp_difference/max': 2.0956380367279053, 'sampling/importance_sampling_ratio/min': 0.12299174070358276, 'sampling/importance_sampling_ratio/mean': 1.0169990062713623, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7186378911137581, 'clip_ratio/low_mean': 0.023024425841867924, 'clip_ratio/low_min': 0.023024425841867924, 'clip_ratio/high_mean': 0.06399360997602344, 'clip_ratio/high_max': 0.06399360997602344, 'clip_ratio/region_mean': 0.08701803581789136, 'reward_total_mean': 0.45846500992774963, 'reward_meter_mean': 0.8800153732299805, 'reward_meter_std': 0.16372206807136536, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.788138747215271, 'reward_repeat_soft_std': 0.09071363508701324, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.45846500992774963, 'reward_total_composite_std': 0.04814787209033966, 'epoch': 0.18}

 55%|█████▌    | 1816/3300 [4:28:04<2:59:11,  7.25s/it]INFO 04-13 11:49:02 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1817/3300 [4:28:11<2:59:58,  7.28s/it]2026-04-13 11:49:09,362 | INFO | train_grpo_train | metrics_logged mode=train step=1817

                                                       
{'loss': -0.0054, 'grad_norm': 6.091757297515869, 'learning_rate': 4.496969696969697e-06, 'num_tokens': 3255702.0, 'completions/mean_length': 101.25, 'completions/min_length': 79.0, 'completions/max_length': 119.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.25, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 119.0, 'rewards/meter/mean': 0.931243360042572, 'rewards/meter/std': 0.1261349767446518, 'rewards/count_adherence/mean': 0.5416666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8289350271224976, 'rewards/repeat_soft/std': 0.03682198002934456, 'rewards/judge_quality/mean': 0.6487500071525574, 'rewards/judge_quality/std': 0.2456151396036148, 'rewards/total_composite/mean': 0.6251682043075562, 'rewards/total_composite/std': 0.15108372271060944, 'reward': 0.6251682043075562, 'reward_std': 0.15108372271060944, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09348579496145248, 'sampling/sampling_logp_difference/max': 1.7902698516845703, 'sampling/importance_sampling_ratio/min': 0.16691511869430542, 'sampling/importance_sampling_ratio/mean': 1.0175946950912476, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6182441972196102, 'clip_ratio/low_mean': 0.0432027718052268, 'clip_ratio/low_min': 0.0432027718052268, 'clip_ratio/high_mean': 0.026850812137126923, 'clip_ratio/high_max': 0.026850812137126923, 'clip_ratio/region_mean': 0.07005358394235373, 'reward_total_mean': 0.6251682043075562, 'reward_meter_mean': 0.931243360042572, 'reward_meter_std': 0.1261349767446518, 'reward_count_adherence_mean': 0.5416666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8289350271224976, 'reward_repeat_soft_std': 0.03682198002934456, 'reward_judge_quality_mean': 0.6487500071525574, 'reward_judge_quality_std': 0.2456151396036148, 'reward_total_composite_mean': 0.6251682043075562, 'reward_total_composite_std': 0.15108372271060944, 'epoch': 0.18}

 55%|█████▌    | 1817/3300 [4:28:11<2:59:58,  7.28s/it]INFO 04-13 11:49:09 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1818/3300 [4:28:17<2:51:44,  6.95s/it]2026-04-13 11:49:15,554 | INFO | train_grpo_train | metrics_logged mode=train step=1818

                                                       
{'loss': 0.1082, 'grad_norm': 9.090434074401855, 'learning_rate': 4.493939393939395e-06, 'num_tokens': 3257463.0, 'completions/mean_length': 49.125, 'completions/min_length': 38.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.76106858253479, 'rewards/meter/std': 0.3347349464893341, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.26726123690605164, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9026095867156982, 'rewards/repeat_soft/std': 0.061921849846839905, 'rewards/judge_quality/mean': 0.46875, 'rewards/judge_quality/std': 0.10260014235973358, 'rewards/total_composite/mean': 0.5050399303436279, 'rewards/total_composite/std': 0.09929977357387543, 'reward': 0.5050399303436279, 'reward_std': 0.09929978847503662, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08518665283918381, 'sampling/sampling_logp_difference/max': 1.5348838567733765, 'sampling/importance_sampling_ratio/min': 0.2154807299375534, 'sampling/importance_sampling_ratio/mean': 1.012081503868103, 'sampling/importance_sampling_ratio/max': 1.8807759284973145, 'entropy': 0.6251075603067875, 'clip_ratio/low_mean': 0.043447833973914385, 'clip_ratio/low_min': 0.043447833973914385, 'clip_ratio/high_mean': 0.03728580195456743, 'clip_ratio/high_max': 0.03728580195456743, 'clip_ratio/region_mean': 0.08073363592848182, 'reward_total_mean': 0.5050399303436279, 'reward_meter_mean': 0.76106858253479, 'reward_meter_std': 0.3347349464893341, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.26726123690605164, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9026095867156982, 'reward_repeat_soft_std': 0.061921849846839905, 'reward_judge_quality_mean': 0.46875, 'reward_judge_quality_std': 0.10260014235973358, 'reward_total_composite_mean': 0.5050399303436279, 'reward_total_composite_std': 0.09929977357387543, 'epoch': 0.18}

 55%|█████▌    | 1818/3300 [4:28:17<2:51:44,  6.95s/it]INFO 04-13 11:49:15 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1819/3300 [4:28:24<2:47:47,  6.80s/it]2026-04-13 11:49:21,992 | INFO | train_grpo_train | metrics_logged mode=train step=1819

                                                       
{'loss': -0.0971, 'grad_norm': 11.203838348388672, 'learning_rate': 4.490909090909091e-06, 'num_tokens': 3259080.0, 'completions/mean_length': 41.125, 'completions/min_length': 33.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.6667453646659851, 'rewards/meter/std': 0.3940989077091217, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9571197628974915, 'rewards/repeat_soft/std': 0.044309817254543304, 'rewards/judge_quality/mean': 0.5062499642372131, 'rewards/judge_quality/std': 0.14201989769935608, 'rewards/total_composite/mean': 0.57743239402771, 'rewards/total_composite/std': 0.16943275928497314, 'reward': 0.57743239402771, 'reward_std': 0.16943275928497314, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13686740398406982, 'sampling/sampling_logp_difference/max': 1.297973394393921, 'sampling/importance_sampling_ratio/min': 0.29697665572166443, 'sampling/importance_sampling_ratio/mean': 1.0200828313827515, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9041113927960396, 'clip_ratio/low_mean': 0.05710151046514511, 'clip_ratio/low_min': 0.05710151046514511, 'clip_ratio/high_mean': 0.09784480463713408, 'clip_ratio/high_max': 0.09784480463713408, 'clip_ratio/region_mean': 0.1549463151022792, 'reward_total_mean': 0.57743239402771, 'reward_meter_mean': 0.6667453646659851, 'reward_meter_std': 0.3940989077091217, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9571197628974915, 'reward_repeat_soft_std': 0.044309817254543304, 'reward_judge_quality_mean': 0.5062499642372131, 'reward_judge_quality_std': 0.14201989769935608, 'reward_total_composite_mean': 0.57743239402771, 'reward_total_composite_std': 0.16943275928497314, 'epoch': 0.18}

 55%|█████▌    | 1819/3300 [4:28:24<2:47:47,  6.80s/it]INFO 04-13 11:49:22 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1820/3300 [4:28:30<2:42:51,  6.60s/it]2026-04-13 11:49:28,137 | INFO | train_grpo_train | metrics_logged mode=train step=1820

                                                       
{'loss': 0.0264, 'grad_norm': 12.091035842895508, 'learning_rate': 4.487878787878788e-06, 'num_tokens': 3260612.0, 'completions/mean_length': 35.5, 'completions/min_length': 29.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.5414127707481384, 'rewards/meter/std': 0.3649313747882843, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9394425749778748, 'rewards/repeat_soft/std': 0.05158919468522072, 'rewards/judge_quality/mean': 0.6150000095367432, 'rewards/judge_quality/std': 0.25286927819252014, 'rewards/total_composite/mean': 0.5937775373458862, 'rewards/total_composite/std': 0.23556756973266602, 'reward': 0.5937775373458862, 'reward_std': 0.23556756973266602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10851365327835083, 'sampling/sampling_logp_difference/max': 1.3843728303909302, 'sampling/importance_sampling_ratio/min': 0.3599967360496521, 'sampling/importance_sampling_ratio/mean': 1.012622594833374, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5805025473237038, 'clip_ratio/low_mean': 0.05828921776264906, 'clip_ratio/low_min': 0.05828921776264906, 'clip_ratio/high_mean': 0.028174937702715397, 'clip_ratio/high_max': 0.028174937702715397, 'clip_ratio/region_mean': 0.08646415546536446, 'reward_total_mean': 0.5937775373458862, 'reward_meter_mean': 0.5414127707481384, 'reward_meter_std': 0.3649313747882843, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9394425749778748, 'reward_repeat_soft_std': 0.05158919468522072, 'reward_judge_quality_mean': 0.6150000095367432, 'reward_judge_quality_std': 0.25286927819252014, 'reward_total_composite_mean': 0.5937775373458862, 'reward_total_composite_std': 0.23556756973266602, 'epoch': 0.18}

 55%|█████▌    | 1820/3300 [4:28:30<2:42:51,  6.60s/it]INFO 04-13 11:49:28 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1821/3300 [4:28:36<2:38:10,  6.42s/it]2026-04-13 11:49:34,127 | INFO | train_grpo_train | metrics_logged mode=train step=1821

                                                       
{'loss': 0.0455, 'grad_norm': 17.626693725585938, 'learning_rate': 4.4848484848484855e-06, 'num_tokens': 3261961.0, 'completions/mean_length': 20.625, 'completions/min_length': 18.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.625, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.5819944739341736, 'rewards/meter/std': 0.4677361249923706, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5055780410766602, 'rewards/total_composite/std': 0.12993228435516357, 'reward': 0.5055780410766602, 'reward_std': 0.12993228435516357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1103300154209137, 'sampling/sampling_logp_difference/max': 1.2852458953857422, 'sampling/importance_sampling_ratio/min': 0.2765825688838959, 'sampling/importance_sampling_ratio/mean': 1.0284539461135864, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0949403643608093, 'clip_ratio/low_mean': 0.018483709543943405, 'clip_ratio/low_min': 0.018483709543943405, 'clip_ratio/high_mean': 0.09154396597296, 'clip_ratio/high_max': 0.09154396597296, 'clip_ratio/region_mean': 0.1100276755169034, 'reward_total_mean': 0.5055780410766602, 'reward_meter_mean': 0.5819944739341736, 'reward_meter_std': 0.4677361249923706, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5055780410766602, 'reward_total_composite_std': 0.12993228435516357, 'epoch': 0.18}

 55%|█████▌    | 1821/3300 [4:28:36<2:38:10,  6.42s/it]INFO 04-13 11:49:34 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1822/3300 [4:28:43<2:45:47,  6.73s/it]2026-04-13 11:49:41,597 | INFO | train_grpo_train | metrics_logged mode=train step=1822

                                                       
{'loss': 0.035, 'grad_norm': 5.529888153076172, 'learning_rate': 4.481818181818182e-06, 'num_tokens': 3264335.0, 'completions/mean_length': 130.75, 'completions/min_length': 109.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 130.75, 'completions/min_terminated_length': 109.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.4340297281742096, 'rewards/meter/std': 0.22348229587078094, 'rewards/count_adherence/mean': 0.4375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8360930681228638, 'rewards/repeat_soft/std': 0.02245735377073288, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.331297367811203, 'rewards/total_composite/std': 0.037000950425863266, 'reward': 0.331297367811203, 'reward_std': 0.03700094297528267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10386641323566437, 'sampling/sampling_logp_difference/max': 1.4150023460388184, 'sampling/importance_sampling_ratio/min': 0.24292504787445068, 'sampling/importance_sampling_ratio/mean': 1.0231037139892578, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8596381545066833, 'clip_ratio/low_mean': 0.05849697347730398, 'clip_ratio/low_min': 0.05849697347730398, 'clip_ratio/high_mean': 0.03997781686484814, 'clip_ratio/high_max': 0.03997781686484814, 'clip_ratio/region_mean': 0.09847479034215212, 'reward_total_mean': 0.331297367811203, 'reward_meter_mean': 0.4340297281742096, 'reward_meter_std': 0.22348229587078094, 'reward_count_adherence_mean': 0.4375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8360930681228638, 'reward_repeat_soft_std': 0.02245735377073288, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.331297367811203, 'reward_total_composite_std': 0.037000950425863266, 'epoch': 0.18}

 55%|█████▌    | 1822/3300 [4:28:43<2:45:47,  6.73s/it]INFO 04-13 11:49:41 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1823/3300 [4:28:55<3:22:47,  8.24s/it]2026-04-13 11:49:53,337 | INFO | train_grpo_train | metrics_logged mode=train step=1823

                                                       
{'loss': -0.0888, 'grad_norm': 2.9576637744903564, 'learning_rate': 4.478787878787879e-06, 'num_tokens': 3265815.0, 'completions/mean_length': 97.0, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.6791916489601135, 'rewards/meter/std': 0.44092419743537903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.904679536819458, 'rewards/repeat_soft/std': 0.05452427640557289, 'rewards/judge_quality/mean': 0.5762500166893005, 'rewards/judge_quality/std': 0.31513887643814087, 'rewards/total_composite/mean': 0.6027641892433167, 'rewards/total_composite/std': 0.3282713294029236, 'reward': 0.6027641892433167, 'reward_std': 0.3282712996006012, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10675515979528427, 'sampling/sampling_logp_difference/max': 1.2832794189453125, 'sampling/importance_sampling_ratio/min': 0.2771269977092743, 'sampling/importance_sampling_ratio/mean': 1.016241431236267, 'sampling/importance_sampling_ratio/max': 1.4655214548110962, 'entropy': 0.7538627684116364, 'clip_ratio/low_mean': 0.020707071293145418, 'clip_ratio/low_min': 0.020707071293145418, 'clip_ratio/high_mean': 0.058742116671055555, 'clip_ratio/high_max': 0.058742116671055555, 'clip_ratio/region_mean': 0.07944918796420097, 'reward_total_mean': 0.6027641892433167, 'reward_meter_mean': 0.6791916489601135, 'reward_meter_std': 0.44092419743537903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.904679536819458, 'reward_repeat_soft_std': 0.05452427640557289, 'reward_judge_quality_mean': 0.5762500166893005, 'reward_judge_quality_std': 0.31513887643814087, 'reward_total_composite_mean': 0.6027641892433167, 'reward_total_composite_std': 0.3282713294029236, 'epoch': 0.18}

 55%|█████▌    | 1823/3300 [4:28:55<3:22:47,  8.24s/it]INFO 04-13 11:49:53 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1824/3300 [4:29:01<3:06:06,  7.57s/it]2026-04-13 11:49:59,348 | INFO | train_grpo_train | metrics_logged mode=train step=1824

                                                       
{'loss': 0.0017, 'grad_norm': 18.037513732910156, 'learning_rate': 4.4757575757575765e-06, 'num_tokens': 3267181.0, 'completions/mean_length': 20.75, 'completions/min_length': 19.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.75, 'completions/min_terminated_length': 19.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.8389955163002014, 'rewards/meter/std': 0.2952535152435303, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9601136445999146, 'rewards/repeat_soft/std': 0.006749649066478014, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5754857063293457, 'rewards/total_composite/std': 0.08214115351438522, 'reward': 0.5754857063293457, 'reward_std': 0.08214113116264343, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13900591433048248, 'sampling/sampling_logp_difference/max': 1.2857437133789062, 'sampling/importance_sampling_ratio/min': 0.2764449119567871, 'sampling/importance_sampling_ratio/mean': 1.0151656866073608, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9926154911518097, 'clip_ratio/low_mean': 0.03157894778996706, 'clip_ratio/low_min': 0.03157894778996706, 'clip_ratio/high_mean': 0.08142809476703405, 'clip_ratio/high_max': 0.08142809476703405, 'clip_ratio/region_mean': 0.11300704255700111, 'reward_total_mean': 0.5754857063293457, 'reward_meter_mean': 0.8389955163002014, 'reward_meter_std': 0.2952535152435303, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9601136445999146, 'reward_repeat_soft_std': 0.006749649066478014, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5754857063293457, 'reward_total_composite_std': 0.08214115351438522, 'epoch': 0.18}

 55%|█████▌    | 1824/3300 [4:29:01<3:06:06,  7.57s/it]INFO 04-13 11:49:59 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1825/3300 [4:29:07<2:57:10,  7.21s/it]2026-04-13 11:50:05,712 | INFO | train_grpo_train | metrics_logged mode=train step=1825

                                                       
{'loss': -0.0043, 'grad_norm': 12.696378707885742, 'learning_rate': 4.472727272727273e-06, 'num_tokens': 3268846.0, 'completions/mean_length': 43.125, 'completions/min_length': 33.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.125, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.650660514831543, 'rewards/meter/std': 0.35668233036994934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9241976141929626, 'rewards/repeat_soft/std': 0.05137051269412041, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5163896083831787, 'rewards/total_composite/std': 0.09794284403324127, 'reward': 0.5163896083831787, 'reward_std': 0.09794283658266068, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11815313994884491, 'sampling/sampling_logp_difference/max': 1.5346651077270508, 'sampling/importance_sampling_ratio/min': 0.21552786231040955, 'sampling/importance_sampling_ratio/mean': 1.007631540298462, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8516337424516678, 'clip_ratio/low_mean': 0.044034091755747795, 'clip_ratio/low_min': 0.044034091755747795, 'clip_ratio/high_mean': 0.051652208901941776, 'clip_ratio/high_max': 0.051652208901941776, 'clip_ratio/region_mean': 0.09568630065768957, 'reward_total_mean': 0.5163896083831787, 'reward_meter_mean': 0.650660514831543, 'reward_meter_std': 0.35668233036994934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9241976141929626, 'reward_repeat_soft_std': 0.05137051269412041, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5163896083831787, 'reward_total_composite_std': 0.09794284403324127, 'epoch': 0.18}

 55%|█████▌    | 1825/3300 [4:29:07<2:57:10,  7.21s/it]INFO 04-13 11:50:05 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1826/3300 [4:29:15<3:01:00,  7.37s/it]2026-04-13 11:50:13,467 | INFO | train_grpo_train | metrics_logged mode=train step=1826

                                                       
{'loss': 0.0322, 'grad_norm': 5.049038410186768, 'learning_rate': 4.46969696969697e-06, 'num_tokens': 3271207.0, 'completions/mean_length': 140.125, 'completions/min_length': 96.0, 'completions/max_length': 199.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 140.125, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 199.0, 'rewards/meter/mean': 0.8653736114501953, 'rewards/meter/std': 0.15277306735515594, 'rewards/count_adherence/mean': 0.40625, 'rewards/count_adherence/std': 0.18600596487522125, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6539048552513123, 'rewards/repeat_soft/std': 0.17207203805446625, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.45973342657089233, 'rewards/total_composite/std': 0.1313636749982834, 'reward': 0.45973342657089233, 'reward_std': 0.1313636600971222, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08705134689807892, 'sampling/sampling_logp_difference/max': 1.5911340713500977, 'sampling/importance_sampling_ratio/min': 0.20369447767734528, 'sampling/importance_sampling_ratio/mean': 1.0095009803771973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6451926082372665, 'clip_ratio/low_mean': 0.038877638056874275, 'clip_ratio/low_min': 0.038877638056874275, 'clip_ratio/high_mean': 0.04204982612282038, 'clip_ratio/high_max': 0.04204982612282038, 'clip_ratio/region_mean': 0.08092746417969465, 'reward_total_mean': 0.45973342657089233, 'reward_meter_mean': 0.8653736114501953, 'reward_meter_std': 0.15277306735515594, 'reward_count_adherence_mean': 0.40625, 'reward_count_adherence_std': 0.18600596487522125, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6539048552513123, 'reward_repeat_soft_std': 0.17207203805446625, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.45973342657089233, 'reward_total_composite_std': 0.1313636749982834, 'epoch': 0.18}

 55%|█████▌    | 1826/3300 [4:29:15<3:01:00,  7.37s/it]INFO 04-13 11:50:13 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1827/3300 [4:29:28<3:40:06,  8.97s/it]2026-04-13 11:50:26,138 | INFO | train_grpo_train | metrics_logged mode=train step=1827

                                                       
{'loss': -0.0921, 'grad_norm': 3.373699426651001, 'learning_rate': 4.4666666666666665e-06, 'num_tokens': 3272922.0, 'completions/mean_length': 129.375, 'completions/min_length': 61.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 74.71428680419922, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.6079522371292114, 'rewards/meter/std': 0.27880096435546875, 'rewards/count_adherence/mean': 0.4583333432674408, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8452585935592651, 'rewards/repeat_soft/std': 0.12623430788516998, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.23445606231689453, 'rewards/total_composite/mean': 0.39055413007736206, 'rewards/total_composite/std': 0.2326713651418686, 'reward': 0.39055413007736206, 'reward_std': 0.23267138004302979, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13240386545658112, 'sampling/sampling_logp_difference/max': 1.4575583934783936, 'sampling/importance_sampling_ratio/min': 0.23280400037765503, 'sampling/importance_sampling_ratio/mean': 1.0285807847976685, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0534615963697433, 'clip_ratio/low_mean': 0.04445145092904568, 'clip_ratio/low_min': 0.04445145092904568, 'clip_ratio/high_mean': 0.0673421062529087, 'clip_ratio/high_max': 0.0673421062529087, 'clip_ratio/region_mean': 0.11179355718195438, 'reward_total_mean': 0.39055413007736206, 'reward_meter_mean': 0.6079522371292114, 'reward_meter_std': 0.27880096435546875, 'reward_count_adherence_mean': 0.4583333432674408, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8452585935592651, 'reward_repeat_soft_std': 0.12623430788516998, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.23445606231689453, 'reward_total_composite_mean': 0.39055413007736206, 'reward_total_composite_std': 0.2326713651418686, 'epoch': 0.18}

 55%|█████▌    | 1827/3300 [4:29:28<3:40:06,  8.97s/it]INFO 04-13 11:50:26 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1828/3300 [4:29:34<3:17:25,  8.05s/it]2026-04-13 11:50:32,037 | INFO | train_grpo_train | metrics_logged mode=train step=1828

                                                       
{'loss': -0.0703, 'grad_norm': 16.006153106689453, 'learning_rate': 4.463636363636364e-06, 'num_tokens': 3274493.0, 'completions/mean_length': 36.375, 'completions/min_length': 28.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.7035261392593384, 'rewards/meter/std': 0.3472166955471039, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9747592806816101, 'rewards/repeat_soft/std': 0.03407462313771248, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5427587032318115, 'rewards/total_composite/std': 0.10208763182163239, 'reward': 0.5427587032318115, 'reward_std': 0.10208763182163239, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11395373940467834, 'sampling/sampling_logp_difference/max': 1.2686762809753418, 'sampling/importance_sampling_ratio/min': 0.28120359778404236, 'sampling/importance_sampling_ratio/mean': 0.9948596358299255, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7612007334828377, 'clip_ratio/low_mean': 0.02616995107382536, 'clip_ratio/low_min': 0.02616995107382536, 'clip_ratio/high_mean': 0.08919057063758373, 'clip_ratio/high_max': 0.08919057063758373, 'clip_ratio/region_mean': 0.11536052171140909, 'reward_total_mean': 0.5427587032318115, 'reward_meter_mean': 0.7035261392593384, 'reward_meter_std': 0.3472166955471039, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9747592806816101, 'reward_repeat_soft_std': 0.03407462313771248, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5427587032318115, 'reward_total_composite_std': 0.10208763182163239, 'epoch': 0.18}

 55%|█████▌    | 1828/3300 [4:29:34<3:17:25,  8.05s/it]INFO 04-13 11:50:32 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1829/3300 [4:29:41<3:14:33,  7.94s/it]2026-04-13 11:50:39,725 | INFO | train_grpo_train | metrics_logged mode=train step=1829

                                                       
{'loss': 0.1141, 'grad_norm': 5.330620288848877, 'learning_rate': 4.460606060606061e-06, 'num_tokens': 3276974.0, 'completions/mean_length': 147.125, 'completions/min_length': 98.0, 'completions/max_length': 193.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 147.125, 'completions/min_terminated_length': 98.0, 'completions/max_terminated_length': 193.0, 'rewards/meter/mean': 0.8941610455513, 'rewards/meter/std': 0.08215868473052979, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9671491384506226, 'rewards/repeat_soft/std': 0.022209011018276215, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5016783475875854, 'rewards/total_composite/std': 0.02898823283612728, 'reward': 0.5016783475875854, 'reward_std': 0.028988230973482132, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14371787011623383, 'sampling/sampling_logp_difference/max': 1.7989435195922852, 'sampling/importance_sampling_ratio/min': 0.16547362506389618, 'sampling/importance_sampling_ratio/mean': 1.0169726610183716, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.2326049730181694, 'clip_ratio/low_mean': 0.08393213618546724, 'clip_ratio/low_min': 0.08393213618546724, 'clip_ratio/high_mean': 0.059747302904725075, 'clip_ratio/high_max': 0.059747302904725075, 'clip_ratio/region_mean': 0.14367943909019232, 'reward_total_mean': 0.5016783475875854, 'reward_meter_mean': 0.8941610455513, 'reward_meter_std': 0.08215868473052979, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9671491384506226, 'reward_repeat_soft_std': 0.022209011018276215, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5016783475875854, 'reward_total_composite_std': 0.02898823283612728, 'epoch': 0.18}

 55%|█████▌    | 1829/3300 [4:29:41<3:14:33,  7.94s/it]INFO 04-13 11:50:39 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1830/3300 [4:29:49<3:15:18,  7.97s/it]2026-04-13 11:50:47,769 | INFO | train_grpo_train | metrics_logged mode=train step=1830

                                                       
{'loss': 0.0691, 'grad_norm': 5.620277404785156, 'learning_rate': 4.4575757575757575e-06, 'num_tokens': 3279432.0, 'completions/mean_length': 136.25, 'completions/min_length': 114.0, 'completions/max_length': 162.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 136.25, 'completions/min_terminated_length': 114.0, 'completions/max_terminated_length': 162.0, 'rewards/meter/mean': 0.7948927879333496, 'rewards/meter/std': 0.18234854936599731, 'rewards/count_adherence/mean': 0.5, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.835483968257904, 'rewards/repeat_soft/std': 0.0281214602291584, 'rewards/judge_quality/mean': 0.4950000047683716, 'rewards/judge_quality/std': 0.13887304067611694, 'rewards/total_composite/mean': 0.48247966170310974, 'rewards/total_composite/std': 0.09574178606271744, 'reward': 0.48247966170310974, 'reward_std': 0.09574179351329803, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09292007237672806, 'sampling/sampling_logp_difference/max': 1.4724364280700684, 'sampling/importance_sampling_ratio/min': 0.2293659746646881, 'sampling/importance_sampling_ratio/mean': 1.0040966272354126, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7820024564862251, 'clip_ratio/low_mean': 0.041064705699682236, 'clip_ratio/low_min': 0.041064705699682236, 'clip_ratio/high_mean': 0.031241677701473236, 'clip_ratio/high_max': 0.031241677701473236, 'clip_ratio/region_mean': 0.07230638340115547, 'reward_total_mean': 0.48247966170310974, 'reward_meter_mean': 0.7948927879333496, 'reward_meter_std': 0.18234854936599731, 'reward_count_adherence_mean': 0.5, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.835483968257904, 'reward_repeat_soft_std': 0.0281214602291584, 'reward_judge_quality_mean': 0.4950000047683716, 'reward_judge_quality_std': 0.13887304067611694, 'reward_total_composite_mean': 0.48247966170310974, 'reward_total_composite_std': 0.09574178606271744, 'epoch': 0.18}

 55%|█████▌    | 1830/3300 [4:29:49<3:15:18,  7.97s/it]INFO 04-13 11:50:48 [block_pool.py:378] Successfully reset prefix cache

 55%|█████▌    | 1831/3300 [4:29:57<3:09:30,  7.74s/it]2026-04-13 11:50:54,975 | INFO | train_grpo_train | metrics_logged mode=train step=1831

                                                       
{'loss': -0.0368, 'grad_norm': 8.392786979675293, 'learning_rate': 4.454545454545455e-06, 'num_tokens': 3281491.0, 'completions/mean_length': 86.375, 'completions/min_length': 72.0, 'completions/max_length': 110.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 86.375, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 110.0, 'rewards/meter/mean': 0.8850648999214172, 'rewards/meter/std': 0.1256731152534485, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8781824111938477, 'rewards/repeat_soft/std': 0.02163693681359291, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.603774905204773, 'rewards/total_composite/std': 0.14991037547588348, 'reward': 0.603774905204773, 'reward_std': 0.14991037547588348, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1255534440279007, 'sampling/sampling_logp_difference/max': 1.7929372787475586, 'sampling/importance_sampling_ratio/min': 0.166470468044281, 'sampling/importance_sampling_ratio/mean': 1.0084514617919922, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9133397415280342, 'clip_ratio/low_mean': 0.07230066228657961, 'clip_ratio/low_min': 0.07230066228657961, 'clip_ratio/high_mean': 0.05087065976113081, 'clip_ratio/high_max': 0.05087065976113081, 'clip_ratio/region_mean': 0.12317132204771042, 'reward_total_mean': 0.603774905204773, 'reward_meter_mean': 0.8850648999214172, 'reward_meter_std': 0.1256731152534485, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8781824111938477, 'reward_repeat_soft_std': 0.02163693681359291, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.603774905204773, 'reward_total_composite_std': 0.14991037547588348, 'epoch': 0.18}

 55%|█████▌    | 1831/3300 [4:29:57<3:09:30,  7.74s/it]INFO 04-13 11:50:55 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1832/3300 [4:30:08<3:39:25,  8.97s/it]2026-04-13 11:51:06,817 | INFO | train_grpo_train | metrics_logged mode=train step=1832

                                                       
{'loss': -0.2005, 'grad_norm': 1.9224159717559814, 'learning_rate': 4.451515151515152e-06, 'num_tokens': 3284013.0, 'completions/mean_length': 177.25, 'completions/min_length': 119.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 129.42857360839844, 'completions/min_terminated_length': 119.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.812292218208313, 'rewards/meter/std': 0.18085992336273193, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8280702233314514, 'rewards/repeat_soft/std': 0.0513884536921978, 'rewards/judge_quality/mean': 0.6700000166893005, 'rewards/judge_quality/std': 0.27255406975746155, 'rewards/total_composite/mean': 0.5601329803466797, 'rewards/total_composite/std': 0.2452675700187683, 'reward': 0.5601329803466797, 'reward_std': 0.24526755511760712, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0855712890625, 'sampling/sampling_logp_difference/max': 1.4115657806396484, 'sampling/importance_sampling_ratio/min': 0.2437613159418106, 'sampling/importance_sampling_ratio/mean': 1.0131275653839111, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5874464139342308, 'clip_ratio/low_mean': 0.010576923377811909, 'clip_ratio/low_min': 0.010576923377811909, 'clip_ratio/high_mean': 0.06200599577277899, 'clip_ratio/high_max': 0.06200599577277899, 'clip_ratio/region_mean': 0.0725829191505909, 'reward_total_mean': 0.5601329803466797, 'reward_meter_mean': 0.812292218208313, 'reward_meter_std': 0.18085992336273193, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8280702233314514, 'reward_repeat_soft_std': 0.0513884536921978, 'reward_judge_quality_mean': 0.6700000166893005, 'reward_judge_quality_std': 0.27255406975746155, 'reward_total_composite_mean': 0.5601329803466797, 'reward_total_composite_std': 0.2452675700187683, 'epoch': 0.18}

 56%|█████▌    | 1832/3300 [4:30:09<3:39:25,  8.97s/it]INFO 04-13 11:51:07 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1833/3300 [4:30:16<3:29:18,  8.56s/it]2026-04-13 11:51:14,419 | INFO | train_grpo_train | metrics_logged mode=train step=1833

                                                       
{'loss': 0.0703, 'grad_norm': 5.307923316955566, 'learning_rate': 4.448484848484848e-06, 'num_tokens': 3286692.0, 'completions/mean_length': 160.875, 'completions/min_length': 129.0, 'completions/max_length': 179.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 160.875, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 179.0, 'rewards/meter/mean': 0.8338100910186768, 'rewards/meter/std': 0.20722128450870514, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7941572070121765, 'rewards/repeat_soft/std': 0.034432124346494675, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.45925372838974, 'rewards/total_composite/std': 0.06867709010839462, 'reward': 0.45925372838974, 'reward_std': 0.06867708265781403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10376652330160141, 'sampling/sampling_logp_difference/max': 1.7389531135559082, 'sampling/importance_sampling_ratio/min': 0.1757042557001114, 'sampling/importance_sampling_ratio/mean': 1.016282558441162, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9125862717628479, 'clip_ratio/low_mean': 0.031769067980349064, 'clip_ratio/low_min': 0.031769067980349064, 'clip_ratio/high_mean': 0.04758817981928587, 'clip_ratio/high_max': 0.04758817981928587, 'clip_ratio/region_mean': 0.07935724779963493, 'reward_total_mean': 0.45925372838974, 'reward_meter_mean': 0.8338100910186768, 'reward_meter_std': 0.20722128450870514, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7941572070121765, 'reward_repeat_soft_std': 0.034432124346494675, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.45925372838974, 'reward_total_composite_std': 0.06867709010839462, 'epoch': 0.18}

 56%|█████▌    | 1833/3300 [4:30:16<3:29:18,  8.56s/it]INFO 04-13 11:51:14 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1834/3300 [4:30:22<3:12:05,  7.86s/it]2026-04-13 11:51:20,652 | INFO | train_grpo_train | metrics_logged mode=train step=1834

                                                       
{'loss': -0.0049, 'grad_norm': 11.121776580810547, 'learning_rate': 4.445454545454546e-06, 'num_tokens': 3288381.0, 'completions/mean_length': 48.125, 'completions/min_length': 43.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.8402247428894043, 'rewards/meter/std': 0.3212410509586334, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.99656742811203, 'rewards/repeat_soft/std': 0.005215909332036972, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.1348809152841568, 'rewards/total_composite/mean': 0.6299073696136475, 'rewards/total_composite/std': 0.14270322024822235, 'reward': 0.6299073696136475, 'reward_std': 0.14270322024822235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.143593430519104, 'sampling/sampling_logp_difference/max': 1.4234967231750488, 'sampling/importance_sampling_ratio/min': 0.24087028205394745, 'sampling/importance_sampling_ratio/mean': 1.0094298124313354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.3203355967998505, 'clip_ratio/low_mean': 0.07586692273616791, 'clip_ratio/low_min': 0.07586692273616791, 'clip_ratio/high_mean': 0.05965951643884182, 'clip_ratio/high_max': 0.05965951643884182, 'clip_ratio/region_mean': 0.13552643917500973, 'reward_total_mean': 0.6299073696136475, 'reward_meter_mean': 0.8402247428894043, 'reward_meter_std': 0.3212410509586334, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.99656742811203, 'reward_repeat_soft_std': 0.005215909332036972, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.1348809152841568, 'reward_total_composite_mean': 0.6299073696136475, 'reward_total_composite_std': 0.14270322024822235, 'epoch': 0.18}

 56%|█████▌    | 1834/3300 [4:30:22<3:12:05,  7.86s/it]INFO 04-13 11:51:20 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1835/3300 [4:30:34<3:39:52,  9.01s/it]2026-04-13 11:51:32,330 | INFO | train_grpo_train | metrics_logged mode=train step=1835

                                                       
{'loss': -0.1881, 'grad_norm': 1.684330701828003, 'learning_rate': 4.442424242424243e-06, 'num_tokens': 3291010.0, 'completions/mean_length': 278.625, 'completions/min_length': 149.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 200.83334350585938, 'completions/min_terminated_length': 149.0, 'completions/max_terminated_length': 235.0, 'rewards/meter/mean': 0.635834813117981, 'rewards/meter/std': 0.24058298766613007, 'rewards/count_adherence/mean': 0.5625, 'rewards/count_adherence/std': 0.08625820279121399, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7982053756713867, 'rewards/repeat_soft/std': 0.06978677213191986, 'rewards/judge_quality/mean': 0.34375, 'rewards/judge_quality/std': 0.19427062571048737, 'rewards/total_composite/mean': 0.34005433320999146, 'rewards/total_composite/std': 0.16458089649677277, 'reward': 0.34005433320999146, 'reward_std': 0.16458088159561157, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09030982106924057, 'sampling/sampling_logp_difference/max': 1.9321351051330566, 'sampling/importance_sampling_ratio/min': 0.1961755007505417, 'sampling/importance_sampling_ratio/mean': 1.0259186029434204, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5805070772767067, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.053594366647303104, 'clip_ratio/high_max': 0.053594366647303104, 'clip_ratio/region_mean': 0.06117012444883585, 'reward_total_mean': 0.34005433320999146, 'reward_meter_mean': 0.635834813117981, 'reward_meter_std': 0.24058298766613007, 'reward_count_adherence_mean': 0.5625, 'reward_count_adherence_std': 0.08625820279121399, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7982053756713867, 'reward_repeat_soft_std': 0.06978677213191986, 'reward_judge_quality_mean': 0.34375, 'reward_judge_quality_std': 0.19427062571048737, 'reward_total_composite_mean': 0.34005433320999146, 'reward_total_composite_std': 0.16458089649677277, 'epoch': 0.18}

 56%|█████▌    | 1835/3300 [4:30:34<3:39:52,  9.01s/it]INFO 04-13 11:51:32 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1836/3300 [4:30:40<3:21:40,  8.27s/it]2026-04-13 11:51:38,863 | INFO | train_grpo_train | metrics_logged mode=train step=1836

                                                       
{'loss': 0.0789, 'grad_norm': 7.766142845153809, 'learning_rate': 4.43939393939394e-06, 'num_tokens': 3292684.0, 'completions/mean_length': 47.25, 'completions/min_length': 40.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.25, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8402402400970459, 'rewards/meter/std': 0.21419556438922882, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9650239944458008, 'rewards/repeat_soft/std': 0.03594968095421791, 'rewards/judge_quality/mean': 0.5562499761581421, 'rewards/judge_quality/std': 0.2249404937028885, 'rewards/total_composite/mean': 0.6383761167526245, 'rewards/total_composite/std': 0.12639735639095306, 'reward': 0.6383761167526245, 'reward_std': 0.12639737129211426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1247914731502533, 'sampling/sampling_logp_difference/max': 1.0808696746826172, 'sampling/importance_sampling_ratio/min': 0.3393003046512604, 'sampling/importance_sampling_ratio/mean': 1.0129711627960205, 'sampling/importance_sampling_ratio/max': 1.6747112274169922, 'entropy': 0.9699152186512947, 'clip_ratio/low_mean': 0.07616180088371038, 'clip_ratio/low_min': 0.07616180088371038, 'clip_ratio/high_mean': 0.009146341122686863, 'clip_ratio/high_max': 0.009146341122686863, 'clip_ratio/region_mean': 0.08530814200639725, 'reward_total_mean': 0.6383761167526245, 'reward_meter_mean': 0.8402402400970459, 'reward_meter_std': 0.21419556438922882, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9650239944458008, 'reward_repeat_soft_std': 0.03594968095421791, 'reward_judge_quality_mean': 0.5562499761581421, 'reward_judge_quality_std': 0.2249404937028885, 'reward_total_composite_mean': 0.6383761167526245, 'reward_total_composite_std': 0.12639735639095306, 'epoch': 0.18}

 56%|█████▌    | 1836/3300 [4:30:41<3:21:40,  8.27s/it]INFO 04-13 11:51:39 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1837/3300 [4:30:48<3:17:43,  8.11s/it]2026-04-13 11:51:46,617 | INFO | train_grpo_train | metrics_logged mode=train step=1837

                                                       
{'loss': 0.1382, 'grad_norm': 6.6513566970825195, 'learning_rate': 4.436363636363637e-06, 'num_tokens': 3295231.0, 'completions/mean_length': 126.375, 'completions/min_length': 86.0, 'completions/max_length': 161.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 126.375, 'completions/min_terminated_length': 86.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.801716685295105, 'rewards/meter/std': 0.24808786809444427, 'rewards/count_adherence/mean': 0.6875, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8516167402267456, 'rewards/repeat_soft/std': 0.04234924167394638, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.4964723587036133, 'rewards/total_composite/std': 0.07371412217617035, 'reward': 0.4964723587036133, 'reward_std': 0.07371412217617035, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09954273700714111, 'sampling/sampling_logp_difference/max': 2.240025758743286, 'sampling/importance_sampling_ratio/min': 0.10645575821399689, 'sampling/importance_sampling_ratio/mean': 1.008533239364624, 'sampling/importance_sampling_ratio/max': 1.911268711090088, 'entropy': 0.7372321635484695, 'clip_ratio/low_mean': 0.02968578739091754, 'clip_ratio/low_min': 0.02968578739091754, 'clip_ratio/high_mean': 0.07113707065582275, 'clip_ratio/high_max': 0.07113707065582275, 'clip_ratio/region_mean': 0.1008228580467403, 'reward_total_mean': 0.4964723587036133, 'reward_meter_mean': 0.801716685295105, 'reward_meter_std': 0.24808786809444427, 'reward_count_adherence_mean': 0.6875, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8516167402267456, 'reward_repeat_soft_std': 0.04234924167394638, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.4964723587036133, 'reward_total_composite_std': 0.07371412217617035, 'epoch': 0.18}

 56%|█████▌    | 1837/3300 [4:30:48<3:17:43,  8.11s/it]INFO 04-13 11:51:46 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1838/3300 [4:30:54<3:02:20,  7.48s/it]2026-04-13 11:51:52,656 | INFO | train_grpo_train | metrics_logged mode=train step=1838

                                                       
{'loss': -0.1061, 'grad_norm': 12.099515914916992, 'learning_rate': 4.433333333333334e-06, 'num_tokens': 3296819.0, 'completions/mean_length': 35.5, 'completions/min_length': 31.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.5, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.6132403612136841, 'rewards/meter/std': 0.41704481840133667, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9716784954071045, 'rewards/repeat_soft/std': 0.03539866581559181, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.2822834253311157, 'rewards/total_composite/mean': 0.5992453098297119, 'rewards/total_composite/std': 0.20921917259693146, 'reward': 0.5992453098297119, 'reward_std': 0.20921917259693146, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11490785330533981, 'sampling/sampling_logp_difference/max': 1.231729507446289, 'sampling/importance_sampling_ratio/min': 0.2917875051498413, 'sampling/importance_sampling_ratio/mean': 0.9885565638542175, 'sampling/importance_sampling_ratio/max': 1.5996848344802856, 'entropy': 0.64373379945755, 'clip_ratio/low_mean': 0.046913184225559235, 'clip_ratio/low_min': 0.046913184225559235, 'clip_ratio/high_mean': 0.06361937709152699, 'clip_ratio/high_max': 0.06361937709152699, 'clip_ratio/region_mean': 0.11053256131708622, 'reward_total_mean': 0.5992453098297119, 'reward_meter_mean': 0.6132403612136841, 'reward_meter_std': 0.41704481840133667, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9716784954071045, 'reward_repeat_soft_std': 0.03539866581559181, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.2822834253311157, 'reward_total_composite_mean': 0.5992453098297119, 'reward_total_composite_std': 0.20921917259693146, 'epoch': 0.18}

 56%|█████▌    | 1838/3300 [4:30:54<3:02:20,  7.48s/it]INFO 04-13 11:51:52 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1839/3300 [4:31:01<2:53:21,  7.12s/it]2026-04-13 11:51:58,910 | INFO | train_grpo_train | metrics_logged mode=train step=1839

                                                       
{'loss': 0.0012, 'grad_norm': 13.442041397094727, 'learning_rate': 4.430303030303031e-06, 'num_tokens': 3298290.0, 'completions/mean_length': 39.875, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.875, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9158987402915955, 'rewards/meter/std': 0.17611320316791534, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8697051405906677, 'rewards/repeat_soft/std': 0.03761998936533928, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.6640946865081787, 'rewards/total_composite/std': 0.1221679225564003, 'reward': 0.6640946865081787, 'reward_std': 0.1221679151058197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11774171888828278, 'sampling/sampling_logp_difference/max': 1.8575263023376465, 'sampling/importance_sampling_ratio/min': 0.1560581922531128, 'sampling/importance_sampling_ratio/mean': 1.017078161239624, 'sampling/importance_sampling_ratio/max': 1.6088244915008545, 'entropy': 0.8993172124028206, 'clip_ratio/low_mean': 0.04582105437293649, 'clip_ratio/low_min': 0.04582105437293649, 'clip_ratio/high_mean': 0.019720497075468302, 'clip_ratio/high_max': 0.019720497075468302, 'clip_ratio/region_mean': 0.06554155144840479, 'reward_total_mean': 0.6640946865081787, 'reward_meter_mean': 0.9158987402915955, 'reward_meter_std': 0.17611320316791534, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8697051405906677, 'reward_repeat_soft_std': 0.03761998936533928, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.6640946865081787, 'reward_total_composite_std': 0.1221679225564003, 'epoch': 0.18}

 56%|█████▌    | 1839/3300 [4:31:01<2:53:21,  7.12s/it]INFO 04-13 11:51:59 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1840/3300 [4:31:12<3:25:28,  8.44s/it]2026-04-13 11:52:10,449 | INFO | train_grpo_train | metrics_logged mode=train step=1840

                                                       
{'loss': -0.229, 'grad_norm': 1.546162486076355, 'learning_rate': 4.4272727272727275e-06, 'num_tokens': 3301079.0, 'completions/mean_length': 204.625, 'completions/min_length': 134.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 160.71429443359375, 'completions/min_terminated_length': 134.0, 'completions/max_terminated_length': 207.0, 'rewards/meter/mean': 0.9185236692428589, 'rewards/meter/std': 0.1468735933303833, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7914280891418457, 'rewards/repeat_soft/std': 0.046963345259428024, 'rewards/judge_quality/mean': 0.3824999928474426, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.45604652166366577, 'rewards/total_composite/std': 0.18849195539951324, 'reward': 0.45604652166366577, 'reward_std': 0.18849195539951324, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08804997056722641, 'sampling/sampling_logp_difference/max': 2.1487064361572266, 'sampling/importance_sampling_ratio/min': 0.11663493514060974, 'sampling/importance_sampling_ratio/mean': 1.0134732723236084, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5453835353255272, 'clip_ratio/low_mean': 0.006410256493836641, 'clip_ratio/low_min': 0.006410256493836641, 'clip_ratio/high_mean': 0.06665645446628332, 'clip_ratio/high_max': 0.06665645446628332, 'clip_ratio/region_mean': 0.07306671096011996, 'reward_total_mean': 0.45604652166366577, 'reward_meter_mean': 0.9185236692428589, 'reward_meter_std': 0.1468735933303833, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7914280891418457, 'reward_repeat_soft_std': 0.046963345259428024, 'reward_judge_quality_mean': 0.3824999928474426, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.45604652166366577, 'reward_total_composite_std': 0.18849195539951324, 'epoch': 0.18}

 56%|█████▌    | 1840/3300 [4:31:12<3:25:28,  8.44s/it]INFO 04-13 11:52:10 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1841/3300 [4:31:19<3:11:26,  7.87s/it]2026-04-13 11:52:16,994 | INFO | train_grpo_train | metrics_logged mode=train step=1841

                                                       
{'loss': -0.0022, 'grad_norm': 11.853466987609863, 'learning_rate': 4.424242424242425e-06, 'num_tokens': 3302539.0, 'completions/mean_length': 33.5, 'completions/min_length': 28.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.5, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9478296041488647, 'rewards/meter/std': 0.06476016342639923, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9152984023094177, 'rewards/repeat_soft/std': 0.05110251531004906, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.640056848526001, 'rewards/total_composite/std': 0.11371893435716629, 'reward': 0.640056848526001, 'reward_std': 0.1137189269065857, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1272178441286087, 'sampling/sampling_logp_difference/max': 1.722355842590332, 'sampling/importance_sampling_ratio/min': 0.1786447912454605, 'sampling/importance_sampling_ratio/mean': 1.0080894231796265, 'sampling/importance_sampling_ratio/max': 1.6141676902770996, 'entropy': 0.9794385582208633, 'clip_ratio/low_mean': 0.08459542435593903, 'clip_ratio/low_min': 0.08459542435593903, 'clip_ratio/high_mean': 0.007575757801532745, 'clip_ratio/high_max': 0.007575757801532745, 'clip_ratio/region_mean': 0.09217118215747178, 'reward_total_mean': 0.640056848526001, 'reward_meter_mean': 0.9478296041488647, 'reward_meter_std': 0.06476016342639923, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9152984023094177, 'reward_repeat_soft_std': 0.05110251531004906, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.640056848526001, 'reward_total_composite_std': 0.11371893435716629, 'epoch': 0.18}

 56%|█████▌    | 1841/3300 [4:31:19<3:11:26,  7.87s/it]INFO 04-13 11:52:17 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1842/3300 [4:31:30<3:39:43,  9.04s/it]2026-04-13 11:52:28,749 | INFO | train_grpo_train | metrics_logged mode=train step=1842

                                                       
{'loss': -0.2157, 'grad_norm': 1.6094402074813843, 'learning_rate': 4.421212121212122e-06, 'num_tokens': 3305073.0, 'completions/mean_length': 206.75, 'completions/min_length': 131.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 163.1428680419922, 'completions/min_terminated_length': 131.0, 'completions/max_terminated_length': 182.0, 'rewards/meter/mean': 0.837173581123352, 'rewards/meter/std': 0.2746213674545288, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.853219747543335, 'rewards/repeat_soft/std': 0.07812196016311646, 'rewards/judge_quality/mean': 0.4737499952316284, 'rewards/judge_quality/std': 0.254611074924469, 'rewards/total_composite/mean': 0.521868109703064, 'rewards/total_composite/std': 0.2212262898683548, 'reward': 0.521868109703064, 'reward_std': 0.2212262898683548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11141830682754517, 'sampling/sampling_logp_difference/max': 1.8805809020996094, 'sampling/importance_sampling_ratio/min': 0.15250149369239807, 'sampling/importance_sampling_ratio/mean': 1.0141204595565796, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7322415187954903, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.0984646612778306, 'clip_ratio/high_max': 0.0984646612778306, 'clip_ratio/region_mean': 0.0984646612778306, 'reward_total_mean': 0.521868109703064, 'reward_meter_mean': 0.837173581123352, 'reward_meter_std': 0.2746213674545288, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.853219747543335, 'reward_repeat_soft_std': 0.07812196016311646, 'reward_judge_quality_mean': 0.4737499952316284, 'reward_judge_quality_std': 0.254611074924469, 'reward_total_composite_mean': 0.521868109703064, 'reward_total_composite_std': 0.2212262898683548, 'epoch': 0.19}

 56%|█████▌    | 1842/3300 [4:31:30<3:39:43,  9.04s/it]INFO 04-13 11:52:29 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1843/3300 [4:31:37<3:22:49,  8.35s/it]2026-04-13 11:52:35,499 | INFO | train_grpo_train | metrics_logged mode=train step=1843

                                                       
{'loss': 0.0401, 'grad_norm': 9.5546236038208, 'learning_rate': 4.418181818181818e-06, 'num_tokens': 3306753.0, 'completions/mean_length': 34.0, 'completions/min_length': 32.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.0, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.937764048576355, 'rewards/meter/std': 0.04446383938193321, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8251459002494812, 'rewards/repeat_soft/std': 0.07597748935222626, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5821202993392944, 'rewards/total_composite/std': 0.01404713373631239, 'reward': 0.5821202993392944, 'reward_std': 0.014047140255570412, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05521784722805023, 'sampling/sampling_logp_difference/max': 1.006291151046753, 'sampling/importance_sampling_ratio/min': 0.42414647340774536, 'sampling/importance_sampling_ratio/mean': 1.0107085704803467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34332825243473053, 'clip_ratio/low_mean': 0.024603174766525626, 'clip_ratio/low_min': 0.024603174766525626, 'clip_ratio/high_mean': 0.04203570680692792, 'clip_ratio/high_max': 0.04203570680692792, 'clip_ratio/region_mean': 0.06663888157345355, 'reward_total_mean': 0.5821202993392944, 'reward_meter_mean': 0.937764048576355, 'reward_meter_std': 0.04446383938193321, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8251459002494812, 'reward_repeat_soft_std': 0.07597748935222626, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5821202993392944, 'reward_total_composite_std': 0.01404713373631239, 'epoch': 0.19}

 56%|█████▌    | 1843/3300 [4:31:37<3:22:49,  8.35s/it]INFO 04-13 11:52:35 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1844/3300 [4:31:44<3:13:51,  7.99s/it]2026-04-13 11:52:42,624 | INFO | train_grpo_train | metrics_logged mode=train step=1844

                                                       
{'loss': 0.0632, 'grad_norm': 6.3035569190979, 'learning_rate': 4.415151515151516e-06, 'num_tokens': 3309227.0, 'completions/mean_length': 117.25, 'completions/min_length': 89.0, 'completions/max_length': 142.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 117.25, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.9056618809700012, 'rewards/meter/std': 0.11452306807041168, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7821943759918213, 'rewards/repeat_soft/std': 0.04733280465006828, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5445857048034668, 'rewards/total_composite/std': 0.07488405704498291, 'reward': 0.5445857048034668, 'reward_std': 0.07488404214382172, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08216746896505356, 'sampling/sampling_logp_difference/max': 1.2165040969848633, 'sampling/importance_sampling_ratio/min': 0.29626408219337463, 'sampling/importance_sampling_ratio/mean': 1.0209424495697021, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6338837668299675, 'clip_ratio/low_mean': 0.06343623623251915, 'clip_ratio/low_min': 0.06343623623251915, 'clip_ratio/high_mean': 0.01123595517128706, 'clip_ratio/high_max': 0.01123595517128706, 'clip_ratio/region_mean': 0.07467219140380621, 'reward_total_mean': 0.5445857048034668, 'reward_meter_mean': 0.9056618809700012, 'reward_meter_std': 0.11452306807041168, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7821943759918213, 'reward_repeat_soft_std': 0.04733280465006828, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5445857048034668, 'reward_total_composite_std': 0.07488405704498291, 'epoch': 0.19}

 56%|█████▌    | 1844/3300 [4:31:44<3:13:51,  7.99s/it]INFO 04-13 11:52:42 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1845/3300 [4:31:56<3:39:06,  9.04s/it]2026-04-13 11:52:54,107 | INFO | train_grpo_train | metrics_logged mode=train step=1845

                                                       
{'loss': -0.0873, 'grad_norm': 3.8071818351745605, 'learning_rate': 4.412121212121213e-06, 'num_tokens': 3310756.0, 'completions/mean_length': 97.125, 'completions/min_length': 30.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.85714340209961, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.7326558232307434, 'rewards/meter/std': 0.3586273789405823, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9280301928520203, 'rewards/repeat_soft/std': 0.04183439910411835, 'rewards/judge_quality/mean': 0.7487499713897705, 'rewards/judge_quality/std': 0.332154780626297, 'rewards/total_composite/mean': 0.7463199496269226, 'rewards/total_composite/std': 0.22930696606636047, 'reward': 0.7463199496269226, 'reward_std': 0.22930695116519928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12542758882045746, 'sampling/sampling_logp_difference/max': 1.4823064804077148, 'sampling/importance_sampling_ratio/min': 0.2271132618188858, 'sampling/importance_sampling_ratio/mean': 1.0407088994979858, 'sampling/importance_sampling_ratio/max': 1.644066572189331, 'entropy': 0.9438460394740105, 'clip_ratio/low_mean': 0.013095238711684942, 'clip_ratio/low_min': 0.013095238711684942, 'clip_ratio/high_mean': 0.03941104235127568, 'clip_ratio/high_max': 0.03941104235127568, 'clip_ratio/region_mean': 0.052506281062960625, 'reward_total_mean': 0.7463199496269226, 'reward_meter_mean': 0.7326558232307434, 'reward_meter_std': 0.3586273789405823, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9280301928520203, 'reward_repeat_soft_std': 0.04183439910411835, 'reward_judge_quality_mean': 0.7487499713897705, 'reward_judge_quality_std': 0.332154780626297, 'reward_total_composite_mean': 0.7463199496269226, 'reward_total_composite_std': 0.22930696606636047, 'epoch': 0.19}

 56%|█████▌    | 1845/3300 [4:31:56<3:39:06,  9.04s/it]INFO 04-13 11:52:54 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1846/3300 [4:32:08<4:00:51,  9.94s/it]2026-04-13 11:53:06,163 | INFO | train_grpo_train | metrics_logged mode=train step=1846

                                                       
{'loss': -0.2039, 'grad_norm': 2.956387996673584, 'learning_rate': 4.409090909090909e-06, 'num_tokens': 3313146.0, 'completions/mean_length': 162.75, 'completions/min_length': 82.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 112.85714721679688, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.6697490215301514, 'rewards/meter/std': 0.22099152207374573, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7872285842895508, 'rewards/repeat_soft/std': 0.07616162300109863, 'rewards/judge_quality/mean': 0.6237499713897705, 'rewards/judge_quality/std': 0.33907175064086914, 'rewards/total_composite/mean': 0.5034215450286865, 'rewards/total_composite/std': 0.2599618434906006, 'reward': 0.5034215450286865, 'reward_std': 0.2599618434906006, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08168012648820877, 'sampling/sampling_logp_difference/max': 2.0867412090301514, 'sampling/importance_sampling_ratio/min': 0.12409086525440216, 'sampling/importance_sampling_ratio/mean': 1.0142415761947632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49081745743751526, 'clip_ratio/low_mean': 0.01892037340439856, 'clip_ratio/low_min': 0.01892037340439856, 'clip_ratio/high_mean': 0.04009432205930352, 'clip_ratio/high_max': 0.04009432205930352, 'clip_ratio/region_mean': 0.05901469546370208, 'reward_total_mean': 0.5034215450286865, 'reward_meter_mean': 0.6697490215301514, 'reward_meter_std': 0.22099152207374573, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7872285842895508, 'reward_repeat_soft_std': 0.07616162300109863, 'reward_judge_quality_mean': 0.6237499713897705, 'reward_judge_quality_std': 0.33907175064086914, 'reward_total_composite_mean': 0.5034215450286865, 'reward_total_composite_std': 0.2599618434906006, 'epoch': 0.19}

 56%|█████▌    | 1846/3300 [4:32:08<4:00:51,  9.94s/it]INFO 04-13 11:53:06 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1847/3300 [4:32:14<3:37:11,  8.97s/it]2026-04-13 11:53:12,869 | INFO | train_grpo_train | metrics_logged mode=train step=1847

                                                       
{'loss': 0.0214, 'grad_norm': 5.1473188400268555, 'learning_rate': 4.4060606060606066e-06, 'num_tokens': 3315028.0, 'completions/mean_length': 74.25, 'completions/min_length': 60.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.25, 'completions/min_terminated_length': 60.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8777709007263184, 'rewards/meter/std': 0.18060046434402466, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8241331577301025, 'rewards/repeat_soft/std': 0.03560381755232811, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5443388223648071, 'rewards/total_composite/std': 0.12876400351524353, 'reward': 0.5443388223648071, 'reward_std': 0.12876400351524353, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05746019259095192, 'sampling/sampling_logp_difference/max': 1.274916410446167, 'sampling/importance_sampling_ratio/min': 0.2794543206691742, 'sampling/importance_sampling_ratio/mean': 1.001806378364563, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.26211889274418354, 'clip_ratio/low_mean': 0.023142279242165387, 'clip_ratio/low_min': 0.023142279242165387, 'clip_ratio/high_mean': 0.012660256586968899, 'clip_ratio/high_max': 0.012660256586968899, 'clip_ratio/region_mean': 0.035802535829134285, 'reward_total_mean': 0.5443388223648071, 'reward_meter_mean': 0.8777709007263184, 'reward_meter_std': 0.18060046434402466, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8241331577301025, 'reward_repeat_soft_std': 0.03560381755232811, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5443388223648071, 'reward_total_composite_std': 0.12876400351524353, 'epoch': 0.19}

 56%|█████▌    | 1847/3300 [4:32:15<3:37:11,  8.97s/it]INFO 04-13 11:53:13 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1848/3300 [4:32:20<3:13:04,  7.98s/it]2026-04-13 11:53:18,535 | INFO | train_grpo_train | metrics_logged mode=train step=1848

                                                       
{'loss': -0.0234, 'grad_norm': 13.755486488342285, 'learning_rate': 4.403030303030304e-06, 'num_tokens': 3316386.0, 'completions/mean_length': 16.75, 'completions/min_length': 15.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.75, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.4936433434486389, 'rewards/meter/std': 0.49729251861572266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9505281448364258, 'rewards/repeat_soft/std': 0.017931431531906128, 'rewards/judge_quality/mean': 0.48624998331069946, 'rewards/judge_quality/std': 0.1755755990743637, 'rewards/total_composite/mean': 0.5175467133522034, 'rewards/total_composite/std': 0.2085137963294983, 'reward': 0.5175467133522034, 'reward_std': 0.2085137963294983, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07921064645051956, 'sampling/sampling_logp_difference/max': 0.8258018493652344, 'sampling/importance_sampling_ratio/min': 0.43788373470306396, 'sampling/importance_sampling_ratio/mean': 1.0038883686065674, 'sampling/importance_sampling_ratio/max': 1.4596023559570312, 'entropy': 0.6067915931344032, 'clip_ratio/low_mean': 0.047222224064171314, 'clip_ratio/low_min': 0.047222224064171314, 'clip_ratio/high_mean': 0.045833335258066654, 'clip_ratio/high_max': 0.045833335258066654, 'clip_ratio/region_mean': 0.09305555932223797, 'reward_total_mean': 0.5175467133522034, 'reward_meter_mean': 0.4936433434486389, 'reward_meter_std': 0.49729251861572266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9505281448364258, 'reward_repeat_soft_std': 0.017931431531906128, 'reward_judge_quality_mean': 0.48624998331069946, 'reward_judge_quality_std': 0.1755755990743637, 'reward_total_composite_mean': 0.5175467133522034, 'reward_total_composite_std': 0.2085137963294983, 'epoch': 0.19}

 56%|█████▌    | 1848/3300 [4:32:20<3:13:04,  7.98s/it]INFO 04-13 11:53:18 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1849/3300 [4:32:27<3:01:39,  7.51s/it]2026-04-13 11:53:24,955 | INFO | train_grpo_train | metrics_logged mode=train step=1849

                                                       
{'loss': 0.0543, 'grad_norm': 11.871057510375977, 'learning_rate': 4.4e-06, 'num_tokens': 3318082.0, 'completions/mean_length': 50.0, 'completions/min_length': 37.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.9819462299346924, 'rewards/meter/std': 0.009022936224937439, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.2314550280570984, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9770676493644714, 'rewards/repeat_soft/std': 0.019794698804616928, 'rewards/judge_quality/mean': 0.5862500071525574, 'rewards/judge_quality/std': 0.22984081506729126, 'rewards/total_composite/mean': 0.6959177255630493, 'rewards/total_composite/std': 0.15842773020267487, 'reward': 0.6959177255630493, 'reward_std': 0.15842771530151367, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14248159527778625, 'sampling/sampling_logp_difference/max': 1.2809314727783203, 'sampling/importance_sampling_ratio/min': 0.2777784466743469, 'sampling/importance_sampling_ratio/mean': 1.0329411029815674, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1133587881922722, 'clip_ratio/low_mean': 0.0994670232757926, 'clip_ratio/low_min': 0.0994670232757926, 'clip_ratio/high_mean': 0.05312500149011612, 'clip_ratio/high_max': 0.05312500149011612, 'clip_ratio/region_mean': 0.15259202476590872, 'reward_total_mean': 0.6959177255630493, 'reward_meter_mean': 0.9819462299346924, 'reward_meter_std': 0.009022936224937439, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.2314550280570984, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9770676493644714, 'reward_repeat_soft_std': 0.019794698804616928, 'reward_judge_quality_mean': 0.5862500071525574, 'reward_judge_quality_std': 0.22984081506729126, 'reward_total_composite_mean': 0.6959177255630493, 'reward_total_composite_std': 0.15842773020267487, 'epoch': 0.19}

 56%|█████▌    | 1849/3300 [4:32:27<3:01:39,  7.51s/it]INFO 04-13 11:53:25 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1850/3300 [4:32:38<3:27:59,  8.61s/it]2026-04-13 11:53:36,104 | INFO | train_grpo_train | metrics_logged mode=train step=1850

                                                       
{'loss': -0.2215, 'grad_norm': 1.7200616598129272, 'learning_rate': 4.3969696969696975e-06, 'num_tokens': 3320144.0, 'completions/mean_length': 225.75, 'completions/min_length': 83.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 130.33334350585938, 'completions/min_terminated_length': 83.0, 'completions/max_terminated_length': 184.0, 'rewards/meter/mean': 0.9502502679824829, 'rewards/meter/std': 0.07333400100469589, 'rewards/count_adherence/mean': 0.59375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8547098636627197, 'rewards/repeat_soft/std': 0.04857774078845978, 'rewards/judge_quality/mean': 0.3399999737739563, 'rewards/judge_quality/std': 0.15052290260791779, 'rewards/total_composite/mean': 0.3778553009033203, 'rewards/total_composite/std': 0.2338632047176361, 'reward': 0.3778553009033203, 'reward_std': 0.2338632047176361, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11060090363025665, 'sampling/sampling_logp_difference/max': 1.5649824142456055, 'sampling/importance_sampling_ratio/min': 0.2090916782617569, 'sampling/importance_sampling_ratio/mean': 1.007638692855835, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.569040559232235, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07817493844777346, 'clip_ratio/high_max': 0.07817493844777346, 'clip_ratio/region_mean': 0.07817493844777346, 'reward_total_mean': 0.3778553009033203, 'reward_meter_mean': 0.9502502679824829, 'reward_meter_std': 0.07333400100469589, 'reward_count_adherence_mean': 0.59375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8547098636627197, 'reward_repeat_soft_std': 0.04857774078845978, 'reward_judge_quality_mean': 0.3399999737739563, 'reward_judge_quality_std': 0.15052290260791779, 'reward_total_composite_mean': 0.3778553009033203, 'reward_total_composite_std': 0.2338632047176361, 'epoch': 0.19}

 56%|█████▌    | 1850/3300 [4:32:38<3:27:59,  8.61s/it]INFO 04-13 11:53:36 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:18,  2.28s/it]

 30%|███       | 3/10 [00:10<00:25,  3.62s/it]

 40%|████      | 4/10 [00:13<00:21,  3.64s/it]

 50%|█████     | 5/10 [00:18<00:20,  4.11s/it]

 60%|██████    | 6/10 [00:27<00:22,  5.67s/it]

 70%|███████   | 7/10 [00:36<00:20,  6.75s/it]

 80%|████████  | 8/10 [00:42<00:13,  6.65s/it]

 90%|█████████ | 9/10 [00:48<00:06,  6.15s/it]

100%|██████████| 10/10 [00:51<00:00,  5.36s/it]2026-04-13 11:54:32,513 | INFO | train_grpo_train | metrics_logged mode=eval step=1850

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 56.3018, 'eval_samples_per_second': 1.421, 'eval_steps_per_second': 0.178, 'eval_num_tokens': 3320144.0, 'eval_completions/mean_length': 102.0125, 'eval_completions/min_length': 29.9, 'eval_completions/max_length': 240.2, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 91.89464416503907, 'eval_completions/min_terminated_length': 29.9, 'eval_completions/max_terminated_length': 170.3, 'eval_rewards/meter/mean': 0.8248995125293732, 'eval_rewards/meter/std': 0.23458105847239494, 'eval_rewards/count_adherence/mean': 0.8335416734218597, 'eval_rewards/count_adherence/std': 0.13577206507325174, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.8762265861034393, 'eval_rewards/repeat_soft/std': 0.09761511832475663, 'eval_rewards/judge_quality/mean': 0.4736250102519989, 'eval_rewards/judge_quality/std': 0.20385925192385912, 'eval_rewards/total_composite/mean': 0.5502476394176483, 'eval_rewards/total_composite/std': 0.1676765900105238, 'eval_reward': 0.5502476394176483, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.06305473819375038, 'eval_sampling/sampling_logp_difference/max': 1.197375774383545, 'eval_sampling/importance_sampling_ratio/min': 0.31693738251924514, 'eval_sampling/importance_sampling_ratio/mean': 1.0191887259483337, 'eval_sampling/importance_sampling_ratio/max': 1.457780361175537, 'eval_entropy': 0.6957984805107117, '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': 0.5502476394176483, 'eval_reward_meter_mean': 0.8248995125293732, 'eval_reward_meter_std': 0.23458105847239494, 'eval_reward_count_adherence_mean': 0.8335416734218597, 'eval_reward_count_adherence_std': 0.13577206507325174, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.8762265861034393, 'eval_reward_repeat_soft_std': 0.09761511832475663, 'eval_reward_judge_quality_mean': 0.4736250102519989, 'eval_reward_judge_quality_std': 0.20385925192385912, 'eval_reward_total_composite_mean': 0.5502476394176483, 'eval_reward_total_composite_std': 0.1676765900105238, 'epoch': 0.19}

 56%|█████▌    | 1850/3300 [4:33:34<3:27:59,  8.61s/it]

100%|██████████| 10/10 [00:51<00:00,  5.36s/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(

 56%|█████▌    | 1851/3300 [4:33:44<10:24:36, 25.86s/it]2026-04-13 11:54:42,289 | INFO | train_grpo_train | metrics_logged mode=train step=1851

                                                        
{'loss': 0.0096, 'grad_norm': 5.400481224060059, 'learning_rate': 4.393939393939394e-06, 'num_tokens': 3322238.0, 'completions/mean_length': 95.75, 'completions/min_length': 94.0, 'completions/max_length': 101.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 95.75, 'completions/min_terminated_length': 94.0, 'completions/max_terminated_length': 101.0, 'rewards/meter/mean': 0.9656529426574707, 'rewards/meter/std': 0.008796771056950092, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7203428745269775, 'rewards/repeat_soft/std': 0.05236854776740074, 'rewards/judge_quality/mean': 0.3987500071525574, 'rewards/judge_quality/std': 0.06010407209396362, 'rewards/total_composite/mean': 0.5083029270172119, 'rewards/total_composite/std': 0.03721180930733681, 'reward': 0.5083029270172119, 'reward_std': 0.037211816757917404, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057944923639297485, 'sampling/sampling_logp_difference/max': 1.7273783683776855, 'sampling/importance_sampling_ratio/min': 0.17774979770183563, 'sampling/importance_sampling_ratio/mean': 1.0042781829833984, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.29475096240639687, 'clip_ratio/low_mean': 0.007894736714661121, 'clip_ratio/low_min': 0.007894736714661121, 'clip_ratio/high_mean': 0.0548994611017406, 'clip_ratio/high_max': 0.0548994611017406, 'clip_ratio/region_mean': 0.06279419781640172, 'reward_total_mean': 0.5083029270172119, 'reward_meter_mean': 0.9656529426574707, 'reward_meter_std': 0.008796771056950092, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7203428745269775, 'reward_repeat_soft_std': 0.05236854776740074, 'reward_judge_quality_mean': 0.3987500071525574, 'reward_judge_quality_std': 0.06010407209396362, 'reward_total_composite_mean': 0.5083029270172119, 'reward_total_composite_std': 0.03721180930733681, 'epoch': 0.19}

 56%|█████▌    | 1851/3300 [4:33:44<10:24:36, 25.86s/it]INFO 04-13 11:54:42 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1852/3300 [4:33:50<8:02:03, 19.98s/it] 2026-04-13 11:54:48,481 | INFO | train_grpo_train | metrics_logged mode=train step=1852

                                                       
{'loss': -0.0237, 'grad_norm': 10.641679763793945, 'learning_rate': 4.390909090909091e-06, 'num_tokens': 3323871.0, 'completions/mean_length': 41.125, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9228650331497192, 'rewards/meter/std': 0.10433200746774673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.895309329032898, 'rewards/repeat_soft/std': 0.07829222828149796, 'rewards/judge_quality/mean': 0.5387499928474426, 'rewards/judge_quality/std': 0.24485784769058228, 'rewards/total_composite/mean': 0.6497100591659546, 'rewards/total_composite/std': 0.13361245393753052, 'reward': 0.6497100591659546, 'reward_std': 0.13361245393753052, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12940683960914612, 'sampling/sampling_logp_difference/max': 1.7096335887908936, 'sampling/importance_sampling_ratio/min': 0.18093208968639374, 'sampling/importance_sampling_ratio/mean': 1.009081482887268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7946780137717724, 'clip_ratio/low_mean': 0.08262495556846261, 'clip_ratio/low_min': 0.08262495556846261, 'clip_ratio/high_mean': 0.03240481577813625, 'clip_ratio/high_max': 0.03240481577813625, 'clip_ratio/region_mean': 0.11502977134659886, 'reward_total_mean': 0.6497100591659546, 'reward_meter_mean': 0.9228650331497192, 'reward_meter_std': 0.10433200746774673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.895309329032898, 'reward_repeat_soft_std': 0.07829222828149796, 'reward_judge_quality_mean': 0.5387499928474426, 'reward_judge_quality_std': 0.24485784769058228, 'reward_total_composite_mean': 0.6497100591659546, 'reward_total_composite_std': 0.13361245393753052, 'epoch': 0.19}

 56%|█████▌    | 1852/3300 [4:33:50<8:02:03, 19.98s/it]INFO 04-13 11:54:48 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1853/3300 [4:34:02<7:02:33, 17.52s/it]2026-04-13 11:55:00,281 | INFO | train_grpo_train | metrics_logged mode=train step=1853

                                                       
{'loss': -0.2299, 'grad_norm': 1.8711302280426025, 'learning_rate': 4.387878787878788e-06, 'num_tokens': 3326861.0, 'completions/mean_length': 248.75, 'completions/min_length': 192.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 211.1428680419922, 'completions/min_terminated_length': 192.0, 'completions/max_terminated_length': 244.0, 'rewards/meter/mean': 0.9305763244628906, 'rewards/meter/std': 0.14726680517196655, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8247259855270386, 'rewards/repeat_soft/std': 0.057971034198999405, 'rewards/judge_quality/mean': 0.29374998807907104, 'rewards/judge_quality/std': 0.1399936079978943, 'rewards/total_composite/mean': 0.4310820698738098, 'rewards/total_composite/std': 0.1934940665960312, 'reward': 0.4310820698738098, 'reward_std': 0.19349405169487, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10495075583457947, 'sampling/sampling_logp_difference/max': 1.4805593490600586, 'sampling/importance_sampling_ratio/min': 0.22751040756702423, 'sampling/importance_sampling_ratio/mean': 1.0249090194702148, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8760777711868286, 'clip_ratio/low_mean': 0.01909655425697565, 'clip_ratio/low_min': 0.01909655425697565, 'clip_ratio/high_mean': 0.059721045196056366, 'clip_ratio/high_max': 0.059721045196056366, 'clip_ratio/region_mean': 0.07881759945303202, 'reward_total_mean': 0.4310820698738098, 'reward_meter_mean': 0.9305763244628906, 'reward_meter_std': 0.14726680517196655, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8247259855270386, 'reward_repeat_soft_std': 0.057971034198999405, 'reward_judge_quality_mean': 0.29374998807907104, 'reward_judge_quality_std': 0.1399936079978943, 'reward_total_composite_mean': 0.4310820698738098, 'reward_total_composite_std': 0.1934940665960312, 'epoch': 0.19}

 56%|█████▌    | 1853/3300 [4:34:02<7:02:33, 17.52s/it]INFO 04-13 11:55:00 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1854/3300 [4:34:09<5:45:15, 14.33s/it]2026-04-13 11:55:07,157 | INFO | train_grpo_train | metrics_logged mode=train step=1854

                                                       
{'loss': -0.0435, 'grad_norm': 6.689228534698486, 'learning_rate': 4.384848484848485e-06, 'num_tokens': 3328859.0, 'completions/mean_length': 83.75, 'completions/min_length': 52.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 83.75, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8792723417282104, 'rewards/meter/std': 0.24548029899597168, 'rewards/count_adherence/mean': 0.7083333730697632, 'rewards/count_adherence/std': 0.11785111576318741, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7887636423110962, 'rewards/repeat_soft/std': 0.053252626210451126, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5860437750816345, 'rewards/total_composite/std': 0.16476160287857056, 'reward': 0.5860437750816345, 'reward_std': 0.16476160287857056, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08537229150533676, 'sampling/sampling_logp_difference/max': 1.5889043807983398, 'sampling/importance_sampling_ratio/min': 0.20414915680885315, 'sampling/importance_sampling_ratio/mean': 1.0225580930709839, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7083901986479759, 'clip_ratio/low_mean': 0.040940255392342806, 'clip_ratio/low_min': 0.040940255392342806, 'clip_ratio/high_mean': 0.022865363396704197, 'clip_ratio/high_max': 0.022865363396704197, 'clip_ratio/region_mean': 0.063805618789047, 'reward_total_mean': 0.5860437750816345, 'reward_meter_mean': 0.8792723417282104, 'reward_meter_std': 0.24548029899597168, 'reward_count_adherence_mean': 0.7083333730697632, 'reward_count_adherence_std': 0.11785111576318741, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7887636423110962, 'reward_repeat_soft_std': 0.053252626210451126, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5860437750816345, 'reward_total_composite_std': 0.16476160287857056, 'epoch': 0.19}

 56%|█████▌    | 1854/3300 [4:34:09<5:45:15, 14.33s/it]INFO 04-13 11:55:07 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1855/3300 [4:34:20<5:22:34, 13.39s/it]2026-04-13 11:55:18,386 | INFO | train_grpo_train | metrics_logged mode=train step=1855

                                                       
{'loss': -0.1029, 'grad_norm': 1.8869119882583618, 'learning_rate': 4.381818181818182e-06, 'num_tokens': 3330471.0, 'completions/mean_length': 166.5, 'completions/min_length': 40.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 51.333335876464844, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.5134392976760864, 'rewards/meter/std': 0.44988754391670227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.98429274559021, 'rewards/repeat_soft/std': 0.0269598551094532, 'rewards/judge_quality/mean': 0.4350000023841858, 'rewards/judge_quality/std': 0.29611775279045105, 'rewards/total_composite/mean': 0.428304523229599, 'rewards/total_composite/std': 0.3222726583480835, 'reward': 0.428304523229599, 'reward_std': 0.3222726285457611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14015433192253113, 'sampling/sampling_logp_difference/max': 1.8146681785583496, 'sampling/importance_sampling_ratio/min': 0.1628919541835785, 'sampling/importance_sampling_ratio/mean': 1.0194242000579834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5727350860834122, 'clip_ratio/low_mean': 0.03384146373718977, 'clip_ratio/low_min': 0.03384146373718977, 'clip_ratio/high_mean': 0.06243182811886072, 'clip_ratio/high_max': 0.06243182811886072, 'clip_ratio/region_mean': 0.09627329185605049, 'reward_total_mean': 0.428304523229599, 'reward_meter_mean': 0.5134392976760864, 'reward_meter_std': 0.44988754391670227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.98429274559021, 'reward_repeat_soft_std': 0.0269598551094532, 'reward_judge_quality_mean': 0.4350000023841858, 'reward_judge_quality_std': 0.29611775279045105, 'reward_total_composite_mean': 0.428304523229599, 'reward_total_composite_std': 0.3222726583480835, 'epoch': 0.19}

 56%|█████▌    | 1855/3300 [4:34:20<5:22:34, 13.39s/it]INFO 04-13 11:55:18 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▌    | 1856/3300 [4:34:27<4:34:59, 11.43s/it]2026-04-13 11:55:25,215 | INFO | train_grpo_train | metrics_logged mode=train step=1856

                                                       
{'loss': -0.0603, 'grad_norm': 12.850541114807129, 'learning_rate': 4.378787878787879e-06, 'num_tokens': 3332075.0, 'completions/mean_length': 36.5, 'completions/min_length': 27.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5566971302032471, 'rewards/meter/std': 0.46029070019721985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9029535055160522, 'rewards/repeat_soft/std': 0.053695887327194214, 'rewards/judge_quality/mean': 0.8575000166893005, 'rewards/judge_quality/std': 0.176776722073555, 'rewards/total_composite/mean': 0.6285346150398254, 'rewards/total_composite/std': 0.2570481300354004, 'reward': 0.6285346150398254, 'reward_std': 0.257048100233078, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09701843559741974, 'sampling/sampling_logp_difference/max': 1.2670092582702637, 'sampling/importance_sampling_ratio/min': 0.28167277574539185, 'sampling/importance_sampling_ratio/mean': 1.0066252946853638, 'sampling/importance_sampling_ratio/max': 1.831130862236023, 'entropy': 0.7173832096159458, 'clip_ratio/low_mean': 0.04563816008158028, 'clip_ratio/low_min': 0.04563816008158028, 'clip_ratio/high_mean': 0.04727564286440611, 'clip_ratio/high_max': 0.04727564286440611, 'clip_ratio/region_mean': 0.09291380294598639, 'reward_total_mean': 0.6285346150398254, 'reward_meter_mean': 0.5566971302032471, 'reward_meter_std': 0.46029070019721985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9029535055160522, 'reward_repeat_soft_std': 0.053695887327194214, 'reward_judge_quality_mean': 0.8575000166893005, 'reward_judge_quality_std': 0.176776722073555, 'reward_total_composite_mean': 0.6285346150398254, 'reward_total_composite_std': 0.2570481300354004, 'epoch': 0.19}

 56%|█████▌    | 1856/3300 [4:34:27<4:34:59, 11.43s/it]INFO 04-13 11:55:25 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1857/3300 [4:34:33<3:57:04,  9.86s/it]2026-04-13 11:55:31,415 | INFO | train_grpo_train | metrics_logged mode=train step=1857

                                                       
{'loss': -0.0377, 'grad_norm': 10.985547065734863, 'learning_rate': 4.375757575757576e-06, 'num_tokens': 3333626.0, 'completions/mean_length': 33.875, 'completions/min_length': 29.0, 'completions/max_length': 37.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.875, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.616787314414978, 'rewards/meter/std': 0.42414337396621704, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9226333498954773, 'rewards/repeat_soft/std': 0.03281796723604202, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.7072337865829468, 'rewards/total_composite/std': 0.25289255380630493, 'reward': 0.7072337865829468, 'reward_std': 0.25289252400398254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08271818608045578, 'sampling/sampling_logp_difference/max': 1.549966812133789, 'sampling/importance_sampling_ratio/min': 0.21225501596927643, 'sampling/importance_sampling_ratio/mean': 0.9723013043403625, 'sampling/importance_sampling_ratio/max': 1.7442371845245361, 'entropy': 0.27876042760908604, 'clip_ratio/low_mean': 0.008342602755874395, 'clip_ratio/low_min': 0.008342602755874395, 'clip_ratio/high_mean': 0.052414074540138245, 'clip_ratio/high_max': 0.052414074540138245, 'clip_ratio/region_mean': 0.06075667729601264, 'reward_total_mean': 0.7072337865829468, 'reward_meter_mean': 0.616787314414978, 'reward_meter_std': 0.42414337396621704, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9226333498954773, 'reward_repeat_soft_std': 0.03281796723604202, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.7072337865829468, 'reward_total_composite_std': 0.25289255380630493, 'epoch': 0.19}

 56%|█████▋    | 1857/3300 [4:34:33<3:57:04,  9.86s/it]INFO 04-13 11:55:31 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1858/3300 [4:34:42<3:47:02,  9.45s/it]2026-04-13 11:55:39,885 | INFO | train_grpo_train | metrics_logged mode=train step=1858

                                                       
{'loss': -0.0044, 'grad_norm': 5.345325946807861, 'learning_rate': 4.372727272727273e-06, 'num_tokens': 3336330.0, 'completions/mean_length': 157.0, 'completions/min_length': 137.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 157.0, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.8315615057945251, 'rewards/meter/std': 0.23402608931064606, 'rewards/count_adherence/mean': 0.53125, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8731966018676758, 'rewards/repeat_soft/std': 0.04789191856980324, 'rewards/judge_quality/mean': 0.4612500071525574, 'rewards/judge_quality/std': 0.19467465579509735, 'rewards/total_composite/mean': 0.4668791592121124, 'rewards/total_composite/std': 0.06063517555594444, 'reward': 0.4668791592121124, 'reward_std': 0.060635171830654144, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10826198756694794, 'sampling/sampling_logp_difference/max': 2.527419090270996, 'sampling/importance_sampling_ratio/min': 0.07986488193273544, 'sampling/importance_sampling_ratio/mean': 1.015867829322815, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.742377758026123, 'clip_ratio/low_mean': 0.04996927920728922, 'clip_ratio/low_min': 0.04996927920728922, 'clip_ratio/high_mean': 0.05791372526437044, 'clip_ratio/high_max': 0.05791372526437044, 'clip_ratio/region_mean': 0.10788300447165966, 'reward_total_mean': 0.4668791592121124, 'reward_meter_mean': 0.8315615057945251, 'reward_meter_std': 0.23402608931064606, 'reward_count_adherence_mean': 0.53125, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8731966018676758, 'reward_repeat_soft_std': 0.04789191856980324, 'reward_judge_quality_mean': 0.4612500071525574, 'reward_judge_quality_std': 0.19467465579509735, 'reward_total_composite_mean': 0.4668791592121124, 'reward_total_composite_std': 0.06063517555594444, 'epoch': 0.19}

 56%|█████▋    | 1858/3300 [4:34:42<3:47:02,  9.45s/it]INFO 04-13 11:55:40 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1859/3300 [4:34:49<3:34:18,  8.92s/it]2026-04-13 11:55:47,589 | INFO | train_grpo_train | metrics_logged mode=train step=1859

                                                       
{'loss': 0.0957, 'grad_norm': 7.484099388122559, 'learning_rate': 4.36969696969697e-06, 'num_tokens': 3338346.0, 'completions/mean_length': 63.0, 'completions/min_length': 52.0, 'completions/max_length': 83.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.0, 'completions/min_terminated_length': 52.0, 'completions/max_terminated_length': 83.0, 'rewards/meter/mean': 0.9657058119773865, 'rewards/meter/std': 0.005547062493860722, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7004228830337524, 'rewards/repeat_soft/std': 0.06686412543058395, 'rewards/judge_quality/mean': 0.29249998927116394, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.46350228786468506, 'rewards/total_composite/std': 0.06039566174149513, 'reward': 0.46350228786468506, 'reward_std': 0.06039566174149513, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0521131195127964, 'sampling/sampling_logp_difference/max': 1.4458332061767578, 'sampling/importance_sampling_ratio/min': 0.23554974794387817, 'sampling/importance_sampling_ratio/mean': 0.9977630972862244, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2415139451622963, 'clip_ratio/low_mean': 0.006274366518482566, 'clip_ratio/low_min': 0.006274366518482566, 'clip_ratio/high_mean': 0.0413695378229022, 'clip_ratio/high_max': 0.0413695378229022, 'clip_ratio/region_mean': 0.04764390434138477, 'reward_total_mean': 0.46350228786468506, 'reward_meter_mean': 0.9657058119773865, 'reward_meter_std': 0.005547062493860722, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7004228830337524, 'reward_repeat_soft_std': 0.06686412543058395, 'reward_judge_quality_mean': 0.29249998927116394, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.46350228786468506, 'reward_total_composite_std': 0.06039566174149513, 'epoch': 0.19}

 56%|█████▋    | 1859/3300 [4:34:49<3:34:18,  8.92s/it]INFO 04-13 11:55:47 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1860/3300 [4:34:56<3:22:16,  8.43s/it]2026-04-13 11:55:54,873 | INFO | train_grpo_train | metrics_logged mode=train step=1860

                                                       
{'loss': -0.0241, 'grad_norm': 5.998904228210449, 'learning_rate': 4.366666666666667e-06, 'num_tokens': 3340688.0, 'completions/mean_length': 108.75, 'completions/min_length': 82.0, 'completions/max_length': 127.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.8752556443214417, 'rewards/meter/std': 0.22284598648548126, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7929337024688721, 'rewards/repeat_soft/std': 0.03223147615790367, 'rewards/judge_quality/mean': 0.4362500011920929, 'rewards/judge_quality/std': 0.12916629016399384, 'rewards/total_composite/mean': 0.5178137421607971, 'rewards/total_composite/std': 0.10374221205711365, 'reward': 0.5178137421607971, 'reward_std': 0.10374220460653305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08187668025493622, 'sampling/sampling_logp_difference/max': 2.881627082824707, 'sampling/importance_sampling_ratio/min': 0.05604349821805954, 'sampling/importance_sampling_ratio/mean': 1.0006663799285889, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5625111572444439, 'clip_ratio/low_mean': 0.036750719882547855, 'clip_ratio/low_min': 0.036750719882547855, 'clip_ratio/high_mean': 0.0497631561011076, 'clip_ratio/high_max': 0.0497631561011076, 'clip_ratio/region_mean': 0.08651387598365545, 'reward_total_mean': 0.5178137421607971, 'reward_meter_mean': 0.8752556443214417, 'reward_meter_std': 0.22284598648548126, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7929337024688721, 'reward_repeat_soft_std': 0.03223147615790367, 'reward_judge_quality_mean': 0.4362500011920929, 'reward_judge_quality_std': 0.12916629016399384, 'reward_total_composite_mean': 0.5178137421607971, 'reward_total_composite_std': 0.10374221205711365, 'epoch': 0.19}

 56%|█████▋    | 1860/3300 [4:34:57<3:22:16,  8.43s/it]INFO 04-13 11:55:55 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1861/3300 [4:35:04<3:16:21,  8.19s/it]2026-04-13 11:56:02,495 | INFO | train_grpo_train | metrics_logged mode=train step=1861

                                                       
{'loss': 0.0764, 'grad_norm': 7.427359580993652, 'learning_rate': 4.363636363636364e-06, 'num_tokens': 3343149.0, 'completions/mean_length': 129.625, 'completions/min_length': 110.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 129.625, 'completions/min_terminated_length': 110.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.9266027212142944, 'rewards/meter/std': 0.07922521233558655, 'rewards/count_adherence/mean': 0.7749999761581421, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8284742832183838, 'rewards/repeat_soft/std': 0.07496636360883713, 'rewards/judge_quality/mean': 0.35624998807907104, 'rewards/judge_quality/std': 0.08798335492610931, 'rewards/total_composite/mean': 0.4970102608203888, 'rewards/total_composite/std': 0.07223185896873474, 'reward': 0.4970102608203888, 'reward_std': 0.07223183661699295, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10745149105787277, 'sampling/sampling_logp_difference/max': 1.8511078357696533, 'sampling/importance_sampling_ratio/min': 0.18094448745250702, 'sampling/importance_sampling_ratio/mean': 1.0111578702926636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6480537839233875, 'clip_ratio/low_mean': 0.03165066707879305, 'clip_ratio/low_min': 0.03165066707879305, 'clip_ratio/high_mean': 0.05648232623934746, 'clip_ratio/high_max': 0.05648232623934746, 'clip_ratio/region_mean': 0.0881329933181405, 'reward_total_mean': 0.4970102608203888, 'reward_meter_mean': 0.9266027212142944, 'reward_meter_std': 0.07922521233558655, 'reward_count_adherence_mean': 0.7749999761581421, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8284742832183838, 'reward_repeat_soft_std': 0.07496636360883713, 'reward_judge_quality_mean': 0.35624998807907104, 'reward_judge_quality_std': 0.08798335492610931, 'reward_total_composite_mean': 0.4970102608203888, 'reward_total_composite_std': 0.07223185896873474, 'epoch': 0.19}

 56%|█████▋    | 1861/3300 [4:35:04<3:16:21,  8.19s/it]INFO 04-13 11:56:02 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1862/3300 [4:35:16<3:41:41,  9.25s/it]2026-04-13 11:56:14,214 | INFO | train_grpo_train | metrics_logged mode=train step=1862

                                                       
{'loss': -0.1356, 'grad_norm': 2.285029888153076, 'learning_rate': 4.36060606060606e-06, 'num_tokens': 3344944.0, 'completions/mean_length': 108.375, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 50.71428680419922, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8659183979034424, 'rewards/meter/std': 0.2861400246620178, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8929972648620605, 'rewards/repeat_soft/std': 0.02832763083279133, 'rewards/judge_quality/mean': 0.5649999976158142, 'rewards/judge_quality/std': 0.32062438130378723, 'rewards/total_composite/mean': 0.6402934193611145, 'rewards/total_composite/std': 0.2991735339164734, 'reward': 0.6402934193611145, 'reward_std': 0.2991735339164734, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10415394604206085, 'sampling/sampling_logp_difference/max': 1.1604008674621582, 'sampling/importance_sampling_ratio/min': 0.3133605420589447, 'sampling/importance_sampling_ratio/mean': 1.0106756687164307, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7009025365114212, 'clip_ratio/low_mean': 0.056871891021728516, 'clip_ratio/low_min': 0.056871891021728516, 'clip_ratio/high_mean': 0.03592592664062977, 'clip_ratio/high_max': 0.03592592664062977, 'clip_ratio/region_mean': 0.09279781766235828, 'reward_total_mean': 0.6402934193611145, 'reward_meter_mean': 0.8659183979034424, 'reward_meter_std': 0.2861400246620178, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8929972648620605, 'reward_repeat_soft_std': 0.02832763083279133, 'reward_judge_quality_mean': 0.5649999976158142, 'reward_judge_quality_std': 0.32062438130378723, 'reward_total_composite_mean': 0.6402934193611145, 'reward_total_composite_std': 0.2991735339164734, 'epoch': 0.19}

 56%|█████▋    | 1862/3300 [4:35:16<3:41:41,  9.25s/it]INFO 04-13 11:56:14 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1863/3300 [4:35:25<3:42:38,  9.30s/it]2026-04-13 11:56:23,635 | INFO | train_grpo_train | metrics_logged mode=train step=1863

                                                       
{'loss': 0.0731, 'grad_norm': 4.1012701988220215, 'learning_rate': 4.3575757575757576e-06, 'num_tokens': 3348524.0, 'completions/mean_length': 238.5, 'completions/min_length': 220.0, 'completions/max_length': 260.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 238.5, 'completions/min_terminated_length': 220.0, 'completions/max_terminated_length': 260.0, 'rewards/meter/mean': 0.9505437612533569, 'rewards/meter/std': 0.042091019451618195, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.08908706158399582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6943099498748779, 'rewards/repeat_soft/std': 0.03186754137277603, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5136449337005615, 'rewards/total_composite/std': 0.02377447858452797, 'reward': 0.5136449337005615, 'reward_std': 0.023774469271302223, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06255806237459183, 'sampling/sampling_logp_difference/max': 1.5884747505187988, 'sampling/importance_sampling_ratio/min': 0.20423687994480133, 'sampling/importance_sampling_ratio/mean': 1.001752495765686, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.368648037314415, 'clip_ratio/low_mean': 0.022642745170742273, 'clip_ratio/low_min': 0.022642745170742273, 'clip_ratio/high_mean': 0.02703581377863884, 'clip_ratio/high_max': 0.02703581377863884, 'clip_ratio/region_mean': 0.04967855894938111, 'reward_total_mean': 0.5136449337005615, 'reward_meter_mean': 0.9505437612533569, 'reward_meter_std': 0.042091019451618195, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.08908706158399582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6943099498748779, 'reward_repeat_soft_std': 0.03186754137277603, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5136449337005615, 'reward_total_composite_std': 0.02377447858452797, 'epoch': 0.19}

 56%|█████▋    | 1863/3300 [4:35:25<3:42:38,  9.30s/it]INFO 04-13 11:56:23 [block_pool.py:378] Successfully reset prefix cache

 56%|█████▋    | 1864/3300 [4:35:33<3:28:08,  8.70s/it]2026-04-13 11:56:30,919 | INFO | train_grpo_train | metrics_logged mode=train step=1864

                                                       
{'loss': -0.0471, 'grad_norm': 6.634945392608643, 'learning_rate': 4.354545454545455e-06, 'num_tokens': 3350888.0, 'completions/mean_length': 122.5, 'completions/min_length': 108.0, 'completions/max_length': 139.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 122.5, 'completions/min_terminated_length': 108.0, 'completions/max_terminated_length': 139.0, 'rewards/meter/mean': 0.8727961182594299, 'rewards/meter/std': 0.24367918074131012, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9506475329399109, 'rewards/repeat_soft/std': 0.019312597811222076, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5545367002487183, 'rewards/total_composite/std': 0.14379402995109558, 'reward': 0.5545367002487183, 'reward_std': 0.1437940150499344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12057143449783325, 'sampling/sampling_logp_difference/max': 1.7719669342041016, 'sampling/importance_sampling_ratio/min': 0.16999828815460205, 'sampling/importance_sampling_ratio/mean': 1.0171279907226562, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9396882504224777, 'clip_ratio/low_mean': 0.08226289879530668, 'clip_ratio/low_min': 0.08226289879530668, 'clip_ratio/high_mean': 0.008992806077003479, 'clip_ratio/high_max': 0.008992806077003479, 'clip_ratio/region_mean': 0.09125570487231016, 'reward_total_mean': 0.5545367002487183, 'reward_meter_mean': 0.8727961182594299, 'reward_meter_std': 0.24367918074131012, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9506475329399109, 'reward_repeat_soft_std': 0.019312597811222076, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5545367002487183, 'reward_total_composite_std': 0.14379402995109558, 'epoch': 0.19}

 56%|█████▋    | 1864/3300 [4:35:33<3:28:08,  8.70s/it]INFO 04-13 11:56:31 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1865/3300 [4:35:39<3:13:18,  8.08s/it]2026-04-13 11:56:37,562 | INFO | train_grpo_train | metrics_logged mode=train step=1865

                                                       
{'loss': 0.1142, 'grad_norm': 13.94418716430664, 'learning_rate': 4.351515151515152e-06, 'num_tokens': 3352545.0, 'completions/mean_length': 34.125, 'completions/min_length': 27.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.125, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.5891563892364502, 'rewards/meter/std': 0.42681676149368286, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9475993514060974, 'rewards/repeat_soft/std': 0.04067608714103699, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.6522279977798462, 'rewards/total_composite/std': 0.23976483941078186, 'reward': 0.6522279977798462, 'reward_std': 0.23976485431194305, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10949678719043732, 'sampling/sampling_logp_difference/max': 1.8282232284545898, 'sampling/importance_sampling_ratio/min': 0.16069884598255157, 'sampling/importance_sampling_ratio/mean': 1.0080785751342773, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5568082258105278, 'clip_ratio/low_mean': 0.052935092244297266, 'clip_ratio/low_min': 0.052935092244297266, 'clip_ratio/high_mean': 0.03388047218322754, 'clip_ratio/high_max': 0.03388047218322754, 'clip_ratio/region_mean': 0.0868155644275248, 'reward_total_mean': 0.6522279977798462, 'reward_meter_mean': 0.5891563892364502, 'reward_meter_std': 0.42681676149368286, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9475993514060974, 'reward_repeat_soft_std': 0.04067608714103699, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.6522279977798462, 'reward_total_composite_std': 0.23976483941078186, 'epoch': 0.19}

 57%|█████▋    | 1865/3300 [4:35:39<3:13:18,  8.08s/it]INFO 04-13 11:56:37 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1866/3300 [4:35:47<3:11:59,  8.03s/it]2026-04-13 11:56:45,490 | INFO | train_grpo_train | metrics_logged mode=train step=1866

                                                       
{'loss': -0.0036, 'grad_norm': 6.008486747741699, 'learning_rate': 4.348484848484849e-06, 'num_tokens': 3355173.0, 'completions/mean_length': 154.5, 'completions/min_length': 143.0, 'completions/max_length': 172.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 154.5, 'completions/min_terminated_length': 143.0, 'completions/max_terminated_length': 172.0, 'rewards/meter/mean': 0.8012826442718506, 'rewards/meter/std': 0.2174297273159027, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8598213195800781, 'rewards/repeat_soft/std': 0.027137139812111855, 'rewards/judge_quality/mean': 0.5824999809265137, 'rewards/judge_quality/std': 0.23260943591594696, 'rewards/total_composite/mean': 0.5909258723258972, 'rewards/total_composite/std': 0.16364625096321106, 'reward': 0.5909258723258972, 'reward_std': 0.16364626586437225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08679886162281036, 'sampling/sampling_logp_difference/max': 1.854860782623291, 'sampling/importance_sampling_ratio/min': 0.1564747393131256, 'sampling/importance_sampling_ratio/mean': 1.002786636352539, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5204651392996311, 'clip_ratio/low_mean': 0.04906986514106393, 'clip_ratio/low_min': 0.04906986514106393, 'clip_ratio/high_mean': 0.033117182552814484, 'clip_ratio/high_max': 0.033117182552814484, 'clip_ratio/region_mean': 0.08218704769387841, 'reward_total_mean': 0.5909258723258972, 'reward_meter_mean': 0.8012826442718506, 'reward_meter_std': 0.2174297273159027, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8598213195800781, 'reward_repeat_soft_std': 0.027137139812111855, 'reward_judge_quality_mean': 0.5824999809265137, 'reward_judge_quality_std': 0.23260943591594696, 'reward_total_composite_mean': 0.5909258723258972, 'reward_total_composite_std': 0.16364625096321106, 'epoch': 0.19}

 57%|█████▋    | 1866/3300 [4:35:47<3:11:59,  8.03s/it]INFO 04-13 11:56:45 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1867/3300 [4:35:53<2:58:45,  7.48s/it]2026-04-13 11:56:51,698 | INFO | train_grpo_train | metrics_logged mode=train step=1867

                                                       
{'loss': -0.0284, 'grad_norm': 11.911911964416504, 'learning_rate': 4.345454545454546e-06, 'num_tokens': 3356930.0, 'completions/mean_length': 39.625, 'completions/min_length': 29.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.6499167680740356, 'rewards/meter/std': 0.30164313316345215, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9578688144683838, 'rewards/repeat_soft/std': 0.04335197061300278, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.52110755443573, 'rewards/total_composite/std': 0.08668576180934906, 'reward': 0.52110755443573, 'reward_std': 0.08668576180934906, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10192485898733139, 'sampling/sampling_logp_difference/max': 0.9830365180969238, 'sampling/importance_sampling_ratio/min': 0.37417319416999817, 'sampling/importance_sampling_ratio/mean': 1.0202760696411133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6632933504879475, 'clip_ratio/low_mean': 0.04292092937976122, 'clip_ratio/low_min': 0.04292092937976122, 'clip_ratio/high_mean': 0.05193390650674701, 'clip_ratio/high_max': 0.05193390650674701, 'clip_ratio/region_mean': 0.09485483588650823, 'reward_total_mean': 0.52110755443573, 'reward_meter_mean': 0.6499167680740356, 'reward_meter_std': 0.30164313316345215, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9578688144683838, 'reward_repeat_soft_std': 0.04335197061300278, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.52110755443573, 'reward_total_composite_std': 0.08668576180934906, 'epoch': 0.19}

 57%|█████▋    | 1867/3300 [4:35:53<2:58:45,  7.48s/it]INFO 04-13 11:56:51 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1868/3300 [4:36:00<2:54:06,  7.30s/it]2026-04-13 11:56:58,544 | INFO | train_grpo_train | metrics_logged mode=train step=1868

                                                       
{'loss': 0.0075, 'grad_norm': 10.356154441833496, 'learning_rate': 4.342424242424243e-06, 'num_tokens': 3358867.0, 'completions/mean_length': 61.125, 'completions/min_length': 56.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7613774538040161, 'rewards/meter/std': 0.31797435879707336, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.995015025138855, 'rewards/repeat_soft/std': 0.007305712439119816, 'rewards/judge_quality/mean': 0.7362500429153442, 'rewards/judge_quality/std': 0.25376805663108826, 'rewards/total_composite/mean': 0.6909343004226685, 'rewards/total_composite/std': 0.1758246123790741, 'reward': 0.6909343004226685, 'reward_std': 0.1758246123790741, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11673297733068466, 'sampling/sampling_logp_difference/max': 1.9016571044921875, 'sampling/importance_sampling_ratio/min': 0.14932097494602203, 'sampling/importance_sampling_ratio/mean': 0.991946280002594, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7148168757557869, 'clip_ratio/low_mean': 0.05922755692154169, 'clip_ratio/low_min': 0.05922755692154169, 'clip_ratio/high_mean': 0.046878415159881115, 'clip_ratio/high_max': 0.046878415159881115, 'clip_ratio/region_mean': 0.1061059720814228, 'reward_total_mean': 0.6909343004226685, 'reward_meter_mean': 0.7613774538040161, 'reward_meter_std': 0.31797435879707336, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.995015025138855, 'reward_repeat_soft_std': 0.007305712439119816, 'reward_judge_quality_mean': 0.7362500429153442, 'reward_judge_quality_std': 0.25376805663108826, 'reward_total_composite_mean': 0.6909343004226685, 'reward_total_composite_std': 0.1758246123790741, 'epoch': 0.19}

 57%|█████▋    | 1868/3300 [4:36:00<2:54:06,  7.30s/it]INFO 04-13 11:56:58 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1869/3300 [4:36:06<2:43:26,  6.85s/it]2026-04-13 11:57:04,373 | INFO | train_grpo_train | metrics_logged mode=train step=1869

                                                       
{'loss': -0.0939, 'grad_norm': 21.15143585205078, 'learning_rate': 4.33939393939394e-06, 'num_tokens': 3360190.0, 'completions/mean_length': 21.375, 'completions/min_length': 17.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.375, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8383228778839111, 'rewards/meter/std': 0.27383944392204285, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9510330557823181, 'rewards/repeat_soft/std': 0.027066897600889206, 'rewards/judge_quality/mean': 0.4437500238418579, 'rewards/judge_quality/std': 0.2084594964981079, 'rewards/total_composite/mean': 0.5913371443748474, 'rewards/total_composite/std': 0.15817147493362427, 'reward': 0.5913371443748474, 'reward_std': 0.15817148983478546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13284669816493988, 'sampling/sampling_logp_difference/max': 1.202202320098877, 'sampling/importance_sampling_ratio/min': 0.30053162574768066, 'sampling/importance_sampling_ratio/mean': 1.0145890712738037, 'sampling/importance_sampling_ratio/max': 1.8619064092636108, 'entropy': 1.180416114628315, 'clip_ratio/low_mean': 0.03137254947796464, 'clip_ratio/low_min': 0.03137254947796464, 'clip_ratio/high_mean': 0.06869431119412184, 'clip_ratio/high_max': 0.06869431119412184, 'clip_ratio/region_mean': 0.10006686067208648, 'reward_total_mean': 0.5913371443748474, 'reward_meter_mean': 0.8383228778839111, 'reward_meter_std': 0.27383944392204285, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9510330557823181, 'reward_repeat_soft_std': 0.027066897600889206, 'reward_judge_quality_mean': 0.4437500238418579, 'reward_judge_quality_std': 0.2084594964981079, 'reward_total_composite_mean': 0.5913371443748474, 'reward_total_composite_std': 0.15817147493362427, 'epoch': 0.19}

 57%|█████▋    | 1869/3300 [4:36:06<2:43:26,  6.85s/it]INFO 04-13 11:57:04 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1870/3300 [4:36:12<2:36:58,  6.59s/it]2026-04-13 11:57:10,330 | INFO | train_grpo_train | metrics_logged mode=train step=1870

                                                       
{'loss': 0.0627, 'grad_norm': 15.947404861450195, 'learning_rate': 4.336363636363637e-06, 'num_tokens': 3361583.0, 'completions/mean_length': 32.125, 'completions/min_length': 24.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 32.125, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.5034498572349548, 'rewards/meter/std': 0.43890899419784546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.938712477684021, 'rewards/repeat_soft/std': 0.04584834352135658, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.478248655796051, 'rewards/total_composite/std': 0.11615147441625595, 'reward': 0.478248655796051, 'reward_std': 0.11615145951509476, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08563661575317383, 'sampling/sampling_logp_difference/max': 1.4977467060089111, 'sampling/importance_sampling_ratio/min': 0.2236335128545761, 'sampling/importance_sampling_ratio/mean': 0.9934297204017639, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4042210765182972, 'clip_ratio/low_mean': 0.02282913215458393, 'clip_ratio/low_min': 0.02282913215458393, 'clip_ratio/high_mean': 0.04942481033504009, 'clip_ratio/high_max': 0.04942481033504009, 'clip_ratio/region_mean': 0.07225394248962402, 'reward_total_mean': 0.478248655796051, 'reward_meter_mean': 0.5034498572349548, 'reward_meter_std': 0.43890899419784546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.938712477684021, 'reward_repeat_soft_std': 0.04584834352135658, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.478248655796051, 'reward_total_composite_std': 0.11615147441625595, 'epoch': 0.19}

 57%|█████▋    | 1870/3300 [4:36:12<2:36:58,  6.59s/it]INFO 04-13 11:57:10 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1871/3300 [4:36:18<2:35:36,  6.53s/it]2026-04-13 11:57:16,749 | INFO | train_grpo_train | metrics_logged mode=train step=1871

                                                       
{'loss': 0.036, 'grad_norm': 8.205052375793457, 'learning_rate': 4.333333333333334e-06, 'num_tokens': 3363452.0, 'completions/mean_length': 60.625, 'completions/min_length': 51.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 60.625, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.926069974899292, 'rewards/meter/std': 0.14992724359035492, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9523845911026001, 'rewards/repeat_soft/std': 0.03608405217528343, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.8966474533081055, 'rewards/total_composite/std': 0.09090722352266312, 'reward': 0.8966474533081055, 'reward_std': 0.09090722352266312, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11477474868297577, 'sampling/sampling_logp_difference/max': 3.223674774169922, 'sampling/importance_sampling_ratio/min': 0.03980850428342819, 'sampling/importance_sampling_ratio/mean': 1.0129293203353882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6815303079783916, 'clip_ratio/low_mean': 0.015384615398943424, 'clip_ratio/low_min': 0.015384615398943424, 'clip_ratio/high_mean': 0.1016759155318141, 'clip_ratio/high_max': 0.1016759155318141, 'clip_ratio/region_mean': 0.11706053093075752, 'reward_total_mean': 0.8966474533081055, 'reward_meter_mean': 0.926069974899292, 'reward_meter_std': 0.14992724359035492, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9523845911026001, 'reward_repeat_soft_std': 0.03608405217528343, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.8966474533081055, 'reward_total_composite_std': 0.09090722352266312, 'epoch': 0.19}

 57%|█████▋    | 1871/3300 [4:36:18<2:35:36,  6.53s/it]INFO 04-13 11:57:16 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1872/3300 [4:36:35<3:48:58,  9.62s/it]2026-04-13 11:57:33,559 | INFO | train_grpo_train | metrics_logged mode=train step=1872

                                                       
{'loss': -0.0684, 'grad_norm': 2.9684641361236572, 'learning_rate': 4.330303030303031e-06, 'num_tokens': 3365012.0, 'completions/mean_length': 89.0, 'completions/min_length': 25.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 28.571430206298828, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.7443959712982178, 'rewards/meter/std': 0.4385094940662384, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9671874642372131, 'rewards/repeat_soft/std': 0.013258260674774647, 'rewards/judge_quality/mean': 0.5024999976158142, 'rewards/judge_quality/std': 0.2886792719364166, 'rewards/total_composite/mean': 0.5870416164398193, 'rewards/total_composite/std': 0.3021523356437683, 'reward': 0.5870416164398193, 'reward_std': 0.3021523058414459, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10582827031612396, 'sampling/sampling_logp_difference/max': 0.9363492727279663, 'sampling/importance_sampling_ratio/min': 0.39205652475357056, 'sampling/importance_sampling_ratio/mean': 0.9953686594963074, 'sampling/importance_sampling_ratio/max': 1.7579275369644165, 'entropy': 0.5372766740620136, 'clip_ratio/low_mean': 0.02016128972172737, 'clip_ratio/low_min': 0.02016128972172737, 'clip_ratio/high_mean': 0.08254634728655219, 'clip_ratio/high_max': 0.08254634728655219, 'clip_ratio/region_mean': 0.10270763700827956, 'reward_total_mean': 0.5870416164398193, 'reward_meter_mean': 0.7443959712982178, 'reward_meter_std': 0.4385094940662384, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9671874642372131, 'reward_repeat_soft_std': 0.013258260674774647, 'reward_judge_quality_mean': 0.5024999976158142, 'reward_judge_quality_std': 0.2886792719364166, 'reward_total_composite_mean': 0.5870416164398193, 'reward_total_composite_std': 0.3021523356437683, 'epoch': 0.19}

 57%|█████▋    | 1872/3300 [4:36:35<3:48:58,  9.62s/it]INFO 04-13 11:57:33 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1873/3300 [4:36:47<4:01:56, 10.17s/it]2026-04-13 11:57:45,023 | INFO | train_grpo_train | metrics_logged mode=train step=1873

                                                       
{'loss': -0.1429, 'grad_norm': 1.6155564785003662, 'learning_rate': 4.327272727272728e-06, 'num_tokens': 3366714.0, 'completions/mean_length': 109.75, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.28571701049805, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9725117683410645, 'rewards/meter/std': 0.025253908708691597, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.888059139251709, 'rewards/repeat_soft/std': 0.11275843530893326, 'rewards/judge_quality/mean': 0.38499999046325684, 'rewards/judge_quality/std': 0.13617216050624847, 'rewards/total_composite/mean': 0.530485212802887, 'rewards/total_composite/std': 0.21530216932296753, 'reward': 0.530485212802887, 'reward_std': 0.21530215442180634, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09288246929645538, 'sampling/sampling_logp_difference/max': 0.9357633590698242, 'sampling/importance_sampling_ratio/min': 0.3922863006591797, 'sampling/importance_sampling_ratio/mean': 1.0329862833023071, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5780495665967464, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06867086002603173, 'clip_ratio/high_max': 0.06867086002603173, 'clip_ratio/region_mean': 0.06867086002603173, 'reward_total_mean': 0.530485212802887, 'reward_meter_mean': 0.9725117683410645, 'reward_meter_std': 0.025253908708691597, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.888059139251709, 'reward_repeat_soft_std': 0.11275843530893326, 'reward_judge_quality_mean': 0.38499999046325684, 'reward_judge_quality_std': 0.13617216050624847, 'reward_total_composite_mean': 0.530485212802887, 'reward_total_composite_std': 0.21530216932296753, 'epoch': 0.19}

 57%|█████▋    | 1873/3300 [4:36:47<4:01:56, 10.17s/it]INFO 04-13 11:57:45 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1874/3300 [4:36:54<3:39:16,  9.23s/it]2026-04-13 11:57:52,050 | INFO | train_grpo_train | metrics_logged mode=train step=1874

                                                       
{'loss': 0.078, 'grad_norm': 5.653331279754639, 'learning_rate': 4.324242424242425e-06, 'num_tokens': 3368907.0, 'completions/mean_length': 99.125, 'completions/min_length': 63.0, 'completions/max_length': 138.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.125, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 138.0, 'rewards/meter/mean': 0.6673324108123779, 'rewards/meter/std': 0.34372588992118835, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.806655764579773, 'rewards/repeat_soft/std': 0.03883816674351692, 'rewards/judge_quality/mean': 0.5450000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.5225415825843811, 'rewards/total_composite/std': 0.16865621507167816, 'reward': 0.5225415825843811, 'reward_std': 0.16865621507167816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08662410825490952, 'sampling/sampling_logp_difference/max': 2.349956750869751, 'sampling/importance_sampling_ratio/min': 0.09537328779697418, 'sampling/importance_sampling_ratio/mean': 1.0094082355499268, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4720325320959091, 'clip_ratio/low_mean': 0.03813416324555874, 'clip_ratio/low_min': 0.03813416324555874, 'clip_ratio/high_mean': 0.043572402093559504, 'clip_ratio/high_max': 0.043572402093559504, 'clip_ratio/region_mean': 0.08170656533911824, 'reward_total_mean': 0.5225415825843811, 'reward_meter_mean': 0.6673324108123779, 'reward_meter_std': 0.34372588992118835, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.806655764579773, 'reward_repeat_soft_std': 0.03883816674351692, 'reward_judge_quality_mean': 0.5450000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.5225415825843811, 'reward_total_composite_std': 0.16865621507167816, 'epoch': 0.19}

 57%|█████▋    | 1874/3300 [4:36:54<3:39:16,  9.23s/it]INFO 04-13 11:57:52 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1875/3300 [4:36:59<3:14:26,  8.19s/it]2026-04-13 11:57:57,805 | INFO | train_grpo_train | metrics_logged mode=train step=1875

                                                       
{'loss': 0.0816, 'grad_norm': 16.872835159301758, 'learning_rate': 4.321212121212121e-06, 'num_tokens': 3370271.0, 'completions/mean_length': 20.5, 'completions/min_length': 14.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 20.5, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.760849118232727, 'rewards/meter/std': 0.3263269066810608, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9567934274673462, 'rewards/repeat_soft/std': 0.016140474006533623, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.555090606212616, 'rewards/total_composite/std': 0.09060538560152054, 'reward': 0.555090606212616, 'reward_std': 0.09060538560152054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12364717572927475, 'sampling/sampling_logp_difference/max': 0.842437744140625, 'sampling/importance_sampling_ratio/min': 0.4306594133377075, 'sampling/importance_sampling_ratio/mean': 1.0259215831756592, 'sampling/importance_sampling_ratio/max': 1.9881998300552368, 'entropy': 0.7779849320650101, 'clip_ratio/low_mean': 0.05576923117041588, 'clip_ratio/low_min': 0.05576923117041588, 'clip_ratio/high_mean': 0.06890453398227692, 'clip_ratio/high_max': 0.06890453398227692, 'clip_ratio/region_mean': 0.1246737651526928, 'reward_total_mean': 0.555090606212616, 'reward_meter_mean': 0.760849118232727, 'reward_meter_std': 0.3263269066810608, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9567934274673462, 'reward_repeat_soft_std': 0.016140474006533623, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.555090606212616, 'reward_total_composite_std': 0.09060538560152054, 'epoch': 0.19}

 57%|█████▋    | 1875/3300 [4:37:00<3:14:26,  8.19s/it]INFO 04-13 11:57:58 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1876/3300 [4:37:07<3:11:12,  8.06s/it]2026-04-13 11:58:05,547 | INFO | train_grpo_train | metrics_logged mode=train step=1876

                                                       
{'loss': 0.0623, 'grad_norm': 6.934006690979004, 'learning_rate': 4.3181818181818185e-06, 'num_tokens': 3372924.0, 'completions/mean_length': 149.625, 'completions/min_length': 129.0, 'completions/max_length': 186.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 149.625, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 186.0, 'rewards/meter/mean': 0.9845083951950073, 'rewards/meter/std': 0.008187457919120789, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8939747214317322, 'rewards/repeat_soft/std': 0.06038915738463402, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5255609750747681, 'rewards/total_composite/std': 0.05376099422574043, 'reward': 0.5255609750747681, 'reward_std': 0.053760990500450134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1152753233909607, 'sampling/sampling_logp_difference/max': 5.3994598388671875, 'sampling/importance_sampling_ratio/min': 0.004519021604210138, 'sampling/importance_sampling_ratio/mean': 1.0056995153427124, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7588530033826828, 'clip_ratio/low_mean': 0.020154332742094994, 'clip_ratio/low_min': 0.020154332742094994, 'clip_ratio/high_mean': 0.09387489315122366, 'clip_ratio/high_max': 0.09387489315122366, 'clip_ratio/region_mean': 0.11402922589331865, 'reward_total_mean': 0.5255609750747681, 'reward_meter_mean': 0.9845083951950073, 'reward_meter_std': 0.008187457919120789, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8939747214317322, 'reward_repeat_soft_std': 0.06038915738463402, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5255609750747681, 'reward_total_composite_std': 0.05376099422574043, 'epoch': 0.19}

 57%|█████▋    | 1876/3300 [4:37:07<3:11:12,  8.06s/it]INFO 04-13 11:58:05 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1877/3300 [4:37:14<3:02:29,  7.69s/it]2026-04-13 11:58:12,442 | INFO | train_grpo_train | metrics_logged mode=train step=1877

                                                       
{'loss': 0.0987, 'grad_norm': 7.469276428222656, 'learning_rate': 4.315151515151516e-06, 'num_tokens': 3375096.0, 'completions/mean_length': 102.5, 'completions/min_length': 91.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.5, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.992778480052948, 'rewards/meter/std': 0.0032708081416785717, 'rewards/count_adherence/mean': 0.6666666865348816, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8459296226501465, 'rewards/repeat_soft/std': 0.05445515364408493, 'rewards/judge_quality/mean': 0.32624998688697815, 'rewards/judge_quality/std': 0.11350739002227783, 'rewards/total_composite/mean': 0.4707384705543518, 'rewards/total_composite/std': 0.07509089261293411, 'reward': 0.4707384705543518, 'reward_std': 0.07509089261293411, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09624933451414108, 'sampling/sampling_logp_difference/max': 1.7367115020751953, 'sampling/importance_sampling_ratio/min': 0.1760985553264618, 'sampling/importance_sampling_ratio/mean': 1.0022552013397217, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.560472272336483, 'clip_ratio/low_mean': 0.03613204089924693, 'clip_ratio/low_min': 0.03613204089924693, 'clip_ratio/high_mean': 0.04738752171397209, 'clip_ratio/high_max': 0.04738752171397209, 'clip_ratio/region_mean': 0.08351956261321902, 'reward_total_mean': 0.4707384705543518, 'reward_meter_mean': 0.992778480052948, 'reward_meter_std': 0.0032708081416785717, 'reward_count_adherence_mean': 0.6666666865348816, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8459296226501465, 'reward_repeat_soft_std': 0.05445515364408493, 'reward_judge_quality_mean': 0.32624998688697815, 'reward_judge_quality_std': 0.11350739002227783, 'reward_total_composite_mean': 0.4707384705543518, 'reward_total_composite_std': 0.07509089261293411, 'epoch': 0.19}

 57%|█████▋    | 1877/3300 [4:37:14<3:02:29,  7.69s/it]INFO 04-13 11:58:12 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1878/3300 [4:37:21<2:56:15,  7.44s/it]2026-04-13 11:58:19,239 | INFO | train_grpo_train | metrics_logged mode=train step=1878

                                                       
{'loss': -0.0117, 'grad_norm': 6.692818641662598, 'learning_rate': 4.312121212121212e-06, 'num_tokens': 3376917.0, 'completions/mean_length': 59.625, 'completions/min_length': 54.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.625, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.893966794013977, 'rewards/meter/std': 0.19910527765750885, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7298498153686523, 'rewards/repeat_soft/std': 0.05829789489507675, 'rewards/judge_quality/mean': 0.36000001430511475, 'rewards/judge_quality/std': 0.1302744299173355, 'rewards/total_composite/mean': 0.5063621997833252, 'rewards/total_composite/std': 0.08025981485843658, 'reward': 0.5063621997833252, 'reward_std': 0.08025981485843658, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05902862176299095, 'sampling/sampling_logp_difference/max': 1.0365785360336304, 'sampling/importance_sampling_ratio/min': 0.3546660840511322, 'sampling/importance_sampling_ratio/mean': 1.0158969163894653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30942083708941936, 'clip_ratio/low_mean': 0.027048735413700342, 'clip_ratio/low_min': 0.027048735413700342, 'clip_ratio/high_mean': 0.031222944846376777, 'clip_ratio/high_max': 0.031222944846376777, 'clip_ratio/region_mean': 0.05827168026007712, 'reward_total_mean': 0.5063621997833252, 'reward_meter_mean': 0.893966794013977, 'reward_meter_std': 0.19910527765750885, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7298498153686523, 'reward_repeat_soft_std': 0.05829789489507675, 'reward_judge_quality_mean': 0.36000001430511475, 'reward_judge_quality_std': 0.1302744299173355, 'reward_total_composite_mean': 0.5063621997833252, 'reward_total_composite_std': 0.08025981485843658, 'epoch': 0.19}

 57%|█████▋    | 1878/3300 [4:37:21<2:56:15,  7.44s/it]INFO 04-13 11:58:19 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1879/3300 [4:37:28<2:56:09,  7.44s/it]2026-04-13 11:58:26,682 | INFO | train_grpo_train | metrics_logged mode=train step=1879

                                                       
{'loss': 0.0134, 'grad_norm': 6.908302307128906, 'learning_rate': 4.309090909090909e-06, 'num_tokens': 3379144.0, 'completions/mean_length': 99.375, 'completions/min_length': 69.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 99.375, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9821760058403015, 'rewards/meter/std': 0.0065070525743067265, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7847424745559692, 'rewards/repeat_soft/std': 0.08276831358671188, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.5845522880554199, 'rewards/total_composite/std': 0.06448714435100555, 'reward': 0.5845522880554199, 'reward_std': 0.06448714435100555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09772232174873352, 'sampling/sampling_logp_difference/max': 1.08424711227417, 'sampling/importance_sampling_ratio/min': 0.3381562829017639, 'sampling/importance_sampling_ratio/mean': 1.0168607234954834, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5596269220113754, 'clip_ratio/low_mean': 0.077846416272223, 'clip_ratio/low_min': 0.077846416272223, 'clip_ratio/high_mean': 0.010714286006987095, 'clip_ratio/high_max': 0.010714286006987095, 'clip_ratio/region_mean': 0.08856070227921009, 'reward_total_mean': 0.5845522880554199, 'reward_meter_mean': 0.9821760058403015, 'reward_meter_std': 0.0065070525743067265, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7847424745559692, 'reward_repeat_soft_std': 0.08276831358671188, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.5845522880554199, 'reward_total_composite_std': 0.06448714435100555, 'epoch': 0.19}

 57%|█████▋    | 1879/3300 [4:37:28<2:56:09,  7.44s/it]INFO 04-13 11:58:26 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1880/3300 [4:37:35<2:52:24,  7.28s/it]2026-04-13 11:58:33,627 | INFO | train_grpo_train | metrics_logged mode=train step=1880

                                                       
{'loss': 0.2086, 'grad_norm': 14.22120189666748, 'learning_rate': 4.306060606060607e-06, 'num_tokens': 3380772.0, 'completions/mean_length': 43.5, 'completions/min_length': 33.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.8323148488998413, 'rewards/meter/std': 0.3391997814178467, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8871290683746338, 'rewards/repeat_soft/std': 0.05498763918876648, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.5602912902832031, 'rewards/total_composite/std': 0.08942002058029175, 'reward': 0.5602912902832031, 'reward_std': 0.08942001312971115, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11437718570232391, 'sampling/sampling_logp_difference/max': 1.2057747840881348, 'sampling/importance_sampling_ratio/min': 0.29945987462997437, 'sampling/importance_sampling_ratio/mean': 1.0102508068084717, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6011449582874775, 'clip_ratio/low_mean': 0.017480531707406044, 'clip_ratio/low_min': 0.017480531707406044, 'clip_ratio/high_mean': 0.08124872390180826, 'clip_ratio/high_max': 0.08124872390180826, 'clip_ratio/region_mean': 0.0987292556092143, 'reward_total_mean': 0.5602912902832031, 'reward_meter_mean': 0.8323148488998413, 'reward_meter_std': 0.3391997814178467, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8871290683746338, 'reward_repeat_soft_std': 0.05498763918876648, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.5602912902832031, 'reward_total_composite_std': 0.08942002058029175, 'epoch': 0.19}

 57%|█████▋    | 1880/3300 [4:37:35<2:52:24,  7.28s/it]INFO 04-13 11:58:33 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1881/3300 [4:37:47<3:20:59,  8.50s/it]2026-04-13 11:58:44,937 | INFO | train_grpo_train | metrics_logged mode=train step=1881

                                                       
{'loss': -0.1174, 'grad_norm': 2.324276924133301, 'learning_rate': 4.303030303030303e-06, 'num_tokens': 3382480.0, 'completions/mean_length': 106.5, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.57143020629883, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.6712371110916138, 'rewards/meter/std': 0.39124536514282227, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9916514158248901, 'rewards/repeat_soft/std': 0.009813602082431316, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.47794660925865173, 'rewards/total_composite/std': 0.21686643362045288, 'reward': 0.47794660925865173, 'reward_std': 0.21686643362045288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10020927339792252, 'sampling/sampling_logp_difference/max': 2.10355281829834, 'sampling/importance_sampling_ratio/min': 0.1220221295952797, 'sampling/importance_sampling_ratio/mean': 1.0279405117034912, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5319827161729336, 'clip_ratio/low_mean': 0.021701388992369175, 'clip_ratio/low_min': 0.021701388992369175, 'clip_ratio/high_mean': 0.04760117130354047, 'clip_ratio/high_max': 0.04760117130354047, 'clip_ratio/region_mean': 0.06930256029590964, 'reward_total_mean': 0.47794660925865173, 'reward_meter_mean': 0.6712371110916138, 'reward_meter_std': 0.39124536514282227, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9916514158248901, 'reward_repeat_soft_std': 0.009813602082431316, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.47794660925865173, 'reward_total_composite_std': 0.21686643362045288, 'epoch': 0.19}

 57%|█████▋    | 1881/3300 [4:37:47<3:20:59,  8.50s/it]INFO 04-13 11:58:45 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1882/3300 [4:37:53<3:04:05,  7.79s/it]2026-04-13 11:58:51,068 | INFO | train_grpo_train | metrics_logged mode=train step=1882

                                                       
{'loss': 0.0771, 'grad_norm': 12.32876205444336, 'learning_rate': 4.3e-06, 'num_tokens': 3384013.0, 'completions/mean_length': 28.625, 'completions/min_length': 22.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.625, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.9835982322692871, 'rewards/meter/std': 0.014550337567925453, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9614755511283875, 'rewards/repeat_soft/std': 0.002003318862989545, 'rewards/judge_quality/mean': 0.39750000834465027, 'rewards/judge_quality/std': 0.10110107809305191, 'rewards/total_composite/mean': 0.5981491804122925, 'rewards/total_composite/std': 0.06445374339818954, 'reward': 0.5981491804122925, 'reward_std': 0.06445373594760895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10443581640720367, 'sampling/sampling_logp_difference/max': 1.2602341175079346, 'sampling/importance_sampling_ratio/min': 0.28358760476112366, 'sampling/importance_sampling_ratio/mean': 1.012134075164795, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6408927664160728, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.12095134239643812, 'clip_ratio/high_max': 0.12095134239643812, 'clip_ratio/region_mean': 0.12852710019797087, 'reward_total_mean': 0.5981491804122925, 'reward_meter_mean': 0.9835982322692871, 'reward_meter_std': 0.014550337567925453, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9614755511283875, 'reward_repeat_soft_std': 0.002003318862989545, 'reward_judge_quality_mean': 0.39750000834465027, 'reward_judge_quality_std': 0.10110107809305191, 'reward_total_composite_mean': 0.5981491804122925, 'reward_total_composite_std': 0.06445374339818954, 'epoch': 0.19}

 57%|█████▋    | 1882/3300 [4:37:53<3:04:05,  7.79s/it]INFO 04-13 11:58:51 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1883/3300 [4:38:04<3:29:46,  8.88s/it]2026-04-13 11:59:02,509 | INFO | train_grpo_train | metrics_logged mode=train step=1883

                                                       
{'loss': -0.212, 'grad_norm': 1.3718756437301636, 'learning_rate': 4.296969696969698e-06, 'num_tokens': 3386258.0, 'completions/mean_length': 176.625, 'completions/min_length': 97.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 128.71429443359375, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 161.0, 'rewards/meter/mean': 0.9235599040985107, 'rewards/meter/std': 0.10048312693834305, 'rewards/count_adherence/mean': 0.78125, 'rewards/count_adherence/std': 0.24775780737400055, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7810044884681702, 'rewards/repeat_soft/std': 0.10195527225732803, 'rewards/judge_quality/mean': 0.398749977350235, 'rewards/judge_quality/std': 0.15733835101127625, 'rewards/total_composite/mean': 0.4790325462818146, 'rewards/total_composite/std': 0.1957475244998932, 'reward': 0.4790325462818146, 'reward_std': 0.195747509598732, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0761503279209137, 'sampling/sampling_logp_difference/max': 2.0421500205993652, 'sampling/importance_sampling_ratio/min': 0.12974944710731506, 'sampling/importance_sampling_ratio/mean': 1.0136713981628418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39990518242120743, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.061085117515176535, 'clip_ratio/high_max': 0.061085117515176535, 'clip_ratio/region_mean': 0.061085117515176535, 'reward_total_mean': 0.4790325462818146, 'reward_meter_mean': 0.9235599040985107, 'reward_meter_std': 0.10048312693834305, 'reward_count_adherence_mean': 0.78125, 'reward_count_adherence_std': 0.24775780737400055, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7810044884681702, 'reward_repeat_soft_std': 0.10195527225732803, 'reward_judge_quality_mean': 0.398749977350235, 'reward_judge_quality_std': 0.15733835101127625, 'reward_total_composite_mean': 0.4790325462818146, 'reward_total_composite_std': 0.1957475244998932, 'epoch': 0.19}

 57%|█████▋    | 1883/3300 [4:38:04<3:29:46,  8.88s/it]INFO 04-13 11:59:02 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1884/3300 [4:38:10<3:09:57,  8.05s/it]2026-04-13 11:59:08,616 | INFO | train_grpo_train | metrics_logged mode=train step=1884

                                                       
{'loss': 0.0481, 'grad_norm': 12.55371379852295, 'learning_rate': 4.293939393939394e-06, 'num_tokens': 3387991.0, 'completions/mean_length': 42.625, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9154489636421204, 'rewards/meter/std': 0.09620413929224014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8638527989387512, 'rewards/repeat_soft/std': 0.03854406252503395, 'rewards/judge_quality/mean': 0.9200000166893005, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.8770164251327515, 'rewards/total_composite/std': 0.054819393903017044, 'reward': 0.8770164251327515, 'reward_std': 0.05481937527656555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08091326057910919, 'sampling/sampling_logp_difference/max': 2.1089088916778564, 'sampling/importance_sampling_ratio/min': 0.12137032300233841, 'sampling/importance_sampling_ratio/mean': 1.006975769996643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43506034091115, 'clip_ratio/low_mean': 0.01957590552046895, 'clip_ratio/low_min': 0.01957590552046895, 'clip_ratio/high_mean': 0.0516329575330019, 'clip_ratio/high_max': 0.0516329575330019, 'clip_ratio/region_mean': 0.07120886305347085, 'reward_total_mean': 0.8770164251327515, 'reward_meter_mean': 0.9154489636421204, 'reward_meter_std': 0.09620413929224014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8638527989387512, 'reward_repeat_soft_std': 0.03854406252503395, 'reward_judge_quality_mean': 0.9200000166893005, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.8770164251327515, 'reward_total_composite_std': 0.054819393903017044, 'epoch': 0.19}

 57%|█████▋    | 1884/3300 [4:38:10<3:09:57,  8.05s/it]INFO 04-13 11:59:08 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1885/3300 [4:38:16<2:53:38,  7.36s/it]2026-04-13 11:59:14,374 | INFO | train_grpo_train | metrics_logged mode=train step=1885

                                                       
{'loss': -0.0067, 'grad_norm': 17.774818420410156, 'learning_rate': 4.290909090909091e-06, 'num_tokens': 3389345.0, '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': 0.8867471218109131, 'rewards/meter/std': 0.17445634305477142, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9166569113731384, 'rewards/repeat_soft/std': 0.10989454388618469, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.22403763234615326, 'rewards/total_composite/mean': 0.5397394299507141, 'rewards/total_composite/std': 0.26202690601348877, 'reward': 0.5397394299507141, 'reward_std': 0.26202690601348877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.19745789468288422, 'sampling/sampling_logp_difference/max': 1.2882919311523438, 'sampling/importance_sampling_ratio/min': 0.2757413685321808, 'sampling/importance_sampling_ratio/mean': 1.0029056072235107, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.4838181287050247, 'clip_ratio/low_mean': 0.07601686753332615, 'clip_ratio/low_min': 0.07601686753332615, 'clip_ratio/high_mean': 0.1273423619568348, 'clip_ratio/high_max': 0.1273423619568348, 'clip_ratio/region_mean': 0.20335922949016094, 'reward_total_mean': 0.5397394299507141, 'reward_meter_mean': 0.8867471218109131, 'reward_meter_std': 0.17445634305477142, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9166569113731384, 'reward_repeat_soft_std': 0.10989454388618469, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.22403763234615326, 'reward_total_composite_mean': 0.5397394299507141, 'reward_total_composite_std': 0.26202690601348877, 'epoch': 0.19}

 57%|█████▋    | 1885/3300 [4:38:16<2:53:38,  7.36s/it]INFO 04-13 11:59:14 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1886/3300 [4:38:23<2:47:30,  7.11s/it]2026-04-13 11:59:20,899 | INFO | train_grpo_train | metrics_logged mode=train step=1886

                                                       
{'loss': -0.0556, 'grad_norm': 10.604535102844238, 'learning_rate': 4.287878787878788e-06, 'num_tokens': 3390990.0, 'completions/mean_length': 35.625, 'completions/min_length': 30.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.625, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.9604357481002808, 'rewards/meter/std': 0.0025081527419388294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8992391228675842, 'rewards/repeat_soft/std': 0.05777277797460556, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.6040985584259033, 'rewards/total_composite/std': 0.014271872118115425, 'reward': 0.6040985584259033, 'reward_std': 0.014271865598857403, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06473870575428009, 'sampling/sampling_logp_difference/max': 1.0003001689910889, 'sampling/importance_sampling_ratio/min': 0.3677690625190735, 'sampling/importance_sampling_ratio/mean': 1.0079113245010376, 'sampling/importance_sampling_ratio/max': 1.7900887727737427, 'entropy': 0.3560235071927309, 'clip_ratio/low_mean': 0.04741950798779726, 'clip_ratio/low_min': 0.04741950798779726, 'clip_ratio/high_mean': 0.035813594702631235, 'clip_ratio/high_max': 0.035813594702631235, 'clip_ratio/region_mean': 0.0832331026904285, 'reward_total_mean': 0.6040985584259033, 'reward_meter_mean': 0.9604357481002808, 'reward_meter_std': 0.0025081527419388294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8992391228675842, 'reward_repeat_soft_std': 0.05777277797460556, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.6040985584259033, 'reward_total_composite_std': 0.014271872118115425, 'epoch': 0.19}

 57%|█████▋    | 1886/3300 [4:38:23<2:47:30,  7.11s/it]INFO 04-13 11:59:21 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1887/3300 [4:38:29<2:40:22,  6.81s/it]2026-04-13 11:59:26,992 | INFO | train_grpo_train | metrics_logged mode=train step=1887

                                                       
{'loss': 0.0415, 'grad_norm': 9.261543273925781, 'learning_rate': 4.284848484848485e-06, 'num_tokens': 3392507.0, 'completions/mean_length': 44.625, 'completions/min_length': 31.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.6726040244102478, 'rewards/meter/std': 0.40645524859428406, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.790650486946106, 'rewards/repeat_soft/std': 0.07165295630693436, 'rewards/judge_quality/mean': 0.4099999964237213, 'rewards/judge_quality/std': 0.06633248925209045, 'rewards/total_composite/mean': 0.4995495676994324, 'rewards/total_composite/std': 0.11973252892494202, 'reward': 0.4995495676994324, 'reward_std': 0.11973252147436142, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07547628879547119, 'sampling/sampling_logp_difference/max': 1.427231788635254, 'sampling/importance_sampling_ratio/min': 0.2399723082780838, 'sampling/importance_sampling_ratio/mean': 1.0029417276382446, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39059338346123695, 'clip_ratio/low_mean': 0.027143463492393494, 'clip_ratio/low_min': 0.027143463492393494, 'clip_ratio/high_mean': 0.04254913330078125, 'clip_ratio/high_max': 0.04254913330078125, 'clip_ratio/region_mean': 0.06969259679317474, 'reward_total_mean': 0.4995495676994324, 'reward_meter_mean': 0.6726040244102478, 'reward_meter_std': 0.40645524859428406, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.790650486946106, 'reward_repeat_soft_std': 0.07165295630693436, 'reward_judge_quality_mean': 0.4099999964237213, 'reward_judge_quality_std': 0.06633248925209045, 'reward_total_composite_mean': 0.4995495676994324, 'reward_total_composite_std': 0.11973252892494202, 'epoch': 0.19}

 57%|█████▋    | 1887/3300 [4:38:29<2:40:22,  6.81s/it]INFO 04-13 11:59:27 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1888/3300 [4:38:36<2:42:44,  6.92s/it]2026-04-13 11:59:34,161 | INFO | train_grpo_train | metrics_logged mode=train step=1888

                                                       
{'loss': 0.0792, 'grad_norm': 7.557681083679199, 'learning_rate': 4.281818181818182e-06, 'num_tokens': 3394379.0, 'completions/mean_length': 72.0, 'completions/min_length': 55.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 72.0, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8011589646339417, 'rewards/meter/std': 0.33095604181289673, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.830003023147583, 'rewards/repeat_soft/std': 0.04706054553389549, 'rewards/judge_quality/mean': 0.41499999165534973, 'rewards/judge_quality/std': 0.14520922303199768, 'rewards/total_composite/mean': 0.532202959060669, 'rewards/total_composite/std': 0.1358167976140976, 'reward': 0.532202959060669, 'reward_std': 0.1358168125152588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11180947721004486, 'sampling/sampling_logp_difference/max': 1.8365421295166016, 'sampling/importance_sampling_ratio/min': 0.15936754643917084, 'sampling/importance_sampling_ratio/mean': 1.0054874420166016, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.697334386408329, 'clip_ratio/low_mean': 0.044057718478143215, 'clip_ratio/low_min': 0.044057718478143215, 'clip_ratio/high_mean': 0.037179597187787294, 'clip_ratio/high_max': 0.037179597187787294, 'clip_ratio/region_mean': 0.08123731566593051, 'reward_total_mean': 0.532202959060669, 'reward_meter_mean': 0.8011589646339417, 'reward_meter_std': 0.33095604181289673, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.830003023147583, 'reward_repeat_soft_std': 0.04706054553389549, 'reward_judge_quality_mean': 0.41499999165534973, 'reward_judge_quality_std': 0.14520922303199768, 'reward_total_composite_mean': 0.532202959060669, 'reward_total_composite_std': 0.1358167976140976, 'epoch': 0.19}

 57%|█████▋    | 1888/3300 [4:38:36<2:42:44,  6.92s/it]INFO 04-13 11:59:34 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1889/3300 [4:38:48<3:17:22,  8.39s/it]2026-04-13 11:59:46,006 | INFO | train_grpo_train | metrics_logged mode=train step=1889

                                                       
{'loss': -0.1033, 'grad_norm': 3.4962470531463623, 'learning_rate': 4.278787878787879e-06, 'num_tokens': 3395937.0, 'completions/mean_length': 96.75, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.42857360839844, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5064377784729004, 'rewards/meter/std': 0.45250165462493896, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9447498321533203, 'rewards/repeat_soft/std': 0.08484742045402527, 'rewards/judge_quality/mean': 0.3737499713897705, 'rewards/judge_quality/std': 0.13081474602222443, 'rewards/total_composite/mean': 0.436312198638916, 'rewards/total_composite/std': 0.20593100786209106, 'reward': 0.436312198638916, 'reward_std': 0.20593099296092987, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15999017655849457, 'sampling/sampling_logp_difference/max': 1.2710683345794678, 'sampling/importance_sampling_ratio/min': 0.28053176403045654, 'sampling/importance_sampling_ratio/mean': 1.0161256790161133, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.784225694835186, 'clip_ratio/low_mean': 0.053742311894893646, 'clip_ratio/low_min': 0.053742311894893646, 'clip_ratio/high_mean': 0.07266347017139196, 'clip_ratio/high_max': 0.07266347017139196, 'clip_ratio/region_mean': 0.1264057820662856, 'reward_total_mean': 0.436312198638916, 'reward_meter_mean': 0.5064377784729004, 'reward_meter_std': 0.45250165462493896, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9447498321533203, 'reward_repeat_soft_std': 0.08484742045402527, 'reward_judge_quality_mean': 0.3737499713897705, 'reward_judge_quality_std': 0.13081474602222443, 'reward_total_composite_mean': 0.436312198638916, 'reward_total_composite_std': 0.20593100786209106, 'epoch': 0.19}

 57%|█████▋    | 1889/3300 [4:38:48<3:17:22,  8.39s/it]INFO 04-13 11:59:46 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1890/3300 [4:38:55<3:09:26,  8.06s/it]2026-04-13 11:59:53,293 | INFO | train_grpo_train | metrics_logged mode=train step=1890

                                                       
{'loss': -0.0124, 'grad_norm': 5.567215442657471, 'learning_rate': 4.275757575757576e-06, 'num_tokens': 3398548.0, 'completions/mean_length': 138.375, 'completions/min_length': 122.0, 'completions/max_length': 149.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.375, 'completions/min_terminated_length': 122.0, 'completions/max_terminated_length': 149.0, 'rewards/meter/mean': 0.9934046268463135, 'rewards/meter/std': 0.0010323392925783992, 'rewards/count_adherence/mean': 0.75, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7519139051437378, 'rewards/repeat_soft/std': 0.0979333445429802, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.5743303298950195, 'rewards/total_composite/std': 0.12184125930070877, 'reward': 0.5743303298950195, 'reward_std': 0.12184125185012817, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06597357243299484, 'sampling/sampling_logp_difference/max': 1.602780818939209, 'sampling/importance_sampling_ratio/min': 0.2013358473777771, 'sampling/importance_sampling_ratio/mean': 1.0086263418197632, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.35575790889561176, 'clip_ratio/low_mean': 0.04292728519067168, 'clip_ratio/low_min': 0.04292728519067168, 'clip_ratio/high_mean': 0.014261744916439056, 'clip_ratio/high_max': 0.014261744916439056, 'clip_ratio/region_mean': 0.05718903010711074, 'reward_total_mean': 0.5743303298950195, 'reward_meter_mean': 0.9934046268463135, 'reward_meter_std': 0.0010323392925783992, 'reward_count_adherence_mean': 0.75, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7519139051437378, 'reward_repeat_soft_std': 0.0979333445429802, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.5743303298950195, 'reward_total_composite_std': 0.12184125930070877, 'epoch': 0.19}

 57%|█████▋    | 1890/3300 [4:38:55<3:09:26,  8.06s/it]INFO 04-13 11:59:53 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1891/3300 [4:39:01<2:55:31,  7.47s/it]2026-04-13 11:59:59,402 | INFO | train_grpo_train | metrics_logged mode=train step=1891

                                                       
{'loss': 0.0069, 'grad_norm': 11.446745872497559, 'learning_rate': 4.272727272727273e-06, 'num_tokens': 3400145.0, 'completions/mean_length': 33.625, 'completions/min_length': 29.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.625, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.9276425838470459, 'rewards/meter/std': 0.09783067554235458, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8701931238174438, 'rewards/repeat_soft/std': 0.053935930132865906, 'rewards/judge_quality/mean': 0.42750000953674316, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.5884767770767212, 'rewards/total_composite/std': 0.03028133139014244, 'reward': 0.5884767770767212, 'reward_std': 0.03028133697807789, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08166030049324036, 'sampling/sampling_logp_difference/max': 1.5011775493621826, 'sampling/importance_sampling_ratio/min': 0.290174663066864, 'sampling/importance_sampling_ratio/mean': 0.9986169338226318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30929396115243435, 'clip_ratio/low_mean': 0.007195723708719015, 'clip_ratio/low_min': 0.007195723708719015, 'clip_ratio/high_mean': 0.03407761291600764, 'clip_ratio/high_max': 0.03407761291600764, 'clip_ratio/region_mean': 0.04127333662472665, 'reward_total_mean': 0.5884767770767212, 'reward_meter_mean': 0.9276425838470459, 'reward_meter_std': 0.09783067554235458, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8701931238174438, 'reward_repeat_soft_std': 0.053935930132865906, 'reward_judge_quality_mean': 0.42750000953674316, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.5884767770767212, 'reward_total_composite_std': 0.03028133139014244, 'epoch': 0.19}

 57%|█████▋    | 1891/3300 [4:39:01<2:55:31,  7.47s/it]INFO 04-13 11:59:59 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1892/3300 [4:39:08<2:49:53,  7.24s/it]2026-04-13 12:00:06,098 | INFO | train_grpo_train | metrics_logged mode=train step=1892

                                                       
{'loss': 0.1411, 'grad_norm': 9.459722518920898, 'learning_rate': 4.2696969696969695e-06, 'num_tokens': 3402135.0, 'completions/mean_length': 78.75, 'completions/min_length': 49.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.75, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.8605332374572754, 'rewards/meter/std': 0.20299570262432098, 'rewards/count_adherence/mean': 0.7916666865348816, 'rewards/count_adherence/std': 0.17251639068126678, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8585145473480225, 'rewards/repeat_soft/std': 0.030557211488485336, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.5244558453559875, 'rewards/total_composite/std': 0.06855473667383194, 'reward': 0.5244558453559875, 'reward_std': 0.06855473667383194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09535474330186844, 'sampling/sampling_logp_difference/max': 3.2019381523132324, 'sampling/importance_sampling_ratio/min': 0.04068327695131302, 'sampling/importance_sampling_ratio/mean': 1.0037459135055542, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.468066081404686, 'clip_ratio/low_mean': 0.03889781702309847, 'clip_ratio/low_min': 0.03889781702309847, 'clip_ratio/high_mean': 0.039728786796331406, 'clip_ratio/high_max': 0.039728786796331406, 'clip_ratio/region_mean': 0.07862660381942987, 'reward_total_mean': 0.5244558453559875, 'reward_meter_mean': 0.8605332374572754, 'reward_meter_std': 0.20299570262432098, 'reward_count_adherence_mean': 0.7916666865348816, 'reward_count_adherence_std': 0.17251639068126678, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8585145473480225, 'reward_repeat_soft_std': 0.030557211488485336, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.5244558453559875, 'reward_total_composite_std': 0.06855473667383194, 'epoch': 0.19}

 57%|█████▋    | 1892/3300 [4:39:08<2:49:53,  7.24s/it]INFO 04-13 12:00:06 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1893/3300 [4:39:14<2:46:06,  7.08s/it]2026-04-13 12:00:12,807 | INFO | train_grpo_train | metrics_logged mode=train step=1893

                                                       
{'loss': 0.0269, 'grad_norm': 10.69278621673584, 'learning_rate': 4.266666666666668e-06, 'num_tokens': 3403761.0, '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': 0.9858983755111694, 'rewards/meter/std': 0.004938791040331125, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9576983451843262, 'rewards/repeat_soft/std': 0.021132754161953926, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6529911756515503, 'rewards/total_composite/std': 0.11349096894264221, 'reward': 0.6529911756515503, 'reward_std': 0.11349096894264221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08822016417980194, 'sampling/sampling_logp_difference/max': 1.7165651321411133, 'sampling/importance_sampling_ratio/min': 0.17968228459358215, 'sampling/importance_sampling_ratio/mean': 1.0064579248428345, 'sampling/importance_sampling_ratio/max': 1.6938608884811401, 'entropy': 0.4698624052107334, 'clip_ratio/low_mean': 0.06939548999071121, 'clip_ratio/low_min': 0.06939548999071121, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.08075912669301033, 'reward_total_mean': 0.6529911756515503, 'reward_meter_mean': 0.9858983755111694, 'reward_meter_std': 0.004938791040331125, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9576983451843262, 'reward_repeat_soft_std': 0.021132754161953926, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6529911756515503, 'reward_total_composite_std': 0.11349096894264221, 'epoch': 0.19}

 57%|█████▋    | 1893/3300 [4:39:15<2:46:06,  7.08s/it]INFO 04-13 12:00:13 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1894/3300 [4:39:22<2:47:28,  7.15s/it]2026-04-13 12:00:20,114 | INFO | train_grpo_train | metrics_logged mode=train step=1894

                                                       
{'loss': 0.0732, 'grad_norm': 5.861888885498047, 'learning_rate': 4.263636363636364e-06, 'num_tokens': 3406006.0, 'completions/mean_length': 112.625, 'completions/min_length': 79.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.625, 'completions/min_terminated_length': 79.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.9630520939826965, 'rewards/meter/std': 0.07279664278030396, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7933350801467896, 'rewards/repeat_soft/std': 0.11016028374433517, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.5295034646987915, 'rewards/total_composite/std': 0.055732253938913345, 'reward': 0.5295034646987915, 'reward_std': 0.05573224276304245, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09980825334787369, 'sampling/sampling_logp_difference/max': 4.133489608764648, 'sampling/importance_sampling_ratio/min': 0.016026854515075684, 'sampling/importance_sampling_ratio/mean': 1.0049759149551392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4927608408033848, 'clip_ratio/low_mean': 0.02559055434539914, 'clip_ratio/low_min': 0.02559055434539914, 'clip_ratio/high_mean': 0.07266552560031414, 'clip_ratio/high_max': 0.07266552560031414, 'clip_ratio/region_mean': 0.09825607994571328, 'reward_total_mean': 0.5295034646987915, 'reward_meter_mean': 0.9630520939826965, 'reward_meter_std': 0.07279664278030396, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7933350801467896, 'reward_repeat_soft_std': 0.11016028374433517, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.5295034646987915, 'reward_total_composite_std': 0.055732253938913345, 'epoch': 0.19}

 57%|█████▋    | 1894/3300 [4:39:22<2:47:28,  7.15s/it]INFO 04-13 12:00:20 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1895/3300 [4:39:29<2:46:59,  7.13s/it]2026-04-13 12:00:27,207 | INFO | train_grpo_train | metrics_logged mode=train step=1895

                                                       
{'loss': 0.0533, 'grad_norm': 11.755945205688477, 'learning_rate': 4.260606060606061e-06, 'num_tokens': 3408047.0, 'completions/mean_length': 89.125, 'completions/min_length': 77.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.125, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.5747484564781189, 'rewards/meter/std': 0.36859145760536194, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.13363061845302582, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9119118452072144, 'rewards/repeat_soft/std': 0.03609064221382141, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.07869470119476318, 'rewards/total_composite/mean': 0.4490486979484558, 'rewards/total_composite/std': 0.10694235563278198, 'reward': 0.4490486979484558, 'reward_std': 0.10694236308336258, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14104905724525452, 'sampling/sampling_logp_difference/max': 3.7344534397125244, 'sampling/importance_sampling_ratio/min': 0.023886224254965782, 'sampling/importance_sampling_ratio/mean': 0.983735203742981, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.559051401913166, 'clip_ratio/low_mean': 0.04991707857698202, 'clip_ratio/low_min': 0.04991707857698202, 'clip_ratio/high_mean': 0.06310982350260019, 'clip_ratio/high_max': 0.06310982350260019, 'clip_ratio/region_mean': 0.11302690207958221, 'reward_total_mean': 0.4490486979484558, 'reward_meter_mean': 0.5747484564781189, 'reward_meter_std': 0.36859145760536194, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.13363061845302582, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9119118452072144, 'reward_repeat_soft_std': 0.03609064221382141, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.07869470119476318, 'reward_total_composite_mean': 0.4490486979484558, 'reward_total_composite_std': 0.10694235563278198, 'epoch': 0.19}

 57%|█████▋    | 1895/3300 [4:39:29<2:46:59,  7.13s/it]INFO 04-13 12:00:27 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1896/3300 [4:39:40<3:17:02,  8.42s/it]2026-04-13 12:00:38,633 | INFO | train_grpo_train | metrics_logged mode=train step=1896

                                                       
{'loss': -0.0977, 'grad_norm': 1.9929245710372925, 'learning_rate': 4.2575757575757585e-06, 'num_tokens': 3409574.0, 'completions/mean_length': 92.875, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.0, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.8006120324134827, 'rewards/meter/std': 0.2018144726753235, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9413082003593445, 'rewards/repeat_soft/std': 0.053778115659952164, 'rewards/judge_quality/mean': 0.3774999976158142, 'rewards/judge_quality/std': 0.13274572789669037, 'rewards/total_composite/mean': 0.4947419762611389, 'rewards/total_composite/std': 0.20789597928524017, 'reward': 0.4947419762611389, 'reward_std': 0.20789596438407898, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14270998537540436, 'sampling/sampling_logp_difference/max': 1.1320117712020874, 'sampling/importance_sampling_ratio/min': 0.3223840296268463, 'sampling/importance_sampling_ratio/mean': 1.0156162977218628, 'sampling/importance_sampling_ratio/max': 1.8814842700958252, 'entropy': 0.9908416122198105, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.10267268819734454, 'clip_ratio/high_max': 0.10267268819734454, 'clip_ratio/region_mean': 0.11024844599887729, 'reward_total_mean': 0.4947419762611389, 'reward_meter_mean': 0.8006120324134827, 'reward_meter_std': 0.2018144726753235, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9413082003593445, 'reward_repeat_soft_std': 0.053778115659952164, 'reward_judge_quality_mean': 0.3774999976158142, 'reward_judge_quality_std': 0.13274572789669037, 'reward_total_composite_mean': 0.4947419762611389, 'reward_total_composite_std': 0.20789597928524017, 'epoch': 0.19}

 57%|█████▋    | 1896/3300 [4:39:40<3:17:02,  8.42s/it]INFO 04-13 12:00:38 [block_pool.py:378] Successfully reset prefix cache

 57%|█████▋    | 1897/3300 [4:39:46<2:59:32,  7.68s/it]2026-04-13 12:00:44,576 | INFO | train_grpo_train | metrics_logged mode=train step=1897

                                                       
{'loss': -0.0055, 'grad_norm': 21.212984085083008, 'learning_rate': 4.254545454545455e-06, 'num_tokens': 3410863.0, 'completions/mean_length': 21.125, 'completions/min_length': 17.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9038698673248291, 'rewards/meter/std': 0.1721266359090805, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9348776340484619, 'rewards/repeat_soft/std': 0.054454438388347626, 'rewards/judge_quality/mean': 0.48250001668930054, 'rewards/judge_quality/std': 0.1767767071723938, 'rewards/total_composite/mean': 0.6266239881515503, 'rewards/total_composite/std': 0.13112255930900574, 'reward': 0.6266239881515503, 'reward_std': 0.13112254440784454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15069088339805603, 'sampling/sampling_logp_difference/max': 1.5056264400482178, 'sampling/importance_sampling_ratio/min': 0.34185534715652466, 'sampling/importance_sampling_ratio/mean': 1.0511457920074463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.155650481581688, 'clip_ratio/low_mean': 0.12694388814270496, 'clip_ratio/low_min': 0.12694388814270496, 'clip_ratio/high_mean': 0.010869565419852734, 'clip_ratio/high_max': 0.010869565419852734, 'clip_ratio/region_mean': 0.1378134535625577, 'reward_total_mean': 0.6266239881515503, 'reward_meter_mean': 0.9038698673248291, 'reward_meter_std': 0.1721266359090805, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9348776340484619, 'reward_repeat_soft_std': 0.054454438388347626, 'reward_judge_quality_mean': 0.48250001668930054, 'reward_judge_quality_std': 0.1767767071723938, 'reward_total_composite_mean': 0.6266239881515503, 'reward_total_composite_std': 0.13112255930900574, 'epoch': 0.19}

 57%|█████▋    | 1897/3300 [4:39:46<2:59:32,  7.68s/it]INFO 04-13 12:00:44 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1898/3300 [4:39:54<3:02:05,  7.79s/it]2026-04-13 12:00:52,659 | INFO | train_grpo_train | metrics_logged mode=train step=1898

                                                       
{'loss': 0.0936, 'grad_norm': 6.053118705749512, 'learning_rate': 4.251515151515152e-06, 'num_tokens': 3413535.0, 'completions/mean_length': 146.0, 'completions/min_length': 126.0, 'completions/max_length': 173.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 146.0, 'completions/min_terminated_length': 126.0, 'completions/max_terminated_length': 173.0, 'rewards/meter/mean': 0.7404102087020874, 'rewards/meter/std': 0.2980844974517822, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8465932607650757, 'rewards/repeat_soft/std': 0.0765487477183342, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.528233528137207, 'rewards/total_composite/std': 0.10594785958528519, 'reward': 0.528233528137207, 'reward_std': 0.10594785958528519, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09265933930873871, 'sampling/sampling_logp_difference/max': 1.8282966613769531, 'sampling/importance_sampling_ratio/min': 0.16068704426288605, 'sampling/importance_sampling_ratio/mean': 1.0120998620986938, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.529948141425848, 'clip_ratio/low_mean': 0.03747152350842953, 'clip_ratio/low_min': 0.03747152350842953, 'clip_ratio/high_mean': 0.030135025968775153, 'clip_ratio/high_max': 0.030135025968775153, 'clip_ratio/region_mean': 0.06760654947720468, 'reward_total_mean': 0.528233528137207, 'reward_meter_mean': 0.7404102087020874, 'reward_meter_std': 0.2980844974517822, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8465932607650757, 'reward_repeat_soft_std': 0.0765487477183342, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.528233528137207, 'reward_total_composite_std': 0.10594785958528519, 'epoch': 0.19}

 58%|█████▊    | 1898/3300 [4:39:54<3:02:05,  7.79s/it]INFO 04-13 12:00:52 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1899/3300 [4:40:01<2:51:22,  7.34s/it]2026-04-13 12:00:58,915 | INFO | train_grpo_train | metrics_logged mode=train step=1899

                                                       
{'loss': 0.1288, 'grad_norm': 13.905967712402344, 'learning_rate': 4.248484848484849e-06, 'num_tokens': 3415019.0, 'completions/mean_length': 41.5, 'completions/min_length': 33.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.8378007411956787, 'rewards/meter/std': 0.32178792357444763, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9815511107444763, 'rewards/repeat_soft/std': 0.011156021617352962, 'rewards/judge_quality/mean': 0.42374998331069946, 'rewards/judge_quality/std': 0.1524970829486847, 'rewards/total_composite/mean': 0.5784533023834229, 'rewards/total_composite/std': 0.13378991186618805, 'reward': 0.5784533023834229, 'reward_std': 0.13378991186618805, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13036690652370453, 'sampling/sampling_logp_difference/max': 1.4765377044677734, 'sampling/importance_sampling_ratio/min': 0.2284272164106369, 'sampling/importance_sampling_ratio/mean': 1.0019536018371582, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8005569577217102, 'clip_ratio/low_mean': 0.03787206578999758, 'clip_ratio/low_min': 0.03787206578999758, 'clip_ratio/high_mean': 0.1173378536477685, 'clip_ratio/high_max': 0.1173378536477685, 'clip_ratio/region_mean': 0.15520991943776608, 'reward_total_mean': 0.5784533023834229, 'reward_meter_mean': 0.8378007411956787, 'reward_meter_std': 0.32178792357444763, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9815511107444763, 'reward_repeat_soft_std': 0.011156021617352962, 'reward_judge_quality_mean': 0.42374998331069946, 'reward_judge_quality_std': 0.1524970829486847, 'reward_total_composite_mean': 0.5784533023834229, 'reward_total_composite_std': 0.13378991186618805, 'epoch': 0.19}

 58%|█████▊    | 1899/3300 [4:40:01<2:51:22,  7.34s/it]INFO 04-13 12:00:59 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1900/3300 [4:40:07<2:43:11,  6.99s/it]2026-04-13 12:01:05,118 | INFO | train_grpo_train | metrics_logged mode=train step=1900

                                                       
{'loss': -0.0636, 'grad_norm': 12.112760543823242, 'learning_rate': 4.245454545454546e-06, 'num_tokens': 3416585.0, 'completions/mean_length': 41.75, 'completions/min_length': 32.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.8482105135917664, 'rewards/meter/std': 0.23347781598567963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9181222915649414, 'rewards/repeat_soft/std': 0.0497940294444561, 'rewards/judge_quality/mean': 0.7950000166893005, 'rewards/judge_quality/std': 0.23145504295825958, 'rewards/total_composite/mean': 0.7733848094940186, 'rewards/total_composite/std': 0.18588495254516602, 'reward': 0.7733848094940186, 'reward_std': 0.18588493764400482, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14247700572013855, 'sampling/sampling_logp_difference/max': 0.990262508392334, 'sampling/importance_sampling_ratio/min': 0.38564333319664, 'sampling/importance_sampling_ratio/mean': 1.0247104167938232, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.1424989625811577, 'clip_ratio/low_mean': 0.056029041297733784, 'clip_ratio/low_min': 0.056029041297733784, 'clip_ratio/high_mean': 0.08261819859035313, 'clip_ratio/high_max': 0.08261819859035313, 'clip_ratio/region_mean': 0.13864723988808692, 'reward_total_mean': 0.7733848094940186, 'reward_meter_mean': 0.8482105135917664, 'reward_meter_std': 0.23347781598567963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9181222915649414, 'reward_repeat_soft_std': 0.0497940294444561, 'reward_judge_quality_mean': 0.7950000166893005, 'reward_judge_quality_std': 0.23145504295825958, 'reward_total_composite_mean': 0.7733848094940186, 'reward_total_composite_std': 0.18588495254516602, 'epoch': 0.19}

 58%|█████▊    | 1900/3300 [4:40:07<2:43:11,  6.99s/it]INFO 04-13 12:01:05 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.77s/it]

 30%|███       | 3/10 [00:08<00:21,  3.06s/it]

 40%|████      | 4/10 [00:12<00:20,  3.37s/it]

 50%|█████     | 5/10 [00:17<00:20,  4.01s/it]

 60%|██████    | 6/10 [00:22<00:16,  4.24s/it]

 70%|███████   | 7/10 [00:26<00:12,  4.26s/it]

 80%|████████  | 8/10 [00:30<00:08,  4.08s/it]

 90%|█████████ | 9/10 [00:34<00:04,  4.05s/it]

100%|██████████| 10/10 [00:37<00:00,  3.92s/it]2026-04-13 12:01:47,608 | INFO | train_grpo_train | metrics_logged mode=eval step=1900

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 42.3837, 'eval_samples_per_second': 1.888, 'eval_steps_per_second': 0.236, 'eval_num_tokens': 3416585.0, 'eval_completions/mean_length': 78.225, 'eval_completions/min_length': 32.6, 'eval_completions/max_length': 141.3, 'eval_completions/clipped_ratio': 0.0, 'eval_completions/mean_terminated_length': 78.225, 'eval_completions/min_terminated_length': 32.6, 'eval_completions/max_terminated_length': 141.3, 'eval_rewards/meter/mean': 0.8480098903179168, 'eval_rewards/meter/std': 0.23681166246533394, 'eval_rewards/count_adherence/mean': 0.9772916555404663, 'eval_rewards/count_adherence/std': 0.05044138841331005, 'eval_rewards/hard_gate/mean': 1.0, 'eval_rewards/hard_gate/std': 0.0, 'eval_rewards/repeat_soft/mean': 0.8693194091320038, 'eval_rewards/repeat_soft/std': 0.08963314779102802, 'eval_rewards/judge_quality/mean': 0.46762500107288363, 'eval_rewards/judge_quality/std': 0.12215339820832014, 'eval_rewards/total_composite/mean': 0.5825744986534118, 'eval_rewards/total_composite/std': 0.10912723653018475, 'eval_reward': 0.5825744986534118, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.05554373003542423, 'eval_sampling/sampling_logp_difference/max': 0.9246527910232544, 'eval_sampling/importance_sampling_ratio/min': 0.40136079490184784, 'eval_sampling/importance_sampling_ratio/mean': 1.0137291550636292, 'eval_sampling/importance_sampling_ratio/max': 1.4351726770401, 'eval_entropy': 0.5898070454597473, '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': 0.5825744986534118, 'eval_reward_meter_mean': 0.8480098903179168, 'eval_reward_meter_std': 0.23681166246533394, 'eval_reward_count_adherence_mean': 0.9772916555404663, 'eval_reward_count_adherence_std': 0.05044138841331005, 'eval_reward_hard_gate_mean': 1.0, 'eval_reward_hard_gate_std': 0.0, 'eval_reward_repeat_soft_mean': 0.8693194091320038, 'eval_reward_repeat_soft_std': 0.08963314779102802, 'eval_reward_judge_quality_mean': 0.46762500107288363, 'eval_reward_judge_quality_std': 0.12215339820832014, 'eval_reward_total_composite_mean': 0.5825744986534118, 'eval_reward_total_composite_std': 0.10912723653018475, 'epoch': 0.19}

 58%|█████▊    | 1900/3300 [4:40:49<2:43:11,  6.99s/it]

100%|██████████| 10/10 [00:37<00:00,  3.92s/it]

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

 58%|█████▊    | 1901/3300 [4:41:04<8:36:29, 22.15s/it]2026-04-13 12:02:02,638 | INFO | train_grpo_train | metrics_logged mode=train step=1901

                                                       
{'loss': -0.1149, 'grad_norm': 1.9661214351654053, 'learning_rate': 4.242424242424243e-06, 'num_tokens': 3418105.0, 'completions/mean_length': 162.0, 'completions/min_length': 35.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 45.333335876464844, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.6698601245880127, 'rewards/meter/std': 0.41053760051727295, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9652444124221802, 'rewards/repeat_soft/std': 0.03090461529791355, 'rewards/judge_quality/mean': 0.45249998569488525, 'rewards/judge_quality/std': 0.33065950870513916, 'rewards/total_composite/mean': 0.5056458115577698, 'rewards/total_composite/std': 0.36281538009643555, 'reward': 0.5056458115577698, 'reward_std': 0.36281538009643555, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11610380560159683, 'sampling/sampling_logp_difference/max': 1.0769240856170654, 'sampling/importance_sampling_ratio/min': 0.34064167737960815, 'sampling/importance_sampling_ratio/mean': 1.004833459854126, 'sampling/importance_sampling_ratio/max': 1.8720961809158325, 'entropy': 0.564405545592308, 'clip_ratio/low_mean': 0.01785714365541935, 'clip_ratio/low_min': 0.01785714365541935, 'clip_ratio/high_mean': 0.0837968117557466, 'clip_ratio/high_max': 0.0837968117557466, 'clip_ratio/region_mean': 0.10165395541116595, 'reward_total_mean': 0.5056458115577698, 'reward_meter_mean': 0.6698601245880127, 'reward_meter_std': 0.41053760051727295, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9652444124221802, 'reward_repeat_soft_std': 0.03090461529791355, 'reward_judge_quality_mean': 0.45249998569488525, 'reward_judge_quality_std': 0.33065950870513916, 'reward_total_composite_mean': 0.5056458115577698, 'reward_total_composite_std': 0.36281538009643555, 'epoch': 0.19}

 58%|█████▊    | 1901/3300 [4:41:04<8:36:29, 22.15s/it]INFO 04-13 12:02:02 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1902/3300 [4:41:16<7:25:19, 19.11s/it]2026-04-13 12:02:14,716 | INFO | train_grpo_train | metrics_logged mode=train step=1902

                                                       
{'loss': -0.066, 'grad_norm': 4.033031463623047, 'learning_rate': 4.2393939393939395e-06, 'num_tokens': 3419486.0, 'completions/mean_length': 84.625, 'completions/min_length': 16.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.571430206298828, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.7202394008636475, 'rewards/meter/std': 0.4406243860721588, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.945844292640686, 'rewards/repeat_soft/std': 0.04052875190973282, 'rewards/judge_quality/mean': 0.7524999976158142, 'rewards/judge_quality/std': 0.3345252573490143, 'rewards/total_composite/mean': 0.6909451484680176, 'rewards/total_composite/std': 0.3505527675151825, 'reward': 0.6909451484680176, 'reward_std': 0.3505527675151825, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.18114034831523895, 'sampling/sampling_logp_difference/max': 1.3069181442260742, 'sampling/importance_sampling_ratio/min': 0.30163708329200745, 'sampling/importance_sampling_ratio/mean': 1.0173735618591309, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0979121401906013, 'clip_ratio/low_mean': 0.0633680559694767, 'clip_ratio/low_min': 0.0633680559694767, 'clip_ratio/high_mean': 0.14102468639612198, 'clip_ratio/high_max': 0.14102468639612198, 'clip_ratio/region_mean': 0.20439274236559868, 'reward_total_mean': 0.6909451484680176, 'reward_meter_mean': 0.7202394008636475, 'reward_meter_std': 0.4406243860721588, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.945844292640686, 'reward_repeat_soft_std': 0.04052875190973282, 'reward_judge_quality_mean': 0.7524999976158142, 'reward_judge_quality_std': 0.3345252573490143, 'reward_total_composite_mean': 0.6909451484680176, 'reward_total_composite_std': 0.3505527675151825, 'epoch': 0.19}

 58%|█████▊    | 1902/3300 [4:41:16<7:25:19, 19.11s/it]INFO 04-13 12:02:14 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1903/3300 [4:41:22<5:54:50, 15.24s/it]2026-04-13 12:02:20,852 | INFO | train_grpo_train | metrics_logged mode=train step=1903

                                                       
{'loss': -0.0389, 'grad_norm': 14.084986686706543, 'learning_rate': 4.236363636363637e-06, 'num_tokens': 3421074.0, 'completions/mean_length': 25.5, 'completions/min_length': 18.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.5, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.6613826751708984, 'rewards/meter/std': 0.43064865469932556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8781934976577759, 'rewards/repeat_soft/std': 0.09914916008710861, 'rewards/judge_quality/mean': 0.49000000953674316, 'rewards/judge_quality/std': 0.1742740124464035, 'rewards/total_composite/mean': 0.5169066190719604, 'rewards/total_composite/std': 0.1330171823501587, 'reward': 0.5169066190719604, 'reward_std': 0.13301719725131989, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11851590126752853, 'sampling/sampling_logp_difference/max': 1.0182857513427734, 'sampling/importance_sampling_ratio/min': 0.3612136244773865, 'sampling/importance_sampling_ratio/mean': 1.0146753787994385, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7701932080090046, 'clip_ratio/low_mean': 0.034722222946584225, 'clip_ratio/low_min': 0.034722222946584225, 'clip_ratio/high_mean': 0.08670286927372217, 'clip_ratio/high_max': 0.08670286927372217, 'clip_ratio/region_mean': 0.1214250922203064, 'reward_total_mean': 0.5169066190719604, 'reward_meter_mean': 0.6613826751708984, 'reward_meter_std': 0.43064865469932556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8781934976577759, 'reward_repeat_soft_std': 0.09914916008710861, 'reward_judge_quality_mean': 0.49000000953674316, 'reward_judge_quality_std': 0.1742740124464035, 'reward_total_composite_mean': 0.5169066190719604, 'reward_total_composite_std': 0.1330171823501587, 'epoch': 0.19}

 58%|█████▊    | 1903/3300 [4:41:23<5:54:50, 15.24s/it]INFO 04-13 12:02:21 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1904/3300 [4:41:29<4:52:22, 12.57s/it]2026-04-13 12:02:27,185 | INFO | train_grpo_train | metrics_logged mode=train step=1904

                                                       
{'loss': 0.0592, 'grad_norm': 10.84548282623291, 'learning_rate': 4.233333333333334e-06, 'num_tokens': 3422843.0, 'completions/mean_length': 51.125, 'completions/min_length': 43.0, 'completions/max_length': 63.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 63.0, 'rewards/meter/mean': 0.6600973606109619, 'rewards/meter/std': 0.31269407272338867, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9792370796203613, 'rewards/repeat_soft/std': 0.014737443998456001, 'rewards/judge_quality/mean': 0.4312500059604645, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5339018106460571, 'rewards/total_composite/std': 0.09232258051633835, 'reward': 0.5339018106460571, 'reward_std': 0.09232258796691895, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15003661811351776, 'sampling/sampling_logp_difference/max': 2.6863467693328857, 'sampling/importance_sampling_ratio/min': 0.06812937557697296, 'sampling/importance_sampling_ratio/mean': 1.0140804052352905, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9993136450648308, 'clip_ratio/low_mean': 0.054302156902849674, 'clip_ratio/low_min': 0.054302156902849674, 'clip_ratio/high_mean': 0.10764239821583033, 'clip_ratio/high_max': 0.10764239821583033, 'clip_ratio/region_mean': 0.16194455511868, 'reward_total_mean': 0.5339018106460571, 'reward_meter_mean': 0.6600973606109619, 'reward_meter_std': 0.31269407272338867, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9792370796203613, 'reward_repeat_soft_std': 0.014737443998456001, 'reward_judge_quality_mean': 0.4312500059604645, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5339018106460571, 'reward_total_composite_std': 0.09232258051633835, 'epoch': 0.19}

 58%|█████▊    | 1904/3300 [4:41:29<4:52:22, 12.57s/it]INFO 04-13 12:02:27 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1905/3300 [4:41:40<4:45:59, 12.30s/it]2026-04-13 12:02:38,855 | INFO | train_grpo_train | metrics_logged mode=train step=1905

                                                       
{'loss': -0.1045, 'grad_norm': 2.3378593921661377, 'learning_rate': 4.2303030303030304e-06, 'num_tokens': 3424337.0, 'completions/mean_length': 96.75, 'completions/min_length': 32.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.42857360839844, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8729835748672485, 'rewards/meter/std': 0.30031493306159973, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8887836933135986, 'rewards/repeat_soft/std': 0.07469533383846283, 'rewards/judge_quality/mean': 0.4399999976158142, 'rewards/judge_quality/std': 0.23439893126487732, 'rewards/total_composite/mean': 0.5661696791648865, 'rewards/total_composite/std': 0.2565822899341583, 'reward': 0.5661696791648865, 'reward_std': 0.2565822899341583, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10371766239404678, 'sampling/sampling_logp_difference/max': 1.1465868949890137, 'sampling/importance_sampling_ratio/min': 0.31771931052207947, 'sampling/importance_sampling_ratio/mean': 1.0026928186416626, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5314153209328651, 'clip_ratio/low_mean': 0.015625, 'clip_ratio/low_min': 0.015625, 'clip_ratio/high_mean': 0.07597637944854796, 'clip_ratio/high_max': 0.07597637944854796, 'clip_ratio/region_mean': 0.09160137944854796, 'reward_total_mean': 0.5661696791648865, 'reward_meter_mean': 0.8729835748672485, 'reward_meter_std': 0.30031493306159973, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8887836933135986, 'reward_repeat_soft_std': 0.07469533383846283, 'reward_judge_quality_mean': 0.4399999976158142, 'reward_judge_quality_std': 0.23439893126487732, 'reward_total_composite_mean': 0.5661696791648865, 'reward_total_composite_std': 0.2565822899341583, 'epoch': 0.19}

 58%|█████▊    | 1905/3300 [4:41:41<4:45:59, 12.30s/it]INFO 04-13 12:02:39 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1906/3300 [4:41:47<4:04:22, 10.52s/it]2026-04-13 12:02:45,217 | INFO | train_grpo_train | metrics_logged mode=train step=1906

                                                       
{'loss': 0.04, 'grad_norm': 12.90833854675293, 'learning_rate': 4.227272727272728e-06, 'num_tokens': 3425960.0, 'completions/mean_length': 53.875, 'completions/min_length': 45.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.875, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8947017192840576, 'rewards/meter/std': 0.2500592768192291, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9868673086166382, 'rewards/repeat_soft/std': 0.01419753022491932, 'rewards/judge_quality/mean': 0.4937500059604645, 'rewards/judge_quality/std': 0.1728696972131729, 'rewards/total_composite/mean': 0.6106759309768677, 'rewards/total_composite/std': 0.03996795788407326, 'reward': 0.6106759309768677, 'reward_std': 0.039967965334653854, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12730588018894196, 'sampling/sampling_logp_difference/max': 2.330036163330078, 'sampling/importance_sampling_ratio/min': 0.09729223698377609, 'sampling/importance_sampling_ratio/mean': 1.0107167959213257, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7010943591594696, 'clip_ratio/low_mean': 0.008620689623057842, 'clip_ratio/low_min': 0.008620689623057842, 'clip_ratio/high_mean': 0.09743543481454253, 'clip_ratio/high_max': 0.09743543481454253, 'clip_ratio/region_mean': 0.10605612443760037, 'reward_total_mean': 0.6106759309768677, 'reward_meter_mean': 0.8947017192840576, 'reward_meter_std': 0.2500592768192291, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9868673086166382, 'reward_repeat_soft_std': 0.01419753022491932, 'reward_judge_quality_mean': 0.4937500059604645, 'reward_judge_quality_std': 0.1728696972131729, 'reward_total_composite_mean': 0.6106759309768677, 'reward_total_composite_std': 0.03996795788407326, 'epoch': 0.19}

 58%|█████▊    | 1906/3300 [4:41:47<4:04:22, 10.52s/it]INFO 04-13 12:02:45 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1907/3300 [4:41:58<4:08:07, 10.69s/it]2026-04-13 12:02:56,298 | INFO | train_grpo_train | metrics_logged mode=train step=1907

                                                       
{'loss': -0.0754, 'grad_norm': 1.5282032489776611, 'learning_rate': 4.224242424242425e-06, 'num_tokens': 3427283.0, 'completions/mean_length': 148.375, 'completions/min_length': 20.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 27.166667938232422, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6922097206115723, 'rewards/meter/std': 0.35771870613098145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9556818008422852, 'rewards/repeat_soft/std': 0.01928471215069294, 'rewards/judge_quality/mean': 0.3137499690055847, 'rewards/judge_quality/std': 0.17492344975471497, 'rewards/total_composite/mean': 0.3942892253398895, 'rewards/total_composite/std': 0.25641167163848877, 'reward': 0.3942892253398895, 'reward_std': 0.25641167163848877, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.17665165662765503, 'sampling/sampling_logp_difference/max': 1.8716812133789062, 'sampling/importance_sampling_ratio/min': 0.1538647562265396, 'sampling/importance_sampling_ratio/mean': 1.0526965856552124, 'sampling/importance_sampling_ratio/max': 1.782524824142456, 'entropy': 1.2915838807821274, 'clip_ratio/low_mean': 0.011904762126505375, 'clip_ratio/low_min': 0.011904762126505375, 'clip_ratio/high_mean': 0.10561077040620148, 'clip_ratio/high_max': 0.10561077040620148, 'clip_ratio/region_mean': 0.11751553253270686, 'reward_total_mean': 0.3942892253398895, 'reward_meter_mean': 0.6922097206115723, 'reward_meter_std': 0.35771870613098145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9556818008422852, 'reward_repeat_soft_std': 0.01928471215069294, 'reward_judge_quality_mean': 0.3137499690055847, 'reward_judge_quality_std': 0.17492344975471497, 'reward_total_composite_mean': 0.3942892253398895, 'reward_total_composite_std': 0.25641167163848877, 'epoch': 0.19}

 58%|█████▊    | 1907/3300 [4:41:58<4:08:07, 10.69s/it]INFO 04-13 12:02:56 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1908/3300 [4:42:05<3:41:44,  9.56s/it]2026-04-13 12:03:03,219 | INFO | train_grpo_train | metrics_logged mode=train step=1908

                                                       
{'loss': 0.0469, 'grad_norm': 11.930583953857422, 'learning_rate': 4.221212121212121e-06, 'num_tokens': 3429356.0, 'completions/mean_length': 85.125, 'completions/min_length': 76.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.125, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.5915040969848633, 'rewards/meter/std': 0.2694706320762634, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.982235312461853, 'rewards/repeat_soft/std': 0.009068382903933525, 'rewards/judge_quality/mean': 0.5199999809265137, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.5378624200820923, 'rewards/total_composite/std': 0.08761601895093918, 'reward': 0.5378624200820923, 'reward_std': 0.08761601150035858, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1465836614370346, 'sampling/sampling_logp_difference/max': 1.7468070983886719, 'sampling/importance_sampling_ratio/min': 0.17432966828346252, 'sampling/importance_sampling_ratio/mean': 0.9981339573860168, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7508384957909584, 'clip_ratio/low_mean': 0.06788656301796436, 'clip_ratio/low_min': 0.06788656301796436, 'clip_ratio/high_mean': 0.06018355302512646, 'clip_ratio/high_max': 0.06018355302512646, 'clip_ratio/region_mean': 0.12807011604309082, 'reward_total_mean': 0.5378624200820923, 'reward_meter_mean': 0.5915040969848633, 'reward_meter_std': 0.2694706320762634, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.982235312461853, 'reward_repeat_soft_std': 0.009068382903933525, 'reward_judge_quality_mean': 0.5199999809265137, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.5378624200820923, 'reward_total_composite_std': 0.08761601895093918, 'epoch': 0.19}

 58%|█████▊    | 1908/3300 [4:42:05<3:41:44,  9.56s/it]INFO 04-13 12:03:03 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1909/3300 [4:42:11<3:18:08,  8.55s/it]2026-04-13 12:03:09,414 | INFO | train_grpo_train | metrics_logged mode=train step=1909

                                                       
{'loss': 0.1083, 'grad_norm': 11.530652046203613, 'learning_rate': 4.218181818181819e-06, 'num_tokens': 3431028.0, 'completions/mean_length': 46.0, 'completions/min_length': 41.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.0, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9786093831062317, 'rewards/meter/std': 0.01941728964447975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9682313203811646, 'rewards/repeat_soft/std': 0.038088828325271606, 'rewards/judge_quality/mean': 0.5850000381469727, 'rewards/judge_quality/std': 0.2946668863296509, 'rewards/total_composite/mean': 0.7170112133026123, 'rewards/total_composite/std': 0.18612056970596313, 'reward': 0.7170112133026123, 'reward_std': 0.18612055480480194, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1269945502281189, 'sampling/sampling_logp_difference/max': 1.3092458248138428, 'sampling/importance_sampling_ratio/min': 0.3732983469963074, 'sampling/importance_sampling_ratio/mean': 1.0166714191436768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9716865196824074, 'clip_ratio/low_mean': 0.0770115265622735, 'clip_ratio/low_min': 0.0770115265622735, 'clip_ratio/high_mean': 0.05192000512033701, 'clip_ratio/high_max': 0.05192000512033701, 'clip_ratio/region_mean': 0.1289315316826105, 'reward_total_mean': 0.7170112133026123, 'reward_meter_mean': 0.9786093831062317, 'reward_meter_std': 0.01941728964447975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9682313203811646, 'reward_repeat_soft_std': 0.038088828325271606, 'reward_judge_quality_mean': 0.5850000381469727, 'reward_judge_quality_std': 0.2946668863296509, 'reward_total_composite_mean': 0.7170112133026123, 'reward_total_composite_std': 0.18612056970596313, 'epoch': 0.19}

 58%|█████▊    | 1909/3300 [4:42:11<3:18:08,  8.55s/it]INFO 04-13 12:03:09 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1910/3300 [4:42:18<3:09:12,  8.17s/it]2026-04-13 12:03:16,701 | INFO | train_grpo_train | metrics_logged mode=train step=1910

                                                       
{'loss': 0.0465, 'grad_norm': 6.4757304191589355, 'learning_rate': 4.215151515151515e-06, 'num_tokens': 3433133.0, 'completions/mean_length': 103.125, 'completions/min_length': 76.0, 'completions/max_length': 133.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.125, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.9347220063209534, 'rewards/meter/std': 0.1163896769285202, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8798757791519165, 'rewards/repeat_soft/std': 0.07860218733549118, 'rewards/judge_quality/mean': 0.45749998092651367, 'rewards/judge_quality/std': 0.10606604069471359, 'rewards/total_composite/mean': 0.6112393736839294, 'rewards/total_composite/std': 0.08730139583349228, 'reward': 0.6112393736839294, 'reward_std': 0.08730139583349228, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11001424491405487, 'sampling/sampling_logp_difference/max': 1.6104984283447266, 'sampling/importance_sampling_ratio/min': 0.19978800415992737, 'sampling/importance_sampling_ratio/mean': 1.0204784870147705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8050570487976074, 'clip_ratio/low_mean': 0.06305000837892294, 'clip_ratio/low_min': 0.06305000837892294, 'clip_ratio/high_mean': 0.03493107855319977, 'clip_ratio/high_max': 0.03493107855319977, 'clip_ratio/region_mean': 0.09798108693212271, 'reward_total_mean': 0.6112393736839294, 'reward_meter_mean': 0.9347220063209534, 'reward_meter_std': 0.1163896769285202, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8798757791519165, 'reward_repeat_soft_std': 0.07860218733549118, 'reward_judge_quality_mean': 0.45749998092651367, 'reward_judge_quality_std': 0.10606604069471359, 'reward_total_composite_mean': 0.6112393736839294, 'reward_total_composite_std': 0.08730139583349228, 'epoch': 0.19}

 58%|█████▊    | 1910/3300 [4:42:18<3:09:12,  8.17s/it]INFO 04-13 12:03:16 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1911/3300 [4:42:25<2:59:09,  7.74s/it]2026-04-13 12:03:23,430 | INFO | train_grpo_train | metrics_logged mode=train step=1911

                                                       
{'loss': 0.0613, 'grad_norm': 6.657021522521973, 'learning_rate': 4.212121212121212e-06, 'num_tokens': 3435058.0, 'completions/mean_length': 64.625, 'completions/min_length': 55.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 64.625, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.954378068447113, 'rewards/meter/std': 0.08675800263881683, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8774073719978333, 'rewards/repeat_soft/std': 0.0204410869628191, 'rewards/judge_quality/mean': 0.5362499952316284, 'rewards/judge_quality/std': 0.1524970978498459, 'rewards/total_composite/mean': 0.6668540239334106, 'rewards/total_composite/std': 0.10783916711807251, 'reward': 0.6668540239334106, 'reward_std': 0.10783915966749191, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07967215776443481, 'sampling/sampling_logp_difference/max': 1.5013203620910645, 'sampling/importance_sampling_ratio/min': 0.2228357493877411, 'sampling/importance_sampling_ratio/mean': 1.0027366876602173, 'sampling/importance_sampling_ratio/max': 1.8892652988433838, 'entropy': 0.4966052621603012, 'clip_ratio/low_mean': 0.04203710425645113, 'clip_ratio/low_min': 0.04203710425645113, 'clip_ratio/high_mean': 0.03393878135830164, 'clip_ratio/high_max': 0.03393878135830164, 'clip_ratio/region_mean': 0.07597588561475277, 'reward_total_mean': 0.6668540239334106, 'reward_meter_mean': 0.954378068447113, 'reward_meter_std': 0.08675800263881683, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8774073719978333, 'reward_repeat_soft_std': 0.0204410869628191, 'reward_judge_quality_mean': 0.5362499952316284, 'reward_judge_quality_std': 0.1524970978498459, 'reward_total_composite_mean': 0.6668540239334106, 'reward_total_composite_std': 0.10783916711807251, 'epoch': 0.19}

 58%|█████▊    | 1911/3300 [4:42:25<2:59:09,  7.74s/it]INFO 04-13 12:03:23 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1912/3300 [4:42:33<3:01:39,  7.85s/it]2026-04-13 12:03:31,548 | INFO | train_grpo_train | metrics_logged mode=train step=1912

                                                       
{'loss': 0.0134, 'grad_norm': 7.404308319091797, 'learning_rate': 4.2090909090909095e-06, 'num_tokens': 3437690.0, 'completions/mean_length': 142.0, 'completions/min_length': 129.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.0, 'completions/min_terminated_length': 129.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.991479754447937, 'rewards/meter/std': 0.003022808115929365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7146003842353821, 'rewards/repeat_soft/std': 0.1012619212269783, 'rewards/judge_quality/mean': 0.2462500035762787, 'rewards/judge_quality/std': 0.08348438143730164, 'rewards/total_composite/mean': 0.4659019708633423, 'rewards/total_composite/std': 0.060872793197631836, 'reward': 0.4659019708633423, 'reward_std': 0.06087278574705124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05452094227075577, 'sampling/sampling_logp_difference/max': 1.2917299270629883, 'sampling/importance_sampling_ratio/min': 0.2747949957847595, 'sampling/importance_sampling_ratio/mean': 1.0096286535263062, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3699205368757248, 'clip_ratio/low_mean': 0.026885909494012594, 'clip_ratio/low_min': 0.026885909494012594, 'clip_ratio/high_mean': 0.023657962679862976, 'clip_ratio/high_max': 0.023657962679862976, 'clip_ratio/region_mean': 0.05054387217387557, 'reward_total_mean': 0.4659019708633423, 'reward_meter_mean': 0.991479754447937, 'reward_meter_std': 0.003022808115929365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7146003842353821, 'reward_repeat_soft_std': 0.1012619212269783, 'reward_judge_quality_mean': 0.2462500035762787, 'reward_judge_quality_std': 0.08348438143730164, 'reward_total_composite_mean': 0.4659019708633423, 'reward_total_composite_std': 0.060872793197631836, 'epoch': 0.19}

 58%|█████▊    | 1912/3300 [4:42:33<3:01:39,  7.85s/it]INFO 04-13 12:03:31 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1913/3300 [4:42:40<2:54:07,  7.53s/it]2026-04-13 12:03:38,338 | INFO | train_grpo_train | metrics_logged mode=train step=1913

                                                       
{'loss': 0.0689, 'grad_norm': 6.058612823486328, 'learning_rate': 4.206060606060606e-06, 'num_tokens': 3439994.0, 'completions/mean_length': 108.0, 'completions/min_length': 89.0, 'completions/max_length': 124.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.0, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.9075185656547546, 'rewards/meter/std': 0.2047087401151657, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8628815412521362, 'rewards/repeat_soft/std': 0.03367576375603676, 'rewards/judge_quality/mean': 0.41999998688697815, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.577184796333313, 'rewards/total_composite/std': 0.057645779103040695, 'reward': 0.577184796333313, 'reward_std': 0.057645782828330994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08772403001785278, 'sampling/sampling_logp_difference/max': 1.0531864166259766, 'sampling/importance_sampling_ratio/min': 0.34882447123527527, 'sampling/importance_sampling_ratio/mean': 1.0191258192062378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6932972818613052, 'clip_ratio/low_mean': 0.010080644860863686, 'clip_ratio/low_min': 0.010080644860863686, 'clip_ratio/high_mean': 0.06751623563468456, 'clip_ratio/high_max': 0.06751623563468456, 'clip_ratio/region_mean': 0.07759688049554825, 'reward_total_mean': 0.577184796333313, 'reward_meter_mean': 0.9075185656547546, 'reward_meter_std': 0.2047087401151657, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8628815412521362, 'reward_repeat_soft_std': 0.03367576375603676, 'reward_judge_quality_mean': 0.41999998688697815, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.577184796333313, 'reward_total_composite_std': 0.057645779103040695, 'epoch': 0.19}

 58%|█████▊    | 1913/3300 [4:42:40<2:54:07,  7.53s/it]INFO 04-13 12:03:38 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1914/3300 [4:42:47<2:48:41,  7.30s/it]2026-04-13 12:03:45,099 | INFO | train_grpo_train | metrics_logged mode=train step=1914

                                                       
{'loss': 0.0568, 'grad_norm': 23.330875396728516, 'learning_rate': 4.203030303030303e-06, 'num_tokens': 3441599.0, 'completions/mean_length': 48.625, 'completions/min_length': 45.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.625, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.618243396282196, 'rewards/meter/std': 0.3767756223678589, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9622788429260254, 'rewards/repeat_soft/std': 0.0360732264816761, 'rewards/judge_quality/mean': 0.8200000524520874, 'rewards/judge_quality/std': 0.19272483885288239, 'rewards/total_composite/mean': 0.6511150002479553, 'rewards/total_composite/std': 0.19084851443767548, 'reward': 0.6511150002479553, 'reward_std': 0.19084849953651428, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08347024023532867, 'sampling/sampling_logp_difference/max': 2.4505133628845215, 'sampling/importance_sampling_ratio/min': 0.1143178939819336, 'sampling/importance_sampling_ratio/mean': 1.0063972473144531, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2730312719941139, 'clip_ratio/low_mean': 0.030320503283292055, 'clip_ratio/low_min': 0.030320503283292055, 'clip_ratio/high_mean': 0.026296769035980105, 'clip_ratio/high_max': 0.026296769035980105, 'clip_ratio/region_mean': 0.05661727231927216, 'reward_total_mean': 0.6511150002479553, 'reward_meter_mean': 0.618243396282196, 'reward_meter_std': 0.3767756223678589, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9622788429260254, 'reward_repeat_soft_std': 0.0360732264816761, 'reward_judge_quality_mean': 0.8200000524520874, 'reward_judge_quality_std': 0.19272483885288239, 'reward_total_composite_mean': 0.6511150002479553, 'reward_total_composite_std': 0.19084851443767548, 'epoch': 0.19}

 58%|█████▊    | 1914/3300 [4:42:47<2:48:41,  7.30s/it]INFO 04-13 12:03:45 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1915/3300 [4:42:55<2:58:29,  7.73s/it]2026-04-13 12:03:53,830 | INFO | train_grpo_train | metrics_logged mode=train step=1915

                                                       
{'loss': -0.0828, 'grad_norm': 11.6658353805542, 'learning_rate': 4.2000000000000004e-06, 'num_tokens': 3444450.0, 'completions/mean_length': 159.375, 'completions/min_length': 128.0, 'completions/max_length': 198.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.375, 'completions/min_terminated_length': 128.0, 'completions/max_terminated_length': 198.0, 'rewards/meter/mean': 0.9716150760650635, 'rewards/meter/std': 0.01307368092238903, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7041897773742676, 'rewards/repeat_soft/std': 0.06158751994371414, 'rewards/judge_quality/mean': 0.5950000286102295, 'rewards/judge_quality/std': 0.19820626080036163, 'rewards/total_composite/mean': 0.6813640594482422, 'rewards/total_composite/std': 0.12323393672704697, 'reward': 0.6813640594482422, 'reward_std': 0.12323394417762756, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07488817721605301, 'sampling/sampling_logp_difference/max': 1.7577056884765625, 'sampling/importance_sampling_ratio/min': 0.1724400371313095, 'sampling/importance_sampling_ratio/mean': 1.002464771270752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3908664211630821, 'clip_ratio/low_mean': 0.03722066944465041, 'clip_ratio/low_min': 0.03722066944465041, 'clip_ratio/high_mean': 0.03210883028805256, 'clip_ratio/high_max': 0.03210883028805256, 'clip_ratio/region_mean': 0.06932949973270297, 'reward_total_mean': 0.6813640594482422, 'reward_meter_mean': 0.9716150760650635, 'reward_meter_std': 0.01307368092238903, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7041897773742676, 'reward_repeat_soft_std': 0.06158751994371414, 'reward_judge_quality_mean': 0.5950000286102295, 'reward_judge_quality_std': 0.19820626080036163, 'reward_total_composite_mean': 0.6813640594482422, 'reward_total_composite_std': 0.12323393672704697, 'epoch': 0.19}

 58%|█████▊    | 1915/3300 [4:42:56<2:58:29,  7.73s/it]INFO 04-13 12:03:54 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1916/3300 [4:43:03<2:55:52,  7.62s/it]2026-04-13 12:04:01,227 | INFO | train_grpo_train | metrics_logged mode=train step=1916

                                                       
{'loss': -0.0325, 'grad_norm': 6.624622821807861, 'learning_rate': 4.196969696969697e-06, 'num_tokens': 3446921.0, 'completions/mean_length': 112.875, 'completions/min_length': 91.0, 'completions/max_length': 131.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 112.875, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 131.0, 'rewards/meter/mean': 0.8507556915283203, 'rewards/meter/std': 0.18038310110569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8237466812133789, 'rewards/repeat_soft/std': 0.03414364531636238, 'rewards/judge_quality/mean': 0.5774999856948853, 'rewards/judge_quality/std': 0.22461079061031342, 'rewards/total_composite/mean': 0.6472747325897217, 'rewards/total_composite/std': 0.15863634645938873, 'reward': 0.6472747325897217, 'reward_std': 0.15863634645938873, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.095945343375206, 'sampling/sampling_logp_difference/max': 1.421422004699707, 'sampling/importance_sampling_ratio/min': 0.2413705438375473, 'sampling/importance_sampling_ratio/mean': 1.008117914199829, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4990440383553505, 'clip_ratio/low_mean': 0.07097942568361759, 'clip_ratio/low_min': 0.07097942568361759, 'clip_ratio/high_mean': 0.02008723607286811, 'clip_ratio/high_max': 0.02008723607286811, 'clip_ratio/region_mean': 0.0910666617564857, 'reward_total_mean': 0.6472747325897217, 'reward_meter_mean': 0.8507556915283203, 'reward_meter_std': 0.18038310110569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8237466812133789, 'reward_repeat_soft_std': 0.03414364531636238, 'reward_judge_quality_mean': 0.5774999856948853, 'reward_judge_quality_std': 0.22461079061031342, 'reward_total_composite_mean': 0.6472747325897217, 'reward_total_composite_std': 0.15863634645938873, 'epoch': 0.19}

 58%|█████▊    | 1916/3300 [4:43:03<2:55:52,  7.62s/it]INFO 04-13 12:04:01 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1917/3300 [4:43:09<2:45:19,  7.17s/it]2026-04-13 12:04:07,326 | INFO | train_grpo_train | metrics_logged mode=train step=1917

                                                       
{'loss': -0.0559, 'grad_norm': 17.377117156982422, 'learning_rate': 4.193939393939394e-06, 'num_tokens': 3448367.0, 'completions/mean_length': 33.75, 'completions/min_length': 26.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 33.75, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.457536518573761, 'rewards/meter/std': 0.35414940118789673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9779925346374512, 'rewards/repeat_soft/std': 0.0339190848171711, 'rewards/judge_quality/mean': 0.4650000035762787, 'rewards/judge_quality/std': 0.10392305999994278, 'rewards/total_composite/mean': 0.47729986906051636, 'rewards/total_composite/std': 0.09718886762857437, 'reward': 0.47729986906051636, 'reward_std': 0.09718887507915497, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14621400833129883, 'sampling/sampling_logp_difference/max': 1.2445707321166992, 'sampling/importance_sampling_ratio/min': 0.28806453943252563, 'sampling/importance_sampling_ratio/mean': 1.0058984756469727, 'sampling/importance_sampling_ratio/max': 1.7201218605041504, 'entropy': 1.0507483892142773, 'clip_ratio/low_mean': 0.03479864727705717, 'clip_ratio/low_min': 0.03479864727705717, 'clip_ratio/high_mean': 0.05114082945510745, 'clip_ratio/high_max': 0.05114082945510745, 'clip_ratio/region_mean': 0.08593947673216462, 'reward_total_mean': 0.47729986906051636, 'reward_meter_mean': 0.457536518573761, 'reward_meter_std': 0.35414940118789673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9779925346374512, 'reward_repeat_soft_std': 0.0339190848171711, 'reward_judge_quality_mean': 0.4650000035762787, 'reward_judge_quality_std': 0.10392305999994278, 'reward_total_composite_mean': 0.47729986906051636, 'reward_total_composite_std': 0.09718886762857437, 'epoch': 0.19}

 58%|█████▊    | 1917/3300 [4:43:09<2:45:19,  7.17s/it]INFO 04-13 12:04:07 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1918/3300 [4:43:16<2:42:12,  7.04s/it]2026-04-13 12:04:14,076 | INFO | train_grpo_train | metrics_logged mode=train step=1918

                                                       
{'loss': 0.0014, 'grad_norm': 18.626001358032227, 'learning_rate': 4.190909090909091e-06, 'num_tokens': 3449780.0, 'completions/mean_length': 26.625, 'completions/min_length': 24.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8700626492500305, 'rewards/meter/std': 0.3448854684829712, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9449080228805542, 'rewards/repeat_soft/std': 0.022345291450619698, 'rewards/judge_quality/mean': 0.4312499761581421, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.5840761661529541, 'rewards/total_composite/std': 0.09695003926753998, 'reward': 0.5840761661529541, 'reward_std': 0.09695003926753998, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09097051620483398, 'sampling/sampling_logp_difference/max': 0.8367805480957031, 'sampling/importance_sampling_ratio/min': 0.43310266733169556, 'sampling/importance_sampling_ratio/mean': 1.0069423913955688, 'sampling/importance_sampling_ratio/max': 1.6343611478805542, 'entropy': 0.47952786087989807, 'clip_ratio/low_mean': 0.019999999552965164, 'clip_ratio/low_min': 0.019999999552965164, 'clip_ratio/high_mean': 0.07230510842055082, 'clip_ratio/high_max': 0.07230510842055082, 'clip_ratio/region_mean': 0.09230510797351599, 'reward_total_mean': 0.5840761661529541, 'reward_meter_mean': 0.8700626492500305, 'reward_meter_std': 0.3448854684829712, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9449080228805542, 'reward_repeat_soft_std': 0.022345291450619698, 'reward_judge_quality_mean': 0.4312499761581421, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.5840761661529541, 'reward_total_composite_std': 0.09695003926753998, 'epoch': 0.19}

 58%|█████▊    | 1918/3300 [4:43:16<2:42:12,  7.04s/it]INFO 04-13 12:04:14 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1919/3300 [4:43:22<2:37:18,  6.83s/it]2026-04-13 12:04:20,416 | INFO | train_grpo_train | metrics_logged mode=train step=1919

                                                       
{'loss': -0.0942, 'grad_norm': 7.9500651359558105, 'learning_rate': 4.187878787878788e-06, 'num_tokens': 3451401.0, 'completions/mean_length': 43.625, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9833869934082031, 'rewards/meter/std': 0.004617537371814251, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9158543348312378, 'rewards/repeat_soft/std': 0.05400705710053444, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4356558918952942, 'rewards/total_composite/std': 0.01236470602452755, 'reward': 0.4356558918952942, 'reward_std': 0.012364705093204975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07290530204772949, 'sampling/sampling_logp_difference/max': 1.3190256357192993, 'sampling/importance_sampling_ratio/min': 0.2673957049846649, 'sampling/importance_sampling_ratio/mean': 1.0044238567352295, 'sampling/importance_sampling_ratio/max': 1.6072239875793457, 'entropy': 0.4352567121386528, 'clip_ratio/low_mean': 0.03689136798493564, 'clip_ratio/low_min': 0.03689136798493564, 'clip_ratio/high_mean': 0.04079236835241318, 'clip_ratio/high_max': 0.04079236835241318, 'clip_ratio/region_mean': 0.07768373633734882, 'reward_total_mean': 0.4356558918952942, 'reward_meter_mean': 0.9833869934082031, 'reward_meter_std': 0.004617537371814251, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9158543348312378, 'reward_repeat_soft_std': 0.05400705710053444, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4356558918952942, 'reward_total_composite_std': 0.01236470602452755, 'epoch': 0.19}

 58%|█████▊    | 1919/3300 [4:43:22<2:37:18,  6.83s/it]INFO 04-13 12:04:20 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1920/3300 [4:43:30<2:46:13,  7.23s/it]2026-04-13 12:04:28,556 | INFO | train_grpo_train | metrics_logged mode=train step=1920

                                                       
{'loss': 0.0382, 'grad_norm': 5.887547016143799, 'learning_rate': 4.184848484848485e-06, 'num_tokens': 3454200.0, 'completions/mean_length': 143.875, 'completions/min_length': 133.0, 'completions/max_length': 160.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.875, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 160.0, 'rewards/meter/mean': 0.987758219242096, 'rewards/meter/std': 0.005048134829849005, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8466187715530396, 'rewards/repeat_soft/std': 0.06207213178277016, 'rewards/judge_quality/mean': 0.11999999731779099, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4040379524230957, 'rewards/total_composite/std': 0.009467762894928455, 'reward': 0.4040379524230957, 'reward_std': 0.00946776196360588, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09985199570655823, 'sampling/sampling_logp_difference/max': 1.9229249954223633, 'sampling/importance_sampling_ratio/min': 0.14617876708507538, 'sampling/importance_sampling_ratio/mean': 0.9935129284858704, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5678234249353409, 'clip_ratio/low_mean': 0.03725255560129881, 'clip_ratio/low_min': 0.03725255560129881, 'clip_ratio/high_mean': 0.03946106508374214, 'clip_ratio/high_max': 0.03946106508374214, 'clip_ratio/region_mean': 0.07671362068504095, 'reward_total_mean': 0.4040379524230957, 'reward_meter_mean': 0.987758219242096, 'reward_meter_std': 0.005048134829849005, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8466187715530396, 'reward_repeat_soft_std': 0.06207213178277016, 'reward_judge_quality_mean': 0.11999999731779099, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4040379524230957, 'reward_total_composite_std': 0.009467762894928455, 'epoch': 0.19}

 58%|█████▊    | 1920/3300 [4:43:30<2:46:13,  7.23s/it]INFO 04-13 12:04:28 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1921/3300 [4:43:41<3:14:08,  8.45s/it]2026-04-13 12:04:39,842 | INFO | train_grpo_train | metrics_logged mode=train step=1921

                                                       
{'loss': -0.0847, 'grad_norm': 1.4573488235473633, 'learning_rate': 4.181818181818182e-06, 'num_tokens': 3455565.0, 'completions/mean_length': 84.625, 'completions/min_length': 15.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.571430206298828, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.6292486190795898, 'rewards/meter/std': 0.39336225390434265, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9447200298309326, 'rewards/repeat_soft/std': 0.02195359766483307, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.35830748081207275, 'rewards/total_composite/std': 0.14822307229042053, 'reward': 0.35830748081207275, 'reward_std': 0.14822307229042053, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15679016709327698, 'sampling/sampling_logp_difference/max': 1.3619756698608398, 'sampling/importance_sampling_ratio/min': 0.25615420937538147, 'sampling/importance_sampling_ratio/mean': 1.0062261819839478, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0058701038360596, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.14873274974524975, 'clip_ratio/high_max': 0.14873274974524975, 'clip_ratio/region_mean': 0.14873274974524975, 'reward_total_mean': 0.35830748081207275, 'reward_meter_mean': 0.6292486190795898, 'reward_meter_std': 0.39336225390434265, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9447200298309326, 'reward_repeat_soft_std': 0.02195359766483307, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.35830748081207275, 'reward_total_composite_std': 0.14822307229042053, 'epoch': 0.19}

 58%|█████▊    | 1921/3300 [4:43:42<3:14:08,  8.45s/it]INFO 04-13 12:04:40 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1922/3300 [4:43:53<3:33:12,  9.28s/it]2026-04-13 12:04:51,097 | INFO | train_grpo_train | metrics_logged mode=train step=1922

                                                       
{'loss': -0.1185, 'grad_norm': 2.2207748889923096, 'learning_rate': 4.1787878787878795e-06, 'num_tokens': 3457154.0, 'completions/mean_length': 99.625, 'completions/min_length': 37.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.71428680419922, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.7216843366622925, 'rewards/meter/std': 0.35830414295196533, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9028103351593018, 'rewards/repeat_soft/std': 0.09586292505264282, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.3565751016139984, 'rewards/total_composite/std': 0.15102575719356537, 'reward': 0.3565751016139984, 'reward_std': 0.15102575719356537, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13488441705703735, 'sampling/sampling_logp_difference/max': 2.1499409675598145, 'sampling/importance_sampling_ratio/min': 0.11649102717638016, 'sampling/importance_sampling_ratio/mean': 1.0255366563796997, 'sampling/importance_sampling_ratio/max': 1.8868846893310547, 'entropy': 0.8553041182458401, 'clip_ratio/low_mean': 0.013513513840734959, 'clip_ratio/low_min': 0.013513513840734959, 'clip_ratio/high_mean': 0.07870935089886189, 'clip_ratio/high_max': 0.07870935089886189, 'clip_ratio/region_mean': 0.09222286473959684, 'reward_total_mean': 0.3565751016139984, 'reward_meter_mean': 0.7216843366622925, 'reward_meter_std': 0.35830414295196533, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9028103351593018, 'reward_repeat_soft_std': 0.09586292505264282, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.3565751016139984, 'reward_total_composite_std': 0.15102575719356537, 'epoch': 0.19}

 58%|█████▊    | 1922/3300 [4:43:53<3:33:12,  9.28s/it]INFO 04-13 12:04:51 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1923/3300 [4:43:59<3:14:07,  8.46s/it]2026-04-13 12:04:57,621 | INFO | train_grpo_train | metrics_logged mode=train step=1923

                                                       
{'loss': -0.0912, 'grad_norm': 11.489913940429688, 'learning_rate': 4.175757575757576e-06, 'num_tokens': 3458823.0, 'completions/mean_length': 52.625, 'completions/min_length': 38.0, 'completions/max_length': 75.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 75.0, 'rewards/meter/mean': 0.5990158915519714, 'rewards/meter/std': 0.3955284059047699, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9780142307281494, 'rewards/repeat_soft/std': 0.01682078093290329, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.406122088432312, 'rewards/total_composite/std': 0.03803851455450058, 'reward': 0.406122088432312, 'reward_std': 0.03803851455450058, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13839095830917358, 'sampling/sampling_logp_difference/max': 1.4287137985229492, 'sampling/importance_sampling_ratio/min': 0.23961691558361053, 'sampling/importance_sampling_ratio/mean': 1.0154194831848145, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0459632501006126, 'clip_ratio/low_mean': 0.07959857024252415, 'clip_ratio/low_min': 0.07959857024252415, 'clip_ratio/high_mean': 0.0758630963973701, 'clip_ratio/high_max': 0.0758630963973701, 'clip_ratio/region_mean': 0.15546166663989425, 'reward_total_mean': 0.406122088432312, 'reward_meter_mean': 0.5990158915519714, 'reward_meter_std': 0.3955284059047699, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9780142307281494, 'reward_repeat_soft_std': 0.01682078093290329, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.406122088432312, 'reward_total_composite_std': 0.03803851455450058, 'epoch': 0.19}

 58%|█████▊    | 1923/3300 [4:43:59<3:14:07,  8.46s/it]INFO 04-13 12:04:57 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1924/3300 [4:44:11<3:35:38,  9.40s/it]2026-04-13 12:05:09,223 | INFO | train_grpo_train | metrics_logged mode=train step=1924

                                                       
{'loss': -0.1086, 'grad_norm': 1.5600149631500244, 'learning_rate': 4.172727272727273e-06, 'num_tokens': 3460324.0, 'completions/mean_length': 93.625, 'completions/min_length': 29.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.85714340209961, 'completions/min_terminated_length': 29.0, 'completions/max_terminated_length': 37.0, 'rewards/meter/mean': 0.6899369955062866, 'rewards/meter/std': 0.3944839835166931, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.90090012550354, 'rewards/repeat_soft/std': 0.13305890560150146, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.35865187644958496, 'rewards/total_composite/std': 0.1468174010515213, 'reward': 0.35865187644958496, 'reward_std': 0.1468174010515213, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05273240804672241, 'sampling/sampling_logp_difference/max': 0.9871878027915955, 'sampling/importance_sampling_ratio/min': 0.37262311577796936, 'sampling/importance_sampling_ratio/mean': 1.01171875, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.23229003325104713, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.03230845835059881, 'clip_ratio/high_max': 0.03230845835059881, 'clip_ratio/region_mean': 0.03230845835059881, 'reward_total_mean': 0.35865187644958496, 'reward_meter_mean': 0.6899369955062866, 'reward_meter_std': 0.3944839835166931, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.90090012550354, 'reward_repeat_soft_std': 0.13305890560150146, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.35865187644958496, 'reward_total_composite_std': 0.1468174010515213, 'epoch': 0.19}

 58%|█████▊    | 1924/3300 [4:44:11<3:35:38,  9.40s/it]INFO 04-13 12:05:09 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1925/3300 [4:44:17<3:13:31,  8.44s/it]2026-04-13 12:05:15,445 | INFO | train_grpo_train | metrics_logged mode=train step=1925

                                                       
{'loss': 0.0353, 'grad_norm': 15.171744346618652, 'learning_rate': 4.1696969696969705e-06, 'num_tokens': 3461805.0, 'completions/mean_length': 39.125, 'completions/min_length': 34.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.125, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.9467993974685669, 'rewards/meter/std': 0.08223433792591095, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9046837091445923, 'rewards/repeat_soft/std': 0.0497312992811203, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.42801550030708313, 'rewards/total_composite/std': 0.010744020342826843, 'reward': 0.42801550030708313, 'reward_std': 0.010744019411504269, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12122054398059845, 'sampling/sampling_logp_difference/max': 1.4478511810302734, 'sampling/importance_sampling_ratio/min': 0.354824960231781, 'sampling/importance_sampling_ratio/mean': 1.0006885528564453, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.939532570540905, 'clip_ratio/low_mean': 0.030678459908813238, 'clip_ratio/low_min': 0.030678459908813238, 'clip_ratio/high_mean': 0.052022610791027546, 'clip_ratio/high_max': 0.052022610791027546, 'clip_ratio/region_mean': 0.08270107069984078, 'reward_total_mean': 0.42801550030708313, 'reward_meter_mean': 0.9467993974685669, 'reward_meter_std': 0.08223433792591095, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9046837091445923, 'reward_repeat_soft_std': 0.0497312992811203, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.42801550030708313, 'reward_total_composite_std': 0.010744020342826843, 'epoch': 0.19}

 58%|█████▊    | 1925/3300 [4:44:17<3:13:31,  8.44s/it]INFO 04-13 12:05:15 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1926/3300 [4:44:28<3:33:18,  9.31s/it]2026-04-13 12:05:26,781 | INFO | train_grpo_train | metrics_logged mode=train step=1926

                                                       
{'loss': -0.1535, 'grad_norm': 1.6101049184799194, 'learning_rate': 4.166666666666667e-06, 'num_tokens': 3463584.0, 'completions/mean_length': 115.375, 'completions/min_length': 50.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.71428680419922, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8866132497787476, 'rewards/meter/std': 0.24672026932239532, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9512063264846802, 'rewards/repeat_soft/std': 0.038331110030412674, 'rewards/judge_quality/mean': 0.14125001430511475, 'rewards/judge_quality/std': 0.038335926830768585, 'rewards/total_composite/mean': 0.3850158452987671, 'rewards/total_composite/std': 0.15585671365261078, 'reward': 0.3850158452987671, 'reward_std': 0.15585669875144958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11420281231403351, 'sampling/sampling_logp_difference/max': 2.4861817359924316, 'sampling/importance_sampling_ratio/min': 0.08322714269161224, 'sampling/importance_sampling_ratio/mean': 1.0202020406723022, 'sampling/importance_sampling_ratio/max': 1.828543782234192, 'entropy': 0.798855647444725, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09341429825872183, 'clip_ratio/high_max': 0.09341429825872183, 'clip_ratio/region_mean': 0.09341429825872183, 'reward_total_mean': 0.3850158452987671, 'reward_meter_mean': 0.8866132497787476, 'reward_meter_std': 0.24672026932239532, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9512063264846802, 'reward_repeat_soft_std': 0.038331110030412674, 'reward_judge_quality_mean': 0.14125001430511475, 'reward_judge_quality_std': 0.038335926830768585, 'reward_total_composite_mean': 0.3850158452987671, 'reward_total_composite_std': 0.15585671365261078, 'epoch': 0.19}

 58%|█████▊    | 1926/3300 [4:44:29<3:33:18,  9.31s/it]INFO 04-13 12:05:27 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1927/3300 [4:44:34<3:09:32,  8.28s/it]2026-04-13 12:05:32,664 | INFO | train_grpo_train | metrics_logged mode=train step=1927

                                                       
{'loss': 0.0036, 'grad_norm': 11.964699745178223, 'learning_rate': 4.163636363636364e-06, 'num_tokens': 3465041.0, 'completions/mean_length': 23.125, 'completions/min_length': 20.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.994118332862854, 'rewards/meter/std': 0.0013978873612359166, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9560492038726807, 'rewards/repeat_soft/std': 0.010533427819609642, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.44033390283584595, 'rewards/total_composite/std': 0.001578116905875504, 'reward': 0.44033390283584595, 'reward_std': 0.0015781193505972624, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10783885419368744, 'sampling/sampling_logp_difference/max': 0.7501993179321289, 'sampling/importance_sampling_ratio/min': 0.4722723960876465, 'sampling/importance_sampling_ratio/mean': 1.0312190055847168, 'sampling/importance_sampling_ratio/max': 1.8718894720077515, 'entropy': 0.8328215926885605, 'clip_ratio/low_mean': 0.04190476145595312, 'clip_ratio/low_min': 0.04190476145595312, 'clip_ratio/high_mean': 0.08828983781859279, 'clip_ratio/high_max': 0.08828983781859279, 'clip_ratio/region_mean': 0.1301945992745459, 'reward_total_mean': 0.44033390283584595, 'reward_meter_mean': 0.994118332862854, 'reward_meter_std': 0.0013978873612359166, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9560492038726807, 'reward_repeat_soft_std': 0.010533427819609642, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.44033390283584595, 'reward_total_composite_std': 0.001578116905875504, 'epoch': 0.19}

 58%|█████▊    | 1927/3300 [4:44:34<3:09:32,  8.28s/it]INFO 04-13 12:05:32 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1928/3300 [4:44:40<2:53:13,  7.58s/it]2026-04-13 12:05:38,580 | INFO | train_grpo_train | metrics_logged mode=train step=1928

                                                       
{'loss': 0.0531, 'grad_norm': 19.441036224365234, 'learning_rate': 4.160606060606061e-06, 'num_tokens': 3466389.0, 'completions/mean_length': 23.5, 'completions/min_length': 20.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.5, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9700110554695129, 'rewards/meter/std': 0.028271492570638657, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9606921672821045, 'rewards/repeat_soft/std': 0.005113361403346062, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43867990374565125, 'rewards/total_composite/std': 0.003153893630951643, 'reward': 0.43867990374565125, 'reward_std': 0.003153891069814563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13899697363376617, 'sampling/sampling_logp_difference/max': 1.5593090057373047, 'sampling/importance_sampling_ratio/min': 0.21028132736682892, 'sampling/importance_sampling_ratio/mean': 1.0141392946243286, 'sampling/importance_sampling_ratio/max': 1.9066965579986572, 'entropy': 0.9456418007612228, 'clip_ratio/low_mean': 0.0297413794323802, 'clip_ratio/low_min': 0.0297413794323802, 'clip_ratio/high_mean': 0.08298542629927397, 'clip_ratio/high_max': 0.08298542629927397, 'clip_ratio/region_mean': 0.11272680573165417, 'reward_total_mean': 0.43867990374565125, 'reward_meter_mean': 0.9700110554695129, 'reward_meter_std': 0.028271492570638657, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9606921672821045, 'reward_repeat_soft_std': 0.005113361403346062, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43867990374565125, 'reward_total_composite_std': 0.003153893630951643, 'epoch': 0.19}

 58%|█████▊    | 1928/3300 [4:44:40<2:53:13,  7.58s/it]INFO 04-13 12:05:38 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1929/3300 [4:44:46<2:43:11,  7.14s/it]2026-04-13 12:05:44,705 | INFO | train_grpo_train | metrics_logged mode=train step=1929

                                                       
{'loss': 0.0178, 'grad_norm': 14.826068878173828, 'learning_rate': 4.157575757575758e-06, 'num_tokens': 3468025.0, 'completions/mean_length': 50.5, 'completions/min_length': 34.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.5, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9803265333175659, 'rewards/meter/std': 0.021168824285268784, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.88011634349823, 'rewards/repeat_soft/std': 0.043118324130773544, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.4252009093761444, 'rewards/total_composite/std': 0.012539559043943882, 'reward': 0.4252009093761444, 'reward_std': 0.012539563700556755, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11705910414457321, 'sampling/sampling_logp_difference/max': 1.5417990684509277, 'sampling/importance_sampling_ratio/min': 0.21399575471878052, 'sampling/importance_sampling_ratio/mean': 1.0303473472595215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7311799600720406, 'clip_ratio/low_mean': 0.026612902991473675, 'clip_ratio/low_min': 0.026612902991473675, 'clip_ratio/high_mean': 0.07826747675426304, 'clip_ratio/high_max': 0.07826747675426304, 'clip_ratio/region_mean': 0.10488037974573672, 'reward_total_mean': 0.4252009093761444, 'reward_meter_mean': 0.9803265333175659, 'reward_meter_std': 0.021168824285268784, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.88011634349823, 'reward_repeat_soft_std': 0.043118324130773544, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.4252009093761444, 'reward_total_composite_std': 0.012539559043943882, 'epoch': 0.19}

 58%|█████▊    | 1929/3300 [4:44:46<2:43:11,  7.14s/it]INFO 04-13 12:05:44 [block_pool.py:378] Successfully reset prefix cache

 58%|█████▊    | 1930/3300 [4:45:00<3:29:33,  9.18s/it]2026-04-13 12:05:58,641 | INFO | train_grpo_train | metrics_logged mode=train step=1930

                                                       
{'loss': 0.0846, 'grad_norm': 11.098047256469727, 'learning_rate': 4.154545454545455e-06, 'num_tokens': 3469790.0, 'completions/mean_length': 61.625, 'completions/min_length': 44.0, 'completions/max_length': 70.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.7586449980735779, 'rewards/meter/std': 0.34413576126098633, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9118528366088867, 'rewards/repeat_soft/std': 0.034493397921323776, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4107457995414734, 'rewards/total_composite/std': 0.03100598230957985, 'reward': 0.4107457995414734, 'reward_std': 0.031005987897515297, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11681599169969559, 'sampling/sampling_logp_difference/max': 1.6404461860656738, 'sampling/importance_sampling_ratio/min': 0.19389350712299347, 'sampling/importance_sampling_ratio/mean': 1.006695032119751, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5450751818716526, 'clip_ratio/low_mean': 0.024801588617265224, 'clip_ratio/low_min': 0.024801588617265224, 'clip_ratio/high_mean': 0.09530821582302451, 'clip_ratio/high_max': 0.09530821582302451, 'clip_ratio/region_mean': 0.12010980444028974, 'reward_total_mean': 0.4107457995414734, 'reward_meter_mean': 0.7586449980735779, 'reward_meter_std': 0.34413576126098633, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9118528366088867, 'reward_repeat_soft_std': 0.034493397921323776, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4107457995414734, 'reward_total_composite_std': 0.03100598230957985, 'epoch': 0.19}

 58%|█████▊    | 1930/3300 [4:45:00<3:29:33,  9.18s/it]INFO 04-13 12:05:58 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1931/3300 [4:45:07<3:09:26,  8.30s/it]2026-04-13 12:06:04,904 | INFO | train_grpo_train | metrics_logged mode=train step=1931

                                                       
{'loss': 0.0894, 'grad_norm': 7.9793477058410645, 'learning_rate': 4.151515151515152e-06, 'num_tokens': 3471453.0, 'completions/mean_length': 58.875, 'completions/min_length': 41.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.875, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9689643383026123, 'rewards/meter/std': 0.022389547899365425, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8962827920913696, 'rewards/repeat_soft/std': 0.040072936564683914, 'rewards/judge_quality/mean': 0.32500001788139343, 'rewards/judge_quality/std': 0.27485060691833496, 'rewards/total_composite/mean': 0.5389706492424011, 'rewards/total_composite/std': 0.17612947523593903, 'reward': 0.5389706492424011, 'reward_std': 0.17612949013710022, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09707507491111755, 'sampling/sampling_logp_difference/max': 1.8834095001220703, 'sampling/importance_sampling_ratio/min': 0.15207073092460632, 'sampling/importance_sampling_ratio/mean': 1.0107446908950806, 'sampling/importance_sampling_ratio/max': 1.7325270175933838, 'entropy': 0.5837307199835777, 'clip_ratio/low_mean': 0.05217541428282857, 'clip_ratio/low_min': 0.05217541428282857, 'clip_ratio/high_mean': 0.03508944879285991, 'clip_ratio/high_max': 0.03508944879285991, 'clip_ratio/region_mean': 0.08726486307568848, 'reward_total_mean': 0.5389706492424011, 'reward_meter_mean': 0.9689643383026123, 'reward_meter_std': 0.022389547899365425, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8962827920913696, 'reward_repeat_soft_std': 0.040072936564683914, 'reward_judge_quality_mean': 0.32500001788139343, 'reward_judge_quality_std': 0.27485060691833496, 'reward_total_composite_mean': 0.5389706492424011, 'reward_total_composite_std': 0.17612947523593903, 'epoch': 0.19}

 59%|█████▊    | 1931/3300 [4:45:07<3:09:26,  8.30s/it]INFO 04-13 12:06:05 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1932/3300 [4:45:14<3:02:01,  7.98s/it]2026-04-13 12:06:12,138 | INFO | train_grpo_train | metrics_logged mode=train step=1932

                                                       
{'loss': 0.0345, 'grad_norm': 7.466383457183838, 'learning_rate': 4.148484848484849e-06, 'num_tokens': 3473801.0, 'completions/mean_length': 114.5, 'completions/min_length': 93.0, 'completions/max_length': 143.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 114.5, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 143.0, 'rewards/meter/mean': 0.82825767993927, 'rewards/meter/std': 0.2862468361854553, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9518787860870361, 'rewards/repeat_soft/std': 0.024463597685098648, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.4324488639831543, 'rewards/total_composite/std': 0.0321044884622097, 'reward': 0.4324488639831543, 'reward_std': 0.0321044959127903, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11769791692495346, 'sampling/sampling_logp_difference/max': 1.4505510330200195, 'sampling/importance_sampling_ratio/min': 0.23444108664989471, 'sampling/importance_sampling_ratio/mean': 1.0082027912139893, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8597783148288727, 'clip_ratio/low_mean': 0.04211760591715574, 'clip_ratio/low_min': 0.04211760591715574, 'clip_ratio/high_mean': 0.0752584207803011, 'clip_ratio/high_max': 0.0752584207803011, 'clip_ratio/region_mean': 0.11737602669745684, 'reward_total_mean': 0.4324488639831543, 'reward_meter_mean': 0.82825767993927, 'reward_meter_std': 0.2862468361854553, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9518787860870361, 'reward_repeat_soft_std': 0.024463597685098648, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.4324488639831543, 'reward_total_composite_std': 0.0321044884622097, 'epoch': 0.19}

 59%|█████▊    | 1932/3300 [4:45:14<3:02:01,  7.98s/it]INFO 04-13 12:06:12 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1933/3300 [4:45:25<3:25:55,  9.04s/it]2026-04-13 12:06:23,651 | INFO | train_grpo_train | metrics_logged mode=train step=1933

                                                       
{'loss': -0.1958, 'grad_norm': 1.8619639873504639, 'learning_rate': 4.145454545454546e-06, 'num_tokens': 3475994.0, 'completions/mean_length': 160.125, 'completions/min_length': 96.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 109.85714721679688, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 142.0, 'rewards/meter/mean': 0.8007317781448364, 'rewards/meter/std': 0.31228408217430115, 'rewards/count_adherence/mean': 0.925000011920929, 'rewards/count_adherence/std': 0.2121320217847824, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8632869720458984, 'rewards/repeat_soft/std': 0.08443199098110199, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.047790467739105225, 'rewards/total_composite/mean': 0.402592271566391, 'rewards/total_composite/std': 0.07177911698818207, 'reward': 0.402592271566391, 'reward_std': 0.07177910208702087, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10312580317258835, 'sampling/sampling_logp_difference/max': 1.4721665382385254, 'sampling/importance_sampling_ratio/min': 0.22942787408828735, 'sampling/importance_sampling_ratio/mean': 1.009140968322754, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5213595479726791, 'clip_ratio/low_mean': 0.006433823611587286, 'clip_ratio/low_min': 0.006433823611587286, 'clip_ratio/high_mean': 0.0725303990766406, 'clip_ratio/high_max': 0.0725303990766406, 'clip_ratio/region_mean': 0.07896422268822789, 'reward_total_mean': 0.402592271566391, 'reward_meter_mean': 0.8007317781448364, 'reward_meter_std': 0.31228408217430115, 'reward_count_adherence_mean': 0.925000011920929, 'reward_count_adherence_std': 0.2121320217847824, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8632869720458984, 'reward_repeat_soft_std': 0.08443199098110199, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.047790467739105225, 'reward_total_composite_mean': 0.402592271566391, 'reward_total_composite_std': 0.07177911698818207, 'epoch': 0.19}

 59%|█████▊    | 1933/3300 [4:45:25<3:25:55,  9.04s/it]INFO 04-13 12:06:23 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1934/3300 [4:45:33<3:18:00,  8.70s/it]2026-04-13 12:06:31,561 | INFO | train_grpo_train | metrics_logged mode=train step=1934

                                                       
{'loss': 0.0111, 'grad_norm': 7.697418212890625, 'learning_rate': 4.142424242424243e-06, 'num_tokens': 3477957.0, 'completions/mean_length': 82.375, 'completions/min_length': 70.0, 'completions/max_length': 93.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.375, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 93.0, 'rewards/meter/mean': 0.7867481708526611, 'rewards/meter/std': 0.3093531131744385, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9399130344390869, 'rewards/repeat_soft/std': 0.030503612011671066, 'rewards/judge_quality/mean': 0.21000000834465027, 'rewards/judge_quality/std': 0.0975411981344223, 'rewards/total_composite/mean': 0.45291727781295776, 'rewards/total_composite/std': 0.08042343705892563, 'reward': 0.45291727781295776, 'reward_std': 0.08042343705892563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10768397152423859, 'sampling/sampling_logp_difference/max': 1.8874082565307617, 'sampling/importance_sampling_ratio/min': 0.15146386623382568, 'sampling/importance_sampling_ratio/mean': 1.0010343790054321, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6115433350205421, 'clip_ratio/low_mean': 0.07161152362823486, 'clip_ratio/low_min': 0.07161152362823486, 'clip_ratio/high_mean': 0.0290178582072258, 'clip_ratio/high_max': 0.0290178582072258, 'clip_ratio/region_mean': 0.10062938183546066, 'reward_total_mean': 0.45291727781295776, 'reward_meter_mean': 0.7867481708526611, 'reward_meter_std': 0.3093531131744385, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9399130344390869, 'reward_repeat_soft_std': 0.030503612011671066, 'reward_judge_quality_mean': 0.21000000834465027, 'reward_judge_quality_std': 0.0975411981344223, 'reward_total_composite_mean': 0.45291727781295776, 'reward_total_composite_std': 0.08042343705892563, 'epoch': 0.19}

 59%|█████▊    | 1934/3300 [4:45:33<3:18:00,  8.70s/it]INFO 04-13 12:06:31 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1935/3300 [4:45:40<3:04:24,  8.11s/it]2026-04-13 12:06:38,271 | INFO | train_grpo_train | metrics_logged mode=train step=1935

                                                       
{'loss': -0.0081, 'grad_norm': 9.461777687072754, 'learning_rate': 4.13939393939394e-06, 'num_tokens': 3479615.0, 'completions/mean_length': 57.25, 'completions/min_length': 49.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.25, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7915312647819519, 'rewards/meter/std': 0.3439336121082306, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8316245079040527, 'rewards/repeat_soft/std': 0.066276915371418, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.401917964220047, 'rewards/total_composite/std': 0.032592639327049255, 'reward': 0.401917964220047, 'reward_std': 0.032592643052339554, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0932314544916153, 'sampling/sampling_logp_difference/max': 1.5124073028564453, 'sampling/importance_sampling_ratio/min': 0.2203788310289383, 'sampling/importance_sampling_ratio/mean': 1.012587308883667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6603519469499588, 'clip_ratio/low_mean': 0.025438006035983562, 'clip_ratio/low_min': 0.025438006035983562, 'clip_ratio/high_mean': 0.06806070310994983, 'clip_ratio/high_max': 0.06806070310994983, 'clip_ratio/region_mean': 0.09349870914593339, 'reward_total_mean': 0.401917964220047, 'reward_meter_mean': 0.7915312647819519, 'reward_meter_std': 0.3439336121082306, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8316245079040527, 'reward_repeat_soft_std': 0.066276915371418, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.401917964220047, 'reward_total_composite_std': 0.032592639327049255, 'epoch': 0.19}

 59%|█████▊    | 1935/3300 [4:45:40<3:04:24,  8.11s/it]INFO 04-13 12:06:38 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1936/3300 [4:45:46<2:52:01,  7.57s/it]2026-04-13 12:06:44,572 | INFO | train_grpo_train | metrics_logged mode=train step=1936

                                                       
{'loss': 0.0203, 'grad_norm': 9.021432876586914, 'learning_rate': 4.136363636363637e-06, 'num_tokens': 3481209.0, 'completions/mean_length': 42.25, 'completions/min_length': 36.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.7995381355285645, 'rewards/meter/std': 0.18903721868991852, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9265373349189758, 'rewards/repeat_soft/std': 0.04161253198981285, 'rewards/judge_quality/mean': 0.24625001847743988, 'rewards/judge_quality/std': 0.2722361087799072, 'rewards/total_composite/mean': 0.42288756370544434, 'rewards/total_composite/std': 0.23189647495746613, 'reward': 0.42288756370544434, 'reward_std': 0.23189647495746613, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12788273394107819, 'sampling/sampling_logp_difference/max': 1.9591566324234009, 'sampling/importance_sampling_ratio/min': 0.14097726345062256, 'sampling/importance_sampling_ratio/mean': 0.9938132762908936, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8986107259988785, 'clip_ratio/low_mean': 0.042058270424604416, 'clip_ratio/low_min': 0.042058270424604416, 'clip_ratio/high_mean': 0.07521497923880816, 'clip_ratio/high_max': 0.07521497923880816, 'clip_ratio/region_mean': 0.11727324966341257, 'reward_total_mean': 0.42288756370544434, 'reward_meter_mean': 0.7995381355285645, 'reward_meter_std': 0.18903721868991852, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9265373349189758, 'reward_repeat_soft_std': 0.04161253198981285, 'reward_judge_quality_mean': 0.24625001847743988, 'reward_judge_quality_std': 0.2722361087799072, 'reward_total_composite_mean': 0.42288756370544434, 'reward_total_composite_std': 0.23189647495746613, 'epoch': 0.19}

 59%|█████▊    | 1936/3300 [4:45:46<2:52:01,  7.57s/it]INFO 04-13 12:06:44 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1937/3300 [4:45:53<2:45:57,  7.31s/it]2026-04-13 12:06:51,269 | INFO | train_grpo_train | metrics_logged mode=train step=1937

                                                       
{'loss': 0.0372, 'grad_norm': 12.0358247756958, 'learning_rate': 4.133333333333333e-06, 'num_tokens': 3482851.0, 'completions/mean_length': 49.25, 'completions/min_length': 39.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.25, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7147437930107117, 'rewards/meter/std': 0.32077938318252563, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8967220783233643, 'rewards/repeat_soft/std': 0.0788450762629509, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4041958451271057, 'rewards/total_composite/std': 0.03756310045719147, 'reward': 0.4041958451271057, 'reward_std': 0.03756310045719147, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10153629630804062, 'sampling/sampling_logp_difference/max': 1.3987922668457031, 'sampling/importance_sampling_ratio/min': 0.246894970536232, 'sampling/importance_sampling_ratio/mean': 1.0087206363677979, 'sampling/importance_sampling_ratio/max': 1.8293730020523071, 'entropy': 0.7433633059263229, 'clip_ratio/low_mean': 0.03178792679682374, 'clip_ratio/low_min': 0.03178792679682374, 'clip_ratio/high_mean': 0.06418938236311078, 'clip_ratio/high_max': 0.06418938236311078, 'clip_ratio/region_mean': 0.09597730915993452, 'reward_total_mean': 0.4041958451271057, 'reward_meter_mean': 0.7147437930107117, 'reward_meter_std': 0.32077938318252563, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8967220783233643, 'reward_repeat_soft_std': 0.0788450762629509, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4041958451271057, 'reward_total_composite_std': 0.03756310045719147, 'epoch': 0.19}

 59%|█████▊    | 1937/3300 [4:45:53<2:45:57,  7.31s/it]INFO 04-13 12:06:51 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▊    | 1938/3300 [4:46:04<3:14:10,  8.55s/it]2026-04-13 12:07:02,738 | INFO | train_grpo_train | metrics_logged mode=train step=1938

                                                       
{'loss': -0.0854, 'grad_norm': 1.6866087913513184, 'learning_rate': 4.1303030303030305e-06, 'num_tokens': 3484174.0, 'completions/mean_length': 84.375, 'completions/min_length': 18.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 23.285715103149414, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.26195597648620605, 'rewards/meter/std': 0.3709193170070648, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9555385112762451, 'rewards/repeat_soft/std': 0.019451912492513657, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.3258320391178131, 'rewards/total_composite/std': 0.13631708920001984, 'reward': 0.3258320391178131, 'reward_std': 0.13631707429885864, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11393966525793076, 'sampling/sampling_logp_difference/max': 1.4023149013519287, 'sampling/importance_sampling_ratio/min': 0.24602676928043365, 'sampling/importance_sampling_ratio/mean': 1.0280745029449463, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6536348387598991, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10601353459060192, 'clip_ratio/high_max': 0.10601353459060192, 'clip_ratio/region_mean': 0.10601353459060192, 'reward_total_mean': 0.3258320391178131, 'reward_meter_mean': 0.26195597648620605, 'reward_meter_std': 0.3709193170070648, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9555385112762451, 'reward_repeat_soft_std': 0.019451912492513657, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.3258320391178131, 'reward_total_composite_std': 0.13631708920001984, 'epoch': 0.19}

 59%|█████▊    | 1938/3300 [4:46:04<3:14:10,  8.55s/it]INFO 04-13 12:07:02 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1939/3300 [4:46:11<2:58:23,  7.86s/it]2026-04-13 12:07:08,999 | INFO | train_grpo_train | metrics_logged mode=train step=1939

                                                       
{'loss': 0.1062, 'grad_norm': 23.812334060668945, 'learning_rate': 4.127272727272728e-06, 'num_tokens': 3485617.0, 'completions/mean_length': 25.375, 'completions/min_length': 21.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.375, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9654932618141174, 'rewards/meter/std': 0.06134282797574997, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9411230087280273, 'rewards/repeat_soft/std': 0.03368011489510536, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43530404567718506, 'rewards/total_composite/std': 0.009656086564064026, 'reward': 0.43530404567718506, 'reward_std': 0.00965608935803175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.15782931447029114, 'sampling/sampling_logp_difference/max': 2.1210851669311523, 'sampling/importance_sampling_ratio/min': 0.11990144103765488, 'sampling/importance_sampling_ratio/mean': 1.0210922956466675, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0095732435584068, 'clip_ratio/low_mean': 0.03354978375136852, 'clip_ratio/low_min': 0.03354978375136852, 'clip_ratio/high_mean': 0.11721561383455992, 'clip_ratio/high_max': 0.11721561383455992, 'clip_ratio/region_mean': 0.15076539758592844, 'reward_total_mean': 0.43530404567718506, 'reward_meter_mean': 0.9654932618141174, 'reward_meter_std': 0.06134282797574997, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9411230087280273, 'reward_repeat_soft_std': 0.03368011489510536, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43530404567718506, 'reward_total_composite_std': 0.009656086564064026, 'epoch': 0.19}

 59%|█████▉    | 1939/3300 [4:46:11<2:58:23,  7.86s/it]INFO 04-13 12:07:09 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1940/3300 [4:46:23<3:29:22,  9.24s/it]2026-04-13 12:07:21,443 | INFO | train_grpo_train | metrics_logged mode=train step=1940

                                                       
{'loss': -0.1536, 'grad_norm': 1.700808048248291, 'learning_rate': 4.124242424242424e-06, 'num_tokens': 3487447.0, 'completions/mean_length': 114.75, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 58.000003814697266, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 69.0, 'rewards/meter/mean': 0.9692917466163635, 'rewards/meter/std': 0.04168885573744774, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9311715364456177, 'rewards/repeat_soft/std': 0.030505457893013954, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.3798644542694092, 'rewards/total_composite/std': 0.15350237488746643, 'reward': 0.3798644542694092, 'reward_std': 0.15350237488746643, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1226680725812912, 'sampling/sampling_logp_difference/max': 1.3897569179534912, 'sampling/importance_sampling_ratio/min': 0.2491358518600464, 'sampling/importance_sampling_ratio/mean': 1.0107020139694214, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8404737338423729, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10910185892134905, 'clip_ratio/high_max': 0.10910185892134905, 'clip_ratio/region_mean': 0.10910185892134905, 'reward_total_mean': 0.3798644542694092, 'reward_meter_mean': 0.9692917466163635, 'reward_meter_std': 0.04168885573744774, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9311715364456177, 'reward_repeat_soft_std': 0.030505457893013954, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.3798644542694092, 'reward_total_composite_std': 0.15350237488746643, 'epoch': 0.19}

 59%|█████▉    | 1940/3300 [4:46:23<3:29:22,  9.24s/it]INFO 04-13 12:07:21 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1941/3300 [4:46:30<3:14:17,  8.58s/it]2026-04-13 12:07:28,473 | INFO | train_grpo_train | metrics_logged mode=train step=1941

                                                       
{'loss': -0.01, 'grad_norm': 7.8766961097717285, 'learning_rate': 4.1212121212121215e-06, 'num_tokens': 3489273.0, 'completions/mean_length': 52.25, 'completions/min_length': 43.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.6973185539245605, 'rewards/meter/std': 0.2894766330718994, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9258058071136475, 'rewards/repeat_soft/std': 0.03156572952866554, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.406859427690506, 'rewards/total_composite/std': 0.02906632423400879, 'reward': 0.406859427690506, 'reward_std': 0.02906632237136364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08351286500692368, 'sampling/sampling_logp_difference/max': 1.7317934036254883, 'sampling/importance_sampling_ratio/min': 0.17696675658226013, 'sampling/importance_sampling_ratio/mean': 1.0031673908233643, 'sampling/importance_sampling_ratio/max': 1.802559494972229, 'entropy': 0.5053677409887314, 'clip_ratio/low_mean': 0.03865512926131487, 'clip_ratio/low_min': 0.03865512926131487, 'clip_ratio/high_mean': 0.030179704073816538, 'clip_ratio/high_max': 0.030179704073816538, 'clip_ratio/region_mean': 0.0688348333351314, 'reward_total_mean': 0.406859427690506, 'reward_meter_mean': 0.6973185539245605, 'reward_meter_std': 0.2894766330718994, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9258058071136475, 'reward_repeat_soft_std': 0.03156572952866554, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.406859427690506, 'reward_total_composite_std': 0.02906632423400879, 'epoch': 0.19}

 59%|█████▉    | 1941/3300 [4:46:30<3:14:17,  8.58s/it]INFO 04-13 12:07:28 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1942/3300 [4:46:37<3:02:12,  8.05s/it]2026-04-13 12:07:35,297 | INFO | train_grpo_train | metrics_logged mode=train step=1942

                                                       
{'loss': 0.0491, 'grad_norm': 16.240028381347656, 'learning_rate': 4.118181818181819e-06, 'num_tokens': 3490902.0, 'completions/mean_length': 29.625, 'completions/min_length': 25.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.625, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.8388985395431519, 'rewards/meter/std': 0.33876854181289673, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9396454691886902, 'rewards/repeat_soft/std': 0.030232569202780724, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4227394461631775, 'rewards/total_composite/std': 0.03190704807639122, 'reward': 0.4227394461631775, 'reward_std': 0.03190704807639122, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08028630167245865, 'sampling/sampling_logp_difference/max': 2.378615379333496, 'sampling/importance_sampling_ratio/min': 0.09267881512641907, 'sampling/importance_sampling_ratio/mean': 1.0173338651657104, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40816619247198105, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.052720814011991024, 'clip_ratio/high_max': 0.052720814011991024, 'clip_ratio/region_mean': 0.06029657181352377, 'reward_total_mean': 0.4227394461631775, 'reward_meter_mean': 0.8388985395431519, 'reward_meter_std': 0.33876854181289673, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9396454691886902, 'reward_repeat_soft_std': 0.030232569202780724, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4227394461631775, 'reward_total_composite_std': 0.03190704807639122, 'epoch': 0.2}

 59%|█████▉    | 1942/3300 [4:46:37<3:02:12,  8.05s/it]INFO 04-13 12:07:35 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1943/3300 [4:46:45<3:03:41,  8.12s/it]2026-04-13 12:07:43,590 | INFO | train_grpo_train | metrics_logged mode=train step=1943

                                                       
{'loss': 0.0331, 'grad_norm': 6.2539472579956055, 'learning_rate': 4.115151515151515e-06, 'num_tokens': 3492864.0, 'completions/mean_length': 91.25, 'completions/min_length': 69.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.25, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.8011467456817627, 'rewards/meter/std': 0.2510037422180176, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9019615650177002, 'rewards/repeat_soft/std': 0.07098674029111862, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4202427864074707, 'rewards/total_composite/std': 0.028576964512467384, 'reward': 0.4202427864074707, 'reward_std': 0.028576962649822235, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09535551816225052, 'sampling/sampling_logp_difference/max': 1.388596534729004, 'sampling/importance_sampling_ratio/min': 0.24942512810230255, 'sampling/importance_sampling_ratio/mean': 1.0110878944396973, 'sampling/importance_sampling_ratio/max': 1.6919692754745483, 'entropy': 0.7175753563642502, 'clip_ratio/low_mean': 0.04605818632990122, 'clip_ratio/low_min': 0.04605818632990122, 'clip_ratio/high_mean': 0.05231527891010046, 'clip_ratio/high_max': 0.05231527891010046, 'clip_ratio/region_mean': 0.09837346524000168, 'reward_total_mean': 0.4202427864074707, 'reward_meter_mean': 0.8011467456817627, 'reward_meter_std': 0.2510037422180176, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9019615650177002, 'reward_repeat_soft_std': 0.07098674029111862, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4202427864074707, 'reward_total_composite_std': 0.028576964512467384, 'epoch': 0.2}

 59%|█████▉    | 1943/3300 [4:46:45<3:03:41,  8.12s/it]INFO 04-13 12:07:43 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1944/3300 [4:46:53<3:03:51,  8.14s/it]2026-04-13 12:07:51,746 | INFO | train_grpo_train | metrics_logged mode=train step=1944

                                                       
{'loss': 0.0599, 'grad_norm': 6.8244500160217285, 'learning_rate': 4.112121212121212e-06, 'num_tokens': 3495360.0, 'completions/mean_length': 141.0, 'completions/min_length': 120.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 141.0, 'completions/min_terminated_length': 120.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.8854571580886841, 'rewards/meter/std': 0.15358369052410126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9386004209518433, 'rewards/repeat_soft/std': 0.02755219116806984, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.43835583329200745, 'rewards/total_composite/std': 0.024016335606575012, 'reward': 0.43835583329200745, 'reward_std': 0.024016333743929863, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11116247624158859, 'sampling/sampling_logp_difference/max': 1.5638446807861328, 'sampling/importance_sampling_ratio/min': 0.20932970941066742, 'sampling/importance_sampling_ratio/mean': 1.0016816854476929, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4846718981862068, 'clip_ratio/low_mean': 0.026041666977107525, 'clip_ratio/low_min': 0.026041666977107525, 'clip_ratio/high_mean': 0.07818500325083733, 'clip_ratio/high_max': 0.07818500325083733, 'clip_ratio/region_mean': 0.10422667022794485, 'reward_total_mean': 0.43835583329200745, 'reward_meter_mean': 0.8854571580886841, 'reward_meter_std': 0.15358369052410126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9386004209518433, 'reward_repeat_soft_std': 0.02755219116806984, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.43835583329200745, 'reward_total_composite_std': 0.024016335606575012, 'epoch': 0.2}

 59%|█████▉    | 1944/3300 [4:46:53<3:03:51,  8.14s/it]INFO 04-13 12:07:51 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1945/3300 [4:47:00<2:53:14,  7.67s/it]2026-04-13 12:07:58,331 | INFO | train_grpo_train | metrics_logged mode=train step=1945

                                                       
{'loss': -0.0001, 'grad_norm': 10.8930082321167, 'learning_rate': 4.10909090909091e-06, 'num_tokens': 3497337.0, 'completions/mean_length': 84.125, 'completions/min_length': 76.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.125, 'completions/min_terminated_length': 76.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.5784173011779785, 'rewards/meter/std': 0.40984830260276794, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9831427335739136, 'rewards/repeat_soft/std': 0.017384203150868416, 'rewards/judge_quality/mean': 0.21000000834465027, 'rewards/judge_quality/std': 0.0975411981344223, 'rewards/total_composite/mean': 0.41480445861816406, 'rewards/total_composite/std': 0.0438450463116169, 'reward': 0.41480445861816406, 'reward_std': 0.043845050036907196, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14168818295001984, 'sampling/sampling_logp_difference/max': 2.2974629402160645, 'sampling/importance_sampling_ratio/min': 0.1005135253071785, 'sampling/importance_sampling_ratio/mean': 0.9871100187301636, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4833308830857277, 'clip_ratio/low_mean': 0.04418520908802748, 'clip_ratio/low_min': 0.04418520908802748, 'clip_ratio/high_mean': 0.068550162948668, 'clip_ratio/high_max': 0.068550162948668, 'clip_ratio/region_mean': 0.11273537203669548, 'reward_total_mean': 0.41480445861816406, 'reward_meter_mean': 0.5784173011779785, 'reward_meter_std': 0.40984830260276794, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9831427335739136, 'reward_repeat_soft_std': 0.017384203150868416, 'reward_judge_quality_mean': 0.21000000834465027, 'reward_judge_quality_std': 0.0975411981344223, 'reward_total_composite_mean': 0.41480445861816406, 'reward_total_composite_std': 0.0438450463116169, 'epoch': 0.2}

 59%|█████▉    | 1945/3300 [4:47:00<2:53:14,  7.67s/it]INFO 04-13 12:07:58 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1946/3300 [4:47:07<2:45:59,  7.36s/it]2026-04-13 12:08:04,959 | INFO | train_grpo_train | metrics_logged mode=train step=1946

                                                       
{'loss': 0.0256, 'grad_norm': 9.013469696044922, 'learning_rate': 4.106060606060606e-06, 'num_tokens': 3499330.0, 'completions/mean_length': 67.125, 'completions/min_length': 56.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 67.125, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.5188466310501099, 'rewards/meter/std': 0.29012858867645264, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9260605573654175, 'rewards/repeat_soft/std': 0.05419487506151199, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.3905407786369324, 'rewards/total_composite/std': 0.029915479943156242, 'reward': 0.3905407786369324, 'reward_std': 0.02991548366844654, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1178271472454071, 'sampling/sampling_logp_difference/max': 3.939318895339966, 'sampling/importance_sampling_ratio/min': 0.019461465999484062, 'sampling/importance_sampling_ratio/mean': 0.9934802651405334, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5234301798045635, 'clip_ratio/low_mean': 0.061407689936459064, 'clip_ratio/low_min': 0.061407689936459064, 'clip_ratio/high_mean': 0.045002466067671776, 'clip_ratio/high_max': 0.045002466067671776, 'clip_ratio/region_mean': 0.10641015600413084, 'reward_total_mean': 0.3905407786369324, 'reward_meter_mean': 0.5188466310501099, 'reward_meter_std': 0.29012858867645264, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9260605573654175, 'reward_repeat_soft_std': 0.05419487506151199, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.3905407786369324, 'reward_total_composite_std': 0.029915479943156242, 'epoch': 0.2}

 59%|█████▉    | 1946/3300 [4:47:07<2:45:59,  7.36s/it]INFO 04-13 12:08:05 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1947/3300 [4:47:13<2:39:04,  7.05s/it]2026-04-13 12:08:11,328 | INFO | train_grpo_train | metrics_logged mode=train step=1947

                                                       
{'loss': 0.139, 'grad_norm': 13.41064167022705, 'learning_rate': 4.103030303030303e-06, 'num_tokens': 3501115.0, 'completions/mean_length': 49.125, 'completions/min_length': 41.0, 'completions/max_length': 66.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 49.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 66.0, 'rewards/meter/mean': 0.7810739278793335, 'rewards/meter/std': 0.32668501138687134, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9738366603851318, 'rewards/repeat_soft/std': 0.03120327740907669, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.42223018407821655, 'rewards/total_composite/std': 0.030131608247756958, 'reward': 0.42223018407821655, 'reward_std': 0.030131608247756958, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14453963935375214, 'sampling/sampling_logp_difference/max': 1.6577138900756836, 'sampling/importance_sampling_ratio/min': 0.19057415425777435, 'sampling/importance_sampling_ratio/mean': 1.0178121328353882, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0398504585027695, 'clip_ratio/low_mean': 0.05733144469559193, 'clip_ratio/low_min': 0.05733144469559193, 'clip_ratio/high_mean': 0.11275254096835852, 'clip_ratio/high_max': 0.11275254096835852, 'clip_ratio/region_mean': 0.17008398566395044, 'reward_total_mean': 0.42223018407821655, 'reward_meter_mean': 0.7810739278793335, 'reward_meter_std': 0.32668501138687134, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9738366603851318, 'reward_repeat_soft_std': 0.03120327740907669, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.42223018407821655, 'reward_total_composite_std': 0.030131608247756958, 'epoch': 0.2}

 59%|█████▉    | 1947/3300 [4:47:13<2:39:04,  7.05s/it]INFO 04-13 12:08:11 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1948/3300 [4:47:21<2:44:14,  7.29s/it]2026-04-13 12:08:19,146 | INFO | train_grpo_train | metrics_logged mode=train step=1948

                                                       
{'loss': 0.0731, 'grad_norm': 12.654583930969238, 'learning_rate': 4.1e-06, 'num_tokens': 3502794.0, 'completions/mean_length': 46.875, 'completions/min_length': 39.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.7894902229309082, 'rewards/meter/std': 0.3574954867362976, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9605759978294373, 'rewards/repeat_soft/std': 0.03400946035981178, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4210616946220398, 'rewards/total_composite/std': 0.03653131425380707, 'reward': 0.4210616946220398, 'reward_std': 0.03653131425380707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11745037883520126, 'sampling/sampling_logp_difference/max': 1.2075090408325195, 'sampling/importance_sampling_ratio/min': 0.2989410161972046, 'sampling/importance_sampling_ratio/mean': 0.9976826310157776, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.599471140652895, 'clip_ratio/low_mean': 0.023413733579218388, 'clip_ratio/low_min': 0.023413733579218388, 'clip_ratio/high_mean': 0.09726464096456766, 'clip_ratio/high_max': 0.09726464096456766, 'clip_ratio/region_mean': 0.12067837454378605, 'reward_total_mean': 0.4210616946220398, 'reward_meter_mean': 0.7894902229309082, 'reward_meter_std': 0.3574954867362976, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9605759978294373, 'reward_repeat_soft_std': 0.03400946035981178, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4210616946220398, 'reward_total_composite_std': 0.03653131425380707, 'epoch': 0.2}

 59%|█████▉    | 1948/3300 [4:47:21<2:44:14,  7.29s/it]INFO 04-13 12:08:19 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1949/3300 [4:47:28<2:41:31,  7.17s/it]2026-04-13 12:08:26,041 | INFO | train_grpo_train | metrics_logged mode=train step=1949

                                                       
{'loss': -0.0375, 'grad_norm': 10.614028930664062, 'learning_rate': 4.096969696969697e-06, 'num_tokens': 3504563.0, 'completions/mean_length': 46.125, 'completions/min_length': 38.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.125, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.853850781917572, 'rewards/meter/std': 0.20310230553150177, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9574228525161743, 'rewards/repeat_soft/std': 0.015529298223555088, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4341689348220825, 'rewards/total_composite/std': 0.018940472975373268, 'reward': 0.4341689348220825, 'reward_std': 0.018940474838018417, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08881442248821259, 'sampling/sampling_logp_difference/max': 1.837799072265625, 'sampling/importance_sampling_ratio/min': 0.1591673642396927, 'sampling/importance_sampling_ratio/mean': 0.9945177435874939, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4520632103085518, 'clip_ratio/low_mean': 0.024074074113741517, 'clip_ratio/low_min': 0.024074074113741517, 'clip_ratio/high_mean': 0.06723985262215137, 'clip_ratio/high_max': 0.06723985262215137, 'clip_ratio/region_mean': 0.09131392673589289, 'reward_total_mean': 0.4341689348220825, 'reward_meter_mean': 0.853850781917572, 'reward_meter_std': 0.20310230553150177, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9574228525161743, 'reward_repeat_soft_std': 0.015529298223555088, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4341689348220825, 'reward_total_composite_std': 0.018940472975373268, 'epoch': 0.2}

 59%|█████▉    | 1949/3300 [4:47:28<2:41:31,  7.17s/it]INFO 04-13 12:08:26 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1950/3300 [4:47:34<2:38:49,  7.06s/it]2026-04-13 12:08:32,838 | INFO | train_grpo_train | metrics_logged mode=train step=1950

                                                       
{'loss': -0.0324, 'grad_norm': 6.639441013336182, 'learning_rate': 4.093939393939394e-06, 'num_tokens': 3506541.0, 'completions/mean_length': 77.25, 'completions/min_length': 61.0, 'completions/max_length': 95.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 77.25, 'completions/min_terminated_length': 61.0, 'completions/max_terminated_length': 95.0, 'rewards/meter/mean': 0.856492817401886, 'rewards/meter/std': 0.23869705200195312, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.86268150806427, 'rewards/repeat_soft/std': 0.03397032245993614, 'rewards/judge_quality/mean': 0.23625001311302185, 'rewards/judge_quality/std': 0.13265825808048248, 'rewards/total_composite/mean': 0.46340858936309814, 'rewards/total_composite/std': 0.08909754455089569, 'reward': 0.46340858936309814, 'reward_std': 0.0890975371003151, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08762527257204056, 'sampling/sampling_logp_difference/max': 2.2014412879943848, 'sampling/importance_sampling_ratio/min': 0.11064357310533524, 'sampling/importance_sampling_ratio/mean': 1.025205373764038, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4552963748574257, 'clip_ratio/low_mean': 0.06273894617334008, 'clip_ratio/low_min': 0.06273894617334008, 'clip_ratio/high_mean': 0.0182926831766963, 'clip_ratio/high_max': 0.0182926831766963, 'clip_ratio/region_mean': 0.08103162935003638, 'reward_total_mean': 0.46340858936309814, 'reward_meter_mean': 0.856492817401886, 'reward_meter_std': 0.23869705200195312, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.86268150806427, 'reward_repeat_soft_std': 0.03397032245993614, 'reward_judge_quality_mean': 0.23625001311302185, 'reward_judge_quality_std': 0.13265825808048248, 'reward_total_composite_mean': 0.46340858936309814, 'reward_total_composite_std': 0.08909754455089569, 'epoch': 0.2}

 59%|█████▉    | 1950/3300 [4:47:35<2:38:49,  7.06s/it]INFO 04-13 12:08:33 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.86s/it]

 30%|███       | 3/10 [00:08<00:21,  3.11s/it]

 40%|████      | 4/10 [00:12<00:20,  3.35s/it]

 50%|█████     | 5/10 [00:17<00:19,  3.85s/it]

 60%|██████    | 6/10 [00:25<00:21,  5.46s/it]

 70%|███████   | 7/10 [00:30<00:15,  5.05s/it]

 80%|████████  | 8/10 [00:34<00:09,  4.92s/it]

 90%|█████████ | 9/10 [00:38<00:04,  4.68s/it]

100%|██████████| 10/10 [00:42<00:00,  4.44s/it]2026-04-13 12:09:20,344 | INFO | train_grpo_train | metrics_logged mode=eval step=1950

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 47.3851, 'eval_samples_per_second': 1.688, 'eval_steps_per_second': 0.211, 'eval_num_tokens': 3506541.0, 'eval_completions/mean_length': 87.3375, 'eval_completions/min_length': 34.0, 'eval_completions/max_length': 179.1, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 82.48571472167968, 'eval_completions/min_terminated_length': 34.0, 'eval_completions/max_terminated_length': 145.7, 'eval_rewards/meter/mean': 0.8456712782382965, 'eval_rewards/meter/std': 0.25767267104238273, 'eval_rewards/count_adherence/mean': 0.990625, 'eval_rewards/count_adherence/std': 0.02651650384068489, 'eval_rewards/hard_gate/mean': 0.9875, 'eval_rewards/hard_gate/std': 0.03535533845424652, 'eval_rewards/repeat_soft/mean': 0.8660417795181274, 'eval_rewards/repeat_soft/std': 0.10166475251317024, 'eval_rewards/judge_quality/mean': 0.1618750050663948, 'eval_rewards/judge_quality/std': 0.028590949438512325, 'eval_rewards/total_composite/mean': 0.41445557177066805, 'eval_rewards/total_composite/std': 0.04768334738910198, 'eval_reward': 0.41445557177066805, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.04655205011367798, 'eval_sampling/sampling_logp_difference/max': 0.9743308305740357, 'eval_sampling/importance_sampling_ratio/min': 0.38960790634155273, 'eval_sampling/importance_sampling_ratio/mean': 1.0100993633270263, 'eval_sampling/importance_sampling_ratio/max': 1.427927815914154, 'eval_entropy': 0.5144751250743866, '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': 0.41445557177066805, 'eval_reward_meter_mean': 0.8456712782382965, 'eval_reward_meter_std': 0.25767267104238273, 'eval_reward_count_adherence_mean': 0.990625, 'eval_reward_count_adherence_std': 0.02651650384068489, 'eval_reward_hard_gate_mean': 0.9875, 'eval_reward_hard_gate_std': 0.03535533845424652, 'eval_reward_repeat_soft_mean': 0.8660417795181274, 'eval_reward_repeat_soft_std': 0.10166475251317024, 'eval_reward_judge_quality_mean': 0.1618750050663948, 'eval_reward_judge_quality_std': 0.028590949438512325, 'eval_reward_total_composite_mean': 0.41445557177066805, 'eval_reward_total_composite_std': 0.04768334738910198, 'epoch': 0.2}

 59%|█████▉    | 1950/3300 [4:48:22<2:38:49,  7.06s/it]

100%|██████████| 10/10 [00:42<00:00,  4.44s/it]

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

 59%|█████▉    | 1951/3300 [4:48:32<8:18:51, 22.19s/it]2026-04-13 12:09:30,327 | INFO | train_grpo_train | metrics_logged mode=train step=1951

                                                       
{'loss': -0.038, 'grad_norm': 10.133533477783203, 'learning_rate': 4.0909090909090915e-06, 'num_tokens': 3508186.0, 'completions/mean_length': 52.625, 'completions/min_length': 43.0, 'completions/max_length': 77.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.625, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 77.0, 'rewards/meter/mean': 0.9263145923614502, 'rewards/meter/std': 0.11202561110258102, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9856852889060974, 'rewards/repeat_soft/std': 0.017451290041208267, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4454079866409302, 'rewards/total_composite/std': 0.016052085906267166, 'reward': 0.4454079866409302, 'reward_std': 0.016052084043622017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11850112676620483, 'sampling/sampling_logp_difference/max': 1.7425165176391602, 'sampling/importance_sampling_ratio/min': 0.17507925629615784, 'sampling/importance_sampling_ratio/mean': 0.9920020699501038, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7089521065354347, 'clip_ratio/low_mean': 0.07087081763893366, 'clip_ratio/low_min': 0.07087081763893366, 'clip_ratio/high_mean': 0.05990118533372879, 'clip_ratio/high_max': 0.05990118533372879, 'clip_ratio/region_mean': 0.13077200297266245, 'reward_total_mean': 0.4454079866409302, 'reward_meter_mean': 0.9263145923614502, 'reward_meter_std': 0.11202561110258102, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9856852889060974, 'reward_repeat_soft_std': 0.017451290041208267, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4454079866409302, 'reward_total_composite_std': 0.016052085906267166, 'epoch': 0.2}

 59%|█████▉    | 1951/3300 [4:48:32<8:18:51, 22.19s/it]INFO 04-13 12:09:30 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1952/3300 [4:48:38<6:31:48, 17.44s/it]2026-04-13 12:09:36,683 | INFO | train_grpo_train | metrics_logged mode=train step=1952

                                                       
{'loss': 0.0555, 'grad_norm': 34.25065994262695, 'learning_rate': 4.087878787878789e-06, 'num_tokens': 3509693.0, 'completions/mean_length': 34.375, 'completions/min_length': 28.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.375, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.4445779025554657, 'rewards/meter/std': 0.27867090702056885, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9061572551727295, 'rewards/repeat_soft/std': 0.013661096803843975, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.37926995754241943, 'rewards/total_composite/std': 0.028161024674773216, 'reward': 0.37926995754241943, 'reward_std': 0.028161028400063515, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10101846605539322, 'sampling/sampling_logp_difference/max': 1.9981274604797363, 'sampling/importance_sampling_ratio/min': 0.13558894395828247, 'sampling/importance_sampling_ratio/mean': 0.9924972057342529, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39059606939554214, 'clip_ratio/low_mean': 0.053744684904813766, 'clip_ratio/low_min': 0.053744684904813766, 'clip_ratio/high_mean': 0.06362083368003368, 'clip_ratio/high_max': 0.06362083368003368, 'clip_ratio/region_mean': 0.11736551858484745, 'reward_total_mean': 0.37926995754241943, 'reward_meter_mean': 0.4445779025554657, 'reward_meter_std': 0.27867090702056885, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9061572551727295, 'reward_repeat_soft_std': 0.013661096803843975, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.37926995754241943, 'reward_total_composite_std': 0.028161024674773216, 'epoch': 0.2}

 59%|█████▉    | 1952/3300 [4:48:38<6:31:48, 17.44s/it]INFO 04-13 12:09:36 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1953/3300 [4:48:45<5:18:52, 14.20s/it]2026-04-13 12:09:43,344 | INFO | train_grpo_train | metrics_logged mode=train step=1953

                                                       
{'loss': 0.0071, 'grad_norm': 8.840203285217285, 'learning_rate': 4.084848484848485e-06, 'num_tokens': 3511276.0, 'completions/mean_length': 41.875, 'completions/min_length': 36.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.990298867225647, 'rewards/meter/std': 0.00986767839640379, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.899217426776886, 'rewards/repeat_soft/std': 0.03236125409603119, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4314367473125458, 'rewards/total_composite/std': 0.005005829967558384, 'reward': 0.4314367473125458, 'reward_std': 0.005005823448300362, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09813302010297775, 'sampling/sampling_logp_difference/max': 1.5159664154052734, 'sampling/importance_sampling_ratio/min': 0.21959586441516876, 'sampling/importance_sampling_ratio/mean': 1.0026865005493164, 'sampling/importance_sampling_ratio/max': 1.9148515462875366, 'entropy': 0.546804204583168, 'clip_ratio/low_mean': 0.0374533380381763, 'clip_ratio/low_min': 0.0374533380381763, 'clip_ratio/high_mean': 0.06395576894283295, 'clip_ratio/high_max': 0.06395576894283295, 'clip_ratio/region_mean': 0.10140910698100924, 'reward_total_mean': 0.4314367473125458, 'reward_meter_mean': 0.990298867225647, 'reward_meter_std': 0.00986767839640379, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.899217426776886, 'reward_repeat_soft_std': 0.03236125409603119, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4314367473125458, 'reward_total_composite_std': 0.005005829967558384, 'epoch': 0.2}

 59%|█████▉    | 1953/3300 [4:48:45<5:18:52, 14.20s/it]INFO 04-13 12:09:43 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1954/3300 [4:48:52<4:27:12, 11.91s/it]2026-04-13 12:09:49,980 | INFO | train_grpo_train | metrics_logged mode=train step=1954

                                                       
{'loss': 0.0691, 'grad_norm': 14.853337287902832, 'learning_rate': 4.081818181818182e-06, 'num_tokens': 3513076.0, 'completions/mean_length': 54.0, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.0, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9747487306594849, 'rewards/meter/std': 0.01338912732899189, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.953041136264801, 'rewards/repeat_soft/std': 0.03769413009285927, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.013887306675314903, 'rewards/total_composite/mean': 0.43320900201797485, 'rewards/total_composite/std': 0.011402982287108898, 'reward': 0.43320900201797485, 'reward_std': 0.011402991600334644, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11398535221815109, 'sampling/sampling_logp_difference/max': 2.2634103298187256, 'sampling/importance_sampling_ratio/min': 0.10399521887302399, 'sampling/importance_sampling_ratio/mean': 1.017561912536621, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6152001395821571, 'clip_ratio/low_mean': 0.02483258955180645, 'clip_ratio/low_min': 0.02483258955180645, 'clip_ratio/high_mean': 0.06634244974702597, 'clip_ratio/high_max': 0.06634244974702597, 'clip_ratio/region_mean': 0.09117503929883242, 'reward_total_mean': 0.43320900201797485, 'reward_meter_mean': 0.9747487306594849, 'reward_meter_std': 0.01338912732899189, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.953041136264801, 'reward_repeat_soft_std': 0.03769413009285927, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.013887306675314903, 'reward_total_composite_mean': 0.43320900201797485, 'reward_total_composite_std': 0.011402982287108898, 'epoch': 0.2}

 59%|█████▉    | 1954/3300 [4:48:52<4:27:12, 11.91s/it]INFO 04-13 12:09:50 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1955/3300 [4:48:58<3:52:44, 10.38s/it]2026-04-13 12:09:56,728 | INFO | train_grpo_train | metrics_logged mode=train step=1955

                                                       
{'loss': 0.0469, 'grad_norm': 11.980443000793457, 'learning_rate': 4.07878787878788e-06, 'num_tokens': 3514807.0, 'completions/mean_length': 56.375, 'completions/min_length': 48.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.375, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.84230637550354, 'rewards/meter/std': 0.32354119420051575, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9703394174575806, 'rewards/repeat_soft/std': 0.024043342098593712, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.4507935643196106, 'rewards/total_composite/std': 0.07669742405414581, 'reward': 0.4507935643196106, 'reward_std': 0.07669741660356522, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13434912264347076, 'sampling/sampling_logp_difference/max': 1.508545160293579, 'sampling/importance_sampling_ratio/min': 0.221231609582901, 'sampling/importance_sampling_ratio/mean': 1.013060450553894, 'sampling/importance_sampling_ratio/max': 1.9136455059051514, 'entropy': 0.8411755263805389, 'clip_ratio/low_mean': 0.09615141851827502, 'clip_ratio/low_min': 0.09615141851827502, 'clip_ratio/high_mean': 0.015625, 'clip_ratio/high_max': 0.015625, 'clip_ratio/region_mean': 0.11177641851827502, 'reward_total_mean': 0.4507935643196106, 'reward_meter_mean': 0.84230637550354, 'reward_meter_std': 0.32354119420051575, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9703394174575806, 'reward_repeat_soft_std': 0.024043342098593712, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.4507935643196106, 'reward_total_composite_std': 0.07669742405414581, 'epoch': 0.2}

 59%|█████▉    | 1955/3300 [4:48:58<3:52:44, 10.38s/it]INFO 04-13 12:09:56 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1956/3300 [4:49:05<3:24:37,  9.14s/it]2026-04-13 12:10:02,947 | INFO | train_grpo_train | metrics_logged mode=train step=1956

                                                       
{'loss': -0.0532, 'grad_norm': 15.185262680053711, 'learning_rate': 4.075757575757576e-06, 'num_tokens': 3516375.0, 'completions/mean_length': 50.0, 'completions/min_length': 38.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.950300931930542, 'rewards/meter/std': 0.053850725293159485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9755334854125977, 'rewards/repeat_soft/std': 0.022816307842731476, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43898436427116394, 'rewards/total_composite/std': 0.005826047156006098, 'reward': 0.43898436427116394, 'reward_std': 0.0058260527439415455, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1201016753911972, 'sampling/sampling_logp_difference/max': 1.6800508499145508, 'sampling/importance_sampling_ratio/min': 0.18636450171470642, 'sampling/importance_sampling_ratio/mean': 1.0156158208847046, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8416338339447975, 'clip_ratio/low_mean': 0.04278420936316252, 'clip_ratio/low_min': 0.04278420936316252, 'clip_ratio/high_mean': 0.0748578580096364, 'clip_ratio/high_max': 0.0748578580096364, 'clip_ratio/region_mean': 0.11764206737279892, 'reward_total_mean': 0.43898436427116394, 'reward_meter_mean': 0.950300931930542, 'reward_meter_std': 0.053850725293159485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9755334854125977, 'reward_repeat_soft_std': 0.022816307842731476, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43898436427116394, 'reward_total_composite_std': 0.005826047156006098, 'epoch': 0.2}

 59%|█████▉    | 1956/3300 [4:49:05<3:24:37,  9.14s/it]INFO 04-13 12:10:03 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1957/3300 [4:49:12<3:15:57,  8.75s/it]2026-04-13 12:10:10,813 | INFO | train_grpo_train | metrics_logged mode=train step=1957

                                                       
{'loss': -0.0039, 'grad_norm': 4.864860534667969, 'learning_rate': 4.072727272727273e-06, 'num_tokens': 3519219.0, 'completions/mean_length': 159.5, 'completions/min_length': 84.0, 'completions/max_length': 189.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 159.5, 'completions/min_terminated_length': 84.0, 'completions/max_terminated_length': 189.0, 'rewards/meter/mean': 0.8495699167251587, 'rewards/meter/std': 0.2617471218109131, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.7231045961380005, 'rewards/repeat_soft/std': 0.10974279046058655, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.031052954494953156, 'rewards/total_composite/mean': 0.3280775547027588, 'rewards/total_composite/std': 0.14272281527519226, 'reward': 0.3280775547027588, 'reward_std': 0.14272280037403107, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07233616709709167, 'sampling/sampling_logp_difference/max': 1.684156894683838, 'sampling/importance_sampling_ratio/min': 0.18560084700584412, 'sampling/importance_sampling_ratio/mean': 1.010395884513855, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.45209553465247154, 'clip_ratio/low_mean': 0.016155288321897388, 'clip_ratio/low_min': 0.016155288321897388, 'clip_ratio/high_mean': 0.05875499313697219, 'clip_ratio/high_max': 0.05875499313697219, 'clip_ratio/region_mean': 0.07491028145886958, 'reward_total_mean': 0.3280775547027588, 'reward_meter_mean': 0.8495699167251587, 'reward_meter_std': 0.2617471218109131, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.7231045961380005, 'reward_repeat_soft_std': 0.10974279046058655, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.031052954494953156, 'reward_total_composite_mean': 0.3280775547027588, 'reward_total_composite_std': 0.14272281527519226, 'epoch': 0.2}

 59%|█████▉    | 1957/3300 [4:49:13<3:15:57,  8.75s/it]INFO 04-13 12:10:11 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1958/3300 [4:49:19<2:59:00,  8.00s/it]2026-04-13 12:10:17,070 | INFO | train_grpo_train | metrics_logged mode=train step=1958

                                                       
{'loss': 0.0012, 'grad_norm': 13.044492721557617, 'learning_rate': 4.0696969696969706e-06, 'num_tokens': 3520729.0, 'completions/mean_length': 34.75, 'completions/min_length': 31.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.8093085289001465, 'rewards/meter/std': 0.3347797989845276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9042980074882507, 'rewards/repeat_soft/std': 0.0484519861638546, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4145523011684418, 'rewards/total_composite/std': 0.03301135078072548, 'reward': 0.4145523011684418, 'reward_std': 0.03301134705543518, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09243471175432205, 'sampling/sampling_logp_difference/max': 2.3862814903259277, 'sampling/importance_sampling_ratio/min': 0.09197104722261429, 'sampling/importance_sampling_ratio/mean': 0.9922285676002502, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46040232852101326, 'clip_ratio/low_mean': 0.018939394503831863, 'clip_ratio/low_min': 0.018939394503831863, 'clip_ratio/high_mean': 0.0745106372050941, 'clip_ratio/high_max': 0.0745106372050941, 'clip_ratio/region_mean': 0.09345003170892596, 'reward_total_mean': 0.4145523011684418, 'reward_meter_mean': 0.8093085289001465, 'reward_meter_std': 0.3347797989845276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9042980074882507, 'reward_repeat_soft_std': 0.0484519861638546, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4145523011684418, 'reward_total_composite_std': 0.03301135078072548, 'epoch': 0.2}

 59%|█████▉    | 1958/3300 [4:49:19<2:59:00,  8.00s/it]INFO 04-13 12:10:17 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1959/3300 [4:49:25<2:46:16,  7.44s/it]2026-04-13 12:10:23,187 | INFO | train_grpo_train | metrics_logged mode=train step=1959

                                                       
{'loss': -0.0112, 'grad_norm': 19.16962432861328, 'learning_rate': 4.066666666666667e-06, 'num_tokens': 3522277.0, 'completions/mean_length': 36.5, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.5, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8629001379013062, 'rewards/meter/std': 0.27387601137161255, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9339948892593384, 'rewards/repeat_soft/std': 0.0631670132279396, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.43170130252838135, 'rewards/total_composite/std': 0.027718383818864822, 'reward': 0.43170130252838135, 'reward_std': 0.02771839313209057, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08232957869768143, 'sampling/sampling_logp_difference/max': 2.5419797897338867, 'sampling/importance_sampling_ratio/min': 0.0787104144692421, 'sampling/importance_sampling_ratio/mean': 0.9993715882301331, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34769029915332794, 'clip_ratio/low_mean': 0.018506493885070086, 'clip_ratio/low_min': 0.018506493885070086, 'clip_ratio/high_mean': 0.049481700640171766, 'clip_ratio/high_max': 0.049481700640171766, 'clip_ratio/region_mean': 0.06798819452524185, 'reward_total_mean': 0.43170130252838135, 'reward_meter_mean': 0.8629001379013062, 'reward_meter_std': 0.27387601137161255, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9339948892593384, 'reward_repeat_soft_std': 0.0631670132279396, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.43170130252838135, 'reward_total_composite_std': 0.027718383818864822, 'epoch': 0.2}

 59%|█████▉    | 1959/3300 [4:49:25<2:46:16,  7.44s/it]INFO 04-13 12:10:23 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1960/3300 [4:49:31<2:37:00,  7.03s/it]2026-04-13 12:10:29,267 | INFO | train_grpo_train | metrics_logged mode=train step=1960

                                                       
{'loss': -0.0229, 'grad_norm': 18.765748977661133, 'learning_rate': 4.063636363636364e-06, 'num_tokens': 3523659.0, 'completions/mean_length': 24.75, 'completions/min_length': 18.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.6937103271484375, 'rewards/meter/std': 0.4257435202598572, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9576764702796936, 'rewards/repeat_soft/std': 0.013642913661897182, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.41128823161125183, 'rewards/total_composite/std': 0.04105198010802269, 'reward': 0.41128823161125183, 'reward_std': 0.04105199873447418, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16380755603313446, 'sampling/sampling_logp_difference/max': 2.570831775665283, 'sampling/importance_sampling_ratio/min': 0.07647190988063812, 'sampling/importance_sampling_ratio/mean': 1.0118906497955322, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.050104521214962, 'clip_ratio/low_mean': 0.06570512987673283, 'clip_ratio/low_min': 0.06570512987673283, 'clip_ratio/high_mean': 0.09601280698552728, 'clip_ratio/high_max': 0.09601280698552728, 'clip_ratio/region_mean': 0.1617179368622601, 'reward_total_mean': 0.41128823161125183, 'reward_meter_mean': 0.6937103271484375, 'reward_meter_std': 0.4257435202598572, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9576764702796936, 'reward_repeat_soft_std': 0.013642913661897182, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.41128823161125183, 'reward_total_composite_std': 0.04105198010802269, 'epoch': 0.2}

 59%|█████▉    | 1960/3300 [4:49:31<2:37:00,  7.03s/it]INFO 04-13 12:10:29 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1961/3300 [4:49:37<2:29:45,  6.71s/it]2026-04-13 12:10:35,218 | INFO | train_grpo_train | metrics_logged mode=train step=1961

                                                       
{'loss': 0.0007, 'grad_norm': 29.277746200561523, 'learning_rate': 4.060606060606061e-06, 'num_tokens': 3524995.0, 'completions/mean_length': 16.0, 'completions/min_length': 15.0, 'completions/max_length': 18.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 18.0, 'rewards/meter/mean': 0.7733647227287292, 'rewards/meter/std': 0.34330764412879944, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.24625001847743988, 'rewards/judge_quality/std': 0.2722361087799072, 'rewards/total_composite/mean': 0.4203203320503235, 'rewards/total_composite/std': 0.032099027186632156, 'reward': 0.4203203320503235, 'reward_std': 0.03209902346134186, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11788924038410187, 'sampling/sampling_logp_difference/max': 1.3102974891662598, 'sampling/importance_sampling_ratio/min': 0.2697398066520691, 'sampling/importance_sampling_ratio/mean': 1.0119284391403198, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6030541844666004, 'clip_ratio/low_mean': 0.023958333767950535, 'clip_ratio/low_min': 0.023958333767950535, 'clip_ratio/high_mean': 0.09982638899236917, 'clip_ratio/high_max': 0.09982638899236917, 'clip_ratio/region_mean': 0.12378472276031971, 'reward_total_mean': 0.4203203320503235, 'reward_meter_mean': 0.7733647227287292, 'reward_meter_std': 0.34330764412879944, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.24625001847743988, 'reward_judge_quality_std': 0.2722361087799072, 'reward_total_composite_mean': 0.4203203320503235, 'reward_total_composite_std': 0.032099027186632156, 'epoch': 0.2}

 59%|█████▉    | 1961/3300 [4:49:37<2:29:45,  6.71s/it]INFO 04-13 12:10:35 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1962/3300 [4:49:44<2:32:16,  6.83s/it]2026-04-13 12:10:42,325 | INFO | train_grpo_train | metrics_logged mode=train step=1962

                                                       
{'loss': 0.0365, 'grad_norm': 8.924705505371094, 'learning_rate': 4.057575757575758e-06, 'num_tokens': 3526657.0, 'completions/mean_length': 44.75, 'completions/min_length': 31.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.75, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9242417216300964, 'rewards/meter/std': 0.13994993269443512, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8894854784011841, 'rewards/repeat_soft/std': 0.014142140746116638, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4259297251701355, 'rewards/total_composite/std': 0.015334603376686573, 'reward': 0.4259297251701355, 'reward_std': 0.015334603376686573, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09322226792573929, 'sampling/sampling_logp_difference/max': 1.4590911865234375, 'sampling/importance_sampling_ratio/min': 0.23244743049144745, 'sampling/importance_sampling_ratio/mean': 0.9982884526252747, 'sampling/importance_sampling_ratio/max': 1.669023871421814, 'entropy': 0.5456208884716034, 'clip_ratio/low_mean': 0.010130719048902392, 'clip_ratio/low_min': 0.010130719048902392, 'clip_ratio/high_mean': 0.08445008657872677, 'clip_ratio/high_max': 0.08445008657872677, 'clip_ratio/region_mean': 0.09458080562762916, 'reward_total_mean': 0.4259297251701355, 'reward_meter_mean': 0.9242417216300964, 'reward_meter_std': 0.13994993269443512, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8894854784011841, 'reward_repeat_soft_std': 0.014142140746116638, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4259297251701355, 'reward_total_composite_std': 0.015334603376686573, 'epoch': 0.2}

 59%|█████▉    | 1962/3300 [4:49:44<2:32:16,  6.83s/it]INFO 04-13 12:10:42 [block_pool.py:378] Successfully reset prefix cache

 59%|█████▉    | 1963/3300 [4:49:51<2:32:48,  6.86s/it]2026-04-13 12:10:49,249 | INFO | train_grpo_train | metrics_logged mode=train step=1963

                                                       
{'loss': -0.0048, 'grad_norm': 10.650731086730957, 'learning_rate': 4.054545454545455e-06, 'num_tokens': 3528243.0, 'completions/mean_length': 38.25, 'completions/min_length': 31.0, 'completions/max_length': 41.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.25, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.9466809034347534, 'rewards/meter/std': 0.0385647751390934, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9052823781967163, 'rewards/repeat_soft/std': 0.03730171173810959, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.42809373140335083, 'rewards/total_composite/std': 0.007216813508421183, 'reward': 0.42809373140335083, 'reward_std': 0.007216822821646929, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.057822395116090775, 'sampling/sampling_logp_difference/max': 0.8509960174560547, 'sampling/importance_sampling_ratio/min': 0.42698943614959717, 'sampling/importance_sampling_ratio/mean': 1.0015625953674316, 'sampling/importance_sampling_ratio/max': 1.9589515924453735, 'entropy': 0.2994825579226017, 'clip_ratio/low_mean': 0.019611074589192867, 'clip_ratio/low_min': 0.019611074589192867, 'clip_ratio/high_mean': 0.045088978949934244, 'clip_ratio/high_max': 0.045088978949934244, 'clip_ratio/region_mean': 0.06470005353912711, 'reward_total_mean': 0.42809373140335083, 'reward_meter_mean': 0.9466809034347534, 'reward_meter_std': 0.0385647751390934, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9052823781967163, 'reward_repeat_soft_std': 0.03730171173810959, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.42809373140335083, 'reward_total_composite_std': 0.007216813508421183, 'epoch': 0.2}

 59%|█████▉    | 1963/3300 [4:49:51<2:32:48,  6.86s/it]INFO 04-13 12:10:49 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1964/3300 [4:49:58<2:34:48,  6.95s/it]2026-04-13 12:10:56,448 | INFO | train_grpo_train | metrics_logged mode=train step=1964

                                                       
{'loss': -0.0893, 'grad_norm': 8.149408340454102, 'learning_rate': 4.0515151515151516e-06, 'num_tokens': 3530286.0, 'completions/mean_length': 91.375, 'completions/min_length': 68.0, 'completions/max_length': 107.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 91.375, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 107.0, 'rewards/meter/mean': 0.8493034243583679, 'rewards/meter/std': 0.26503878831863403, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8664325475692749, 'rewards/repeat_soft/std': 0.04042574390769005, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.42457038164138794, 'rewards/total_composite/std': 0.028597533702850342, 'reward': 0.42457038164138794, 'reward_std': 0.02859753742814064, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0866965651512146, 'sampling/sampling_logp_difference/max': 2.093294620513916, 'sampling/importance_sampling_ratio/min': 0.12328030169010162, 'sampling/importance_sampling_ratio/mean': 1.0044177770614624, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5302143059670925, 'clip_ratio/low_mean': 0.02480624755844474, 'clip_ratio/low_min': 0.02480624755844474, 'clip_ratio/high_mean': 0.052668992429971695, 'clip_ratio/high_max': 0.052668992429971695, 'clip_ratio/region_mean': 0.07747523998841643, 'reward_total_mean': 0.42457038164138794, 'reward_meter_mean': 0.8493034243583679, 'reward_meter_std': 0.26503878831863403, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8664325475692749, 'reward_repeat_soft_std': 0.04042574390769005, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.42457038164138794, 'reward_total_composite_std': 0.028597533702850342, 'epoch': 0.2}

 60%|█████▉    | 1964/3300 [4:49:58<2:34:48,  6.95s/it]INFO 04-13 12:10:56 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1965/3300 [4:50:04<2:31:17,  6.80s/it]2026-04-13 12:11:02,876 | INFO | train_grpo_train | metrics_logged mode=train step=1965

                                                       
{'loss': -0.0035, 'grad_norm': 11.396963119506836, 'learning_rate': 4.048484848484849e-06, 'num_tokens': 3532160.0, 'completions/mean_length': 53.25, 'completions/min_length': 42.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.25, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9767234921455383, 'rewards/meter/std': 0.008832106366753578, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9145652055740356, 'rewards/repeat_soft/std': 0.039254020899534225, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.01603567786514759, 'rewards/total_composite/mean': 0.43235844373703003, 'rewards/total_composite/std': 0.009502322413027287, 'reward': 0.43235844373703003, 'reward_std': 0.00950232520699501, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10968401283025742, 'sampling/sampling_logp_difference/max': 1.363420009613037, 'sampling/importance_sampling_ratio/min': 0.2557845115661621, 'sampling/importance_sampling_ratio/mean': 1.0007675886154175, 'sampling/importance_sampling_ratio/max': 1.8909069299697876, 'entropy': 0.6357946395874023, 'clip_ratio/low_mean': 0.04791590222157538, 'clip_ratio/low_min': 0.04791590222157538, 'clip_ratio/high_mean': 0.05404769070446491, 'clip_ratio/high_max': 0.05404769070446491, 'clip_ratio/region_mean': 0.10196359292604029, 'reward_total_mean': 0.43235844373703003, 'reward_meter_mean': 0.9767234921455383, 'reward_meter_std': 0.008832106366753578, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9145652055740356, 'reward_repeat_soft_std': 0.039254020899534225, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.01603567786514759, 'reward_total_composite_mean': 0.43235844373703003, 'reward_total_composite_std': 0.009502322413027287, 'epoch': 0.2}

 60%|█████▉    | 1965/3300 [4:50:05<2:31:17,  6.80s/it]INFO 04-13 12:11:03 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1966/3300 [4:50:11<2:27:14,  6.62s/it]2026-04-13 12:11:09,083 | INFO | train_grpo_train | metrics_logged mode=train step=1966

                                                       
{'loss': 0.0788, 'grad_norm': 18.476335525512695, 'learning_rate': 4.045454545454546e-06, 'num_tokens': 3533534.0, 'completions/mean_length': 24.75, 'completions/min_length': 18.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 24.75, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8567436933517456, 'rewards/meter/std': 0.16110274195671082, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9362522959709167, 'rewards/repeat_soft/std': 0.043951354920864105, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4239703416824341, 'rewards/total_composite/std': 0.019093535840511322, 'reward': 0.4239703416824341, 'reward_std': 0.019093530252575874, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14432452619075775, 'sampling/sampling_logp_difference/max': 1.0622243881225586, 'sampling/importance_sampling_ratio/min': 0.34568601846694946, 'sampling/importance_sampling_ratio/mean': 1.0156381130218506, 'sampling/importance_sampling_ratio/max': 1.820920467376709, 'entropy': 1.0608180686831474, 'clip_ratio/low_mean': 0.05883590504527092, 'clip_ratio/low_min': 0.05883590504527092, 'clip_ratio/high_mean': 0.10791426338255405, 'clip_ratio/high_max': 0.10791426338255405, 'clip_ratio/region_mean': 0.16675016842782497, 'reward_total_mean': 0.4239703416824341, 'reward_meter_mean': 0.8567436933517456, 'reward_meter_std': 0.16110274195671082, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9362522959709167, 'reward_repeat_soft_std': 0.043951354920864105, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4239703416824341, 'reward_total_composite_std': 0.019093535840511322, 'epoch': 0.2}

 60%|█████▉    | 1966/3300 [4:50:11<2:27:14,  6.62s/it]INFO 04-13 12:11:09 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1967/3300 [4:50:17<2:25:58,  6.57s/it]2026-04-13 12:11:15,538 | INFO | train_grpo_train | metrics_logged mode=train step=1967

                                                       
{'loss': -0.0541, 'grad_norm': 7.662782192230225, 'learning_rate': 4.0424242424242425e-06, 'num_tokens': 3535182.0, 'completions/mean_length': 50.0, 'completions/min_length': 40.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 50.0, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.906194806098938, 'rewards/meter/std': 0.13652828335762024, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.757510244846344, 'rewards/repeat_soft/std': 0.11741093546152115, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.013887306675314903, 'rewards/total_composite/mean': 0.39824041724205017, 'rewards/total_composite/std': 0.032019857317209244, 'reward': 0.39824041724205017, 'reward_std': 0.032019857317209244, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07870042324066162, 'sampling/sampling_logp_difference/max': 1.5747723579406738, 'sampling/importance_sampling_ratio/min': 0.20705467462539673, 'sampling/importance_sampling_ratio/mean': 0.9969842433929443, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.32496722787618637, 'clip_ratio/low_mean': 0.00937500037252903, 'clip_ratio/low_min': 0.00937500037252903, 'clip_ratio/high_mean': 0.06114195752888918, 'clip_ratio/high_max': 0.06114195752888918, 'clip_ratio/region_mean': 0.07051695790141821, 'reward_total_mean': 0.39824041724205017, 'reward_meter_mean': 0.906194806098938, 'reward_meter_std': 0.13652828335762024, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.757510244846344, 'reward_repeat_soft_std': 0.11741093546152115, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.013887306675314903, 'reward_total_composite_mean': 0.39824041724205017, 'reward_total_composite_std': 0.032019857317209244, 'epoch': 0.2}

 60%|█████▉    | 1967/3300 [4:50:17<2:25:58,  6.57s/it]INFO 04-13 12:11:15 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1968/3300 [4:50:29<3:01:46,  8.19s/it]2026-04-13 12:11:27,494 | INFO | train_grpo_train | metrics_logged mode=train step=1968

                                                       
{'loss': -0.0975, 'grad_norm': 2.6602158546447754, 'learning_rate': 4.03939393939394e-06, 'num_tokens': 3536667.0, 'completions/mean_length': 99.625, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 40.71428680419922, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9412317276000977, 'rewards/meter/std': 0.09254694730043411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9318528771400452, 'rewards/repeat_soft/std': 0.0765707939863205, 'rewards/judge_quality/mean': 0.22875000536441803, 'rewards/judge_quality/std': 0.1470119059085846, 'rewards/total_composite/mean': 0.43736425042152405, 'rewards/total_composite/std': 0.19943781197071075, 'reward': 0.43736425042152405, 'reward_std': 0.19943781197071075, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12045356631278992, 'sampling/sampling_logp_difference/max': 1.456504225730896, 'sampling/importance_sampling_ratio/min': 0.23304954171180725, 'sampling/importance_sampling_ratio/mean': 0.9907183051109314, 'sampling/importance_sampling_ratio/max': 1.699324369430542, 'entropy': 0.7500576339662075, 'clip_ratio/low_mean': 0.029835973866283894, 'clip_ratio/low_min': 0.029835973866283894, 'clip_ratio/high_mean': 0.06798899755813181, 'clip_ratio/high_max': 0.06798899755813181, 'clip_ratio/region_mean': 0.09782497142441571, 'reward_total_mean': 0.43736425042152405, 'reward_meter_mean': 0.9412317276000977, 'reward_meter_std': 0.09254694730043411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9318528771400452, 'reward_repeat_soft_std': 0.0765707939863205, 'reward_judge_quality_mean': 0.22875000536441803, 'reward_judge_quality_std': 0.1470119059085846, 'reward_total_composite_mean': 0.43736425042152405, 'reward_total_composite_std': 0.19943781197071075, 'epoch': 0.2}

 60%|█████▉    | 1968/3300 [4:50:29<3:01:46,  8.19s/it]INFO 04-13 12:11:27 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1969/3300 [4:50:37<2:57:03,  7.98s/it]2026-04-13 12:11:34,999 | INFO | train_grpo_train | metrics_logged mode=train step=1969

                                                       
{'loss': -0.0055, 'grad_norm': 9.625309944152832, 'learning_rate': 4.036363636363637e-06, 'num_tokens': 3538634.0, 'completions/mean_length': 79.875, 'completions/min_length': 67.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 79.875, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.845735490322113, 'rewards/meter/std': 0.24396738409996033, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8271574974060059, 'rewards/repeat_soft/std': 0.04683251306414604, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.022320717573165894, 'rewards/total_composite/mean': 0.4140516519546509, 'rewards/total_composite/std': 0.035415295511484146, 'reward': 0.4140516519546509, 'reward_std': 0.03541528806090355, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10792321711778641, 'sampling/sampling_logp_difference/max': 1.4553050994873047, 'sampling/importance_sampling_ratio/min': 0.2333291620016098, 'sampling/importance_sampling_ratio/mean': 1.0013035535812378, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6829820536077023, 'clip_ratio/low_mean': 0.03421615529805422, 'clip_ratio/low_min': 0.03421615529805422, 'clip_ratio/high_mean': 0.05739721702411771, 'clip_ratio/high_max': 0.05739721702411771, 'clip_ratio/region_mean': 0.09161337232217193, 'reward_total_mean': 0.4140516519546509, 'reward_meter_mean': 0.845735490322113, 'reward_meter_std': 0.24396738409996033, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8271574974060059, 'reward_repeat_soft_std': 0.04683251306414604, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.022320717573165894, 'reward_total_composite_mean': 0.4140516519546509, 'reward_total_composite_std': 0.035415295511484146, 'epoch': 0.2}

 60%|█████▉    | 1969/3300 [4:50:37<2:57:03,  7.98s/it]INFO 04-13 12:11:35 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1970/3300 [4:50:44<2:51:47,  7.75s/it]2026-04-13 12:11:42,204 | INFO | train_grpo_train | metrics_logged mode=train step=1970

                                                       
{'loss': 0.0563, 'grad_norm': 11.966935157775879, 'learning_rate': 4.033333333333333e-06, 'num_tokens': 3540319.0, 'completions/mean_length': 57.625, 'completions/min_length': 44.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.625, 'completions/min_terminated_length': 44.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.8777778148651123, 'rewards/meter/std': 0.2484196424484253, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9263529777526855, 'rewards/repeat_soft/std': 0.05342525616288185, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4245362877845764, 'rewards/total_composite/std': 0.022866318002343178, 'reward': 0.4245362877845764, 'reward_std': 0.02286631427705288, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1278618425130844, 'sampling/sampling_logp_difference/max': 1.6122455596923828, 'sampling/importance_sampling_ratio/min': 0.19943925738334656, 'sampling/importance_sampling_ratio/mean': 1.0226764678955078, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9058720171451569, 'clip_ratio/low_mean': 0.06446496304124594, 'clip_ratio/low_min': 0.06446496304124594, 'clip_ratio/high_mean': 0.0880760820582509, 'clip_ratio/high_max': 0.0880760820582509, 'clip_ratio/region_mean': 0.15254104509949684, 'reward_total_mean': 0.4245362877845764, 'reward_meter_mean': 0.8777778148651123, 'reward_meter_std': 0.2484196424484253, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9263529777526855, 'reward_repeat_soft_std': 0.05342525616288185, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4245362877845764, 'reward_total_composite_std': 0.022866318002343178, 'epoch': 0.2}

 60%|█████▉    | 1970/3300 [4:50:44<2:51:47,  7.75s/it]INFO 04-13 12:11:42 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1971/3300 [4:50:50<2:42:28,  7.34s/it]2026-04-13 12:11:48,577 | INFO | train_grpo_train | metrics_logged mode=train step=1971

                                                       
{'loss': -0.059, 'grad_norm': 9.930120468139648, 'learning_rate': 4.030303030303031e-06, 'num_tokens': 3542050.0, 'completions/mean_length': 56.375, 'completions/min_length': 43.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.375, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.9330288171768188, 'rewards/meter/std': 0.15199299156665802, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9713982939720154, 'rewards/repeat_soft/std': 0.0330844447016716, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4404628872871399, 'rewards/total_composite/std': 0.013864600099623203, 'reward': 0.4404628872871399, 'reward_std': 0.01386459730565548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10060733556747437, 'sampling/sampling_logp_difference/max': 0.9856727123260498, 'sampling/importance_sampling_ratio/min': 0.3731880784034729, 'sampling/importance_sampling_ratio/mean': 1.014480471611023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6457661688327789, 'clip_ratio/low_mean': 0.04207662586122751, 'clip_ratio/low_min': 0.04207662586122751, 'clip_ratio/high_mean': 0.0700115580111742, 'clip_ratio/high_max': 0.0700115580111742, 'clip_ratio/region_mean': 0.11208818387240171, 'reward_total_mean': 0.4404628872871399, 'reward_meter_mean': 0.9330288171768188, 'reward_meter_std': 0.15199299156665802, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9713982939720154, 'reward_repeat_soft_std': 0.0330844447016716, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4404628872871399, 'reward_total_composite_std': 0.013864600099623203, 'epoch': 0.2}

 60%|█████▉    | 1971/3300 [4:50:50<2:42:28,  7.34s/it]INFO 04-13 12:11:48 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1972/3300 [4:50:57<2:37:55,  7.14s/it]2026-04-13 12:11:55,231 | INFO | train_grpo_train | metrics_logged mode=train step=1972

                                                       
{'loss': -0.0236, 'grad_norm': 13.1968412399292, 'learning_rate': 4.027272727272727e-06, 'num_tokens': 3543993.0, 'completions/mean_length': 61.875, 'completions/min_length': 56.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 61.875, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.7542418241500854, 'rewards/meter/std': 0.39151984453201294, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9484682083129883, 'rewards/repeat_soft/std': 0.04555134102702141, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.41601330041885376, 'rewards/total_composite/std': 0.03888684883713722, 'reward': 0.41601330041885376, 'reward_std': 0.03888684883713722, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11010038107633591, 'sampling/sampling_logp_difference/max': 2.005039691925049, 'sampling/importance_sampling_ratio/min': 0.1346549540758133, 'sampling/importance_sampling_ratio/mean': 1.0076484680175781, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6559870019555092, 'clip_ratio/low_mean': 0.01963529083877802, 'clip_ratio/low_min': 0.01963529083877802, 'clip_ratio/high_mean': 0.07518272940069437, 'clip_ratio/high_max': 0.07518272940069437, 'clip_ratio/region_mean': 0.09481802023947239, 'reward_total_mean': 0.41601330041885376, 'reward_meter_mean': 0.7542418241500854, 'reward_meter_std': 0.39151984453201294, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9484682083129883, 'reward_repeat_soft_std': 0.04555134102702141, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.41601330041885376, 'reward_total_composite_std': 0.03888684883713722, 'epoch': 0.2}

 60%|█████▉    | 1972/3300 [4:50:57<2:37:55,  7.14s/it]INFO 04-13 12:11:55 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1973/3300 [4:51:03<2:31:15,  6.84s/it]2026-04-13 12:12:01,382 | INFO | train_grpo_train | metrics_logged mode=train step=1973

                                                       
{'loss': -0.0072, 'grad_norm': 11.306916236877441, 'learning_rate': 4.024242424242424e-06, 'num_tokens': 3545615.0, 'completions/mean_length': 42.75, 'completions/min_length': 38.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.75, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.8588011860847473, 'rewards/meter/std': 0.21996483206748962, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.926804780960083, 'rewards/repeat_soft/std': 0.06274452805519104, 'rewards/judge_quality/mean': 0.3037499785423279, 'rewards/judge_quality/std': 0.15665589272975922, 'rewards/total_composite/mean': 0.5050768852233887, 'rewards/total_composite/std': 0.1065869852900505, 'reward': 0.5050768852233887, 'reward_std': 0.10658696293830872, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09504842758178711, 'sampling/sampling_logp_difference/max': 1.140629768371582, 'sampling/importance_sampling_ratio/min': 0.3196176588535309, 'sampling/importance_sampling_ratio/mean': 1.0198540687561035, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6223641932010651, 'clip_ratio/low_mean': 0.05050806328654289, 'clip_ratio/low_min': 0.05050806328654289, 'clip_ratio/high_mean': 0.025493421126157045, 'clip_ratio/high_max': 0.025493421126157045, 'clip_ratio/region_mean': 0.07600148441269994, 'reward_total_mean': 0.5050768852233887, 'reward_meter_mean': 0.8588011860847473, 'reward_meter_std': 0.21996483206748962, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.926804780960083, 'reward_repeat_soft_std': 0.06274452805519104, 'reward_judge_quality_mean': 0.3037499785423279, 'reward_judge_quality_std': 0.15665589272975922, 'reward_total_composite_mean': 0.5050768852233887, 'reward_total_composite_std': 0.1065869852900505, 'epoch': 0.2}

 60%|█████▉    | 1973/3300 [4:51:03<2:31:15,  6.84s/it]INFO 04-13 12:12:01 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1974/3300 [4:51:10<2:28:59,  6.74s/it]2026-04-13 12:12:07,900 | INFO | train_grpo_train | metrics_logged mode=train step=1974

                                                       
{'loss': 0.0579, 'grad_norm': 10.476274490356445, 'learning_rate': 4.0212121212121216e-06, 'num_tokens': 3547106.0, 'completions/mean_length': 39.375, 'completions/min_length': 34.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.375, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9804747104644775, 'rewards/meter/std': 0.002423355123028159, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9483873844146729, 'rewards/repeat_soft/std': 0.027986101806163788, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4378544092178345, 'rewards/total_composite/std': 0.004306635819375515, 'reward': 0.4378544092178345, 'reward_std': 0.0043066395446658134, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12802255153656006, 'sampling/sampling_logp_difference/max': 2.76456618309021, 'sampling/importance_sampling_ratio/min': 0.06300342828035355, 'sampling/importance_sampling_ratio/mean': 0.9863287210464478, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7553790509700775, 'clip_ratio/low_mean': 0.02991346619091928, 'clip_ratio/low_min': 0.02991346619091928, 'clip_ratio/high_mean': 0.08429330121725798, 'clip_ratio/high_max': 0.08429330121725798, 'clip_ratio/region_mean': 0.11420676740817726, 'reward_total_mean': 0.4378544092178345, 'reward_meter_mean': 0.9804747104644775, 'reward_meter_std': 0.002423355123028159, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9483873844146729, 'reward_repeat_soft_std': 0.027986101806163788, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4378544092178345, 'reward_total_composite_std': 0.004306635819375515, 'epoch': 0.2}

 60%|█████▉    | 1974/3300 [4:51:10<2:28:59,  6.74s/it]INFO 04-13 12:12:08 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1975/3300 [4:51:16<2:30:24,  6.81s/it]2026-04-13 12:12:14,888 | INFO | train_grpo_train | metrics_logged mode=train step=1975

                                                       
{'loss': 0.0127, 'grad_norm': 7.559450626373291, 'learning_rate': 4.018181818181818e-06, 'num_tokens': 3548907.0, 'completions/mean_length': 55.125, 'completions/min_length': 45.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.977043092250824, 'rewards/meter/std': 0.012994077056646347, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9380739331245422, 'rewards/repeat_soft/std': 0.04333008453249931, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.4601700007915497, 'rewards/total_composite/std': 0.07168570160865784, 'reward': 0.4601700007915497, 'reward_std': 0.07168568670749664, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09898661077022552, 'sampling/sampling_logp_difference/max': 1.4066228866577148, 'sampling/importance_sampling_ratio/min': 0.2449691742658615, 'sampling/importance_sampling_ratio/mean': 1.0076396465301514, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6857492104172707, 'clip_ratio/low_mean': 0.08906186185777187, 'clip_ratio/low_min': 0.08906186185777187, 'clip_ratio/high_mean': 0.010964912362396717, 'clip_ratio/high_max': 0.010964912362396717, 'clip_ratio/region_mean': 0.10002677422016859, 'reward_total_mean': 0.4601700007915497, 'reward_meter_mean': 0.977043092250824, 'reward_meter_std': 0.012994077056646347, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9380739331245422, 'reward_repeat_soft_std': 0.04333008453249931, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.4601700007915497, 'reward_total_composite_std': 0.07168570160865784, 'epoch': 0.2}

 60%|█████▉    | 1975/3300 [4:51:17<2:30:24,  6.81s/it]INFO 04-13 12:12:15 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1976/3300 [4:51:24<2:32:42,  6.92s/it]2026-04-13 12:12:22,045 | INFO | train_grpo_train | metrics_logged mode=train step=1976

                                                       
{'loss': -0.0137, 'grad_norm': 8.489429473876953, 'learning_rate': 4.015151515151515e-06, 'num_tokens': 3550904.0, 'completions/mean_length': 76.625, 'completions/min_length': 67.0, 'completions/max_length': 81.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.625, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 81.0, 'rewards/meter/mean': 0.741714358329773, 'rewards/meter/std': 0.2703160345554352, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7749097347259521, 'rewards/repeat_soft/std': 0.061277952045202255, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.022677872329950333, 'rewards/total_composite/mean': 0.39811843633651733, 'rewards/total_composite/std': 0.03516698628664017, 'reward': 0.39811843633651733, 'reward_std': 0.035166990011930466, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0874904990196228, 'sampling/sampling_logp_difference/max': 1.5693798065185547, 'sampling/importance_sampling_ratio/min': 0.20817424356937408, 'sampling/importance_sampling_ratio/mean': 0.9921249747276306, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39452505111694336, 'clip_ratio/low_mean': 0.042162060271948576, 'clip_ratio/low_min': 0.042162060271948576, 'clip_ratio/high_mean': 0.03482632851228118, 'clip_ratio/high_max': 0.03482632851228118, 'clip_ratio/region_mean': 0.07698838878422976, 'reward_total_mean': 0.39811843633651733, 'reward_meter_mean': 0.741714358329773, 'reward_meter_std': 0.2703160345554352, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7749097347259521, 'reward_repeat_soft_std': 0.061277952045202255, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.022677872329950333, 'reward_total_composite_mean': 0.39811843633651733, 'reward_total_composite_std': 0.03516698628664017, 'epoch': 0.2}

 60%|█████▉    | 1976/3300 [4:51:24<2:32:42,  6.92s/it]INFO 04-13 12:12:22 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1977/3300 [4:51:32<2:38:42,  7.20s/it]2026-04-13 12:12:29,897 | INFO | train_grpo_train | metrics_logged mode=train step=1977

                                                       
{'loss': 0.0824, 'grad_norm': 6.707657814025879, 'learning_rate': 4.0121212121212125e-06, 'num_tokens': 3553438.0, 'completions/mean_length': 138.75, 'completions/min_length': 123.0, 'completions/max_length': 163.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.75, 'completions/min_terminated_length': 123.0, 'completions/max_terminated_length': 163.0, 'rewards/meter/mean': 0.8465820550918579, 'rewards/meter/std': 0.2426604926586151, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7107216119766235, 'rewards/repeat_soft/std': 0.09084786474704742, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.40393969416618347, 'rewards/total_composite/std': 0.042848143726587296, 'reward': 0.40393969416618347, 'reward_std': 0.042848147451877594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06716953217983246, 'sampling/sampling_logp_difference/max': 1.9756016731262207, 'sampling/importance_sampling_ratio/min': 0.13867785036563873, 'sampling/importance_sampling_ratio/mean': 1.003207802772522, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.34012457728385925, 'clip_ratio/low_mean': 0.017128449864685535, 'clip_ratio/low_min': 0.017128449864685535, 'clip_ratio/high_mean': 0.04925884958356619, 'clip_ratio/high_max': 0.04925884958356619, 'clip_ratio/region_mean': 0.06638729944825172, 'reward_total_mean': 0.40393969416618347, 'reward_meter_mean': 0.8465820550918579, 'reward_meter_std': 0.2426604926586151, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7107216119766235, 'reward_repeat_soft_std': 0.09084786474704742, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.40393969416618347, 'reward_total_composite_std': 0.042848143726587296, 'epoch': 0.2}

 60%|█████▉    | 1977/3300 [4:51:32<2:38:42,  7.20s/it]INFO 04-13 12:12:30 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1978/3300 [4:51:38<2:33:20,  6.96s/it]2026-04-13 12:12:36,291 | INFO | train_grpo_train | metrics_logged mode=train step=1978

                                                       
{'loss': 0.0069, 'grad_norm': 11.788993835449219, 'learning_rate': 4.009090909090909e-06, 'num_tokens': 3555132.0, 'completions/mean_length': 51.75, 'completions/min_length': 46.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.6127851605415344, 'rewards/meter/std': 0.410106360912323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8013951778411865, 'rewards/repeat_soft/std': 0.07675564289093018, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.022677872329950333, 'rewards/total_composite/mean': 0.3426607847213745, 'rewards/total_composite/std': 0.14449571073055267, 'reward': 0.3426607847213745, 'reward_std': 0.14449572563171387, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11311350017786026, 'sampling/sampling_logp_difference/max': 1.820281982421875, 'sampling/importance_sampling_ratio/min': 0.1619800627231598, 'sampling/importance_sampling_ratio/mean': 1.0193040370941162, 'sampling/importance_sampling_ratio/max': 1.9086806774139404, 'entropy': 0.7015680745244026, 'clip_ratio/low_mean': 0.04053589981049299, 'clip_ratio/low_min': 0.04053589981049299, 'clip_ratio/high_mean': 0.06337560061365366, 'clip_ratio/high_max': 0.06337560061365366, 'clip_ratio/region_mean': 0.10391150042414665, 'reward_total_mean': 0.3426607847213745, 'reward_meter_mean': 0.6127851605415344, 'reward_meter_std': 0.410106360912323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8013951778411865, 'reward_repeat_soft_std': 0.07675564289093018, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.022677872329950333, 'reward_total_composite_mean': 0.3426607847213745, 'reward_total_composite_std': 0.14449571073055267, 'epoch': 0.2}

 60%|█████▉    | 1978/3300 [4:51:38<2:33:20,  6.96s/it]INFO 04-13 12:12:36 [block_pool.py:378] Successfully reset prefix cache

 60%|█████▉    | 1979/3300 [4:51:44<2:27:46,  6.71s/it]2026-04-13 12:12:42,432 | INFO | train_grpo_train | metrics_logged mode=train step=1979

                                                       
{'loss': 0.005, 'grad_norm': 15.920906066894531, 'learning_rate': 4.006060606060607e-06, 'num_tokens': 3556622.0, 'completions/mean_length': 42.25, 'completions/min_length': 32.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.25, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.6911098957061768, 'rewards/meter/std': 0.3353486657142639, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9723049402236938, 'rewards/repeat_soft/std': 0.025577707216143608, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.41322892904281616, 'rewards/total_composite/std': 0.032315321266651154, 'reward': 0.41322892904281616, 'reward_std': 0.03231532871723175, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11365252733230591, 'sampling/sampling_logp_difference/max': 1.0233488082885742, 'sampling/importance_sampling_ratio/min': 0.35938942432403564, 'sampling/importance_sampling_ratio/mean': 1.0163404941558838, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6912884451448917, 'clip_ratio/low_mean': 0.04192102886736393, 'clip_ratio/low_min': 0.04192102886736393, 'clip_ratio/high_mean': 0.06938481144607067, 'clip_ratio/high_max': 0.06938481144607067, 'clip_ratio/region_mean': 0.1113058403134346, 'reward_total_mean': 0.41322892904281616, 'reward_meter_mean': 0.6911098957061768, 'reward_meter_std': 0.3353486657142639, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9723049402236938, 'reward_repeat_soft_std': 0.025577707216143608, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.41322892904281616, 'reward_total_composite_std': 0.032315321266651154, 'epoch': 0.2}

 60%|█████▉    | 1979/3300 [4:51:44<2:27:46,  6.71s/it]INFO 04-13 12:12:42 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1980/3300 [4:51:51<2:30:08,  6.82s/it]2026-04-13 12:12:49,524 | INFO | train_grpo_train | metrics_logged mode=train step=1980

                                                       
{'loss': 0.0048, 'grad_norm': 12.009119987487793, 'learning_rate': 4.003030303030303e-06, 'num_tokens': 3558327.0, 'completions/mean_length': 58.125, 'completions/min_length': 50.0, 'completions/max_length': 67.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.125, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 67.0, 'rewards/meter/mean': 0.7157129049301147, 'rewards/meter/std': 0.31780123710632324, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9366830587387085, 'rewards/repeat_soft/std': 0.05577899515628815, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.41028445959091187, 'rewards/total_composite/std': 0.024794554337859154, 'reward': 0.41028445959091187, 'reward_std': 0.02479456178843975, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12345890700817108, 'sampling/sampling_logp_difference/max': 1.8352665901184082, 'sampling/importance_sampling_ratio/min': 0.15957096219062805, 'sampling/importance_sampling_ratio/mean': 0.987436830997467, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6226517632603645, 'clip_ratio/low_mean': 0.05017365072853863, 'clip_ratio/low_min': 0.05017365072853863, 'clip_ratio/high_mean': 0.07431062683463097, 'clip_ratio/high_max': 0.07431062683463097, 'clip_ratio/region_mean': 0.1244842775631696, 'reward_total_mean': 0.41028445959091187, 'reward_meter_mean': 0.7157129049301147, 'reward_meter_std': 0.31780123710632324, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9366830587387085, 'reward_repeat_soft_std': 0.05577899515628815, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.41028445959091187, 'reward_total_composite_std': 0.024794554337859154, 'epoch': 0.2}

 60%|██████    | 1980/3300 [4:51:51<2:30:08,  6.82s/it]INFO 04-13 12:12:49 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1981/3300 [4:51:58<2:31:27,  6.89s/it]2026-04-13 12:12:56,565 | INFO | train_grpo_train | metrics_logged mode=train step=1981

                                                       
{'loss': 0.0349, 'grad_norm': 8.917847633361816, 'learning_rate': 4.000000000000001e-06, 'num_tokens': 3560098.0, 'completions/mean_length': 53.375, 'completions/min_length': 41.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.375, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.7698092460632324, 'rewards/meter/std': 0.38456812500953674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8634123206138611, 'rewards/repeat_soft/std': 0.07371919602155685, 'rewards/judge_quality/mean': 0.19500000774860382, 'rewards/judge_quality/std': 0.10392305254936218, 'rewards/total_composite/mean': 0.40846335887908936, 'rewards/total_composite/std': 0.04037882387638092, 'reward': 0.40846335887908936, 'reward_std': 0.04037882387638092, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1242368221282959, 'sampling/sampling_logp_difference/max': 1.6267108917236328, 'sampling/importance_sampling_ratio/min': 0.1965750753879547, 'sampling/importance_sampling_ratio/mean': 1.0145431756973267, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0416819676756859, 'clip_ratio/low_mean': 0.025384615175426006, 'clip_ratio/low_min': 0.025384615175426006, 'clip_ratio/high_mean': 0.09255419671535492, 'clip_ratio/high_max': 0.09255419671535492, 'clip_ratio/region_mean': 0.11793881189078093, 'reward_total_mean': 0.40846335887908936, 'reward_meter_mean': 0.7698092460632324, 'reward_meter_std': 0.38456812500953674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8634123206138611, 'reward_repeat_soft_std': 0.07371919602155685, 'reward_judge_quality_mean': 0.19500000774860382, 'reward_judge_quality_std': 0.10392305254936218, 'reward_total_composite_mean': 0.40846335887908936, 'reward_total_composite_std': 0.04037882387638092, 'epoch': 0.2}

 60%|██████    | 1981/3300 [4:51:58<2:31:27,  6.89s/it]INFO 04-13 12:12:56 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1982/3300 [4:52:05<2:31:19,  6.89s/it]2026-04-13 12:13:03,476 | INFO | train_grpo_train | metrics_logged mode=train step=1982

                                                       
{'loss': 0.0564, 'grad_norm': 16.385469436645508, 'learning_rate': 3.996969696969698e-06, 'num_tokens': 3561620.0, 'completions/mean_length': 28.25, 'completions/min_length': 26.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.25, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.3600154519081116, 'rewards/meter/std': 0.30427849292755127, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.783305287361145, 'rewards/repeat_soft/std': 0.05267034471035004, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.35498178005218506, 'rewards/total_composite/std': 0.03257571905851364, 'reward': 0.35498178005218506, 'reward_std': 0.03257571533322334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11256179958581924, 'sampling/sampling_logp_difference/max': 1.1291608810424805, 'sampling/importance_sampling_ratio/min': 0.3233044445514679, 'sampling/importance_sampling_ratio/mean': 1.009274959564209, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5077512469142675, 'clip_ratio/low_mean': 0.051971325650811195, 'clip_ratio/low_min': 0.051971325650811195, 'clip_ratio/high_mean': 0.07279956713318825, 'clip_ratio/high_max': 0.07279956713318825, 'clip_ratio/region_mean': 0.12477089278399944, 'reward_total_mean': 0.35498178005218506, 'reward_meter_mean': 0.3600154519081116, 'reward_meter_std': 0.30427849292755127, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.783305287361145, 'reward_repeat_soft_std': 0.05267034471035004, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.35498178005218506, 'reward_total_composite_std': 0.03257571905851364, 'epoch': 0.2}

 60%|██████    | 1982/3300 [4:52:05<2:31:19,  6.89s/it]INFO 04-13 12:13:03 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1983/3300 [4:52:17<3:04:06,  8.39s/it]2026-04-13 12:13:15,334 | INFO | train_grpo_train | metrics_logged mode=train step=1983

                                                       
{'loss': -0.2168, 'grad_norm': 1.40635347366333, 'learning_rate': 3.993939393939394e-06, 'num_tokens': 3563997.0, 'completions/mean_length': 163.125, 'completions/min_length': 91.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.28572082519531, 'completions/min_terminated_length': 91.0, 'completions/max_terminated_length': 133.0, 'rewards/meter/mean': 0.82413649559021, 'rewards/meter/std': 0.3011797070503235, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.21380899846553802, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8315623998641968, 'rewards/repeat_soft/std': 0.061598025262355804, 'rewards/judge_quality/mean': 0.16375000774860382, 'rewards/judge_quality/std': 0.04596194624900818, 'rewards/total_composite/mean': 0.37225809693336487, 'rewards/total_composite/std': 0.1516929715871811, 'reward': 0.37225809693336487, 'reward_std': 0.1516929715871811, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07436981052160263, 'sampling/sampling_logp_difference/max': 2.8483803272247314, 'sampling/importance_sampling_ratio/min': 0.05793808400630951, 'sampling/importance_sampling_ratio/mean': 1.0074925422668457, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.39619259908795357, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06025759922340512, 'clip_ratio/high_max': 0.06025759922340512, 'clip_ratio/region_mean': 0.06025759922340512, 'reward_total_mean': 0.37225809693336487, 'reward_meter_mean': 0.82413649559021, 'reward_meter_std': 0.3011797070503235, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.21380899846553802, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8315623998641968, 'reward_repeat_soft_std': 0.061598025262355804, 'reward_judge_quality_mean': 0.16375000774860382, 'reward_judge_quality_std': 0.04596194624900818, 'reward_total_composite_mean': 0.37225809693336487, 'reward_total_composite_std': 0.1516929715871811, 'epoch': 0.2}

 60%|██████    | 1983/3300 [4:52:17<3:04:06,  8.39s/it]INFO 04-13 12:13:15 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1984/3300 [4:52:24<2:56:37,  8.05s/it]2026-04-13 12:13:22,600 | INFO | train_grpo_train | metrics_logged mode=train step=1984

                                                       
{'loss': -0.0194, 'grad_norm': 10.250814437866211, 'learning_rate': 3.990909090909092e-06, 'num_tokens': 3565772.0, 'completions/mean_length': 46.875, 'completions/min_length': 40.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.8747069835662842, 'rewards/meter/std': 0.2825613021850586, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.975144624710083, 'rewards/repeat_soft/std': 0.02640417404472828, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4315556287765503, 'rewards/total_composite/std': 0.027250397950410843, 'reward': 0.4315556287765503, 'reward_std': 0.02725040167570114, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13391144573688507, 'sampling/sampling_logp_difference/max': 2.1814565658569336, 'sampling/importance_sampling_ratio/min': 0.1128769963979721, 'sampling/importance_sampling_ratio/mean': 1.0205755233764648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8451790362596512, 'clip_ratio/low_mean': 0.022619048599153757, 'clip_ratio/low_min': 0.022619048599153757, 'clip_ratio/high_mean': 0.0787371383048594, 'clip_ratio/high_max': 0.0787371383048594, 'clip_ratio/region_mean': 0.10135618690401316, 'reward_total_mean': 0.4315556287765503, 'reward_meter_mean': 0.8747069835662842, 'reward_meter_std': 0.2825613021850586, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.975144624710083, 'reward_repeat_soft_std': 0.02640417404472828, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4315556287765503, 'reward_total_composite_std': 0.027250397950410843, 'epoch': 0.2}

 60%|██████    | 1984/3300 [4:52:24<2:56:37,  8.05s/it]INFO 04-13 12:13:22 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1985/3300 [4:52:31<2:47:41,  7.65s/it]2026-04-13 12:13:29,315 | INFO | train_grpo_train | metrics_logged mode=train step=1985

                                                       
{'loss': 0.0396, 'grad_norm': 8.761999130249023, 'learning_rate': 3.987878787878788e-06, 'num_tokens': 3567525.0, 'completions/mean_length': 57.125, 'completions/min_length': 36.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.125, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.8532387018203735, 'rewards/meter/std': 0.30601271986961365, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9405505657196045, 'rewards/repeat_soft/std': 0.04098353907465935, 'rewards/judge_quality/mean': 0.26125001907348633, 'rewards/judge_quality/std': 0.2665889263153076, 'rewards/total_composite/mean': 0.4926952123641968, 'rewards/total_composite/std': 0.1742052137851715, 'reward': 0.4926952123641968, 'reward_std': 0.1742052137851715, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10535774379968643, 'sampling/sampling_logp_difference/max': 1.3276968002319336, 'sampling/importance_sampling_ratio/min': 0.2650870978832245, 'sampling/importance_sampling_ratio/mean': 0.9956843256950378, 'sampling/importance_sampling_ratio/max': 1.956322431564331, 'entropy': 0.6120221167802811, 'clip_ratio/low_mean': 0.046632946934551, 'clip_ratio/low_min': 0.046632946934551, 'clip_ratio/high_mean': 0.014423076994717121, 'clip_ratio/high_max': 0.014423076994717121, 'clip_ratio/region_mean': 0.06105602392926812, 'reward_total_mean': 0.4926952123641968, 'reward_meter_mean': 0.8532387018203735, 'reward_meter_std': 0.30601271986961365, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9405505657196045, 'reward_repeat_soft_std': 0.04098353907465935, 'reward_judge_quality_mean': 0.26125001907348633, 'reward_judge_quality_std': 0.2665889263153076, 'reward_total_composite_mean': 0.4926952123641968, 'reward_total_composite_std': 0.1742052137851715, 'epoch': 0.2}

 60%|██████    | 1985/3300 [4:52:31<2:47:41,  7.65s/it]INFO 04-13 12:13:29 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1986/3300 [4:52:38<2:40:30,  7.33s/it]2026-04-13 12:13:35,901 | INFO | train_grpo_train | metrics_logged mode=train step=1986

                                                       
{'loss': 0.0726, 'grad_norm': 7.136348724365234, 'learning_rate': 3.984848484848485e-06, 'num_tokens': 3569292.0, 'completions/mean_length': 46.875, 'completions/min_length': 39.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.875, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.9778975248336792, 'rewards/meter/std': 0.011393287219107151, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9715107679367065, 'rewards/repeat_soft/std': 0.03862692788243294, 'rewards/judge_quality/mean': 0.20250001549720764, 'rewards/judge_quality/std': 0.10110107809305191, 'rewards/total_composite/mean': 0.47444748878479004, 'rewards/total_composite/std': 0.06477969139814377, 'reward': 0.47444748878479004, 'reward_std': 0.06477969139814377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12436336278915405, 'sampling/sampling_logp_difference/max': 2.2677464485168457, 'sampling/importance_sampling_ratio/min': 0.10354526340961456, 'sampling/importance_sampling_ratio/mean': 1.0028719902038574, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5972497388720512, 'clip_ratio/low_mean': 0.11246980540454388, 'clip_ratio/low_min': 0.11246980540454388, 'clip_ratio/high_mean': 0.01923076994717121, 'clip_ratio/high_max': 0.01923076994717121, 'clip_ratio/region_mean': 0.1317005753517151, 'reward_total_mean': 0.47444748878479004, 'reward_meter_mean': 0.9778975248336792, 'reward_meter_std': 0.011393287219107151, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9715107679367065, 'reward_repeat_soft_std': 0.03862692788243294, 'reward_judge_quality_mean': 0.20250001549720764, 'reward_judge_quality_std': 0.10110107809305191, 'reward_total_composite_mean': 0.47444748878479004, 'reward_total_composite_std': 0.06477969139814377, 'epoch': 0.2}

 60%|██████    | 1986/3300 [4:52:38<2:40:30,  7.33s/it]INFO 04-13 12:13:36 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1987/3300 [4:52:44<2:34:20,  7.05s/it]2026-04-13 12:13:42,308 | INFO | train_grpo_train | metrics_logged mode=train step=1987

                                                       
{'loss': 0.0538, 'grad_norm': 11.674086570739746, 'learning_rate': 3.9818181818181825e-06, 'num_tokens': 3571052.0, 'completions/mean_length': 57.0, 'completions/min_length': 48.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 57.0, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.8795936107635498, 'rewards/meter/std': 0.15953429043293, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9084420204162598, 'rewards/repeat_soft/std': 0.044956889003515244, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.42602890729904175, 'rewards/total_composite/std': 0.015605825930833817, 'reward': 0.42602890729904175, 'reward_std': 0.015605824068188667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09734667837619781, 'sampling/sampling_logp_difference/max': 2.8360369205474854, 'sampling/importance_sampling_ratio/min': 0.05865767225623131, 'sampling/importance_sampling_ratio/mean': 1.0042951107025146, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5530225820839405, 'clip_ratio/low_mean': 0.06306716450490057, 'clip_ratio/low_min': 0.06306716450490057, 'clip_ratio/high_mean': 0.04177061468362808, 'clip_ratio/high_max': 0.04177061468362808, 'clip_ratio/region_mean': 0.10483777918852866, 'reward_total_mean': 0.42602890729904175, 'reward_meter_mean': 0.8795936107635498, 'reward_meter_std': 0.15953429043293, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9084420204162598, 'reward_repeat_soft_std': 0.044956889003515244, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.42602890729904175, 'reward_total_composite_std': 0.015605825930833817, 'epoch': 0.2}

 60%|██████    | 1987/3300 [4:52:44<2:34:20,  7.05s/it]INFO 04-13 12:13:42 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1988/3300 [4:52:51<2:37:35,  7.21s/it]2026-04-13 12:13:49,877 | INFO | train_grpo_train | metrics_logged mode=train step=1988

                                                       
{'loss': -0.0229, 'grad_norm': 5.305408477783203, 'learning_rate': 3.978787878787879e-06, 'num_tokens': 3573749.0, 'completions/mean_length': 139.125, 'completions/min_length': 118.0, 'completions/max_length': 155.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 139.125, 'completions/min_terminated_length': 118.0, 'completions/max_terminated_length': 155.0, 'rewards/meter/mean': 0.9411613941192627, 'rewards/meter/std': 0.06411188840866089, 'rewards/count_adherence/mean': 0.9750000238418579, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7919429540634155, 'rewards/repeat_soft/std': 0.04169183969497681, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.4193919599056244, 'rewards/total_composite/std': 0.02630588412284851, 'reward': 0.4193919599056244, 'reward_std': 0.02630588412284851, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07875581830739975, 'sampling/sampling_logp_difference/max': 2.4715664386749268, 'sampling/importance_sampling_ratio/min': 0.08445246517658234, 'sampling/importance_sampling_ratio/mean': 1.006574273109436, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48092228919267654, 'clip_ratio/low_mean': 0.016903753392398357, 'clip_ratio/low_min': 0.016903753392398357, 'clip_ratio/high_mean': 0.06237533036619425, 'clip_ratio/high_max': 0.06237533036619425, 'clip_ratio/region_mean': 0.0792790837585926, 'reward_total_mean': 0.4193919599056244, 'reward_meter_mean': 0.9411613941192627, 'reward_meter_std': 0.06411188840866089, 'reward_count_adherence_mean': 0.9750000238418579, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7919429540634155, 'reward_repeat_soft_std': 0.04169183969497681, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.4193919599056244, 'reward_total_composite_std': 0.02630588412284851, 'epoch': 0.2}

 60%|██████    | 1988/3300 [4:52:52<2:37:35,  7.21s/it]INFO 04-13 12:13:50 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1989/3300 [4:52:58<2:36:04,  7.14s/it]2026-04-13 12:13:56,866 | INFO | train_grpo_train | metrics_logged mode=train step=1989

                                                       
{'loss': -0.0049, 'grad_norm': 6.995320796966553, 'learning_rate': 3.975757575757576e-06, 'num_tokens': 3575843.0, 'completions/mean_length': 84.75, 'completions/min_length': 69.0, 'completions/max_length': 111.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.75, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 111.0, 'rewards/meter/mean': 0.899039626121521, 'rewards/meter/std': 0.14101403951644897, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.815096378326416, 'rewards/repeat_soft/std': 0.057892054319381714, 'rewards/judge_quality/mean': 0.1612500101327896, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4156867563724518, 'rewards/total_composite/std': 0.017065096646547318, 'reward': 0.4156867563724518, 'reward_std': 0.017065098509192467, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11169660091400146, 'sampling/sampling_logp_difference/max': 1.1490683555603027, 'sampling/importance_sampling_ratio/min': 0.31693190336227417, 'sampling/importance_sampling_ratio/mean': 1.0052859783172607, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.726189449429512, 'clip_ratio/low_mean': 0.03251077188178897, 'clip_ratio/low_min': 0.03251077188178897, 'clip_ratio/high_mean': 0.07103743311017752, 'clip_ratio/high_max': 0.07103743311017752, 'clip_ratio/region_mean': 0.10354820499196649, 'reward_total_mean': 0.4156867563724518, 'reward_meter_mean': 0.899039626121521, 'reward_meter_std': 0.14101403951644897, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.815096378326416, 'reward_repeat_soft_std': 0.057892054319381714, 'reward_judge_quality_mean': 0.1612500101327896, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4156867563724518, 'reward_total_composite_std': 0.017065096646547318, 'epoch': 0.2}

 60%|██████    | 1989/3300 [4:52:59<2:36:04,  7.14s/it]INFO 04-13 12:13:57 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1990/3300 [4:53:10<3:03:44,  8.42s/it]2026-04-13 12:14:08,268 | INFO | train_grpo_train | metrics_logged mode=train step=1990

                                                       
{'loss': -0.21, 'grad_norm': 1.948974370956421, 'learning_rate': 3.972727272727273e-06, 'num_tokens': 3577916.0, 'completions/mean_length': 209.125, 'completions/min_length': 87.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 108.16667175292969, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 124.0, 'rewards/meter/mean': 0.6804710626602173, 'rewards/meter/std': 0.34284526109695435, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9123061895370483, 'rewards/repeat_soft/std': 0.0672207847237587, 'rewards/judge_quality/mean': 0.1287499964237213, 'rewards/judge_quality/std': 0.05462535098195076, 'rewards/total_composite/mean': 0.30792689323425293, 'rewards/total_composite/std': 0.19347305595874786, 'reward': 0.30792689323425293, 'reward_std': 0.19347302615642548, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11115031689405441, 'sampling/sampling_logp_difference/max': 1.3480873107910156, 'sampling/importance_sampling_ratio/min': 0.259736567735672, 'sampling/importance_sampling_ratio/mean': 1.0218712091445923, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6053452044725418, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07028499897569418, 'clip_ratio/high_max': 0.07028499897569418, 'clip_ratio/region_mean': 0.07028499897569418, 'reward_total_mean': 0.30792689323425293, 'reward_meter_mean': 0.6804710626602173, 'reward_meter_std': 0.34284526109695435, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9123061895370483, 'reward_repeat_soft_std': 0.0672207847237587, 'reward_judge_quality_mean': 0.1287499964237213, 'reward_judge_quality_std': 0.05462535098195076, 'reward_total_composite_mean': 0.30792689323425293, 'reward_total_composite_std': 0.19347305595874786, 'epoch': 0.2}

 60%|██████    | 1990/3300 [4:53:10<3:03:44,  8.42s/it]INFO 04-13 12:14:08 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1991/3300 [4:53:17<2:54:01,  7.98s/it]2026-04-13 12:14:15,226 | INFO | train_grpo_train | metrics_logged mode=train step=1991

                                                       
{'loss': -0.0089, 'grad_norm': 10.88731575012207, 'learning_rate': 3.96969696969697e-06, 'num_tokens': 3579557.0, 'completions/mean_length': 45.125, 'completions/min_length': 40.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.125, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.9664255380630493, 'rewards/meter/std': 0.015298202633857727, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.955539345741272, 'rewards/repeat_soft/std': 0.03646053746342659, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4375573992729187, 'rewards/total_composite/std': 0.005865698214620352, 'reward': 0.4375573992729187, 'reward_std': 0.005865701008588076, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16127146780490875, 'sampling/sampling_logp_difference/max': 1.6951549053192139, 'sampling/importance_sampling_ratio/min': 0.18357078731060028, 'sampling/importance_sampling_ratio/mean': 1.020660400390625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.9290966093540192, 'clip_ratio/low_mean': 0.06766532361507416, 'clip_ratio/low_min': 0.06766532361507416, 'clip_ratio/high_mean': 0.08113236539065838, 'clip_ratio/high_max': 0.08113236539065838, 'clip_ratio/region_mean': 0.14879768900573254, 'reward_total_mean': 0.4375573992729187, 'reward_meter_mean': 0.9664255380630493, 'reward_meter_std': 0.015298202633857727, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.955539345741272, 'reward_repeat_soft_std': 0.03646053746342659, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4375573992729187, 'reward_total_composite_std': 0.005865698214620352, 'epoch': 0.2}

 60%|██████    | 1991/3300 [4:53:17<2:54:01,  7.98s/it]INFO 04-13 12:14:15 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1992/3300 [4:53:24<2:49:28,  7.77s/it]2026-04-13 12:14:22,510 | INFO | train_grpo_train | metrics_logged mode=train step=1992

                                                       
{'loss': 0.0962, 'grad_norm': 8.176782608032227, 'learning_rate': 3.966666666666667e-06, 'num_tokens': 3581503.0, 'completions/mean_length': 80.25, 'completions/min_length': 66.0, 'completions/max_length': 100.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.25, 'completions/min_terminated_length': 66.0, 'completions/max_terminated_length': 100.0, 'rewards/meter/mean': 0.9668633937835693, 'rewards/meter/std': 0.058630939573049545, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7522540092468262, 'rewards/repeat_soft/std': 0.046631213277578354, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.4156605899333954, 'rewards/total_composite/std': 0.037274960428476334, 'reward': 0.4156605899333954, 'reward_std': 0.037274960428476334, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08009006828069687, 'sampling/sampling_logp_difference/max': 1.8451592922210693, 'sampling/importance_sampling_ratio/min': 0.1580001562833786, 'sampling/importance_sampling_ratio/mean': 1.0078073740005493, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5122839771211147, 'clip_ratio/low_mean': 0.004999999888241291, 'clip_ratio/low_min': 0.004999999888241291, 'clip_ratio/high_mean': 0.06474924786016345, 'clip_ratio/high_max': 0.06474924786016345, 'clip_ratio/region_mean': 0.06974924774840474, 'reward_total_mean': 0.4156605899333954, 'reward_meter_mean': 0.9668633937835693, 'reward_meter_std': 0.058630939573049545, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7522540092468262, 'reward_repeat_soft_std': 0.046631213277578354, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.4156605899333954, 'reward_total_composite_std': 0.037274960428476334, 'epoch': 0.2}

 60%|██████    | 1992/3300 [4:53:24<2:49:28,  7.77s/it]INFO 04-13 12:14:22 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1993/3300 [4:53:31<2:42:11,  7.45s/it]2026-04-13 12:14:29,180 | INFO | train_grpo_train | metrics_logged mode=train step=1993

                                                       
{'loss': 0.0327, 'grad_norm': 11.174549102783203, 'learning_rate': 3.963636363636364e-06, 'num_tokens': 3583373.0, 'completions/mean_length': 66.75, 'completions/min_length': 55.0, 'completions/max_length': 78.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 66.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 78.0, 'rewards/meter/mean': 0.8860222697257996, 'rewards/meter/std': 0.2954499423503876, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9293386936187744, 'rewards/repeat_soft/std': 0.05199816823005676, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.438253790140152, 'rewards/total_composite/std': 0.03380537033081055, 'reward': 0.438253790140152, 'reward_std': 0.03380536660552025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11163672804832458, 'sampling/sampling_logp_difference/max': 2.6584410667419434, 'sampling/importance_sampling_ratio/min': 0.07005735486745834, 'sampling/importance_sampling_ratio/mean': 0.9926469922065735, 'sampling/importance_sampling_ratio/max': 1.786298155784607, 'entropy': 0.5487346686422825, 'clip_ratio/low_mean': 0.022989511489868164, 'clip_ratio/low_min': 0.022989511489868164, 'clip_ratio/high_mean': 0.08296406362205744, 'clip_ratio/high_max': 0.08296406362205744, 'clip_ratio/region_mean': 0.1059535751119256, 'reward_total_mean': 0.438253790140152, 'reward_meter_mean': 0.8860222697257996, 'reward_meter_std': 0.2954499423503876, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9293386936187744, 'reward_repeat_soft_std': 0.05199816823005676, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.438253790140152, 'reward_total_composite_std': 0.03380537033081055, 'epoch': 0.2}

 60%|██████    | 1993/3300 [4:53:31<2:42:11,  7.45s/it]INFO 04-13 12:14:29 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1994/3300 [4:53:37<2:36:59,  7.21s/it]2026-04-13 12:14:35,847 | INFO | train_grpo_train | metrics_logged mode=train step=1994

                                                       
{'loss': -0.1128, 'grad_norm': 19.808204650878906, 'learning_rate': 3.960606060606061e-06, 'num_tokens': 3584738.0, 'completions/mean_length': 18.625, 'completions/min_length': 15.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.625, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.17641128599643707, 'rewards/meter/std': 0.30511367321014404, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9604669809341431, 'rewards/repeat_soft/std': 0.005061782896518707, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.36127012968063354, 'rewards/total_composite/std': 0.029887478798627853, 'reward': 0.36127012968063354, 'reward_std': 0.029887471348047256, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13217005133628845, 'sampling/sampling_logp_difference/max': 1.6605939865112305, 'sampling/importance_sampling_ratio/min': 0.19002607464790344, 'sampling/importance_sampling_ratio/mean': 1.0037096738815308, 'sampling/importance_sampling_ratio/max': 1.9682484865188599, 'entropy': 0.8014398291707039, 'clip_ratio/low_mean': 0.10888788476586342, 'clip_ratio/low_min': 0.10888788476586342, 'clip_ratio/high_mean': 0.01923076994717121, 'clip_ratio/high_max': 0.01923076994717121, 'clip_ratio/region_mean': 0.12811865471303463, 'reward_total_mean': 0.36127012968063354, 'reward_meter_mean': 0.17641128599643707, 'reward_meter_std': 0.30511367321014404, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9604669809341431, 'reward_repeat_soft_std': 0.005061782896518707, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.36127012968063354, 'reward_total_composite_std': 0.029887478798627853, 'epoch': 0.2}

 60%|██████    | 1994/3300 [4:53:38<2:36:59,  7.21s/it]INFO 04-13 12:14:36 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1995/3300 [4:53:44<2:33:04,  7.04s/it]2026-04-13 12:14:42,487 | INFO | train_grpo_train | metrics_logged mode=train step=1995

                                                       
{'loss': -0.0115, 'grad_norm': 10.272811889648438, 'learning_rate': 3.957575757575758e-06, 'num_tokens': 3586344.0, 'completions/mean_length': 39.75, 'completions/min_length': 37.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 39.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9224235415458679, 'rewards/meter/std': 0.11386893689632416, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9877801537513733, 'rewards/repeat_soft/std': 0.024328239262104034, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4381033182144165, 'rewards/total_composite/std': 0.010988668538630009, 'reward': 0.4381033182144165, 'reward_std': 0.010988662950694561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08340685069561005, 'sampling/sampling_logp_difference/max': 2.556256055831909, 'sampling/importance_sampling_ratio/min': 0.07759470492601395, 'sampling/importance_sampling_ratio/mean': 1.008427381515503, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.31179263442754745, 'clip_ratio/low_mean': 0.016372141893953085, 'clip_ratio/low_min': 0.016372141893953085, 'clip_ratio/high_mean': 0.0549057312309742, 'clip_ratio/high_max': 0.0549057312309742, 'clip_ratio/region_mean': 0.07127787312492728, 'reward_total_mean': 0.4381033182144165, 'reward_meter_mean': 0.9224235415458679, 'reward_meter_std': 0.11386893689632416, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9877801537513733, 'reward_repeat_soft_std': 0.024328239262104034, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4381033182144165, 'reward_total_composite_std': 0.010988668538630009, 'epoch': 0.2}

 60%|██████    | 1995/3300 [4:53:44<2:33:04,  7.04s/it]INFO 04-13 12:14:42 [block_pool.py:378] Successfully reset prefix cache

 60%|██████    | 1996/3300 [4:53:56<3:02:26,  8.39s/it]2026-04-13 12:14:54,047 | INFO | train_grpo_train | metrics_logged mode=train step=1996

                                                       
{'loss': -0.1921, 'grad_norm': 1.6462302207946777, 'learning_rate': 3.954545454545454e-06, 'num_tokens': 3588464.0, 'completions/mean_length': 149.0, 'completions/min_length': 77.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 97.14286041259766, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9769530296325684, 'rewards/meter/std': 0.0088235167786479, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8973762392997742, 'rewards/repeat_soft/std': 0.033636774867773056, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.04200340807437897, 'rewards/total_composite/mean': 0.3751123547554016, 'rewards/total_composite/std': 0.15231961011886597, 'reward': 0.3751123547554016, 'reward_std': 0.15231962502002716, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09448987245559692, 'sampling/sampling_logp_difference/max': 1.3202617168426514, 'sampling/importance_sampling_ratio/min': 0.2670654058456421, 'sampling/importance_sampling_ratio/mean': 1.007238745689392, 'sampling/importance_sampling_ratio/max': 1.8536213636398315, 'entropy': 0.5517528206110001, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07715445989742875, 'clip_ratio/high_max': 0.07715445989742875, 'clip_ratio/region_mean': 0.07715445989742875, 'reward_total_mean': 0.3751123547554016, 'reward_meter_mean': 0.9769530296325684, 'reward_meter_std': 0.0088235167786479, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8973762392997742, 'reward_repeat_soft_std': 0.033636774867773056, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.04200340807437897, 'reward_total_composite_mean': 0.3751123547554016, 'reward_total_composite_std': 0.15231961011886597, 'epoch': 0.2}

 60%|██████    | 1996/3300 [4:53:56<3:02:26,  8.39s/it]INFO 04-13 12:14:54 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1997/3300 [4:54:02<2:48:17,  7.75s/it]2026-04-13 12:15:00,280 | INFO | train_grpo_train | metrics_logged mode=train step=1997

                                                       
{'loss': 0.0003, 'grad_norm': 10.457941055297852, 'learning_rate': 3.951515151515152e-06, 'num_tokens': 3590367.0, 'completions/mean_length': 51.875, 'completions/min_length': 47.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.875, 'completions/min_terminated_length': 47.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.8365335464477539, 'rewards/meter/std': 0.21539472043514252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8633731603622437, 'rewards/repeat_soft/std': 0.12307754158973694, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.01603567786514759, 'rewards/total_composite/mean': 0.36861011385917664, 'rewards/total_composite/std': 0.15087535977363586, 'reward': 0.36861011385917664, 'reward_std': 0.15087535977363586, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10408765077590942, 'sampling/sampling_logp_difference/max': 1.6148912906646729, 'sampling/importance_sampling_ratio/min': 0.19891229271888733, 'sampling/importance_sampling_ratio/mean': 1.0068790912628174, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4990379549562931, 'clip_ratio/low_mean': 0.0078125, 'clip_ratio/low_min': 0.0078125, 'clip_ratio/high_mean': 0.07741094008088112, 'clip_ratio/high_max': 0.07741094008088112, 'clip_ratio/region_mean': 0.08522344008088112, 'reward_total_mean': 0.36861011385917664, 'reward_meter_mean': 0.8365335464477539, 'reward_meter_std': 0.21539472043514252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8633731603622437, 'reward_repeat_soft_std': 0.12307754158973694, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.01603567786514759, 'reward_total_composite_mean': 0.36861011385917664, 'reward_total_composite_std': 0.15087535977363586, 'epoch': 0.2}

 61%|██████    | 1997/3300 [4:54:02<2:48:17,  7.75s/it]INFO 04-13 12:15:00 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1998/3300 [4:54:09<2:41:44,  7.45s/it]2026-04-13 12:15:07,051 | INFO | train_grpo_train | metrics_logged mode=train step=1998

                                                       
{'loss': 0.0428, 'grad_norm': 5.65458869934082, 'learning_rate': 3.948484848484849e-06, 'num_tokens': 3591879.0, 'completions/mean_length': 69.0, 'completions/min_length': 58.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.0, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.9892727136611938, 'rewards/meter/std': 0.008448584005236626, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7639752626419067, 'rewards/repeat_soft/std': 0.051463864743709564, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4158993363380432, 'rewards/total_composite/std': 0.01525058876723051, 'reward': 0.4158993363380432, 'reward_std': 0.015250589698553085, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07780607044696808, 'sampling/sampling_logp_difference/max': 1.2431490421295166, 'sampling/importance_sampling_ratio/min': 0.28847435116767883, 'sampling/importance_sampling_ratio/mean': 1.0062282085418701, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.650130495429039, 'clip_ratio/low_mean': 0.02444823505356908, 'clip_ratio/low_min': 0.02444823505356908, 'clip_ratio/high_mean': 0.04893648810684681, 'clip_ratio/high_max': 0.04893648810684681, 'clip_ratio/region_mean': 0.07338472316041589, 'reward_total_mean': 0.4158993363380432, 'reward_meter_mean': 0.9892727136611938, 'reward_meter_std': 0.008448584005236626, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7639752626419067, 'reward_repeat_soft_std': 0.051463864743709564, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4158993363380432, 'reward_total_composite_std': 0.01525058876723051, 'epoch': 0.2}

 61%|██████    | 1998/3300 [4:54:09<2:41:44,  7.45s/it]INFO 04-13 12:15:07 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 1999/3300 [4:54:16<2:38:22,  7.30s/it]2026-04-13 12:15:14,007 | INFO | train_grpo_train | metrics_logged mode=train step=1999

                                                       
{'loss': 0.0096, 'grad_norm': 5.741130352020264, 'learning_rate': 3.945454545454545e-06, 'num_tokens': 3594125.0, 'completions/mean_length': 103.75, 'completions/min_length': 87.0, 'completions/max_length': 123.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 123.0, 'rewards/meter/mean': 0.9535378813743591, 'rewards/meter/std': 0.04614168405532837, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7956213355064392, 'rewards/repeat_soft/std': 0.03852888196706772, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.41989007592201233, 'rewards/total_composite/std': 0.022876359522342682, 'reward': 0.41989007592201233, 'reward_std': 0.02287636697292328, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08122720569372177, 'sampling/sampling_logp_difference/max': 1.6818737983703613, 'sampling/importance_sampling_ratio/min': 0.18602506816387177, 'sampling/importance_sampling_ratio/mean': 1.006203055381775, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4709929823875427, 'clip_ratio/low_mean': 0.02482958883047104, 'clip_ratio/low_min': 0.02482958883047104, 'clip_ratio/high_mean': 0.06267943885177374, 'clip_ratio/high_max': 0.06267943885177374, 'clip_ratio/region_mean': 0.08750902768224478, 'reward_total_mean': 0.41989007592201233, 'reward_meter_mean': 0.9535378813743591, 'reward_meter_std': 0.04614168405532837, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7956213355064392, 'reward_repeat_soft_std': 0.03852888196706772, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.41989007592201233, 'reward_total_composite_std': 0.022876359522342682, 'epoch': 0.2}

 61%|██████    | 1999/3300 [4:54:16<2:38:22,  7.30s/it]INFO 04-13 12:15:14 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2000/3300 [4:54:22<2:32:22,  7.03s/it]2026-04-13 12:15:20,404 | INFO | train_grpo_train | metrics_logged mode=train step=2000

                                                       
{'loss': 0.0798, 'grad_norm': 16.039695739746094, 'learning_rate': 3.942424242424243e-06, 'num_tokens': 3595641.0, 'completions/mean_length': 28.5, 'completions/min_length': 23.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.5, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.9656250476837158, 'rewards/meter/std': 0.038042064756155014, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.921393871307373, 'rewards/repeat_soft/std': 0.04221487417817116, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.432357519865036, 'rewards/total_composite/std': 0.006993249524384737, 'reward': 0.432357519865036, 'reward_std': 0.0069932518526911736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09250590950250626, 'sampling/sampling_logp_difference/max': 1.1924142837524414, 'sampling/importance_sampling_ratio/min': 0.3034876883029938, 'sampling/importance_sampling_ratio/mean': 1.0109021663665771, 'sampling/importance_sampling_ratio/max': 1.5811668634414673, 'entropy': 0.5452602319419384, 'clip_ratio/low_mean': 0.03977534594014287, 'clip_ratio/low_min': 0.03977534594014287, 'clip_ratio/high_mean': 0.04168054927140474, 'clip_ratio/high_max': 0.04168054927140474, 'clip_ratio/region_mean': 0.08145589521154761, 'reward_total_mean': 0.432357519865036, 'reward_meter_mean': 0.9656250476837158, 'reward_meter_std': 0.038042064756155014, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.921393871307373, 'reward_repeat_soft_std': 0.04221487417817116, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.432357519865036, 'reward_total_composite_std': 0.006993249524384737, 'epoch': 0.2}

 61%|██████    | 2000/3300 [4:54:22<2:32:22,  7.03s/it]INFO 04-13 12:15:20 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.79s/it]

 30%|███       | 3/10 [00:08<00:21,  3.08s/it]

 40%|████      | 4/10 [00:12<00:20,  3.39s/it]

 50%|█████     | 5/10 [00:17<00:19,  3.86s/it]

 60%|██████    | 6/10 [00:26<00:22,  5.56s/it]

 70%|███████   | 7/10 [00:31<00:16,  5.40s/it]

 80%|████████  | 8/10 [00:35<00:10,  5.00s/it]

 90%|█████████ | 9/10 [00:39<00:04,  4.76s/it]

100%|██████████| 10/10 [00:42<00:00,  4.35s/it]2026-04-13 12:16:12,516 | INFO | train_grpo_train | metrics_logged mode=eval step=2000

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 51.9949, 'eval_samples_per_second': 1.539, 'eval_steps_per_second': 0.192, 'eval_num_tokens': 3595641.0, 'eval_completions/mean_length': 87.6125, 'eval_completions/min_length': 34.1, 'eval_completions/max_length': 222.1, 'eval_completions/clipped_ratio': 0.025, 'eval_completions/mean_terminated_length': 76.95892944335938, 'eval_completions/min_terminated_length': 34.1, 'eval_completions/max_terminated_length': 144.4, 'eval_rewards/meter/mean': 0.8989010989665985, 'eval_rewards/meter/std': 0.193228018283844, 'eval_rewards/count_adherence/mean': 0.9708333313465118, 'eval_rewards/count_adherence/std': 0.045636449754238126, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.8757998764514923, 'eval_rewards/repeat_soft/std': 0.08093975968658924, 'eval_rewards/judge_quality/mean': 0.16025000661611558, 'eval_rewards/judge_quality/std': 0.03135918769985437, 'eval_rewards/total_composite/mean': 0.41270904541015624, 'eval_rewards/total_composite/std': 0.05069283861666918, 'eval_reward': 0.41270904541015624, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.051789015904068945, 'eval_sampling/sampling_logp_difference/max': 1.1813788414001465, 'eval_sampling/importance_sampling_ratio/min': 0.31964881867170336, 'eval_sampling/importance_sampling_ratio/mean': 1.0109478116035462, 'eval_sampling/importance_sampling_ratio/max': 1.473687183856964, 'eval_entropy': 0.5352312296628952, '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': 0.41270904541015624, 'eval_reward_meter_mean': 0.8989010989665985, 'eval_reward_meter_std': 0.193228018283844, 'eval_reward_count_adherence_mean': 0.9708333313465118, 'eval_reward_count_adherence_std': 0.045636449754238126, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.8757998764514923, 'eval_reward_repeat_soft_std': 0.08093975968658924, 'eval_reward_judge_quality_mean': 0.16025000661611558, 'eval_reward_judge_quality_std': 0.03135918769985437, 'eval_reward_total_composite_mean': 0.41270904541015624, 'eval_reward_total_composite_std': 0.05069283861666918, 'epoch': 0.2}

 61%|██████    | 2000/3300 [4:55:14<2:32:22,  7.03s/it]

100%|██████████| 10/10 [00:43<00:00,  4.35s/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(

 61%|██████    | 2001/3300 [4:55:24<8:27:16, 23.43s/it]2026-04-13 12:16:22,092 | INFO | train_grpo_train | metrics_logged mode=train step=2001

                                                       
{'loss': 0.0855, 'grad_norm': 7.101884841918945, 'learning_rate': 3.93939393939394e-06, 'num_tokens': 3597319.0, 'completions/mean_length': 51.75, 'completions/min_length': 42.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 42.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.942493200302124, 'rewards/meter/std': 0.07095000147819519, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9551660418510437, 'rewards/repeat_soft/std': 0.013545302674174309, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.458089143037796, 'rewards/total_composite/std': 0.06420551985502243, 'reward': 0.458089143037796, 'reward_std': 0.06420551240444183, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08462902158498764, 'sampling/sampling_logp_difference/max': 1.5644030570983887, 'sampling/importance_sampling_ratio/min': 0.20921286940574646, 'sampling/importance_sampling_ratio/mean': 1.0072275400161743, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4804040603339672, 'clip_ratio/low_mean': 0.054138449020683765, 'clip_ratio/low_min': 0.054138449020683765, 'clip_ratio/high_mean': 0.011904762126505375, 'clip_ratio/high_max': 0.011904762126505375, 'clip_ratio/region_mean': 0.06604321114718914, 'reward_total_mean': 0.458089143037796, 'reward_meter_mean': 0.942493200302124, 'reward_meter_std': 0.07095000147819519, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9551660418510437, 'reward_repeat_soft_std': 0.013545302674174309, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.458089143037796, 'reward_total_composite_std': 0.06420551985502243, 'epoch': 0.2}

 61%|██████    | 2001/3300 [4:55:24<8:27:16, 23.43s/it]INFO 04-13 12:16:22 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2002/3300 [4:55:31<6:41:23, 18.55s/it]2026-04-13 12:16:29,276 | INFO | train_grpo_train | metrics_logged mode=train step=2002

                                                       
{'loss': 0.0188, 'grad_norm': 8.576906204223633, 'learning_rate': 3.936363636363636e-06, 'num_tokens': 3599182.0, 'completions/mean_length': 70.875, 'completions/min_length': 65.0, 'completions/max_length': 76.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 70.875, 'completions/min_terminated_length': 65.0, 'completions/max_terminated_length': 76.0, 'rewards/meter/mean': 0.8849501609802246, 'rewards/meter/std': 0.2143920212984085, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9152114391326904, 'rewards/repeat_soft/std': 0.05425975099205971, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.022320717573165894, 'rewards/total_composite/mean': 0.42890387773513794, 'rewards/total_composite/std': 0.02358926832675934, 'reward': 0.42890387773513794, 'reward_std': 0.02358926832675934, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09013252705335617, 'sampling/sampling_logp_difference/max': 1.0857977867126465, 'sampling/importance_sampling_ratio/min': 0.33763232827186584, 'sampling/importance_sampling_ratio/mean': 1.0102643966674805, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4903973340988159, 'clip_ratio/low_mean': 0.057537827640771866, 'clip_ratio/low_min': 0.057537827640771866, 'clip_ratio/high_mean': 0.05734657868742943, 'clip_ratio/high_max': 0.05734657868742943, 'clip_ratio/region_mean': 0.1148844063282013, 'reward_total_mean': 0.42890387773513794, 'reward_meter_mean': 0.8849501609802246, 'reward_meter_std': 0.2143920212984085, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9152114391326904, 'reward_repeat_soft_std': 0.05425975099205971, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.022320717573165894, 'reward_total_composite_mean': 0.42890387773513794, 'reward_total_composite_std': 0.02358926832675934, 'epoch': 0.2}

 61%|██████    | 2002/3300 [4:55:31<6:41:23, 18.55s/it]INFO 04-13 12:16:29 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2003/3300 [4:55:43<5:58:38, 16.59s/it]2026-04-13 12:16:41,295 | INFO | train_grpo_train | metrics_logged mode=train step=2003

                                                       
{'loss': -0.1107, 'grad_norm': 2.2421905994415283, 'learning_rate': 3.9333333333333335e-06, 'num_tokens': 3600729.0, 'completions/mean_length': 106.375, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.42857360839844, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.7859811782836914, 'rewards/meter/std': 0.34156039357185364, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8739618062973022, 'rewards/repeat_soft/std': 0.126714825630188, 'rewards/judge_quality/mean': 0.21250000596046448, 'rewards/judge_quality/std': 0.14270348846912384, 'rewards/total_composite/mean': 0.39180296659469604, 'rewards/total_composite/std': 0.17656150460243225, 'reward': 0.39180296659469604, 'reward_std': 0.17656150460243225, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06401461362838745, 'sampling/sampling_logp_difference/max': 1.0067353248596191, 'sampling/importance_sampling_ratio/min': 0.5014667510986328, 'sampling/importance_sampling_ratio/mean': 1.012511968612671, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4097742661833763, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06148374988697469, 'clip_ratio/high_max': 0.06148374988697469, 'clip_ratio/region_mean': 0.06148374988697469, 'reward_total_mean': 0.39180296659469604, 'reward_meter_mean': 0.7859811782836914, 'reward_meter_std': 0.34156039357185364, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8739618062973022, 'reward_repeat_soft_std': 0.126714825630188, 'reward_judge_quality_mean': 0.21250000596046448, 'reward_judge_quality_std': 0.14270348846912384, 'reward_total_composite_mean': 0.39180296659469604, 'reward_total_composite_std': 0.17656150460243225, 'epoch': 0.2}

 61%|██████    | 2003/3300 [4:55:43<5:58:38, 16.59s/it]INFO 04-13 12:16:41 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2004/3300 [4:55:50<4:56:19, 13.72s/it]2026-04-13 12:16:48,301 | INFO | train_grpo_train | metrics_logged mode=train step=2004

                                                       
{'loss': 0.0901, 'grad_norm': 21.47540855407715, 'learning_rate': 3.930303030303031e-06, 'num_tokens': 3602137.0, 'completions/mean_length': 30.0, 'completions/min_length': 25.0, 'completions/max_length': 39.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 25.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.9228441715240479, 'rewards/meter/std': 0.07131259888410568, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9653183221817017, 'rewards/repeat_soft/std': 0.03552493825554848, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43477505445480347, 'rewards/total_composite/std': 0.009522650390863419, 'reward': 0.43477505445480347, 'reward_std': 0.00952264852821827, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10566788166761398, 'sampling/sampling_logp_difference/max': 1.7538261413574219, 'sampling/importance_sampling_ratio/min': 0.17311032116413116, 'sampling/importance_sampling_ratio/mean': 1.012361764907837, 'sampling/importance_sampling_ratio/max': 1.6949011087417603, 'entropy': 0.5947912894189358, 'clip_ratio/low_mean': 0.051282052882015705, 'clip_ratio/low_min': 0.051282052882015705, 'clip_ratio/high_mean': 0.08320492599159479, 'clip_ratio/high_max': 0.08320492599159479, 'clip_ratio/region_mean': 0.1344869788736105, 'reward_total_mean': 0.43477505445480347, 'reward_meter_mean': 0.9228441715240479, 'reward_meter_std': 0.07131259888410568, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9653183221817017, 'reward_repeat_soft_std': 0.03552493825554848, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43477505445480347, 'reward_total_composite_std': 0.009522650390863419, 'epoch': 0.2}

 61%|██████    | 2004/3300 [4:55:50<4:56:19, 13.72s/it]INFO 04-13 12:16:48 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2005/3300 [4:55:58<4:22:41, 12.17s/it]2026-04-13 12:16:56,873 | INFO | train_grpo_train | metrics_logged mode=train step=2005

                                                       
{'loss': -0.0006, 'grad_norm': 5.459510803222656, 'learning_rate': 3.927272727272727e-06, 'num_tokens': 3604963.0, 'completions/mean_length': 166.25, 'completions/min_length': 137.0, 'completions/max_length': 204.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 166.25, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 204.0, 'rewards/meter/mean': 0.9653419852256775, 'rewards/meter/std': 0.03195981681346893, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7965164184570312, 'rewards/repeat_soft/std': 0.05410079285502434, 'rewards/judge_quality/mean': 0.14249999821186066, 'rewards/judge_quality/std': 0.031052954494953156, 'rewards/total_composite/mean': 0.4087391495704651, 'rewards/total_composite/std': 0.019513331353664398, 'reward': 0.4087391495704651, 'reward_std': 0.01951332949101925, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09489702433347702, 'sampling/sampling_logp_difference/max': 1.434617519378662, 'sampling/importance_sampling_ratio/min': 0.23820646107196808, 'sampling/importance_sampling_ratio/mean': 1.0066559314727783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6807519569993019, 'clip_ratio/low_mean': 0.057409160770475864, 'clip_ratio/low_min': 0.057409160770475864, 'clip_ratio/high_mean': 0.04812806844711304, 'clip_ratio/high_max': 0.04812806844711304, 'clip_ratio/region_mean': 0.1055372292175889, 'reward_total_mean': 0.4087391495704651, 'reward_meter_mean': 0.9653419852256775, 'reward_meter_std': 0.03195981681346893, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7965164184570312, 'reward_repeat_soft_std': 0.05410079285502434, 'reward_judge_quality_mean': 0.14249999821186066, 'reward_judge_quality_std': 0.031052954494953156, 'reward_total_composite_mean': 0.4087391495704651, 'reward_total_composite_std': 0.019513331353664398, 'epoch': 0.2}

 61%|██████    | 2005/3300 [4:55:59<4:22:41, 12.17s/it]INFO 04-13 12:16:57 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2006/3300 [4:56:06<3:55:03, 10.90s/it]2026-04-13 12:17:04,790 | INFO | train_grpo_train | metrics_logged mode=train step=2006

                                                       
{'loss': -0.022, 'grad_norm': 7.954293251037598, 'learning_rate': 3.9242424242424244e-06, 'num_tokens': 3607328.0, 'completions/mean_length': 104.625, 'completions/min_length': 92.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 104.625, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9748185873031616, 'rewards/meter/std': 0.014069481752812862, 'rewards/count_adherence/mean': 0.8250000476837158, 'rewards/count_adherence/std': 0.0707106739282608, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8468031883239746, 'rewards/repeat_soft/std': 0.06200995296239853, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.4012604355812073, 'rewards/total_composite/std': 0.027659742161631584, 'reward': 0.4012604355812073, 'reward_std': 0.027659747749567032, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08801953494548798, 'sampling/sampling_logp_difference/max': 2.205092191696167, 'sampling/importance_sampling_ratio/min': 0.11024036258459091, 'sampling/importance_sampling_ratio/mean': 1.0054447650909424, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43186598643660545, 'clip_ratio/low_mean': 0.04139109933748841, 'clip_ratio/low_min': 0.04139109933748841, 'clip_ratio/high_mean': 0.043632741551846266, 'clip_ratio/high_max': 0.043632741551846266, 'clip_ratio/region_mean': 0.08502384088933468, 'reward_total_mean': 0.4012604355812073, 'reward_meter_mean': 0.9748185873031616, 'reward_meter_std': 0.014069481752812862, 'reward_count_adherence_mean': 0.8250000476837158, 'reward_count_adherence_std': 0.0707106739282608, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8468031883239746, 'reward_repeat_soft_std': 0.06200995296239853, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.4012604355812073, 'reward_total_composite_std': 0.027659742161631584, 'epoch': 0.2}

 61%|██████    | 2006/3300 [4:56:07<3:55:03, 10.90s/it]INFO 04-13 12:17:05 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2007/3300 [4:56:14<3:35:21,  9.99s/it]2026-04-13 12:17:12,660 | INFO | train_grpo_train | metrics_logged mode=train step=2007

                                                       
{'loss': 0.0559, 'grad_norm': 6.958494186401367, 'learning_rate': 3.921212121212122e-06, 'num_tokens': 3609830.0, 'completions/mean_length': 109.75, 'completions/min_length': 87.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 109.75, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.9617154002189636, 'rewards/meter/std': 0.04410649091005325, 'rewards/count_adherence/mean': 0.800000011920929, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8770691156387329, 'rewards/repeat_soft/std': 0.04067332670092583, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.01922610029578209, 'rewards/total_composite/mean': 0.38770896196365356, 'rewards/total_composite/std': 0.015122607350349426, 'reward': 0.38770896196365356, 'reward_std': 0.015122603625059128, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09767697006464005, 'sampling/sampling_logp_difference/max': 1.8629543781280518, 'sampling/importance_sampling_ratio/min': 0.15521340072155, 'sampling/importance_sampling_ratio/mean': 1.0098620653152466, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6366865932941437, 'clip_ratio/low_mean': 0.04750001663342118, 'clip_ratio/low_min': 0.04750001663342118, 'clip_ratio/high_mean': 0.04482072498649359, 'clip_ratio/high_max': 0.04482072498649359, 'clip_ratio/region_mean': 0.09232074161991477, 'reward_total_mean': 0.38770896196365356, 'reward_meter_mean': 0.9617154002189636, 'reward_meter_std': 0.04410649091005325, 'reward_count_adherence_mean': 0.800000011920929, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8770691156387329, 'reward_repeat_soft_std': 0.04067332670092583, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.01922610029578209, 'reward_total_composite_mean': 0.38770896196365356, 'reward_total_composite_std': 0.015122607350349426, 'epoch': 0.2}

 61%|██████    | 2007/3300 [4:56:14<3:35:21,  9.99s/it]INFO 04-13 12:17:12 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2008/3300 [4:56:26<3:43:41, 10.39s/it]2026-04-13 12:17:23,984 | INFO | train_grpo_train | metrics_logged mode=train step=2008

                                                       
{'loss': -0.1108, 'grad_norm': 2.0068905353546143, 'learning_rate': 3.918181818181819e-06, 'num_tokens': 3611559.0, 'completions/mean_length': 97.125, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 37.85714340209961, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 41.0, 'rewards/meter/mean': 0.8242834210395813, 'rewards/meter/std': 0.3362806439399719, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.3535533845424652, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9566799998283386, 'rewards/repeat_soft/std': 0.03871471434831619, 'rewards/judge_quality/mean': 0.22750000655651093, 'rewards/judge_quality/std': 0.13593590259552002, 'rewards/total_composite/mean': 0.43579307198524475, 'rewards/total_composite/std': 0.18939152359962463, 'reward': 0.43579307198524475, 'reward_std': 0.18939152359962463, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14889979362487793, 'sampling/sampling_logp_difference/max': 2.3028697967529297, 'sampling/importance_sampling_ratio/min': 0.09997153282165527, 'sampling/importance_sampling_ratio/mean': 1.0114301443099976, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6406805738806725, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10330542549490929, 'clip_ratio/high_max': 0.10330542549490929, 'clip_ratio/region_mean': 0.10330542549490929, 'reward_total_mean': 0.43579307198524475, 'reward_meter_mean': 0.8242834210395813, 'reward_meter_std': 0.3362806439399719, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.3535533845424652, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9566799998283386, 'reward_repeat_soft_std': 0.03871471434831619, 'reward_judge_quality_mean': 0.22750000655651093, 'reward_judge_quality_std': 0.13593590259552002, 'reward_total_composite_mean': 0.43579307198524475, 'reward_total_composite_std': 0.18939152359962463, 'epoch': 0.2}

 61%|██████    | 2008/3300 [4:56:26<3:43:41, 10.39s/it]INFO 04-13 12:17:24 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2009/3300 [4:56:32<3:14:45,  9.05s/it]2026-04-13 12:17:29,912 | INFO | train_grpo_train | metrics_logged mode=train step=2009

                                                       
{'loss': 0.0056, 'grad_norm': 9.960348129272461, 'learning_rate': 3.915151515151515e-06, 'num_tokens': 3613223.0, 'completions/mean_length': 35.0, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.7716435194015503, 'rewards/meter/std': 0.1756240278482437, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8687934279441833, 'rewards/repeat_soft/std': 0.06876115500926971, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4074971079826355, 'rewards/total_composite/std': 0.02326873503625393, 'reward': 0.4074971079826355, 'reward_std': 0.02326873503625393, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11916100978851318, 'sampling/sampling_logp_difference/max': 1.2213225364685059, 'sampling/importance_sampling_ratio/min': 0.2948399782180786, 'sampling/importance_sampling_ratio/mean': 0.9958372116088867, 'sampling/importance_sampling_ratio/max': 1.6207143068313599, 'entropy': 0.8184601366519928, 'clip_ratio/low_mean': 0.039572364650666714, 'clip_ratio/low_min': 0.039572364650666714, 'clip_ratio/high_mean': 0.08438057219609618, 'clip_ratio/high_max': 0.08438057219609618, 'clip_ratio/region_mean': 0.1239529368467629, 'reward_total_mean': 0.4074971079826355, 'reward_meter_mean': 0.7716435194015503, 'reward_meter_std': 0.1756240278482437, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8687934279441833, 'reward_repeat_soft_std': 0.06876115500926971, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4074971079826355, 'reward_total_composite_std': 0.02326873503625393, 'epoch': 0.2}

 61%|██████    | 2009/3300 [4:56:32<3:14:45,  9.05s/it]INFO 04-13 12:17:30 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2010/3300 [4:56:39<3:02:53,  8.51s/it]2026-04-13 12:17:37,152 | INFO | train_grpo_train | metrics_logged mode=train step=2010

                                                       
{'loss': 0.0437, 'grad_norm': 6.988424301147461, 'learning_rate': 3.912121212121213e-06, 'num_tokens': 3615602.0, 'completions/mean_length': 119.375, 'completions/min_length': 105.0, 'completions/max_length': 130.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 119.375, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.9608349204063416, 'rewards/meter/std': 0.051238901913166046, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9251853823661804, 'rewards/repeat_soft/std': 0.04096551612019539, 'rewards/judge_quality/mean': 0.11999999731779099, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4137229025363922, 'rewards/total_composite/std': 0.007810365874320269, 'reward': 0.4137229025363922, 'reward_std': 0.00781036913394928, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09610017389059067, 'sampling/sampling_logp_difference/max': 2.1673481464385986, 'sampling/importance_sampling_ratio/min': 0.11448080092668533, 'sampling/importance_sampling_ratio/mean': 1.0091322660446167, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5776182152330875, 'clip_ratio/low_mean': 0.018450709991157055, 'clip_ratio/low_min': 0.018450709991157055, 'clip_ratio/high_mean': 0.0501365577802062, 'clip_ratio/high_max': 0.0501365577802062, 'clip_ratio/region_mean': 0.06858726777136326, 'reward_total_mean': 0.4137229025363922, 'reward_meter_mean': 0.9608349204063416, 'reward_meter_std': 0.051238901913166046, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9251853823661804, 'reward_repeat_soft_std': 0.04096551612019539, 'reward_judge_quality_mean': 0.11999999731779099, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4137229025363922, 'reward_total_composite_std': 0.007810365874320269, 'epoch': 0.2}

 61%|██████    | 2010/3300 [4:56:39<3:02:53,  8.51s/it]INFO 04-13 12:17:37 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2011/3300 [4:56:45<2:48:24,  7.84s/it]2026-04-13 12:17:43,433 | INFO | train_grpo_train | metrics_logged mode=train step=2011

                                                       
{'loss': 0.1012, 'grad_norm': 15.545039176940918, 'learning_rate': 3.90909090909091e-06, 'num_tokens': 3617019.0, 'completions/mean_length': 21.125, 'completions/min_length': 18.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.125, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.6375752091407776, 'rewards/meter/std': 0.41336917877197266, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40653860569000244, 'rewards/total_composite/std': 0.04030349850654602, 'reward': 0.40653860569000244, 'reward_std': 0.04030349850654602, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08633841574192047, 'sampling/sampling_logp_difference/max': 1.3101986646652222, 'sampling/importance_sampling_ratio/min': 0.2697664499282837, 'sampling/importance_sampling_ratio/mean': 0.9990162253379822, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4785655103623867, 'clip_ratio/low_mean': 0.027083334047347307, 'clip_ratio/low_min': 0.027083334047347307, 'clip_ratio/high_mean': 0.04612573143094778, 'clip_ratio/high_max': 0.04612573143094778, 'clip_ratio/region_mean': 0.07320906547829509, 'reward_total_mean': 0.40653860569000244, 'reward_meter_mean': 0.6375752091407776, 'reward_meter_std': 0.41336917877197266, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40653860569000244, 'reward_total_composite_std': 0.04030349850654602, 'epoch': 0.2}

 61%|██████    | 2011/3300 [4:56:45<2:48:24,  7.84s/it]INFO 04-13 12:17:43 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2012/3300 [4:56:52<2:41:49,  7.54s/it]2026-04-13 12:17:50,260 | INFO | train_grpo_train | metrics_logged mode=train step=2012

                                                       
{'loss': 0.0113, 'grad_norm': 8.610708236694336, 'learning_rate': 3.906060606060606e-06, 'num_tokens': 3619079.0, 'completions/mean_length': 94.5, 'completions/min_length': 69.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 94.5, 'completions/min_terminated_length': 69.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.6477097272872925, 'rewards/meter/std': 0.3077602982521057, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9543507099151611, 'rewards/repeat_soft/std': 0.0374886617064476, 'rewards/judge_quality/mean': 0.17625001072883606, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4186341166496277, 'rewards/total_composite/std': 0.035938192158937454, 'reward': 0.4186341166496277, 'reward_std': 0.035938192158937454, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.127189040184021, 'sampling/sampling_logp_difference/max': 2.5745933055877686, 'sampling/importance_sampling_ratio/min': 0.07618480175733566, 'sampling/importance_sampling_ratio/mean': 1.0038552284240723, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4742015674710274, 'clip_ratio/low_mean': 0.05968614085577428, 'clip_ratio/low_min': 0.05968614085577428, 'clip_ratio/high_mean': 0.06563483458012342, 'clip_ratio/high_max': 0.06563483458012342, 'clip_ratio/region_mean': 0.1253209754358977, 'reward_total_mean': 0.4186341166496277, 'reward_meter_mean': 0.6477097272872925, 'reward_meter_std': 0.3077602982521057, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9543507099151611, 'reward_repeat_soft_std': 0.0374886617064476, 'reward_judge_quality_mean': 0.17625001072883606, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4186341166496277, 'reward_total_composite_std': 0.035938192158937454, 'epoch': 0.2}

 61%|██████    | 2012/3300 [4:56:52<2:41:49,  7.54s/it]INFO 04-13 12:17:50 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2013/3300 [4:56:58<2:30:30,  7.02s/it]2026-04-13 12:17:56,070 | INFO | train_grpo_train | metrics_logged mode=train step=2013

                                                       
{'loss': 0.0202, 'grad_norm': 13.283822059631348, 'learning_rate': 3.9030303030303035e-06, 'num_tokens': 3620447.0, 'completions/mean_length': 16.0, 'completions/min_length': 15.0, 'completions/max_length': 17.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 16.0, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 17.0, 'rewards/meter/mean': 0.9972677230834961, 'rewards/meter/std': 0.0004921433283016086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8947007656097412, 'rewards/repeat_soft/std': 0.0385967381298542, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43143871426582336, 'rewards/total_composite/std': 0.005836715456098318, 'reward': 0.43143871426582336, 'reward_std': 0.005836710799485445, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.051268719136714935, 'sampling/sampling_logp_difference/max': 1.1296977996826172, 'sampling/importance_sampling_ratio/min': 0.32313090562820435, 'sampling/importance_sampling_ratio/mean': 1.0239907503128052, 'sampling/importance_sampling_ratio/max': 1.2789902687072754, 'entropy': 0.35202306509017944, 'clip_ratio/low_mean': 0.007352941203862429, 'clip_ratio/low_min': 0.007352941203862429, 'clip_ratio/high_mean': 0.023958333767950535, 'clip_ratio/high_max': 0.023958333767950535, 'clip_ratio/region_mean': 0.031311274971812963, 'reward_total_mean': 0.43143871426582336, 'reward_meter_mean': 0.9972677230834961, 'reward_meter_std': 0.0004921433283016086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8947007656097412, 'reward_repeat_soft_std': 0.0385967381298542, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43143871426582336, 'reward_total_composite_std': 0.005836715456098318, 'epoch': 0.2}

 61%|██████    | 2013/3300 [4:56:58<2:30:30,  7.02s/it]INFO 04-13 12:17:56 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2014/3300 [4:57:04<2:27:13,  6.87s/it]2026-04-13 12:18:02,592 | INFO | train_grpo_train | metrics_logged mode=train step=2014

                                                       
{'loss': 0.0049, 'grad_norm': 13.62286376953125, 'learning_rate': 3.900000000000001e-06, 'num_tokens': 3621798.0, 'completions/mean_length': 25.875, 'completions/min_length': 23.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.875, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.8583083152770996, 'rewards/meter/std': 0.34245607256889343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9488004446029663, 'rewards/repeat_soft/std': 0.03874802216887474, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.42600512504577637, 'rewards/total_composite/std': 0.03307943791151047, 'reward': 0.42600512504577637, 'reward_std': 0.03307943791151047, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14248394966125488, 'sampling/sampling_logp_difference/max': 1.1403138637542725, 'sampling/importance_sampling_ratio/min': 0.31971868872642517, 'sampling/importance_sampling_ratio/mean': 1.033398151397705, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 1.0974683165550232, 'clip_ratio/low_mean': 0.033518518321216106, 'clip_ratio/low_min': 0.033518518321216106, 'clip_ratio/high_mean': 0.10463297180831432, 'clip_ratio/high_max': 0.10463297180831432, 'clip_ratio/region_mean': 0.13815149012953043, 'reward_total_mean': 0.42600512504577637, 'reward_meter_mean': 0.8583083152770996, 'reward_meter_std': 0.34245607256889343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9488004446029663, 'reward_repeat_soft_std': 0.03874802216887474, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.42600512504577637, 'reward_total_composite_std': 0.03307943791151047, 'epoch': 0.2}

 61%|██████    | 2014/3300 [4:57:04<2:27:13,  6.87s/it]INFO 04-13 12:18:02 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2015/3300 [4:57:16<2:58:34,  8.34s/it]2026-04-13 12:18:14,361 | INFO | train_grpo_train | metrics_logged mode=train step=2015

                                                       
{'loss': -0.1293, 'grad_norm': 1.7553294897079468, 'learning_rate': 3.896969696969697e-06, 'num_tokens': 3623548.0, 'completions/mean_length': 101.75, 'completions/min_length': 34.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 34.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8746063113212585, 'rewards/meter/std': 0.2304375022649765, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9749953150749207, 'rewards/repeat_soft/std': 0.01606718637049198, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.3840146064758301, 'rewards/total_composite/std': 0.1552589386701584, 'reward': 0.3840146064758301, 'reward_std': 0.1552589386701584, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10318049788475037, 'sampling/sampling_logp_difference/max': 1.5672869682312012, 'sampling/importance_sampling_ratio/min': 0.20861037075519562, 'sampling/importance_sampling_ratio/mean': 1.0245134830474854, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5291539952158928, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07312493887729943, 'clip_ratio/high_max': 0.07312493887729943, 'clip_ratio/region_mean': 0.07312493887729943, 'reward_total_mean': 0.3840146064758301, 'reward_meter_mean': 0.8746063113212585, 'reward_meter_std': 0.2304375022649765, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9749953150749207, 'reward_repeat_soft_std': 0.01606718637049198, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.3840146064758301, 'reward_total_composite_std': 0.1552589386701584, 'epoch': 0.2}

 61%|██████    | 2015/3300 [4:57:16<2:58:34,  8.34s/it]INFO 04-13 12:18:14 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2016/3300 [4:57:22<2:45:39,  7.74s/it]2026-04-13 12:18:20,702 | INFO | train_grpo_train | metrics_logged mode=train step=2016

                                                       
{'loss': 0.0261, 'grad_norm': 13.447662353515625, 'learning_rate': 3.8939393939393944e-06, 'num_tokens': 3625177.0, 'completions/mean_length': 34.625, 'completions/min_length': 31.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.625, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.9722774028778076, 'rewards/meter/std': 0.04150884225964546, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8981412649154663, 'rewards/repeat_soft/std': 0.06752867996692657, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4319431781768799, 'rewards/total_composite/std': 0.014848713763058186, 'reward': 0.4319431781768799, 'reward_std': 0.014848712831735611, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10293879359960556, 'sampling/sampling_logp_difference/max': 1.0293388366699219, 'sampling/importance_sampling_ratio/min': 0.3572430908679962, 'sampling/importance_sampling_ratio/mean': 0.9976475834846497, 'sampling/importance_sampling_ratio/max': 1.748008131980896, 'entropy': 0.7245897948741913, 'clip_ratio/low_mean': 0.024854814866557717, 'clip_ratio/low_min': 0.024854814866557717, 'clip_ratio/high_mean': 0.05339118931442499, 'clip_ratio/high_max': 0.05339118931442499, 'clip_ratio/region_mean': 0.07824600418098271, 'reward_total_mean': 0.4319431781768799, 'reward_meter_mean': 0.9722774028778076, 'reward_meter_std': 0.04150884225964546, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8981412649154663, 'reward_repeat_soft_std': 0.06752867996692657, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4319431781768799, 'reward_total_composite_std': 0.014848713763058186, 'epoch': 0.2}

 61%|██████    | 2016/3300 [4:57:22<2:45:39,  7.74s/it]INFO 04-13 12:18:20 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2017/3300 [4:57:30<2:43:19,  7.64s/it]2026-04-13 12:18:28,102 | INFO | train_grpo_train | metrics_logged mode=train step=2017

                                                       
{'loss': -0.0535, 'grad_norm': 8.292669296264648, 'learning_rate': 3.890909090909092e-06, 'num_tokens': 3627474.0, 'completions/mean_length': 108.125, 'completions/min_length': 92.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.125, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.8273603916168213, 'rewards/meter/std': 0.2854452133178711, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8415908217430115, 'rewards/repeat_soft/std': 0.04468485340476036, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.4153048098087311, 'rewards/total_composite/std': 0.04054592177271843, 'reward': 0.4153048098087311, 'reward_std': 0.04054592177271843, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1157137006521225, 'sampling/sampling_logp_difference/max': 2.0424206256866455, 'sampling/importance_sampling_ratio/min': 0.12971433997154236, 'sampling/importance_sampling_ratio/mean': 1.0190651416778564, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6621493026614189, 'clip_ratio/low_mean': 0.04455561004579067, 'clip_ratio/low_min': 0.04455561004579067, 'clip_ratio/high_mean': 0.06098366668447852, 'clip_ratio/high_max': 0.06098366668447852, 'clip_ratio/region_mean': 0.1055392767302692, 'reward_total_mean': 0.4153048098087311, 'reward_meter_mean': 0.8273603916168213, 'reward_meter_std': 0.2854452133178711, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8415908217430115, 'reward_repeat_soft_std': 0.04468485340476036, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.4153048098087311, 'reward_total_composite_std': 0.04054592177271843, 'epoch': 0.2}

 61%|██████    | 2017/3300 [4:57:30<2:43:19,  7.64s/it]INFO 04-13 12:18:28 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2018/3300 [4:57:36<2:35:36,  7.28s/it]2026-04-13 12:18:34,559 | INFO | train_grpo_train | metrics_logged mode=train step=2018

                                                       
{'loss': -0.0132, 'grad_norm': 12.649149894714355, 'learning_rate': 3.887878787878788e-06, 'num_tokens': 3629025.0, 'completions/mean_length': 29.875, 'completions/min_length': 26.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.875, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.9288086891174316, 'rewards/meter/std': 0.05646306276321411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.889045238494873, 'rewards/repeat_soft/std': 0.10175919532775879, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4239156246185303, 'rewards/total_composite/std': 0.013974902220070362, 'reward': 0.4239156246185303, 'reward_std': 0.013974908739328384, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1317238211631775, 'sampling/sampling_logp_difference/max': 1.3933568000793457, 'sampling/importance_sampling_ratio/min': 0.2482406198978424, 'sampling/importance_sampling_ratio/mean': 0.9737530946731567, 'sampling/importance_sampling_ratio/max': 1.6525967121124268, 'entropy': 0.5082592852413654, 'clip_ratio/low_mean': 0.03976034792140126, 'clip_ratio/low_min': 0.03976034792140126, 'clip_ratio/high_mean': 0.11229077726602554, 'clip_ratio/high_max': 0.11229077726602554, 'clip_ratio/region_mean': 0.1520511251874268, 'reward_total_mean': 0.4239156246185303, 'reward_meter_mean': 0.9288086891174316, 'reward_meter_std': 0.05646306276321411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.889045238494873, 'reward_repeat_soft_std': 0.10175919532775879, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4239156246185303, 'reward_total_composite_std': 0.013974902220070362, 'epoch': 0.2}

 61%|██████    | 2018/3300 [4:57:36<2:35:36,  7.28s/it]INFO 04-13 12:18:34 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2019/3300 [4:57:47<3:01:11,  8.49s/it]2026-04-13 12:18:45,844 | INFO | train_grpo_train | metrics_logged mode=train step=2019

                                                       
{'loss': -0.0915, 'grad_norm': 1.7585482597351074, 'learning_rate': 3.884848484848485e-06, 'num_tokens': 3630678.0, 'completions/mean_length': 93.625, 'completions/min_length': 31.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 33.85714340209961, 'completions/min_terminated_length': 31.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.722576379776001, 'rewards/meter/std': 0.35335662961006165, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9477214813232422, 'rewards/repeat_soft/std': 0.039741020649671555, 'rewards/judge_quality/mean': 0.17875000834465027, 'rewards/judge_quality/std': 0.1056189239025116, 'rewards/total_composite/mean': 0.38987091183662415, 'rewards/total_composite/std': 0.1731443852186203, 'reward': 0.38987091183662415, 'reward_std': 0.1731443852186203, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1067752093076706, 'sampling/sampling_logp_difference/max': 1.1385347843170166, 'sampling/importance_sampling_ratio/min': 0.3202879726886749, 'sampling/importance_sampling_ratio/mean': 1.000455379486084, 'sampling/importance_sampling_ratio/max': 1.959191918373108, 'entropy': 0.5118012577295303, 'clip_ratio/low_mean': 0.007352941203862429, 'clip_ratio/low_min': 0.007352941203862429, 'clip_ratio/high_mean': 0.06762545485980809, 'clip_ratio/high_max': 0.06762545485980809, 'clip_ratio/region_mean': 0.07497839606367052, 'reward_total_mean': 0.38987091183662415, 'reward_meter_mean': 0.722576379776001, 'reward_meter_std': 0.35335662961006165, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9477214813232422, 'reward_repeat_soft_std': 0.039741020649671555, 'reward_judge_quality_mean': 0.17875000834465027, 'reward_judge_quality_std': 0.1056189239025116, 'reward_total_composite_mean': 0.38987091183662415, 'reward_total_composite_std': 0.1731443852186203, 'epoch': 0.2}

 61%|██████    | 2019/3300 [4:57:48<3:01:11,  8.49s/it]INFO 04-13 12:18:46 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2020/3300 [4:57:55<2:52:32,  8.09s/it]2026-04-13 12:18:53,009 | INFO | train_grpo_train | metrics_logged mode=train step=2020

                                                       
{'loss': 0.0009, 'grad_norm': 8.50749397277832, 'learning_rate': 3.881818181818182e-06, 'num_tokens': 3632791.0, 'completions/mean_length': 82.125, 'completions/min_length': 59.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 82.125, 'completions/min_terminated_length': 59.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.9119631052017212, 'rewards/meter/std': 0.2121911197900772, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8878892660140991, 'rewards/repeat_soft/std': 0.0242925938218832, 'rewards/judge_quality/mean': 0.17625001072883606, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4374694526195526, 'rewards/total_composite/std': 0.026745209470391273, 'reward': 0.4374694526195526, 'reward_std': 0.026745207607746124, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11416690796613693, 'sampling/sampling_logp_difference/max': 1.469538688659668, 'sampling/importance_sampling_ratio/min': 0.2300315797328949, 'sampling/importance_sampling_ratio/mean': 1.0092949867248535, 'sampling/importance_sampling_ratio/max': 1.9388819932937622, 'entropy': 0.7308895066380501, 'clip_ratio/low_mean': 0.033150406554341316, 'clip_ratio/low_min': 0.033150406554341316, 'clip_ratio/high_mean': 0.09055669419467449, 'clip_ratio/high_max': 0.09055669419467449, 'clip_ratio/region_mean': 0.12370710074901581, 'reward_total_mean': 0.4374694526195526, 'reward_meter_mean': 0.9119631052017212, 'reward_meter_std': 0.2121911197900772, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8878892660140991, 'reward_repeat_soft_std': 0.0242925938218832, 'reward_judge_quality_mean': 0.17625001072883606, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4374694526195526, 'reward_total_composite_std': 0.026745209470391273, 'epoch': 0.2}

 61%|██████    | 2020/3300 [4:57:55<2:52:32,  8.09s/it]INFO 04-13 12:18:53 [block_pool.py:378] Successfully reset prefix cache

 61%|██████    | 2021/3300 [4:58:02<2:47:34,  7.86s/it]2026-04-13 12:19:00,345 | INFO | train_grpo_train | metrics_logged mode=train step=2021

                                                       
{'loss': 0.0381, 'grad_norm': 8.613554000854492, 'learning_rate': 3.878787878787879e-06, 'num_tokens': 3634310.0, 'completions/mean_length': 47.875, 'completions/min_length': 33.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.875, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.6935790777206421, 'rewards/meter/std': 0.365664541721344, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9693573713302612, 'rewards/repeat_soft/std': 0.012484983541071415, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.41753631830215454, 'rewards/total_composite/std': 0.039422620087862015, 'reward': 0.41753631830215454, 'reward_std': 0.03942260518670082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09699514508247375, 'sampling/sampling_logp_difference/max': 2.03543758392334, 'sampling/importance_sampling_ratio/min': 0.13062331080436707, 'sampling/importance_sampling_ratio/mean': 1.0070172548294067, 'sampling/importance_sampling_ratio/max': 1.9344388246536255, 'entropy': 0.5120501779019833, 'clip_ratio/low_mean': 0.03184592956677079, 'clip_ratio/low_min': 0.03184592956677079, 'clip_ratio/high_mean': 0.06646964652463794, 'clip_ratio/high_max': 0.06646964652463794, 'clip_ratio/region_mean': 0.09831557609140873, 'reward_total_mean': 0.41753631830215454, 'reward_meter_mean': 0.6935790777206421, 'reward_meter_std': 0.365664541721344, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9693573713302612, 'reward_repeat_soft_std': 0.012484983541071415, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.41753631830215454, 'reward_total_composite_std': 0.039422620087862015, 'epoch': 0.2}

 61%|██████    | 2021/3300 [4:58:02<2:47:34,  7.86s/it]INFO 04-13 12:19:00 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2022/3300 [4:58:14<3:11:22,  8.98s/it]2026-04-13 12:19:11,942 | INFO | train_grpo_train | metrics_logged mode=train step=2022

                                                       
{'loss': -0.2056, 'grad_norm': 1.7340670824050903, 'learning_rate': 3.875757575757576e-06, 'num_tokens': 3636463.0, 'completions/mean_length': 157.125, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 106.42857360839844, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 129.0, 'rewards/meter/mean': 0.811650276184082, 'rewards/meter/std': 0.2834619879722595, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8670904040336609, 'rewards/repeat_soft/std': 0.04739189147949219, 'rewards/judge_quality/mean': 0.1262499988079071, 'rewards/judge_quality/std': 0.04103570431470871, 'rewards/total_composite/mean': 0.35624122619628906, 'rewards/total_composite/std': 0.14686189591884613, 'reward': 0.35624122619628906, 'reward_std': 0.14686188101768494, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1068727895617485, 'sampling/sampling_logp_difference/max': 3.1300125122070312, 'sampling/importance_sampling_ratio/min': 0.04371724650263786, 'sampling/importance_sampling_ratio/mean': 1.0075139999389648, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4911436326801777, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09701418410986662, 'clip_ratio/high_max': 0.09701418410986662, 'clip_ratio/region_mean': 0.09701418410986662, 'reward_total_mean': 0.35624122619628906, 'reward_meter_mean': 0.811650276184082, 'reward_meter_std': 0.2834619879722595, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8670904040336609, 'reward_repeat_soft_std': 0.04739189147949219, 'reward_judge_quality_mean': 0.1262499988079071, 'reward_judge_quality_std': 0.04103570431470871, 'reward_total_composite_mean': 0.35624122619628906, 'reward_total_composite_std': 0.14686189591884613, 'epoch': 0.2}

 61%|██████▏   | 2022/3300 [4:58:14<3:11:22,  8.98s/it]INFO 04-13 12:19:12 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2023/3300 [4:58:20<2:53:52,  8.17s/it]2026-04-13 12:19:18,217 | INFO | train_grpo_train | metrics_logged mode=train step=2023

                                                       
{'loss': 0.064, 'grad_norm': 10.871502876281738, 'learning_rate': 3.872727272727273e-06, 'num_tokens': 3638197.0, 'completions/mean_length': 51.75, 'completions/min_length': 45.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.9767289161682129, 'rewards/meter/std': 0.020327460020780563, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9791615009307861, 'rewards/repeat_soft/std': 0.02554360218346119, 'rewards/judge_quality/mean': 0.24625001847743988, 'rewards/judge_quality/std': 0.2722361087799072, 'rewards/total_composite/mean': 0.5042101144790649, 'rewards/total_composite/std': 0.1774170845746994, 'reward': 0.5042101144790649, 'reward_std': 0.1774170994758606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12130969017744064, 'sampling/sampling_logp_difference/max': 1.0781450271606445, 'sampling/importance_sampling_ratio/min': 0.34022602438926697, 'sampling/importance_sampling_ratio/mean': 1.0086771249771118, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7564574480056763, 'clip_ratio/low_mean': 0.09432055428624153, 'clip_ratio/low_min': 0.09432055428624153, 'clip_ratio/high_mean': 0.010869565419852734, 'clip_ratio/high_max': 0.010869565419852734, 'clip_ratio/region_mean': 0.10519011970609426, 'reward_total_mean': 0.5042101144790649, 'reward_meter_mean': 0.9767289161682129, 'reward_meter_std': 0.020327460020780563, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9791615009307861, 'reward_repeat_soft_std': 0.02554360218346119, 'reward_judge_quality_mean': 0.24625001847743988, 'reward_judge_quality_std': 0.2722361087799072, 'reward_total_composite_mean': 0.5042101144790649, 'reward_total_composite_std': 0.1774170845746994, 'epoch': 0.2}

 61%|██████▏   | 2023/3300 [4:58:20<2:53:52,  8.17s/it]INFO 04-13 12:19:18 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2024/3300 [4:58:26<2:42:49,  7.66s/it]2026-04-13 12:19:24,682 | INFO | train_grpo_train | metrics_logged mode=train step=2024

                                                       
{'loss': 0.0081, 'grad_norm': 14.9652099609375, 'learning_rate': 3.86969696969697e-06, 'num_tokens': 3639790.0, 'completions/mean_length': 36.125, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8460080623626709, 'rewards/meter/std': 0.3138962984085083, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9143238663673401, 'rewards/repeat_soft/std': 0.0503779835999012, 'rewards/judge_quality/mean': 0.13875001668930054, 'rewards/judge_quality/std': 0.015526479110121727, 'rewards/total_composite/mean': 0.4125252068042755, 'rewards/total_composite/std': 0.030189499258995056, 'reward': 0.4125252068042755, 'reward_std': 0.030189495533704758, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.05745663493871689, 'sampling/sampling_logp_difference/max': 1.122283935546875, 'sampling/importance_sampling_ratio/min': 0.3255354166030884, 'sampling/importance_sampling_ratio/mean': 0.9936563968658447, 'sampling/importance_sampling_ratio/max': 1.621734619140625, 'entropy': 0.26591492630541325, 'clip_ratio/low_mean': 0.017330858390778303, 'clip_ratio/low_min': 0.017330858390778303, 'clip_ratio/high_mean': 0.021464647026732564, 'clip_ratio/high_max': 0.021464647026732564, 'clip_ratio/region_mean': 0.03879550541751087, 'reward_total_mean': 0.4125252068042755, 'reward_meter_mean': 0.8460080623626709, 'reward_meter_std': 0.3138962984085083, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9143238663673401, 'reward_repeat_soft_std': 0.0503779835999012, 'reward_judge_quality_mean': 0.13875001668930054, 'reward_judge_quality_std': 0.015526479110121727, 'reward_total_composite_mean': 0.4125252068042755, 'reward_total_composite_std': 0.030189499258995056, 'epoch': 0.2}

 61%|██████▏   | 2024/3300 [4:58:26<2:42:49,  7.66s/it]INFO 04-13 12:19:24 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2025/3300 [4:58:32<2:33:25,  7.22s/it]2026-04-13 12:19:30,874 | INFO | train_grpo_train | metrics_logged mode=train step=2025

                                                       
{'loss': 0.0021, 'grad_norm': 10.698853492736816, 'learning_rate': 3.866666666666667e-06, 'num_tokens': 3641540.0, 'completions/mean_length': 51.75, 'completions/min_length': 43.0, 'completions/max_length': 59.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.75, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 59.0, 'rewards/meter/mean': 0.6679095029830933, 'rewards/meter/std': 0.29510411620140076, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9887790679931641, 'rewards/repeat_soft/std': 0.016280222684144974, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4143340587615967, 'rewards/total_composite/std': 0.02845674753189087, 'reward': 0.4143340587615967, 'reward_std': 0.028456738218665123, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1274341642856598, 'sampling/sampling_logp_difference/max': 1.6919660568237305, 'sampling/importance_sampling_ratio/min': 0.1841571182012558, 'sampling/importance_sampling_ratio/mean': 1.0127040147781372, 'sampling/importance_sampling_ratio/max': 1.8242231607437134, 'entropy': 0.6411586254835129, 'clip_ratio/low_mean': 0.04926663590595126, 'clip_ratio/low_min': 0.04926663590595126, 'clip_ratio/high_mean': 0.05039568850770593, 'clip_ratio/high_max': 0.05039568850770593, 'clip_ratio/region_mean': 0.09966232441365719, 'reward_total_mean': 0.4143340587615967, 'reward_meter_mean': 0.6679095029830933, 'reward_meter_std': 0.29510411620140076, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9887790679931641, 'reward_repeat_soft_std': 0.016280222684144974, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4143340587615967, 'reward_total_composite_std': 0.02845674753189087, 'epoch': 0.2}

 61%|██████▏   | 2025/3300 [4:58:33<2:33:25,  7.22s/it]INFO 04-13 12:19:31 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2026/3300 [4:58:39<2:27:31,  6.95s/it]2026-04-13 12:19:37,195 | INFO | train_grpo_train | metrics_logged mode=train step=2026

                                                       
{'loss': -0.0773, 'grad_norm': 11.50498104095459, 'learning_rate': 3.863636363636364e-06, 'num_tokens': 3642999.0, 'completions/mean_length': 35.375, 'completions/min_length': 26.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.375, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.5798919796943665, 'rewards/meter/std': 0.43411311507225037, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9656744003295898, 'rewards/repeat_soft/std': 0.04045727103948593, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40139061212539673, 'rewards/total_composite/std': 0.03712591901421547, 'reward': 0.40139061212539673, 'reward_std': 0.03712591901421547, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12781719863414764, 'sampling/sampling_logp_difference/max': 1.2028522491455078, 'sampling/importance_sampling_ratio/min': 0.3003363609313965, 'sampling/importance_sampling_ratio/mean': 1.0064102411270142, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6434668265283108, 'clip_ratio/low_mean': 0.058067780919373035, 'clip_ratio/low_min': 0.058067780919373035, 'clip_ratio/high_mean': 0.05798420403152704, 'clip_ratio/high_max': 0.05798420403152704, 'clip_ratio/region_mean': 0.11605198495090008, 'reward_total_mean': 0.40139061212539673, 'reward_meter_mean': 0.5798919796943665, 'reward_meter_std': 0.43411311507225037, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9656744003295898, 'reward_repeat_soft_std': 0.04045727103948593, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40139061212539673, 'reward_total_composite_std': 0.03712591901421547, 'epoch': 0.2}

 61%|██████▏   | 2026/3300 [4:58:39<2:27:31,  6.95s/it]INFO 04-13 12:19:37 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2027/3300 [4:58:45<2:23:47,  6.78s/it]2026-04-13 12:19:43,572 | INFO | train_grpo_train | metrics_logged mode=train step=2027

                                                       
{'loss': 0.0414, 'grad_norm': 11.639140129089355, 'learning_rate': 3.860606060606061e-06, 'num_tokens': 3644690.0, 'completions/mean_length': 58.375, 'completions/min_length': 51.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 58.375, 'completions/min_terminated_length': 51.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.8482519388198853, 'rewards/meter/std': 0.30271652340888977, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9650776386260986, 'rewards/repeat_soft/std': 0.03526902198791504, 'rewards/judge_quality/mean': 0.1612500101327896, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4340474307537079, 'rewards/total_composite/std': 0.030730562284588814, 'reward': 0.4340474307537079, 'reward_std': 0.030730556696653366, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10567151010036469, 'sampling/sampling_logp_difference/max': 1.1417315006256104, 'sampling/importance_sampling_ratio/min': 0.319265753030777, 'sampling/importance_sampling_ratio/mean': 1.0057364702224731, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.69110918790102, 'clip_ratio/low_mean': 0.036418041214346886, 'clip_ratio/low_min': 0.036418041214346886, 'clip_ratio/high_mean': 0.07654706481844187, 'clip_ratio/high_max': 0.07654706481844187, 'clip_ratio/region_mean': 0.11296510603278875, 'reward_total_mean': 0.4340474307537079, 'reward_meter_mean': 0.8482519388198853, 'reward_meter_std': 0.30271652340888977, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9650776386260986, 'reward_repeat_soft_std': 0.03526902198791504, 'reward_judge_quality_mean': 0.1612500101327896, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4340474307537079, 'reward_total_composite_std': 0.030730562284588814, 'epoch': 0.2}

 61%|██████▏   | 2027/3300 [4:58:45<2:23:47,  6.78s/it]INFO 04-13 12:19:43 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2028/3300 [4:58:53<2:32:48,  7.21s/it]2026-04-13 12:19:51,781 | INFO | train_grpo_train | metrics_logged mode=train step=2028

                                                       
{'loss': -0.1285, 'grad_norm': 7.573558807373047, 'learning_rate': 3.857575757575758e-06, 'num_tokens': 3647183.0, 'completions/mean_length': 120.625, 'completions/min_length': 81.0, 'completions/max_length': 171.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 120.625, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 171.0, 'rewards/meter/mean': 0.8907461166381836, 'rewards/meter/std': 0.12185462564229965, 'rewards/count_adherence/mean': 0.875, 'rewards/count_adherence/std': 0.1035098284482956, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8501641154289246, 'rewards/repeat_soft/std': 0.04492751881480217, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.031052954494953156, 'rewards/total_composite/mean': 0.3936898708343506, 'rewards/total_composite/std': 0.02330118790268898, 'reward': 0.3936898708343506, 'reward_std': 0.02330118604004383, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10553734749555588, 'sampling/sampling_logp_difference/max': 1.3691177368164062, 'sampling/importance_sampling_ratio/min': 0.2543312609195709, 'sampling/importance_sampling_ratio/mean': 1.0143228769302368, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6704464554786682, 'clip_ratio/low_mean': 0.05444397870451212, 'clip_ratio/low_min': 0.05444397870451212, 'clip_ratio/high_mean': 0.057956038042902946, 'clip_ratio/high_max': 0.057956038042902946, 'clip_ratio/region_mean': 0.11240001674741507, 'reward_total_mean': 0.3936898708343506, 'reward_meter_mean': 0.8907461166381836, 'reward_meter_std': 0.12185462564229965, 'reward_count_adherence_mean': 0.875, 'reward_count_adherence_std': 0.1035098284482956, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8501641154289246, 'reward_repeat_soft_std': 0.04492751881480217, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.031052954494953156, 'reward_total_composite_mean': 0.3936898708343506, 'reward_total_composite_std': 0.02330118790268898, 'epoch': 0.2}

 61%|██████▏   | 2028/3300 [4:58:54<2:32:48,  7.21s/it]INFO 04-13 12:19:52 [block_pool.py:378] Successfully reset prefix cache

 61%|██████▏   | 2029/3300 [4:59:01<2:36:16,  7.38s/it]2026-04-13 12:19:59,551 | INFO | train_grpo_train | metrics_logged mode=train step=2029

                                                       
{'loss': -0.1184, 'grad_norm': 6.718632698059082, 'learning_rate': 3.8545454545454545e-06, 'num_tokens': 3649393.0, 'completions/mean_length': 118.25, 'completions/min_length': 99.0, 'completions/max_length': 176.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.25, 'completions/min_terminated_length': 99.0, 'completions/max_terminated_length': 176.0, 'rewards/meter/mean': 0.7876110672950745, 'rewards/meter/std': 0.16306620836257935, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7729731798171997, 'rewards/repeat_soft/std': 0.08785118907690048, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.02777460776269436, 'rewards/total_composite/mean': 0.40055200457572937, 'rewards/total_composite/std': 0.027080491185188293, 'reward': 0.40055200457572937, 'reward_std': 0.027080485597252846, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0912838876247406, 'sampling/sampling_logp_difference/max': 2.8005564212799072, 'sampling/importance_sampling_ratio/min': 0.0607762336730957, 'sampling/importance_sampling_ratio/mean': 1.0105650424957275, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4359697103500366, 'clip_ratio/low_mean': 0.0342613086104393, 'clip_ratio/low_min': 0.0342613086104393, 'clip_ratio/high_mean': 0.05099095404148102, 'clip_ratio/high_max': 0.05099095404148102, 'clip_ratio/region_mean': 0.08525226265192032, 'reward_total_mean': 0.40055200457572937, 'reward_meter_mean': 0.7876110672950745, 'reward_meter_std': 0.16306620836257935, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7729731798171997, 'reward_repeat_soft_std': 0.08785118907690048, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.02777460776269436, 'reward_total_composite_mean': 0.40055200457572937, 'reward_total_composite_std': 0.027080491185188293, 'epoch': 0.2}

 61%|██████▏   | 2029/3300 [4:59:01<2:36:16,  7.38s/it]INFO 04-13 12:19:59 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2030/3300 [4:59:08<2:33:11,  7.24s/it]2026-04-13 12:20:06,488 | INFO | train_grpo_train | metrics_logged mode=train step=2030

                                                       
{'loss': -0.0294, 'grad_norm': 13.128202438354492, 'learning_rate': 3.851515151515152e-06, 'num_tokens': 3650949.0, 'completions/mean_length': 41.5, 'completions/min_length': 36.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.5, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8338658809661865, 'rewards/meter/std': 0.3192678391933441, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9496046900749207, 'rewards/repeat_soft/std': 0.04710354283452034, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4313012957572937, 'rewards/total_composite/std': 0.03520743176341057, 'reward': 0.4313012957572937, 'reward_std': 0.03520742803812027, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11196162551641464, 'sampling/sampling_logp_difference/max': 1.4962571859359741, 'sampling/importance_sampling_ratio/min': 0.22396688163280487, 'sampling/importance_sampling_ratio/mean': 1.0043363571166992, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5714318752288818, 'clip_ratio/low_mean': 0.03207236900925636, 'clip_ratio/low_min': 0.03207236900925636, 'clip_ratio/high_mean': 0.08106857095845044, 'clip_ratio/high_max': 0.08106857095845044, 'clip_ratio/region_mean': 0.1131409399677068, 'reward_total_mean': 0.4313012957572937, 'reward_meter_mean': 0.8338658809661865, 'reward_meter_std': 0.3192678391933441, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9496046900749207, 'reward_repeat_soft_std': 0.04710354283452034, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4313012957572937, 'reward_total_composite_std': 0.03520743176341057, 'epoch': 0.2}

 62%|██████▏   | 2030/3300 [4:59:08<2:33:11,  7.24s/it]INFO 04-13 12:20:06 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2031/3300 [4:59:15<2:29:10,  7.05s/it]2026-04-13 12:20:13,089 | INFO | train_grpo_train | metrics_logged mode=train step=2031

                                                       
{'loss': 0.0286, 'grad_norm': 12.468624114990234, 'learning_rate': 3.848484848484848e-06, 'num_tokens': 3652423.0, 'completions/mean_length': 21.25, 'completions/min_length': 17.0, 'completions/max_length': 24.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.25, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 24.0, 'rewards/meter/mean': 0.6922364830970764, 'rewards/meter/std': 0.42275819182395935, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9160686731338501, 'rewards/repeat_soft/std': 0.04868149757385254, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4049033522605896, 'rewards/total_composite/std': 0.046748362481594086, 'reward': 0.4049033522605896, 'reward_std': 0.046748362481594086, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10105178505182266, 'sampling/sampling_logp_difference/max': 1.5017375946044922, 'sampling/importance_sampling_ratio/min': 0.22274281084537506, 'sampling/importance_sampling_ratio/mean': 0.9819271564483643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4016204699873924, 'clip_ratio/low_mean': 0.01711956551298499, 'clip_ratio/low_min': 0.01711956551298499, 'clip_ratio/high_mean': 0.07495747972279787, 'clip_ratio/high_max': 0.07495747972279787, 'clip_ratio/region_mean': 0.09207704523578286, 'reward_total_mean': 0.4049033522605896, 'reward_meter_mean': 0.6922364830970764, 'reward_meter_std': 0.42275819182395935, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9160686731338501, 'reward_repeat_soft_std': 0.04868149757385254, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4049033522605896, 'reward_total_composite_std': 0.046748362481594086, 'epoch': 0.2}

 62%|██████▏   | 2031/3300 [4:59:15<2:29:10,  7.05s/it]INFO 04-13 12:20:13 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2032/3300 [4:59:21<2:22:35,  6.75s/it]2026-04-13 12:20:19,128 | INFO | train_grpo_train | metrics_logged mode=train step=2032

                                                       
{'loss': 0.0103, 'grad_norm': 23.99254035949707, 'learning_rate': 3.8454545454545454e-06, 'num_tokens': 3653854.0, 'completions/mean_length': 17.875, 'completions/min_length': 14.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 17.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.9383523464202881, 'rewards/meter/std': 0.11129981279373169, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9575029015541077, 'rewards/repeat_soft/std': 0.014133882708847523, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4351148009300232, 'rewards/total_composite/std': 0.01072826236486435, 'reward': 0.4351148009300232, 'reward_std': 0.010728267952799797, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08517216145992279, 'sampling/sampling_logp_difference/max': 1.2159996032714844, 'sampling/importance_sampling_ratio/min': 0.2964135706424713, 'sampling/importance_sampling_ratio/mean': 1.00950288772583, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.506832730025053, 'clip_ratio/low_mean': 0.01666666753590107, 'clip_ratio/low_min': 0.01666666753590107, 'clip_ratio/high_mean': 0.11809424310922623, 'clip_ratio/high_max': 0.11809424310922623, 'clip_ratio/region_mean': 0.1347609106451273, 'reward_total_mean': 0.4351148009300232, 'reward_meter_mean': 0.9383523464202881, 'reward_meter_std': 0.11129981279373169, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9575029015541077, 'reward_repeat_soft_std': 0.014133882708847523, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4351148009300232, 'reward_total_composite_std': 0.01072826236486435, 'epoch': 0.2}

 62%|██████▏   | 2032/3300 [4:59:21<2:22:35,  6.75s/it]INFO 04-13 12:20:19 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2033/3300 [4:59:29<2:29:04,  7.06s/it]2026-04-13 12:20:26,913 | INFO | train_grpo_train | metrics_logged mode=train step=2033

                                                       
{'loss': -0.0539, 'grad_norm': 7.2069854736328125, 'learning_rate': 3.842424242424243e-06, 'num_tokens': 3656268.0, 'completions/mean_length': 127.75, 'completions/min_length': 101.0, 'completions/max_length': 156.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 127.75, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 156.0, 'rewards/meter/mean': 0.9627584218978882, 'rewards/meter/std': 0.025766663253307343, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7789667844772339, 'rewards/repeat_soft/std': 0.07983645796775818, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.03207135573029518, 'rewards/total_composite/mean': 0.41072261333465576, 'rewards/total_composite/std': 0.02998325042426586, 'reward': 0.41072261333465576, 'reward_std': 0.029983246698975563, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08343739062547684, 'sampling/sampling_logp_difference/max': 1.908365249633789, 'sampling/importance_sampling_ratio/min': 0.1483226716518402, 'sampling/importance_sampling_ratio/mean': 1.0080602169036865, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43094078078866005, 'clip_ratio/low_mean': 0.03441926138475537, 'clip_ratio/low_min': 0.03441926138475537, 'clip_ratio/high_mean': 0.03305764775723219, 'clip_ratio/high_max': 0.03305764775723219, 'clip_ratio/region_mean': 0.06747690914198756, 'reward_total_mean': 0.41072261333465576, 'reward_meter_mean': 0.9627584218978882, 'reward_meter_std': 0.025766663253307343, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7789667844772339, 'reward_repeat_soft_std': 0.07983645796775818, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.03207135573029518, 'reward_total_composite_mean': 0.41072261333465576, 'reward_total_composite_std': 0.02998325042426586, 'epoch': 0.2}

 62%|██████▏   | 2033/3300 [4:59:29<2:29:04,  7.06s/it]INFO 04-13 12:20:27 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2034/3300 [4:59:35<2:28:11,  7.02s/it]2026-04-13 12:20:33,855 | INFO | train_grpo_train | metrics_logged mode=train step=2034

                                                       
{'loss': 0.0586, 'grad_norm': 7.509989261627197, 'learning_rate': 3.839393939393939e-06, 'num_tokens': 3658045.0, 'completions/mean_length': 52.125, 'completions/min_length': 41.0, 'completions/max_length': 61.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.125, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 61.0, 'rewards/meter/mean': 0.8884906768798828, 'rewards/meter/std': 0.1771586537361145, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8319612741470337, 'rewards/repeat_soft/std': 0.06831370294094086, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.01922610215842724, 'rewards/total_composite/mean': 0.3588881194591522, 'rewards/total_composite/std': 0.14733052253723145, 'reward': 0.3588881194591522, 'reward_std': 0.14733052253723145, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12376967817544937, 'sampling/sampling_logp_difference/max': 1.9423317909240723, 'sampling/importance_sampling_ratio/min': 0.14336925745010376, 'sampling/importance_sampling_ratio/mean': 0.9940099120140076, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5854336768388748, 'clip_ratio/low_mean': 0.008333333767950535, 'clip_ratio/low_min': 0.008333333767950535, 'clip_ratio/high_mean': 0.10637603979557753, 'clip_ratio/high_max': 0.10637603979557753, 'clip_ratio/region_mean': 0.11470937356352806, 'reward_total_mean': 0.3588881194591522, 'reward_meter_mean': 0.8884906768798828, 'reward_meter_std': 0.1771586537361145, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8319612741470337, 'reward_repeat_soft_std': 0.06831370294094086, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.01922610215842724, 'reward_total_composite_mean': 0.3588881194591522, 'reward_total_composite_std': 0.14733052253723145, 'epoch': 0.2}

 62%|██████▏   | 2034/3300 [4:59:36<2:28:11,  7.02s/it]INFO 04-13 12:20:34 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2035/3300 [4:59:42<2:26:49,  6.96s/it]2026-04-13 12:20:40,685 | INFO | train_grpo_train | metrics_logged mode=train step=2035

                                                       
{'loss': -0.0323, 'grad_norm': 7.591353893280029, 'learning_rate': 3.836363636363636e-06, 'num_tokens': 3660138.0, 'completions/mean_length': 74.625, 'completions/min_length': 68.0, 'completions/max_length': 90.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 68.0, 'completions/max_terminated_length': 90.0, 'rewards/meter/mean': 0.9790351986885071, 'rewards/meter/std': 0.03450845554471016, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8338398933410645, 'rewards/repeat_soft/std': 0.08215414732694626, 'rewards/judge_quality/mean': 0.18000000715255737, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4396231174468994, 'rewards/total_composite/std': 0.010331184603273869, 'reward': 0.4396231174468994, 'reward_std': 0.01033118087798357, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08554603159427643, 'sampling/sampling_logp_difference/max': 1.9737772941589355, 'sampling/importance_sampling_ratio/min': 0.13893108069896698, 'sampling/importance_sampling_ratio/mean': 1.0047028064727783, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.428936704993248, 'clip_ratio/low_mean': 0.036277403589338064, 'clip_ratio/low_min': 0.036277403589338064, 'clip_ratio/high_mean': 0.046252098865807056, 'clip_ratio/high_max': 0.046252098865807056, 'clip_ratio/region_mean': 0.08252950245514512, 'reward_total_mean': 0.4396231174468994, 'reward_meter_mean': 0.9790351986885071, 'reward_meter_std': 0.03450845554471016, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8338398933410645, 'reward_repeat_soft_std': 0.08215414732694626, 'reward_judge_quality_mean': 0.18000000715255737, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4396231174468994, 'reward_total_composite_std': 0.010331184603273869, 'epoch': 0.2}

 62%|██████▏   | 2035/3300 [4:59:42<2:26:49,  6.96s/it]INFO 04-13 12:20:40 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2036/3300 [4:59:49<2:27:21,  6.99s/it]2026-04-13 12:20:47,736 | INFO | train_grpo_train | metrics_logged mode=train step=2036

                                                       
{'loss': 0.0618, 'grad_norm': 12.391810417175293, 'learning_rate': 3.833333333333334e-06, 'num_tokens': 3662473.0, 'completions/mean_length': 103.875, 'completions/min_length': 89.0, 'completions/max_length': 121.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 103.875, 'completions/min_terminated_length': 89.0, 'completions/max_terminated_length': 121.0, 'rewards/meter/mean': 0.746888279914856, 'rewards/meter/std': 0.25801926851272583, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8703850507736206, 'rewards/repeat_soft/std': 0.024696068838238716, 'rewards/judge_quality/mean': 0.16875000298023224, 'rewards/judge_quality/std': 0.022320717573165894, 'rewards/total_composite/mean': 0.4135923385620117, 'rewards/total_composite/std': 0.031244825571775436, 'reward': 0.4135923385620117, 'reward_std': 0.031244821846485138, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08883036673069, 'sampling/sampling_logp_difference/max': 1.893946886062622, 'sampling/importance_sampling_ratio/min': 0.15047672390937805, 'sampling/importance_sampling_ratio/mean': 1.0045208930969238, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4564935080707073, 'clip_ratio/low_mean': 0.03486485220491886, 'clip_ratio/low_min': 0.03486485220491886, 'clip_ratio/high_mean': 0.04028744110837579, 'clip_ratio/high_max': 0.04028744110837579, 'clip_ratio/region_mean': 0.07515229331329465, 'reward_total_mean': 0.4135923385620117, 'reward_meter_mean': 0.746888279914856, 'reward_meter_std': 0.25801926851272583, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8703850507736206, 'reward_repeat_soft_std': 0.024696068838238716, 'reward_judge_quality_mean': 0.16875000298023224, 'reward_judge_quality_std': 0.022320717573165894, 'reward_total_composite_mean': 0.4135923385620117, 'reward_total_composite_std': 0.031244825571775436, 'epoch': 0.2}

 62%|██████▏   | 2036/3300 [4:59:49<2:27:21,  6.99s/it]INFO 04-13 12:20:47 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2037/3300 [4:59:56<2:23:10,  6.80s/it]2026-04-13 12:20:54,096 | INFO | train_grpo_train | metrics_logged mode=train step=2037

                                                       
{'loss': -0.0844, 'grad_norm': 14.103687286376953, 'learning_rate': 3.830303030303031e-06, 'num_tokens': 3664098.0, 'completions/mean_length': 37.125, 'completions/min_length': 32.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.8908556699752808, 'rewards/meter/std': 0.20262905955314636, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8498884439468384, 'rewards/repeat_soft/std': 0.12281721830368042, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.013887306675314903, 'rewards/total_composite/mean': 0.40948760509490967, 'rewards/total_composite/std': 0.019680894911289215, 'reward': 0.40948760509490967, 'reward_std': 0.019680896773934364, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09558341652154922, 'sampling/sampling_logp_difference/max': 1.2192243337631226, 'sampling/importance_sampling_ratio/min': 0.2954592704772949, 'sampling/importance_sampling_ratio/mean': 1.0035276412963867, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7432537972927094, 'clip_ratio/low_mean': 0.048447174951434135, 'clip_ratio/low_min': 0.048447174951434135, 'clip_ratio/high_mean': 0.038103504572063684, 'clip_ratio/high_max': 0.038103504572063684, 'clip_ratio/region_mean': 0.08655067952349782, 'reward_total_mean': 0.40948760509490967, 'reward_meter_mean': 0.8908556699752808, 'reward_meter_std': 0.20262905955314636, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8498884439468384, 'reward_repeat_soft_std': 0.12281721830368042, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.013887306675314903, 'reward_total_composite_mean': 0.40948760509490967, 'reward_total_composite_std': 0.019680894911289215, 'epoch': 0.2}

 62%|██████▏   | 2037/3300 [4:59:56<2:23:10,  6.80s/it]INFO 04-13 12:20:54 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2038/3300 [5:00:02<2:20:18,  6.67s/it]2026-04-13 12:21:00,454 | INFO | train_grpo_train | metrics_logged mode=train step=2038

                                                       
{'loss': 0.0111, 'grad_norm': 14.108630180358887, 'learning_rate': 3.827272727272728e-06, 'num_tokens': 3665743.0, 'completions/mean_length': 45.625, 'completions/min_length': 36.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.625, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.6388722658157349, 'rewards/meter/std': 0.42735129594802856, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9902987480163574, 'rewards/repeat_soft/std': 0.009319685399532318, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/total_composite/mean': 0.45695754885673523, 'rewards/total_composite/std': 0.11092603206634521, 'reward': 0.45695754885673523, 'reward_std': 0.11092603206634521, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11713860929012299, 'sampling/sampling_logp_difference/max': 1.2782936096191406, 'sampling/importance_sampling_ratio/min': 0.2785121500492096, 'sampling/importance_sampling_ratio/mean': 1.0123933553695679, 'sampling/importance_sampling_ratio/max': 1.7498754262924194, 'entropy': 0.6255333051085472, 'clip_ratio/low_mean': 0.0956260934472084, 'clip_ratio/low_min': 0.0956260934472084, 'clip_ratio/high_mean': 0.02377136843279004, 'clip_ratio/high_max': 0.02377136843279004, 'clip_ratio/region_mean': 0.11939746187999845, 'reward_total_mean': 0.45695754885673523, 'reward_meter_mean': 0.6388722658157349, 'reward_meter_std': 0.42735129594802856, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9902987480163574, 'reward_repeat_soft_std': 0.009319685399532318, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.45695754885673523, 'reward_total_composite_std': 0.11092603206634521, 'epoch': 0.2}

 62%|██████▏   | 2038/3300 [5:00:02<2:20:18,  6.67s/it]INFO 04-13 12:21:00 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2039/3300 [5:00:08<2:17:18,  6.53s/it]2026-04-13 12:21:06,667 | INFO | train_grpo_train | metrics_logged mode=train step=2039

                                                       
{'loss': -0.006, 'grad_norm': 15.822030067443848, 'learning_rate': 3.8242424242424245e-06, 'num_tokens': 3667123.0, 'completions/mean_length': 19.5, 'completions/min_length': 17.0, 'completions/max_length': 22.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.5, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 22.0, 'rewards/meter/mean': 0.7434380054473877, 'rewards/meter/std': 0.25848403573036194, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9545454382896423, 'rewards/repeat_soft/std': 0.014728941954672337, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.413274347782135, 'rewards/total_composite/std': 0.022209085524082184, 'reward': 0.413274347782135, 'reward_std': 0.022209081798791885, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14968574047088623, 'sampling/sampling_logp_difference/max': 1.307535171508789, 'sampling/importance_sampling_ratio/min': 0.27048593759536743, 'sampling/importance_sampling_ratio/mean': 1.002846598625183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7239391058683395, 'clip_ratio/low_mean': 0.07763783447444439, 'clip_ratio/low_min': 0.07763783447444439, 'clip_ratio/high_mean': 0.05220437468960881, 'clip_ratio/high_max': 0.05220437468960881, 'clip_ratio/region_mean': 0.1298422091640532, 'reward_total_mean': 0.413274347782135, 'reward_meter_mean': 0.7434380054473877, 'reward_meter_std': 0.25848403573036194, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9545454382896423, 'reward_repeat_soft_std': 0.014728941954672337, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.413274347782135, 'reward_total_composite_std': 0.022209085524082184, 'epoch': 0.2}

 62%|██████▏   | 2039/3300 [5:00:08<2:17:18,  6.53s/it]INFO 04-13 12:21:06 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2040/3300 [5:00:15<2:18:34,  6.60s/it]2026-04-13 12:21:13,422 | INFO | train_grpo_train | metrics_logged mode=train step=2040

                                                       
{'loss': 0.0466, 'grad_norm': 7.104791164398193, 'learning_rate': 3.821212121212122e-06, 'num_tokens': 3668966.0, 'completions/mean_length': 54.375, 'completions/min_length': 46.0, 'completions/max_length': 65.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 54.375, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 65.0, 'rewards/meter/mean': 0.9765020608901978, 'rewards/meter/std': 0.016999514773488045, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8627753257751465, 'rewards/repeat_soft/std': 0.033737439662218094, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.43422314524650574, 'rewards/total_composite/std': 0.011782454326748848, 'reward': 0.43422314524650574, 'reward_std': 0.011782454326748848, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10320676863193512, 'sampling/sampling_logp_difference/max': 1.7726988792419434, 'sampling/importance_sampling_ratio/min': 0.169873908162117, 'sampling/importance_sampling_ratio/mean': 1.0088846683502197, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5597806870937347, 'clip_ratio/low_mean': 0.047498751897364855, 'clip_ratio/low_min': 0.047498751897364855, 'clip_ratio/high_mean': 0.058153492864221334, 'clip_ratio/high_max': 0.058153492864221334, 'clip_ratio/region_mean': 0.10565224476158619, 'reward_total_mean': 0.43422314524650574, 'reward_meter_mean': 0.9765020608901978, 'reward_meter_std': 0.016999514773488045, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8627753257751465, 'reward_repeat_soft_std': 0.033737439662218094, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.43422314524650574, 'reward_total_composite_std': 0.011782454326748848, 'epoch': 0.2}

 62%|██████▏   | 2040/3300 [5:00:15<2:18:34,  6.60s/it]INFO 04-13 12:21:13 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2041/3300 [5:00:23<2:24:53,  6.91s/it]2026-04-13 12:21:21,048 | INFO | train_grpo_train | metrics_logged mode=train step=2041

                                                       
{'loss': 0.004, 'grad_norm': 8.011926651000977, 'learning_rate': 3.818181818181819e-06, 'num_tokens': 3671616.0, 'completions/mean_length': 138.25, 'completions/min_length': 127.0, 'completions/max_length': 150.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 138.25, 'completions/min_terminated_length': 127.0, 'completions/max_terminated_length': 150.0, 'rewards/meter/mean': 0.9829698801040649, 'rewards/meter/std': 0.004708987195044756, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9096139073371887, 'rewards/repeat_soft/std': 0.023690275847911835, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.4442504644393921, 'rewards/total_composite/std': 0.01413500215858221, 'reward': 0.4442504644393921, 'reward_std': 0.014134998433291912, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0858384296298027, 'sampling/sampling_logp_difference/max': 2.0158839225769043, 'sampling/importance_sampling_ratio/min': 0.13320261240005493, 'sampling/importance_sampling_ratio/mean': 1.0017367601394653, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3669454865157604, 'clip_ratio/low_mean': 0.022644356824457645, 'clip_ratio/low_min': 0.022644356824457645, 'clip_ratio/high_mean': 0.07045797212049365, 'clip_ratio/high_max': 0.07045797212049365, 'clip_ratio/region_mean': 0.0931023289449513, 'reward_total_mean': 0.4442504644393921, 'reward_meter_mean': 0.9829698801040649, 'reward_meter_std': 0.004708987195044756, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9096139073371887, 'reward_repeat_soft_std': 0.023690275847911835, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.4442504644393921, 'reward_total_composite_std': 0.01413500215858221, 'epoch': 0.21}

 62%|██████▏   | 2041/3300 [5:00:23<2:24:53,  6.91s/it]INFO 04-13 12:21:21 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2042/3300 [5:00:29<2:22:21,  6.79s/it]2026-04-13 12:21:27,577 | INFO | train_grpo_train | metrics_logged mode=train step=2042

                                                       
{'loss': 0.0166, 'grad_norm': 14.088408470153809, 'learning_rate': 3.8151515151515155e-06, 'num_tokens': 3673257.0, 'completions/mean_length': 36.125, 'completions/min_length': 32.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8494052290916443, 'rewards/meter/std': 0.2127019166946411, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9717245101928711, 'rewards/repeat_soft/std': 0.03682416304945946, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4294499158859253, 'rewards/total_composite/std': 0.023655403405427933, 'reward': 0.4294499158859253, 'reward_std': 0.023655403405427933, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09994067251682281, 'sampling/sampling_logp_difference/max': 1.8048105239868164, 'sampling/importance_sampling_ratio/min': 0.16450563073158264, 'sampling/importance_sampling_ratio/mean': 0.9966678023338318, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3427230603992939, 'clip_ratio/low_mean': 0.020891986088827252, 'clip_ratio/low_min': 0.020891986088827252, 'clip_ratio/high_mean': 0.0672743059694767, 'clip_ratio/high_max': 0.0672743059694767, 'clip_ratio/region_mean': 0.08816629205830395, 'reward_total_mean': 0.4294499158859253, 'reward_meter_mean': 0.8494052290916443, 'reward_meter_std': 0.2127019166946411, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9717245101928711, 'reward_repeat_soft_std': 0.03682416304945946, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4294499158859253, 'reward_total_composite_std': 0.023655403405427933, 'epoch': 0.21}

 62%|██████▏   | 2042/3300 [5:00:29<2:22:21,  6.79s/it]INFO 04-13 12:21:27 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2043/3300 [5:00:36<2:20:52,  6.72s/it]2026-04-13 12:21:34,136 | INFO | train_grpo_train | metrics_logged mode=train step=2043

                                                       
{'loss': 0.0335, 'grad_norm': 7.588534832000732, 'learning_rate': 3.8121212121212127e-06, 'num_tokens': 3674964.0, 'completions/mean_length': 59.375, 'completions/min_length': 54.0, 'completions/max_length': 64.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.375, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 64.0, 'rewards/meter/mean': 0.9905557036399841, 'rewards/meter/std': 0.003954746760427952, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9122059345245361, 'rewards/repeat_soft/std': 0.05901812016963959, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.4382306933403015, 'rewards/total_composite/std': 0.011901028454303741, 'reward': 0.4382306933403015, 'reward_std': 0.011901023797690868, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07994444668292999, 'sampling/sampling_logp_difference/max': 1.2873473167419434, 'sampling/importance_sampling_ratio/min': 0.2760019600391388, 'sampling/importance_sampling_ratio/mean': 0.9986298084259033, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41187846660614014, 'clip_ratio/low_mean': 0.037985481321811676, 'clip_ratio/low_min': 0.037985481321811676, 'clip_ratio/high_mean': 0.03362857084721327, 'clip_ratio/high_max': 0.03362857084721327, 'clip_ratio/region_mean': 0.07161405216902494, 'reward_total_mean': 0.4382306933403015, 'reward_meter_mean': 0.9905557036399841, 'reward_meter_std': 0.003954746760427952, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9122059345245361, 'reward_repeat_soft_std': 0.05901812016963959, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.4382306933403015, 'reward_total_composite_std': 0.011901028454303741, 'epoch': 0.21}

 62%|██████▏   | 2043/3300 [5:00:36<2:20:52,  6.72s/it]INFO 04-13 12:21:34 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2044/3300 [5:00:43<2:24:22,  6.90s/it]2026-04-13 12:21:41,450 | INFO | train_grpo_train | metrics_logged mode=train step=2044

                                                       
{'loss': 0.2662, 'grad_norm': 15.66690444946289, 'learning_rate': 3.8090909090909095e-06, 'num_tokens': 3676748.0, 'completions/mean_length': 59.0, 'completions/min_length': 45.0, 'completions/max_length': 104.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 59.0, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 104.0, 'rewards/meter/mean': 0.6437088847160339, 'rewards/meter/std': 0.34713518619537354, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8980014324188232, 'rewards/repeat_soft/std': 0.019471298903226852, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.013887306675314903, 'rewards/total_composite/mean': 0.3829920291900635, 'rewards/total_composite/std': 0.0643611028790474, 'reward': 0.3829920291900635, 'reward_std': 0.0643610954284668, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10684850811958313, 'sampling/sampling_logp_difference/max': 2.7498672008514404, 'sampling/importance_sampling_ratio/min': 0.06393635272979736, 'sampling/importance_sampling_ratio/mean': 1.0027379989624023, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.517706822603941, 'clip_ratio/low_mean': 0.04134615417569876, 'clip_ratio/low_min': 0.04134615417569876, 'clip_ratio/high_mean': 0.06284213112667203, 'clip_ratio/high_max': 0.06284213112667203, 'clip_ratio/region_mean': 0.10418828530237079, 'reward_total_mean': 0.3829920291900635, 'reward_meter_mean': 0.6437088847160339, 'reward_meter_std': 0.34713518619537354, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8980014324188232, 'reward_repeat_soft_std': 0.019471298903226852, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.013887306675314903, 'reward_total_composite_mean': 0.3829920291900635, 'reward_total_composite_std': 0.0643611028790474, 'epoch': 0.21}

 62%|██████▏   | 2044/3300 [5:00:43<2:24:22,  6.90s/it]INFO 04-13 12:21:41 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2045/3300 [5:00:51<2:30:14,  7.18s/it]2026-04-13 12:21:49,293 | INFO | train_grpo_train | metrics_logged mode=train step=2045

                                                       
{'loss': 0.1668, 'grad_norm': 5.523110866546631, 'learning_rate': 3.8060606060606064e-06, 'num_tokens': 3679184.0, 'completions/mean_length': 142.5, 'completions/min_length': 102.0, 'completions/max_length': 180.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 142.5, 'completions/min_terminated_length': 102.0, 'completions/max_terminated_length': 180.0, 'rewards/meter/mean': 0.974332332611084, 'rewards/meter/std': 0.021938016638159752, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.842339277267456, 'rewards/repeat_soft/std': 0.056603413075208664, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.031052954494953156, 'rewards/total_composite/mean': 0.37899357080459595, 'rewards/total_composite/std': 0.03571334108710289, 'reward': 0.37899357080459595, 'reward_std': 0.03571334481239319, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08776529133319855, 'sampling/sampling_logp_difference/max': 2.9584617614746094, 'sampling/importance_sampling_ratio/min': 0.051898688077926636, 'sampling/importance_sampling_ratio/mean': 1.0044498443603516, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3657061904668808, 'clip_ratio/low_mean': 0.04262035293504596, 'clip_ratio/low_min': 0.04262035293504596, 'clip_ratio/high_mean': 0.028448880184441805, 'clip_ratio/high_max': 0.028448880184441805, 'clip_ratio/region_mean': 0.07106923311948776, 'reward_total_mean': 0.37899357080459595, 'reward_meter_mean': 0.974332332611084, 'reward_meter_std': 0.021938016638159752, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.842339277267456, 'reward_repeat_soft_std': 0.056603413075208664, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.031052954494953156, 'reward_total_composite_mean': 0.37899357080459595, 'reward_total_composite_std': 0.03571334108710289, 'epoch': 0.21}

 62%|██████▏   | 2045/3300 [5:00:51<2:30:14,  7.18s/it]INFO 04-13 12:21:49 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2046/3300 [5:00:58<2:30:30,  7.20s/it]2026-04-13 12:21:56,536 | INFO | train_grpo_train | metrics_logged mode=train step=2046

                                                       
{'loss': 0.1164, 'grad_norm': 6.8209147453308105, 'learning_rate': 3.803030303030303e-06, 'num_tokens': 3681523.0, 'completions/mean_length': 115.375, 'completions/min_length': 101.0, 'completions/max_length': 147.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.375, 'completions/min_terminated_length': 101.0, 'completions/max_terminated_length': 147.0, 'rewards/meter/mean': 0.9768247604370117, 'rewards/meter/std': 0.0234832800924778, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8576970100402832, 'rewards/repeat_soft/std': 0.06382130086421967, 'rewards/judge_quality/mean': 0.14250001311302185, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.40669384598731995, 'rewards/total_composite/std': 0.035279784351587296, 'reward': 0.40669384598731995, 'reward_std': 0.035279788076877594, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10315954685211182, 'sampling/sampling_logp_difference/max': 4.479145526885986, 'sampling/importance_sampling_ratio/min': 0.011343101970851421, 'sampling/importance_sampling_ratio/mean': 0.9935485124588013, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4440900795161724, 'clip_ratio/low_mean': 0.0415108953602612, 'clip_ratio/low_min': 0.0415108953602612, 'clip_ratio/high_mean': 0.05447726044803858, 'clip_ratio/high_max': 0.05447726044803858, 'clip_ratio/region_mean': 0.09598815580829978, 'reward_total_mean': 0.40669384598731995, 'reward_meter_mean': 0.9768247604370117, 'reward_meter_std': 0.0234832800924778, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8576970100402832, 'reward_repeat_soft_std': 0.06382130086421967, 'reward_judge_quality_mean': 0.14250001311302185, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.40669384598731995, 'reward_total_composite_std': 0.035279784351587296, 'epoch': 0.21}

 62%|██████▏   | 2046/3300 [5:00:58<2:30:30,  7.20s/it]INFO 04-13 12:21:56 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2047/3300 [5:01:09<2:55:42,  8.41s/it]2026-04-13 12:22:07,775 | INFO | train_grpo_train | metrics_logged mode=train step=2047

                                                       
{'loss': -0.1227, 'grad_norm': 1.9844030141830444, 'learning_rate': 3.8000000000000005e-06, 'num_tokens': 3683165.0, 'completions/mean_length': 100.25, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 41.42857360839844, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8852245807647705, 'rewards/meter/std': 0.30692076683044434, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8760888576507568, 'rewards/repeat_soft/std': 0.06388038396835327, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.41670188307762146, 'rewards/total_composite/std': 0.028256313875317574, 'reward': 0.41670188307762146, 'reward_std': 0.02825631946325302, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07085798680782318, 'sampling/sampling_logp_difference/max': 0.6945812702178955, 'sampling/importance_sampling_ratio/min': 0.5122395157814026, 'sampling/importance_sampling_ratio/mean': 1.0203478336334229, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.38697320967912674, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.05263052717782557, 'clip_ratio/high_max': 0.05263052717782557, 'clip_ratio/region_mean': 0.05263052717782557, 'reward_total_mean': 0.41670188307762146, 'reward_meter_mean': 0.8852245807647705, 'reward_meter_std': 0.30692076683044434, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8760888576507568, 'reward_repeat_soft_std': 0.06388038396835327, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.41670188307762146, 'reward_total_composite_std': 0.028256313875317574, 'epoch': 0.21}

 62%|██████▏   | 2047/3300 [5:01:10<2:55:42,  8.41s/it]INFO 04-13 12:22:08 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2048/3300 [5:01:16<2:43:40,  7.84s/it]2026-04-13 12:22:14,294 | INFO | train_grpo_train | metrics_logged mode=train step=2048

                                                       
{'loss': -0.0277, 'grad_norm': 7.260988235473633, 'learning_rate': 3.7969696969696973e-06, 'num_tokens': 3685202.0, 'completions/mean_length': 74.625, 'completions/min_length': 64.0, 'completions/max_length': 84.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 74.625, 'completions/min_terminated_length': 64.0, 'completions/max_terminated_length': 84.0, 'rewards/meter/mean': 0.9884733557701111, 'rewards/meter/std': 0.009229067713022232, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8281691670417786, 'rewards/repeat_soft/std': 0.028537092730402946, 'rewards/judge_quality/mean': 0.21000000834465027, 'rewards/judge_quality/std': 0.08485280722379684, 'rewards/total_composite/mean': 0.45912399888038635, 'rewards/total_composite/std': 0.05568560212850571, 'reward': 0.45912399888038635, 'reward_std': 0.05568559467792511, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07889800518751144, 'sampling/sampling_logp_difference/max': 1.560380458831787, 'sampling/importance_sampling_ratio/min': 0.2100561261177063, 'sampling/importance_sampling_ratio/mean': 1.0017101764678955, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46210258826613426, 'clip_ratio/low_mean': 0.06502922880463302, 'clip_ratio/low_min': 0.06502922880463302, 'clip_ratio/high_mean': 0.0074404762126505375, 'clip_ratio/high_max': 0.0074404762126505375, 'clip_ratio/region_mean': 0.07246970501728356, 'reward_total_mean': 0.45912399888038635, 'reward_meter_mean': 0.9884733557701111, 'reward_meter_std': 0.009229067713022232, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8281691670417786, 'reward_repeat_soft_std': 0.028537092730402946, 'reward_judge_quality_mean': 0.21000000834465027, 'reward_judge_quality_std': 0.08485280722379684, 'reward_total_composite_mean': 0.45912399888038635, 'reward_total_composite_std': 0.05568560212850571, 'epoch': 0.21}

 62%|██████▏   | 2048/3300 [5:01:16<2:43:40,  7.84s/it]INFO 04-13 12:22:14 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2049/3300 [5:01:22<2:35:06,  7.44s/it]2026-04-13 12:22:20,775 | INFO | train_grpo_train | metrics_logged mode=train step=2049

                                                       
{'loss': -0.0284, 'grad_norm': 10.234697341918945, 'learning_rate': 3.793939393939394e-06, 'num_tokens': 3687043.0, 'completions/mean_length': 63.125, 'completions/min_length': 54.0, 'completions/max_length': 72.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 72.0, 'rewards/meter/mean': 0.2560454308986664, 'rewards/meter/std': 0.2626054286956787, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8586742281913757, 'rewards/repeat_soft/std': 0.03004816733300686, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.35626357793807983, 'rewards/total_composite/std': 0.027345743030309677, 'reward': 0.35626357793807983, 'reward_std': 0.027345741167664528, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0907815545797348, 'sampling/sampling_logp_difference/max': 1.9165549278259277, 'sampling/importance_sampling_ratio/min': 0.1471129208803177, 'sampling/importance_sampling_ratio/mean': 0.9916093945503235, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.33177791349589825, 'clip_ratio/low_mean': 0.04983414337038994, 'clip_ratio/low_min': 0.04983414337038994, 'clip_ratio/high_mean': 0.028859290294349194, 'clip_ratio/high_max': 0.028859290294349194, 'clip_ratio/region_mean': 0.07869343366473913, 'reward_total_mean': 0.35626357793807983, 'reward_meter_mean': 0.2560454308986664, 'reward_meter_std': 0.2626054286956787, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8586742281913757, 'reward_repeat_soft_std': 0.03004816733300686, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.35626357793807983, 'reward_total_composite_std': 0.027345743030309677, 'epoch': 0.21}

 62%|██████▏   | 2049/3300 [5:01:23<2:35:06,  7.44s/it]INFO 04-13 12:22:21 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2050/3300 [5:01:30<2:35:59,  7.49s/it]2026-04-13 12:22:28,374 | INFO | train_grpo_train | metrics_logged mode=train step=2050

                                                       
{'loss': 0.1111, 'grad_norm': 7.519757270812988, 'learning_rate': 3.7909090909090914e-06, 'num_tokens': 3689510.0, 'completions/mean_length': 124.375, 'completions/min_length': 104.0, 'completions/max_length': 154.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 124.375, 'completions/min_terminated_length': 104.0, 'completions/max_terminated_length': 154.0, 'rewards/meter/mean': 0.7648420333862305, 'rewards/meter/std': 0.34862419962882996, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8720245957374573, 'rewards/repeat_soft/std': 0.041297126561403275, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.031052954494953156, 'rewards/total_composite/mean': 0.39393067359924316, 'rewards/total_composite/std': 0.030269678682088852, 'reward': 0.39393067359924316, 'reward_std': 0.030269671231508255, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11565042287111282, 'sampling/sampling_logp_difference/max': 2.13187837600708, 'sampling/importance_sampling_ratio/min': 0.11861427873373032, 'sampling/importance_sampling_ratio/mean': 1.0017586946487427, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5740426108241081, 'clip_ratio/low_mean': 0.02805569628253579, 'clip_ratio/low_min': 0.02805569628253579, 'clip_ratio/high_mean': 0.062328214291483164, 'clip_ratio/high_max': 0.062328214291483164, 'clip_ratio/region_mean': 0.09038391057401896, 'reward_total_mean': 0.39393067359924316, 'reward_meter_mean': 0.7648420333862305, 'reward_meter_std': 0.34862419962882996, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8720245957374573, 'reward_repeat_soft_std': 0.041297126561403275, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.031052954494953156, 'reward_total_composite_mean': 0.39393067359924316, 'reward_total_composite_std': 0.030269678682088852, 'epoch': 0.21}

 62%|██████▏   | 2050/3300 [5:01:30<2:35:59,  7.49s/it]INFO 04-13 12:22:28 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:15,  1.88s/it]

 30%|███       | 3/10 [00:08<00:21,  3.07s/it]

 40%|████      | 4/10 [00:17<00:31,  5.21s/it]

 50%|█████     | 5/10 [00:21<00:24,  4.94s/it]

 60%|██████    | 6/10 [00:30<00:24,  6.21s/it]

 70%|███████   | 7/10 [00:35<00:17,  5.95s/it]

 80%|████████  | 8/10 [00:40<00:10,  5.45s/it]

 90%|█████████ | 9/10 [00:49<00:06,  6.51s/it]

100%|██████████| 10/10 [00:52<00:00,  5.60s/it]2026-04-13 12:23:30,157 | INFO | train_grpo_train | metrics_logged mode=eval step=2050

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 61.6625, 'eval_samples_per_second': 1.297, 'eval_steps_per_second': 0.162, 'eval_num_tokens': 3689510.0, 'eval_completions/mean_length': 104.15, 'eval_completions/min_length': 32.8, 'eval_completions/max_length': 299.5, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 82.56428756713868, 'eval_completions/min_terminated_length': 32.8, 'eval_completions/max_terminated_length': 152.1, 'eval_rewards/meter/mean': 0.8470357298851013, 'eval_rewards/meter/std': 0.210162253677845, 'eval_rewards/count_adherence/mean': 0.9524999976158142, 'eval_rewards/count_adherence/std': 0.10983712449669839, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.1414213538169861, 'eval_rewards/repeat_soft/mean': 0.881467753648758, 'eval_rewards/repeat_soft/std': 0.07334110662341117, 'eval_rewards/judge_quality/mean': 0.15550000816583634, 'eval_rewards/judge_quality/std': 0.028039283119142056, 'eval_rewards/total_composite/mean': 0.39409635961055756, 'eval_rewards/total_composite/std': 0.07721717413514853, 'eval_reward': 0.39409635961055756, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.0478603333234787, 'eval_sampling/sampling_logp_difference/max': 1.110507845878601, 'eval_sampling/importance_sampling_ratio/min': 0.35095261931419375, 'eval_sampling/importance_sampling_ratio/mean': 1.0076323390007018, 'eval_sampling/importance_sampling_ratio/max': 1.4491962432861327, 'eval_entropy': 0.48588710129261015, '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': 0.39409635961055756, 'eval_reward_meter_mean': 0.8470357298851013, 'eval_reward_meter_std': 0.210162253677845, 'eval_reward_count_adherence_mean': 0.9524999976158142, 'eval_reward_count_adherence_std': 0.10983712449669839, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.1414213538169861, 'eval_reward_repeat_soft_mean': 0.881467753648758, 'eval_reward_repeat_soft_std': 0.07334110662341117, 'eval_reward_judge_quality_mean': 0.15550000816583634, 'eval_reward_judge_quality_std': 0.028039283119142056, 'eval_reward_total_composite_mean': 0.39409635961055756, 'eval_reward_total_composite_std': 0.07721717413514853, 'epoch': 0.21}

 62%|██████▏   | 2050/3300 [5:02:32<2:35:59,  7.49s/it]

100%|██████████| 10/10 [00:52<00:00,  5.60s/it]

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

 62%|██████▏   | 2051/3300 [5:02:42<9:16:21, 26.73s/it]2026-04-13 12:23:40,012 | INFO | train_grpo_train | metrics_logged mode=train step=2051

                                                       
{'loss': 0.0993, 'grad_norm': 9.6243257522583, 'learning_rate': 3.7878787878787882e-06, 'num_tokens': 3691213.0, 'completions/mean_length': 55.875, 'completions/min_length': 48.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 55.875, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9860696792602539, 'rewards/meter/std': 0.010559221729636192, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9134825468063354, 'rewards/repeat_soft/std': 0.0551791712641716, 'rewards/judge_quality/mean': 0.23250001668930054, 'rewards/judge_quality/std': 0.1348808854818344, 'rewards/total_composite/mean': 0.4856888949871063, 'rewards/total_composite/std': 0.09246828407049179, 'reward': 0.4856888949871063, 'reward_std': 0.09246828407049179, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10899222642183304, 'sampling/sampling_logp_difference/max': 1.9581661224365234, 'sampling/importance_sampling_ratio/min': 0.14111697673797607, 'sampling/importance_sampling_ratio/mean': 1.0144563913345337, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5422843880951405, 'clip_ratio/low_mean': 0.07558443117886782, 'clip_ratio/low_min': 0.07558443117886782, 'clip_ratio/high_mean': 0.01791666680946946, 'clip_ratio/high_max': 0.01791666680946946, 'clip_ratio/region_mean': 0.09350109798833728, 'reward_total_mean': 0.4856888949871063, 'reward_meter_mean': 0.9860696792602539, 'reward_meter_std': 0.010559221729636192, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9134825468063354, 'reward_repeat_soft_std': 0.0551791712641716, 'reward_judge_quality_mean': 0.23250001668930054, 'reward_judge_quality_std': 0.1348808854818344, 'reward_total_composite_mean': 0.4856888949871063, 'reward_total_composite_std': 0.09246828407049179, 'epoch': 0.21}

 62%|██████▏   | 2051/3300 [5:02:42<9:16:21, 26.73s/it]INFO 04-13 12:23:40 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2052/3300 [5:02:54<7:43:36, 22.29s/it]2026-04-13 12:23:51,934 | INFO | train_grpo_train | metrics_logged mode=train step=2052

                                                       
{'loss': -0.0861, 'grad_norm': 3.488429307937622, 'learning_rate': 3.784848484848485e-06, 'num_tokens': 3693351.0, 'completions/mean_length': 158.25, 'completions/min_length': 85.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 107.71428680419922, 'completions/min_terminated_length': 85.0, 'completions/max_terminated_length': 132.0, 'rewards/meter/mean': 0.7394720911979675, 'rewards/meter/std': 0.34064599871635437, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1157275140285492, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.8292969465255737, 'rewards/repeat_soft/std': 0.09894319623708725, 'rewards/judge_quality/mean': 0.14125001430511475, 'rewards/judge_quality/std': 0.04454132169485092, 'rewards/total_composite/mean': 0.294907808303833, 'rewards/total_composite/std': 0.18495747447013855, 'reward': 0.294907808303833, 'reward_std': 0.18495745956897736, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08565527200698853, 'sampling/sampling_logp_difference/max': 2.172407865524292, 'sampling/importance_sampling_ratio/min': 0.11390302330255508, 'sampling/importance_sampling_ratio/mean': 1.0037480592727661, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4531273767352104, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.06291623646393418, 'clip_ratio/high_max': 0.06291623646393418, 'clip_ratio/region_mean': 0.07049199426546693, 'reward_total_mean': 0.294907808303833, 'reward_meter_mean': 0.7394720911979675, 'reward_meter_std': 0.34064599871635437, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1157275140285492, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.8292969465255737, 'reward_repeat_soft_std': 0.09894319623708725, 'reward_judge_quality_mean': 0.14125001430511475, 'reward_judge_quality_std': 0.04454132169485092, 'reward_total_composite_mean': 0.294907808303833, 'reward_total_composite_std': 0.18495747447013855, 'epoch': 0.21}

 62%|██████▏   | 2052/3300 [5:02:54<7:43:36, 22.29s/it]INFO 04-13 12:23:52 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2053/3300 [5:03:01<6:09:54, 17.80s/it]2026-04-13 12:23:59,264 | INFO | train_grpo_train | metrics_logged mode=train step=2053

                                                       
{'loss': 0.0299, 'grad_norm': 19.44232177734375, 'learning_rate': 3.781818181818182e-06, 'num_tokens': 3694749.0, 'completions/mean_length': 18.75, 'completions/min_length': 17.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.75, 'completions/min_terminated_length': 17.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.9142047166824341, 'rewards/meter/std': 0.13621951639652252, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9537500143051147, 'rewards/repeat_soft/std': 0.02474873699247837, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4321974515914917, 'rewards/total_composite/std': 0.013160732574760914, 'reward': 0.4321974515914917, 'reward_std': 0.013160724192857742, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06499005854129791, 'sampling/sampling_logp_difference/max': 1.1320464611053467, 'sampling/importance_sampling_ratio/min': 0.3223728537559509, 'sampling/importance_sampling_ratio/mean': 1.0129449367523193, 'sampling/importance_sampling_ratio/max': 1.500510811805725, 'entropy': 0.4856715239584446, 'clip_ratio/low_mean': 0.047412280924618244, 'clip_ratio/low_min': 0.047412280924618244, 'clip_ratio/high_mean': 0.03345588315278292, 'clip_ratio/high_max': 0.03345588315278292, 'clip_ratio/region_mean': 0.08086816407740116, 'reward_total_mean': 0.4321974515914917, 'reward_meter_mean': 0.9142047166824341, 'reward_meter_std': 0.13621951639652252, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9537500143051147, 'reward_repeat_soft_std': 0.02474873699247837, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4321974515914917, 'reward_total_composite_std': 0.013160732574760914, 'epoch': 0.21}

 62%|██████▏   | 2053/3300 [5:03:01<6:09:54, 17.80s/it]INFO 04-13 12:23:59 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2054/3300 [5:03:08<5:01:19, 14.51s/it]2026-04-13 12:24:06,098 | INFO | train_grpo_train | metrics_logged mode=train step=2054

                                                       
{'loss': 0.0103, 'grad_norm': 7.740578651428223, 'learning_rate': 3.778787878787879e-06, 'num_tokens': 3696438.0, 'completions/mean_length': 63.125, 'completions/min_length': 54.0, 'completions/max_length': 73.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.125, 'completions/min_terminated_length': 54.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.9650364518165588, 'rewards/meter/std': 0.018497325479984283, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9724728465080261, 'rewards/repeat_soft/std': 0.017807554453611374, 'rewards/judge_quality/mean': 0.16625000536441803, 'rewards/judge_quality/std': 0.035431019961833954, 'rewards/total_composite/mean': 0.4502558708190918, 'rewards/total_composite/std': 0.023599883541464806, 'reward': 0.4502558708190918, 'reward_std': 0.023599879816174507, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0975913405418396, 'sampling/sampling_logp_difference/max': 1.569288969039917, 'sampling/importance_sampling_ratio/min': 0.20819316804409027, 'sampling/importance_sampling_ratio/mean': 1.0246812105178833, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4856261573731899, 'clip_ratio/low_mean': 0.054329875856637955, 'clip_ratio/low_min': 0.054329875856637955, 'clip_ratio/high_mean': 0.0265895277261734, 'clip_ratio/high_max': 0.0265895277261734, 'clip_ratio/region_mean': 0.08091940358281136, 'reward_total_mean': 0.4502558708190918, 'reward_meter_mean': 0.9650364518165588, 'reward_meter_std': 0.018497325479984283, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9724728465080261, 'reward_repeat_soft_std': 0.017807554453611374, 'reward_judge_quality_mean': 0.16625000536441803, 'reward_judge_quality_std': 0.035431019961833954, 'reward_total_composite_mean': 0.4502558708190918, 'reward_total_composite_std': 0.023599883541464806, 'epoch': 0.21}

 62%|██████▏   | 2054/3300 [5:03:08<5:01:19, 14.51s/it]INFO 04-13 12:24:06 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2055/3300 [5:03:15<4:17:41, 12.42s/it]2026-04-13 12:24:13,626 | INFO | train_grpo_train | metrics_logged mode=train step=2055

                                                       
{'loss': 0.1448, 'grad_norm': 7.970420837402344, 'learning_rate': 3.775757575757576e-06, 'num_tokens': 3698532.0, 'completions/mean_length': 88.75, 'completions/min_length': 72.0, 'completions/max_length': 122.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.75, 'completions/min_terminated_length': 72.0, 'completions/max_terminated_length': 122.0, 'rewards/meter/mean': 0.9468092322349548, 'rewards/meter/std': 0.06834986060857773, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9011238217353821, 'rewards/repeat_soft/std': 0.041544847190380096, 'rewards/judge_quality/mean': 0.18000000715255737, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43969523906707764, 'rewards/total_composite/std': 0.019526204094290733, 'reward': 0.43969523906707764, 'reward_std': 0.019526205956935883, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09225019812583923, 'sampling/sampling_logp_difference/max': 2.155517101287842, 'sampling/importance_sampling_ratio/min': 0.1158432736992836, 'sampling/importance_sampling_ratio/mean': 1.0105295181274414, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4102574083954096, 'clip_ratio/low_mean': 0.030017105396836996, 'clip_ratio/low_min': 0.030017105396836996, 'clip_ratio/high_mean': 0.04154863650910556, 'clip_ratio/high_max': 0.04154863650910556, 'clip_ratio/region_mean': 0.07156574190594256, 'reward_total_mean': 0.43969523906707764, 'reward_meter_mean': 0.9468092322349548, 'reward_meter_std': 0.06834986060857773, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9011238217353821, 'reward_repeat_soft_std': 0.041544847190380096, 'reward_judge_quality_mean': 0.18000000715255737, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43969523906707764, 'reward_total_composite_std': 0.019526204094290733, 'epoch': 0.21}

 62%|██████▏   | 2055/3300 [5:03:15<4:17:41, 12.42s/it]INFO 04-13 12:24:13 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2056/3300 [5:03:27<4:12:18, 12.17s/it]2026-04-13 12:24:25,223 | INFO | train_grpo_train | metrics_logged mode=train step=2056

                                                       
{'loss': -0.2314, 'grad_norm': 1.922808051109314, 'learning_rate': 3.772727272727273e-06, 'num_tokens': 3701297.0, 'completions/mean_length': 198.625, 'completions/min_length': 137.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 153.85714721679688, 'completions/min_terminated_length': 137.0, 'completions/max_terminated_length': 167.0, 'rewards/meter/mean': 0.8928427696228027, 'rewards/meter/std': 0.1732001155614853, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.1414213478565216, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9112764596939087, 'rewards/repeat_soft/std': 0.03366226330399513, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.047790467739105225, 'rewards/total_composite/mean': 0.37900832295417786, 'rewards/total_composite/std': 0.1544741690158844, 'reward': 0.37900832295417786, 'reward_std': 0.1544741690158844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11159887164831161, 'sampling/sampling_logp_difference/max': 2.3941075801849365, 'sampling/importance_sampling_ratio/min': 0.1788671761751175, 'sampling/importance_sampling_ratio/mean': 1.0042688846588135, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49233855307102203, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.10130473598837852, 'clip_ratio/high_max': 0.10130473598837852, 'clip_ratio/region_mean': 0.10130473598837852, 'reward_total_mean': 0.37900832295417786, 'reward_meter_mean': 0.8928427696228027, 'reward_meter_std': 0.1732001155614853, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.1414213478565216, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9112764596939087, 'reward_repeat_soft_std': 0.03366226330399513, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.047790467739105225, 'reward_total_composite_mean': 0.37900832295417786, 'reward_total_composite_std': 0.1544741690158844, 'epoch': 0.21}

 62%|██████▏   | 2056/3300 [5:03:27<4:12:18, 12.17s/it]INFO 04-13 12:24:25 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2057/3300 [5:03:33<3:37:12, 10.48s/it]2026-04-13 12:24:31,778 | INFO | train_grpo_train | metrics_logged mode=train step=2057

                                                       
{'loss': 0.0255, 'grad_norm': 11.074159622192383, 'learning_rate': 3.76969696969697e-06, 'num_tokens': 3702852.0, 'completions/mean_length': 41.375, 'completions/min_length': 32.0, 'completions/max_length': 48.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.375, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 48.0, 'rewards/meter/mean': 0.9662160277366638, 'rewards/meter/std': 0.046560220420360565, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9596341848373413, 'rewards/repeat_soft/std': 0.04224960133433342, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43815118074417114, 'rewards/total_composite/std': 0.007245528046041727, 'reward': 0.43815118074417114, 'reward_std': 0.00724552758038044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09605085104703903, 'sampling/sampling_logp_difference/max': 1.3370459079742432, 'sampling/importance_sampling_ratio/min': 0.26262032985687256, 'sampling/importance_sampling_ratio/mean': 1.0031450986862183, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.575896430760622, 'clip_ratio/low_mean': 0.02825218066573143, 'clip_ratio/low_min': 0.02825218066573143, 'clip_ratio/high_mean': 0.08676579128950834, 'clip_ratio/high_max': 0.08676579128950834, 'clip_ratio/region_mean': 0.11501797195523977, 'reward_total_mean': 0.43815118074417114, 'reward_meter_mean': 0.9662160277366638, 'reward_meter_std': 0.046560220420360565, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9596341848373413, 'reward_repeat_soft_std': 0.04224960133433342, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43815118074417114, 'reward_total_composite_std': 0.007245528046041727, 'epoch': 0.21}

 62%|██████▏   | 2057/3300 [5:03:34<3:37:12, 10.48s/it]INFO 04-13 12:24:32 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2058/3300 [5:03:40<3:11:33,  9.25s/it]2026-04-13 12:24:38,154 | INFO | train_grpo_train | metrics_logged mode=train step=2058

                                                       
{'loss': 0.0517, 'grad_norm': 14.113903045654297, 'learning_rate': 3.766666666666667e-06, 'num_tokens': 3704402.0, 'completions/mean_length': 26.75, 'completions/min_length': 23.0, 'completions/max_length': 31.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 26.75, 'completions/min_terminated_length': 23.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.6191296577453613, 'rewards/meter/std': 0.3962414264678955, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9412171840667725, 'rewards/repeat_soft/std': 0.030817119404673576, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40154772996902466, 'rewards/total_composite/std': 0.037522632628679276, 'reward': 0.40154772996902466, 'reward_std': 0.03752264380455017, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11040031164884567, 'sampling/sampling_logp_difference/max': 1.6234591007232666, 'sampling/importance_sampling_ratio/min': 0.2405642867088318, 'sampling/importance_sampling_ratio/mean': 1.016933560371399, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6415188424289227, 'clip_ratio/low_mean': 0.050534188747406006, 'clip_ratio/low_min': 0.050534188747406006, 'clip_ratio/high_mean': 0.06094925617799163, 'clip_ratio/high_max': 0.06094925617799163, 'clip_ratio/region_mean': 0.11148344492539763, 'reward_total_mean': 0.40154772996902466, 'reward_meter_mean': 0.6191296577453613, 'reward_meter_std': 0.3962414264678955, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9412171840667725, 'reward_repeat_soft_std': 0.030817119404673576, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40154772996902466, 'reward_total_composite_std': 0.037522632628679276, 'epoch': 0.21}

 62%|██████▏   | 2058/3300 [5:03:40<3:11:33,  9.25s/it]INFO 04-13 12:24:38 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2059/3300 [5:03:51<3:24:17,  9.88s/it]2026-04-13 12:24:49,481 | INFO | train_grpo_train | metrics_logged mode=train step=2059

                                                       
{'loss': -0.129, 'grad_norm': 1.6379417181015015, 'learning_rate': 3.7636363636363637e-06, 'num_tokens': 3706168.0, 'completions/mean_length': 101.75, 'completions/min_length': 38.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 43.142860412597656, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.8695842027664185, 'rewards/meter/std': 0.35068050026893616, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8757412433624268, 'rewards/repeat_soft/std': 0.07615207135677338, 'rewards/judge_quality/mean': 0.13375000655651093, 'rewards/judge_quality/std': 0.03543102368712425, 'rewards/total_composite/mean': 0.3706586956977844, 'rewards/total_composite/std': 0.15045024454593658, 'reward': 0.3706586956977844, 'reward_std': 0.1504502296447754, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09259120374917984, 'sampling/sampling_logp_difference/max': 1.4251604080200195, 'sampling/importance_sampling_ratio/min': 0.24046988785266876, 'sampling/importance_sampling_ratio/mean': 0.992057740688324, 'sampling/importance_sampling_ratio/max': 1.8282526731491089, 'entropy': 0.35852643474936485, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09153049066662788, 'clip_ratio/high_max': 0.09153049066662788, 'clip_ratio/region_mean': 0.09153049066662788, 'reward_total_mean': 0.3706586956977844, 'reward_meter_mean': 0.8695842027664185, 'reward_meter_std': 0.35068050026893616, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8757412433624268, 'reward_repeat_soft_std': 0.07615207135677338, 'reward_judge_quality_mean': 0.13375000655651093, 'reward_judge_quality_std': 0.03543102368712425, 'reward_total_composite_mean': 0.3706586956977844, 'reward_total_composite_std': 0.15045024454593658, 'epoch': 0.21}

 62%|██████▏   | 2059/3300 [5:03:51<3:24:17,  9.88s/it]INFO 04-13 12:24:49 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2060/3300 [5:03:57<3:00:54,  8.75s/it]2026-04-13 12:24:55,631 | INFO | train_grpo_train | metrics_logged mode=train step=2060

                                                       
{'loss': 0.103, 'grad_norm': 12.3268461227417, 'learning_rate': 3.7606060606060605e-06, 'num_tokens': 3707710.0, 'completions/mean_length': 31.75, 'completions/min_length': 24.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.5625284910202026, 'rewards/meter/std': 0.37623330950737, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9688060283660889, 'rewards/repeat_soft/std': 0.020261460915207863, 'rewards/judge_quality/mean': 0.3425000309944153, 'rewards/judge_quality/std': 0.3564407527446747, 'rewards/total_composite/mean': 0.47080230712890625, 'rewards/total_composite/std': 0.1848701536655426, 'reward': 0.47080230712890625, 'reward_std': 0.1848701536655426, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11296968907117844, 'sampling/sampling_logp_difference/max': 2.78413724899292, 'sampling/importance_sampling_ratio/min': 0.061782367527484894, 'sampling/importance_sampling_ratio/mean': 1.0109044313430786, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4610154442489147, 'clip_ratio/low_mean': 0.08212722046300769, 'clip_ratio/low_min': 0.08212722046300769, 'clip_ratio/high_mean': 0.02083333395421505, 'clip_ratio/high_max': 0.02083333395421505, 'clip_ratio/region_mean': 0.10296055441722274, 'reward_total_mean': 0.47080230712890625, 'reward_meter_mean': 0.5625284910202026, 'reward_meter_std': 0.37623330950737, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9688060283660889, 'reward_repeat_soft_std': 0.020261460915207863, 'reward_judge_quality_mean': 0.3425000309944153, 'reward_judge_quality_std': 0.3564407527446747, 'reward_total_composite_mean': 0.47080230712890625, 'reward_total_composite_std': 0.1848701536655426, 'epoch': 0.21}

 62%|██████▏   | 2060/3300 [5:03:57<3:00:54,  8.75s/it]INFO 04-13 12:24:55 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2061/3300 [5:04:04<2:47:15,  8.10s/it]2026-04-13 12:25:02,188 | INFO | train_grpo_train | metrics_logged mode=train step=2061

                                                       
{'loss': -0.0142, 'grad_norm': 10.08353328704834, 'learning_rate': 3.757575757575758e-06, 'num_tokens': 3709564.0, 'completions/mean_length': 63.75, 'completions/min_length': 55.0, 'completions/max_length': 71.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 63.75, 'completions/min_terminated_length': 55.0, 'completions/max_terminated_length': 71.0, 'rewards/meter/mean': 0.47723761200904846, 'rewards/meter/std': 0.39061424136161804, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9800647497177124, 'rewards/repeat_soft/std': 0.0135366041213274, 'rewards/judge_quality/mean': 0.1912500113248825, 'rewards/judge_quality/std': 0.10507649928331375, 'rewards/total_composite/mean': 0.40029656887054443, 'rewards/total_composite/std': 0.038408465683460236, 'reward': 0.40029656887054443, 'reward_std': 0.03840846195816994, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12142689526081085, 'sampling/sampling_logp_difference/max': 1.6213834285736084, 'sampling/importance_sampling_ratio/min': 0.19762511551380157, 'sampling/importance_sampling_ratio/mean': 0.9930986762046814, 'sampling/importance_sampling_ratio/max': 1.9706506729125977, 'entropy': 0.8012176007032394, 'clip_ratio/low_mean': 0.06417649798095226, 'clip_ratio/low_min': 0.06417649798095226, 'clip_ratio/high_mean': 0.05263974145054817, 'clip_ratio/high_max': 0.05263974145054817, 'clip_ratio/region_mean': 0.11681623943150043, 'reward_total_mean': 0.40029656887054443, 'reward_meter_mean': 0.47723761200904846, 'reward_meter_std': 0.39061424136161804, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9800647497177124, 'reward_repeat_soft_std': 0.0135366041213274, 'reward_judge_quality_mean': 0.1912500113248825, 'reward_judge_quality_std': 0.10507649928331375, 'reward_total_composite_mean': 0.40029656887054443, 'reward_total_composite_std': 0.038408465683460236, 'epoch': 0.21}

 62%|██████▏   | 2061/3300 [5:04:04<2:47:15,  8.10s/it]INFO 04-13 12:25:02 [block_pool.py:378] Successfully reset prefix cache

 62%|██████▏   | 2062/3300 [5:04:10<2:36:28,  7.58s/it]2026-04-13 12:25:08,578 | INFO | train_grpo_train | metrics_logged mode=train step=2062

                                                       
{'loss': 0.0719, 'grad_norm': 10.390212059020996, 'learning_rate': 3.7545454545454546e-06, 'num_tokens': 3711010.0, 'completions/mean_length': 28.75, 'completions/min_length': 24.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 28.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.652949869632721, 'rewards/meter/std': 0.3838459551334381, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9330303072929382, 'rewards/repeat_soft/std': 0.027116809040308, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40361714363098145, 'rewards/total_composite/std': 0.037867337465286255, 'reward': 0.40361714363098145, 'reward_std': 0.03786734119057655, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.06865010410547256, 'sampling/sampling_logp_difference/max': 0.7846975326538086, 'sampling/importance_sampling_ratio/min': 0.456257700920105, 'sampling/importance_sampling_ratio/mean': 1.0090291500091553, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3845411613583565, 'clip_ratio/low_mean': 0.034642857033759356, 'clip_ratio/low_min': 0.034642857033759356, 'clip_ratio/high_mean': 0.03498500818386674, 'clip_ratio/high_max': 0.03498500818386674, 'clip_ratio/region_mean': 0.0696278652176261, 'reward_total_mean': 0.40361714363098145, 'reward_meter_mean': 0.652949869632721, 'reward_meter_std': 0.3838459551334381, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9330303072929382, 'reward_repeat_soft_std': 0.027116809040308, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40361714363098145, 'reward_total_composite_std': 0.037867337465286255, 'epoch': 0.21}

 62%|██████▏   | 2062/3300 [5:04:10<2:36:28,  7.58s/it]INFO 04-13 12:25:08 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2063/3300 [5:04:23<3:08:24,  9.14s/it]2026-04-13 12:25:21,338 | INFO | train_grpo_train | metrics_logged mode=train step=2063

                                                       
{'loss': -0.2095, 'grad_norm': 1.7805317640304565, 'learning_rate': 3.7515151515151515e-06, 'num_tokens': 3713276.0, 'completions/mean_length': 175.25, 'completions/min_length': 113.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 127.14286041259766, 'completions/min_terminated_length': 113.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.761021614074707, 'rewards/meter/std': 0.19159801304340363, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.95693039894104, 'rewards/repeat_soft/std': 0.021961698308587074, 'rewards/judge_quality/mean': 0.15625, 'rewards/judge_quality/std': 0.047790467739105225, 'rewards/total_composite/mean': 0.3775297999382019, 'rewards/total_composite/std': 0.15459077060222626, 'reward': 0.3775297999382019, 'reward_std': 0.15459077060222626, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11826728284358978, 'sampling/sampling_logp_difference/max': 2.1097679138183594, 'sampling/importance_sampling_ratio/min': 0.12126611173152924, 'sampling/importance_sampling_ratio/mean': 1.017488718032837, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6875141113996506, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09829550376161933, 'clip_ratio/high_max': 0.09829550376161933, 'clip_ratio/region_mean': 0.09829550376161933, 'reward_total_mean': 0.3775297999382019, 'reward_meter_mean': 0.761021614074707, 'reward_meter_std': 0.19159801304340363, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.95693039894104, 'reward_repeat_soft_std': 0.021961698308587074, 'reward_judge_quality_mean': 0.15625, 'reward_judge_quality_std': 0.047790467739105225, 'reward_total_composite_mean': 0.3775297999382019, 'reward_total_composite_std': 0.15459077060222626, 'epoch': 0.21}

 63%|██████▎   | 2063/3300 [5:04:23<3:08:24,  9.14s/it]INFO 04-13 12:25:21 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2064/3300 [5:04:30<2:55:56,  8.54s/it]2026-04-13 12:25:28,484 | INFO | train_grpo_train | metrics_logged mode=train step=2064

                                                       
{'loss': 0.1502, 'grad_norm': 19.078561782836914, 'learning_rate': 3.748484848484849e-06, 'num_tokens': 3714667.0, 'completions/mean_length': 18.875, 'completions/min_length': 14.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 18.875, 'completions/min_terminated_length': 14.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.5398879051208496, 'rewards/meter/std': 0.44889894127845764, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9491124749183655, 'rewards/repeat_soft/std': 0.02061731554567814, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3950059413909912, 'rewards/total_composite/std': 0.043626610189676285, 'reward': 0.3950059413909912, 'reward_std': 0.04362659528851509, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10711126774549484, 'sampling/sampling_logp_difference/max': 0.8745580911636353, 'sampling/importance_sampling_ratio/min': 0.41704627871513367, 'sampling/importance_sampling_ratio/mean': 1.0046395063400269, 'sampling/importance_sampling_ratio/max': 1.696731448173523, 'entropy': 0.8246192522346973, 'clip_ratio/low_mean': 0.038082108832895756, 'clip_ratio/low_min': 0.038082108832895756, 'clip_ratio/high_mean': 0.039938801899552345, 'clip_ratio/high_max': 0.039938801899552345, 'clip_ratio/region_mean': 0.0780209107324481, 'reward_total_mean': 0.3950059413909912, 'reward_meter_mean': 0.5398879051208496, 'reward_meter_std': 0.44889894127845764, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9491124749183655, 'reward_repeat_soft_std': 0.02061731554567814, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3950059413909912, 'reward_total_composite_std': 0.043626610189676285, 'epoch': 0.21}

 63%|██████▎   | 2064/3300 [5:04:30<2:55:56,  8.54s/it]INFO 04-13 12:25:28 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2065/3300 [5:04:37<2:44:21,  7.99s/it]2026-04-13 12:25:35,180 | INFO | train_grpo_train | metrics_logged mode=train step=2065

                                                       
{'loss': 0.066, 'grad_norm': 13.049152374267578, 'learning_rate': 3.745454545454546e-06, 'num_tokens': 3716331.0, 'completions/mean_length': 40.0, 'completions/min_length': 30.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.4842578172683716, 'rewards/meter/std': 0.4082338809967041, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9329119920730591, 'rewards/repeat_soft/std': 0.08886377513408661, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.38721537590026855, 'rewards/total_composite/std': 0.033759042620658875, 'reward': 0.38721537590026855, 'reward_std': 0.03375904634594917, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10694070905447006, 'sampling/sampling_logp_difference/max': 1.6021480560302734, 'sampling/importance_sampling_ratio/min': 0.20146329700946808, 'sampling/importance_sampling_ratio/mean': 1.0106390714645386, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49556347355246544, 'clip_ratio/low_mean': 0.0787456389516592, 'clip_ratio/low_min': 0.0787456389516592, 'clip_ratio/high_mean': 0.03409310942515731, 'clip_ratio/high_max': 0.03409310942515731, 'clip_ratio/region_mean': 0.11283874837681651, 'reward_total_mean': 0.38721537590026855, 'reward_meter_mean': 0.4842578172683716, 'reward_meter_std': 0.4082338809967041, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9329119920730591, 'reward_repeat_soft_std': 0.08886377513408661, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.38721537590026855, 'reward_total_composite_std': 0.033759042620658875, 'epoch': 0.21}

 63%|██████▎   | 2065/3300 [5:04:37<2:44:21,  7.99s/it]INFO 04-13 12:25:35 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2066/3300 [5:04:43<2:34:07,  7.49s/it]2026-04-13 12:25:41,521 | INFO | train_grpo_train | metrics_logged mode=train step=2066

                                                       
{'loss': 0.1272, 'grad_norm': 16.58196449279785, 'learning_rate': 3.742424242424243e-06, 'num_tokens': 3717834.0, 'completions/mean_length': 21.875, 'completions/min_length': 18.0, 'completions/max_length': 29.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.875, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 29.0, 'rewards/meter/mean': 0.9853771328926086, 'rewards/meter/std': 0.013587907887995243, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.922544002532959, 'rewards/repeat_soft/std': 0.047796379774808884, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43445590138435364, 'rewards/total_composite/std': 0.006304551847279072, 'reward': 0.43445590138435364, 'reward_std': 0.006304553709924221, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08291006088256836, 'sampling/sampling_logp_difference/max': 1.1904220581054688, 'sampling/importance_sampling_ratio/min': 0.36182835698127747, 'sampling/importance_sampling_ratio/mean': 1.013149619102478, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.48387037962675095, 'clip_ratio/low_mean': 0.013829023111611605, 'clip_ratio/low_min': 0.013829023111611605, 'clip_ratio/high_mean': 0.03735047811642289, 'clip_ratio/high_max': 0.03735047811642289, 'clip_ratio/region_mean': 0.051179501228034496, 'reward_total_mean': 0.43445590138435364, 'reward_meter_mean': 0.9853771328926086, 'reward_meter_std': 0.013587907887995243, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.922544002532959, 'reward_repeat_soft_std': 0.047796379774808884, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43445590138435364, 'reward_total_composite_std': 0.006304551847279072, 'epoch': 0.21}

 63%|██████▎   | 2066/3300 [5:04:43<2:34:07,  7.49s/it]INFO 04-13 12:25:41 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2067/3300 [5:04:49<2:26:32,  7.13s/it]2026-04-13 12:25:47,815 | INFO | train_grpo_train | metrics_logged mode=train step=2067

                                                       
{'loss': 0.0469, 'grad_norm': 14.514013290405273, 'learning_rate': 3.73939393939394e-06, 'num_tokens': 3719274.0, 'completions/mean_length': 35.0, 'completions/min_length': 30.0, 'completions/max_length': 43.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 35.0, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 43.0, 'rewards/meter/mean': 0.527634859085083, 'rewards/meter/std': 0.30777183175086975, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9961965084075928, 'rewards/repeat_soft/std': 0.0050888583064079285, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4008738696575165, 'rewards/total_composite/std': 0.02962041273713112, 'reward': 0.4008738696575165, 'reward_std': 0.02962040901184082, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14774520695209503, 'sampling/sampling_logp_difference/max': 2.5967350006103516, 'sampling/importance_sampling_ratio/min': 0.07451647520065308, 'sampling/importance_sampling_ratio/mean': 0.996990978717804, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6231069229543209, 'clip_ratio/low_mean': 0.05282696336507797, 'clip_ratio/low_min': 0.05282696336507797, 'clip_ratio/high_mean': 0.03769909776747227, 'clip_ratio/high_max': 0.03769909776747227, 'clip_ratio/region_mean': 0.09052606113255024, 'reward_total_mean': 0.4008738696575165, 'reward_meter_mean': 0.527634859085083, 'reward_meter_std': 0.30777183175086975, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9961965084075928, 'reward_repeat_soft_std': 0.0050888583064079285, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4008738696575165, 'reward_total_composite_std': 0.02962041273713112, 'epoch': 0.21}

 63%|██████▎   | 2067/3300 [5:04:50<2:26:32,  7.13s/it]INFO 04-13 12:25:48 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2068/3300 [5:04:57<2:28:06,  7.21s/it]2026-04-13 12:25:55,207 | INFO | train_grpo_train | metrics_logged mode=train step=2068

                                                       
{'loss': 0.048, 'grad_norm': 7.030206203460693, 'learning_rate': 3.736363636363637e-06, 'num_tokens': 3721709.0, 'completions/mean_length': 131.375, 'completions/min_length': 121.0, 'completions/max_length': 144.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 131.375, 'completions/min_terminated_length': 121.0, 'completions/max_terminated_length': 144.0, 'rewards/meter/mean': 0.6101386547088623, 'rewards/meter/std': 0.44752225279808044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9160374402999878, 'rewards/repeat_soft/std': 0.03248806670308113, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.02777460776269436, 'rewards/total_composite/mean': 0.40356409549713135, 'rewards/total_composite/std': 0.050754498690366745, 'reward': 0.40356409549713135, 'reward_std': 0.05075449496507645, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12449008226394653, 'sampling/sampling_logp_difference/max': 4.021718502044678, 'sampling/importance_sampling_ratio/min': 0.017922138795256615, 'sampling/importance_sampling_ratio/mean': 1.0004775524139404, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.549030002206564, 'clip_ratio/low_mean': 0.04585020337253809, 'clip_ratio/low_min': 0.04585020337253809, 'clip_ratio/high_mean': 0.07452918961644173, 'clip_ratio/high_max': 0.07452918961644173, 'clip_ratio/region_mean': 0.12037939298897982, 'reward_total_mean': 0.40356409549713135, 'reward_meter_mean': 0.6101386547088623, 'reward_meter_std': 0.44752225279808044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9160374402999878, 'reward_repeat_soft_std': 0.03248806670308113, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.02777460776269436, 'reward_total_composite_mean': 0.40356409549713135, 'reward_total_composite_std': 0.050754498690366745, 'epoch': 0.21}

 63%|██████▎   | 2068/3300 [5:04:57<2:28:06,  7.21s/it]INFO 04-13 12:25:55 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2069/3300 [5:05:04<2:25:21,  7.09s/it]2026-04-13 12:26:01,991 | INFO | train_grpo_train | metrics_logged mode=train step=2069

                                                       
{'loss': 0.0619, 'grad_norm': 9.696648597717285, 'learning_rate': 3.7333333333333337e-06, 'num_tokens': 3723350.0, 'completions/mean_length': 47.125, 'completions/min_length': 43.0, 'completions/max_length': 53.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.9494589567184448, 'rewards/meter/std': 0.08970160037279129, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9439249038696289, 'rewards/repeat_soft/std': 0.02473514713346958, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.434160977602005, 'rewards/total_composite/std': 0.010766803286969662, 'reward': 0.434160977602005, 'reward_std': 0.010766806080937386, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11273051798343658, 'sampling/sampling_logp_difference/max': 2.734816074371338, 'sampling/importance_sampling_ratio/min': 0.06490594148635864, 'sampling/importance_sampling_ratio/mean': 1.0210504531860352, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5686176270246506, 'clip_ratio/low_mean': 0.030455291271209717, 'clip_ratio/low_min': 0.030455291271209717, 'clip_ratio/high_mean': 0.075993905775249, 'clip_ratio/high_max': 0.075993905775249, 'clip_ratio/region_mean': 0.10644919704645872, 'reward_total_mean': 0.434160977602005, 'reward_meter_mean': 0.9494589567184448, 'reward_meter_std': 0.08970160037279129, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9439249038696289, 'reward_repeat_soft_std': 0.02473514713346958, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.434160977602005, 'reward_total_composite_std': 0.010766803286969662, 'epoch': 0.21}

 63%|██████▎   | 2069/3300 [5:05:04<2:25:21,  7.09s/it]INFO 04-13 12:26:02 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2070/3300 [5:05:10<2:19:06,  6.79s/it]2026-04-13 12:26:08,081 | INFO | train_grpo_train | metrics_logged mode=train step=2070

                                                       
{'loss': -0.0183, 'grad_norm': 12.375938415527344, 'learning_rate': 3.7303030303030306e-06, 'num_tokens': 3724969.0, 'completions/mean_length': 40.375, 'completions/min_length': 35.0, 'completions/max_length': 45.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 40.375, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.8671349883079529, 'rewards/meter/std': 0.3140276074409485, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9792593717575073, 'rewards/repeat_soft/std': 0.020110072568058968, 'rewards/judge_quality/mean': 0.1875, 'rewards/judge_quality/std': 0.1060660108923912, 'rewards/total_composite/mean': 0.45553040504455566, 'rewards/total_composite/std': 0.07938668876886368, 'reward': 0.45553040504455566, 'reward_std': 0.07938668131828308, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11620399355888367, 'sampling/sampling_logp_difference/max': 2.658374071121216, 'sampling/importance_sampling_ratio/min': 0.0700620487332344, 'sampling/importance_sampling_ratio/mean': 1.02402925491333, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7220793217420578, 'clip_ratio/low_mean': 0.08675484010018408, 'clip_ratio/low_min': 0.08675484010018408, 'clip_ratio/high_mean': 0.02222222276031971, 'clip_ratio/high_max': 0.02222222276031971, 'clip_ratio/region_mean': 0.10897706286050379, 'reward_total_mean': 0.45553040504455566, 'reward_meter_mean': 0.8671349883079529, 'reward_meter_std': 0.3140276074409485, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9792593717575073, 'reward_repeat_soft_std': 0.020110072568058968, 'reward_judge_quality_mean': 0.1875, 'reward_judge_quality_std': 0.1060660108923912, 'reward_total_composite_mean': 0.45553040504455566, 'reward_total_composite_std': 0.07938668876886368, 'epoch': 0.21}

 63%|██████▎   | 2070/3300 [5:05:10<2:19:06,  6.79s/it]INFO 04-13 12:26:08 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2071/3300 [5:05:16<2:16:34,  6.67s/it]2026-04-13 12:26:14,485 | INFO | train_grpo_train | metrics_logged mode=train step=2071

                                                       
{'loss': -0.0811, 'grad_norm': 11.69269847869873, 'learning_rate': 3.727272727272728e-06, 'num_tokens': 3726590.0, 'completions/mean_length': 47.625, 'completions/min_length': 37.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.625, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.9296966791152954, 'rewards/meter/std': 0.1094554215669632, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9387052655220032, 'rewards/repeat_soft/std': 0.06813524663448334, 'rewards/judge_quality/mean': 0.23250000178813934, 'rewards/judge_quality/std': 0.1348809003829956, 'rewards/total_composite/mean': 0.4744293987751007, 'rewards/total_composite/std': 0.06318916380405426, 'reward': 0.4744293987751007, 'reward_std': 0.06318917125463486, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10637016594409943, 'sampling/sampling_logp_difference/max': 1.755422592163086, 'sampling/importance_sampling_ratio/min': 0.17283418774604797, 'sampling/importance_sampling_ratio/mean': 0.9959914088249207, 'sampling/importance_sampling_ratio/max': 1.7517482042312622, 'entropy': 0.4855124205350876, 'clip_ratio/low_mean': 0.07618559896945953, 'clip_ratio/low_min': 0.07618559896945953, 'clip_ratio/high_mean': 0.02797619067132473, 'clip_ratio/high_max': 0.02797619067132473, 'clip_ratio/region_mean': 0.10416178964078426, 'reward_total_mean': 0.4744293987751007, 'reward_meter_mean': 0.9296966791152954, 'reward_meter_std': 0.1094554215669632, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9387052655220032, 'reward_repeat_soft_std': 0.06813524663448334, 'reward_judge_quality_mean': 0.23250000178813934, 'reward_judge_quality_std': 0.1348809003829956, 'reward_total_composite_mean': 0.4744293987751007, 'reward_total_composite_std': 0.06318916380405426, 'epoch': 0.21}

 63%|██████▎   | 2071/3300 [5:05:16<2:16:34,  6.67s/it]INFO 04-13 12:26:14 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2072/3300 [5:05:23<2:19:28,  6.81s/it]2026-04-13 12:26:21,638 | INFO | train_grpo_train | metrics_logged mode=train step=2072

                                                       
{'loss': 0.0753, 'grad_norm': 8.115363121032715, 'learning_rate': 3.7242424242424246e-06, 'num_tokens': 3728520.0, 'completions/mean_length': 90.25, 'completions/min_length': 81.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 90.25, 'completions/min_terminated_length': 81.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.7392156720161438, 'rewards/meter/std': 0.23921041190624237, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9616492986679077, 'rewards/repeat_soft/std': 0.045542970299720764, 'rewards/judge_quality/mean': 0.13875000178813934, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.41035062074661255, 'rewards/total_composite/std': 0.025785747915506363, 'reward': 0.41035062074661255, 'reward_std': 0.025785747915506363, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11131595075130463, 'sampling/sampling_logp_difference/max': 1.636796474456787, 'sampling/importance_sampling_ratio/min': 0.24769717454910278, 'sampling/importance_sampling_ratio/mean': 1.0275774002075195, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7617958560585976, 'clip_ratio/low_mean': 0.04378078877925873, 'clip_ratio/low_min': 0.04378078877925873, 'clip_ratio/high_mean': 0.06652565207332373, 'clip_ratio/high_max': 0.06652565207332373, 'clip_ratio/region_mean': 0.11030644085258245, 'reward_total_mean': 0.41035062074661255, 'reward_meter_mean': 0.7392156720161438, 'reward_meter_std': 0.23921041190624237, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9616492986679077, 'reward_repeat_soft_std': 0.045542970299720764, 'reward_judge_quality_mean': 0.13875000178813934, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.41035062074661255, 'reward_total_composite_std': 0.025785747915506363, 'epoch': 0.21}

 63%|██████▎   | 2072/3300 [5:05:23<2:19:28,  6.81s/it]INFO 04-13 12:26:21 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2073/3300 [5:05:30<2:18:55,  6.79s/it]2026-04-13 12:26:28,386 | INFO | train_grpo_train | metrics_logged mode=train step=2073

                                                       
{'loss': 0.097, 'grad_norm': 14.835057258605957, 'learning_rate': 3.7212121212121215e-06, 'num_tokens': 3730112.0, 'completions/mean_length': 38.0, 'completions/min_length': 28.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.0, 'completions/min_terminated_length': 28.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9140802621841431, 'rewards/meter/std': 0.09279971569776535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8278565406799316, 'rewards/repeat_soft/std': 0.16277727484703064, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4241696000099182, 'rewards/total_composite/std': 0.027534764260053635, 'reward': 0.4241696000099182, 'reward_std': 0.027534764260053635, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1308312863111496, 'sampling/sampling_logp_difference/max': 2.29068922996521, 'sampling/importance_sampling_ratio/min': 0.10119669139385223, 'sampling/importance_sampling_ratio/mean': 0.9928392171859741, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.497556971386075, 'clip_ratio/low_mean': 0.04509708611294627, 'clip_ratio/low_min': 0.04509708611294627, 'clip_ratio/high_mean': 0.0766675341874361, 'clip_ratio/high_max': 0.0766675341874361, 'clip_ratio/region_mean': 0.12176462030038238, 'reward_total_mean': 0.4241696000099182, 'reward_meter_mean': 0.9140802621841431, 'reward_meter_std': 0.09279971569776535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8278565406799316, 'reward_repeat_soft_std': 0.16277727484703064, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4241696000099182, 'reward_total_composite_std': 0.027534764260053635, 'epoch': 0.21}

 63%|██████▎   | 2073/3300 [5:05:30<2:18:55,  6.79s/it]INFO 04-13 12:26:28 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2074/3300 [5:05:39<2:30:03,  7.34s/it]2026-04-13 12:26:37,037 | INFO | train_grpo_train | metrics_logged mode=train step=2074

                                                       
{'loss': 0.0815, 'grad_norm': 7.157658100128174, 'learning_rate': 3.7181818181818187e-06, 'num_tokens': 3732910.0, 'completions/mean_length': 168.75, 'completions/min_length': 145.0, 'completions/max_length': 192.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 168.75, 'completions/min_terminated_length': 145.0, 'completions/max_terminated_length': 192.0, 'rewards/meter/mean': 0.8324783444404602, 'rewards/meter/std': 0.18502572178840637, 'rewards/count_adherence/mean': 0.9583333134651184, 'rewards/count_adherence/std': 0.07715168595314026, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7660987973213196, 'rewards/repeat_soft/std': 0.03201473131775856, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.02777460776269436, 'rewards/total_composite/mean': 0.3949919044971466, 'rewards/total_composite/std': 0.03361143544316292, 'reward': 0.3949919044971466, 'reward_std': 0.03361142799258232, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09431398659944534, 'sampling/sampling_logp_difference/max': 4.0692362785339355, 'sampling/importance_sampling_ratio/min': 0.017090434208512306, 'sampling/importance_sampling_ratio/mean': 1.0004268884658813, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49766362458467484, 'clip_ratio/low_mean': 0.025941302068531513, 'clip_ratio/low_min': 0.025941302068531513, 'clip_ratio/high_mean': 0.05744477640837431, 'clip_ratio/high_max': 0.05744477640837431, 'clip_ratio/region_mean': 0.08338607847690582, 'reward_total_mean': 0.3949919044971466, 'reward_meter_mean': 0.8324783444404602, 'reward_meter_std': 0.18502572178840637, 'reward_count_adherence_mean': 0.9583333134651184, 'reward_count_adherence_std': 0.07715168595314026, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7660987973213196, 'reward_repeat_soft_std': 0.03201473131775856, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.02777460776269436, 'reward_total_composite_mean': 0.3949919044971466, 'reward_total_composite_std': 0.03361143544316292, 'epoch': 0.21}

 63%|██████▎   | 2074/3300 [5:05:39<2:30:03,  7.34s/it]INFO 04-13 12:26:37 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2075/3300 [5:05:45<2:23:53,  7.05s/it]2026-04-13 12:26:43,377 | INFO | train_grpo_train | metrics_logged mode=train step=2075

                                                       
{'loss': 0.0437, 'grad_norm': 19.99515151977539, 'learning_rate': 3.7151515151515156e-06, 'num_tokens': 3734292.0, 'completions/mean_length': 22.75, 'completions/min_length': 20.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.75, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.8105221390724182, 'rewards/meter/std': 0.35397836565971375, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9114496111869812, 'rewards/repeat_soft/std': 0.06043412163853645, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.415743350982666, 'rewards/total_composite/std': 0.04004393517971039, 'reward': 0.415743350982666, 'reward_std': 0.04004393890500069, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12008140236139297, 'sampling/sampling_logp_difference/max': 1.7892513275146484, 'sampling/importance_sampling_ratio/min': 0.1670852154493332, 'sampling/importance_sampling_ratio/mean': 1.0087186098098755, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7540407590568066, 'clip_ratio/low_mean': 0.03476190473884344, 'clip_ratio/low_min': 0.03476190473884344, 'clip_ratio/high_mean': 0.08422762854024768, 'clip_ratio/high_max': 0.08422762854024768, 'clip_ratio/region_mean': 0.11898953327909112, 'reward_total_mean': 0.415743350982666, 'reward_meter_mean': 0.8105221390724182, 'reward_meter_std': 0.35397836565971375, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9114496111869812, 'reward_repeat_soft_std': 0.06043412163853645, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.415743350982666, 'reward_total_composite_std': 0.04004393517971039, 'epoch': 0.21}

 63%|██████▎   | 2075/3300 [5:05:45<2:23:53,  7.05s/it]INFO 04-13 12:26:43 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2076/3300 [5:05:52<2:25:23,  7.13s/it]2026-04-13 12:26:50,680 | INFO | train_grpo_train | metrics_logged mode=train step=2076

                                                       
{'loss': 0.1202, 'grad_norm': 8.06759262084961, 'learning_rate': 3.7121212121212124e-06, 'num_tokens': 3736648.0, 'completions/mean_length': 107.5, 'completions/min_length': 82.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 107.5, 'completions/min_terminated_length': 82.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.9633580446243286, 'rewards/meter/std': 0.035794198513031006, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8788341283798218, 'rewards/repeat_soft/std': 0.04080803319811821, 'rewards/judge_quality/mean': 0.13875001668930054, 'rewards/judge_quality/std': 0.02748376503586769, 'rewards/total_composite/mean': 0.40023452043533325, 'rewards/total_composite/std': 0.031872548162937164, 'reward': 0.40023452043533325, 'reward_std': 0.031872548162937164, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11378919333219528, 'sampling/sampling_logp_difference/max': 1.6996946334838867, 'sampling/importance_sampling_ratio/min': 0.18273931741714478, 'sampling/importance_sampling_ratio/mean': 0.9930237531661987, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5079356506466866, 'clip_ratio/low_mean': 0.04586605029180646, 'clip_ratio/low_min': 0.04586605029180646, 'clip_ratio/high_mean': 0.04101017955690622, 'clip_ratio/high_max': 0.04101017955690622, 'clip_ratio/region_mean': 0.08687622984871268, 'reward_total_mean': 0.40023452043533325, 'reward_meter_mean': 0.9633580446243286, 'reward_meter_std': 0.035794198513031006, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8788341283798218, 'reward_repeat_soft_std': 0.04080803319811821, 'reward_judge_quality_mean': 0.13875001668930054, 'reward_judge_quality_std': 0.02748376503586769, 'reward_total_composite_mean': 0.40023452043533325, 'reward_total_composite_std': 0.031872548162937164, 'epoch': 0.21}

 63%|██████▎   | 2076/3300 [5:05:52<2:25:23,  7.13s/it]INFO 04-13 12:26:50 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2077/3300 [5:05:58<2:18:26,  6.79s/it]2026-04-13 12:26:56,679 | INFO | train_grpo_train | metrics_logged mode=train step=2077

                                                       
{'loss': 0.0324, 'grad_norm': 14.313187599182129, 'learning_rate': 3.7090909090909092e-06, 'num_tokens': 3738040.0, 'completions/mean_length': 30.0, 'completions/min_length': 27.0, 'completions/max_length': 33.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 30.0, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 33.0, 'rewards/meter/mean': 0.6820467114448547, 'rewards/meter/std': 0.23005546629428864, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9629801511764526, 'rewards/repeat_soft/std': 0.0550442710518837, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4109466075897217, 'rewards/total_composite/std': 0.025824129581451416, 'reward': 0.4109466075897217, 'reward_std': 0.025824127718806267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08204793930053711, 'sampling/sampling_logp_difference/max': 1.4037678241729736, 'sampling/importance_sampling_ratio/min': 0.24566958844661713, 'sampling/importance_sampling_ratio/mean': 1.0015180110931396, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4720030911266804, 'clip_ratio/low_mean': 0.012449766509234905, 'clip_ratio/low_min': 0.012449766509234905, 'clip_ratio/high_mean': 0.051067705266177654, 'clip_ratio/high_max': 0.051067705266177654, 'clip_ratio/region_mean': 0.06351747177541256, 'reward_total_mean': 0.4109466075897217, 'reward_meter_mean': 0.6820467114448547, 'reward_meter_std': 0.23005546629428864, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9629801511764526, 'reward_repeat_soft_std': 0.0550442710518837, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4109466075897217, 'reward_total_composite_std': 0.025824129581451416, 'epoch': 0.21}

 63%|██████▎   | 2077/3300 [5:05:58<2:18:26,  6.79s/it]INFO 04-13 12:26:56 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2078/3300 [5:06:06<2:25:15,  7.13s/it]2026-04-13 12:27:04,619 | INFO | train_grpo_train | metrics_logged mode=train step=2078

                                                       
{'loss': 0.1028, 'grad_norm': 6.437441825866699, 'learning_rate': 3.7060606060606065e-06, 'num_tokens': 3740218.0, 'completions/mean_length': 101.25, 'completions/min_length': 87.0, 'completions/max_length': 117.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 101.25, 'completions/min_terminated_length': 87.0, 'completions/max_terminated_length': 117.0, 'rewards/meter/mean': 0.9617483615875244, 'rewards/meter/std': 0.0327635258436203, 'rewards/count_adherence/mean': 0.84375, 'rewards/count_adherence/std': 0.12938730418682098, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.809037983417511, 'rewards/repeat_soft/std': 0.07080276310443878, 'rewards/judge_quality/mean': 0.1274999976158142, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.36988019943237305, 'rewards/total_composite/std': 0.03989952802658081, 'reward': 0.36988019943237305, 'reward_std': 0.03989953175187111, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09450139850378036, 'sampling/sampling_logp_difference/max': 2.4037623405456543, 'sampling/importance_sampling_ratio/min': 0.09037727862596512, 'sampling/importance_sampling_ratio/mean': 1.0028011798858643, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41972821950912476, 'clip_ratio/low_mean': 0.05465828627347946, 'clip_ratio/low_min': 0.05465828627347946, 'clip_ratio/high_mean': 0.02143743447959423, 'clip_ratio/high_max': 0.02143743447959423, 'clip_ratio/region_mean': 0.07609572075307369, 'reward_total_mean': 0.36988019943237305, 'reward_meter_mean': 0.9617483615875244, 'reward_meter_std': 0.0327635258436203, 'reward_count_adherence_mean': 0.84375, 'reward_count_adherence_std': 0.12938730418682098, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.809037983417511, 'reward_repeat_soft_std': 0.07080276310443878, 'reward_judge_quality_mean': 0.1274999976158142, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.36988019943237305, 'reward_total_composite_std': 0.03989952802658081, 'epoch': 0.21}

 63%|██████▎   | 2078/3300 [5:06:06<2:25:15,  7.13s/it]INFO 04-13 12:27:04 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2079/3300 [5:06:13<2:21:32,  6.96s/it]2026-04-13 12:27:11,158 | INFO | train_grpo_train | metrics_logged mode=train step=2079

                                                       
{'loss': -0.0068, 'grad_norm': 14.443037986755371, 'learning_rate': 3.7030303030303033e-06, 'num_tokens': 3741818.0, 'completions/mean_length': 52.0, 'completions/min_length': 49.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9573231935501099, 'rewards/meter/std': 0.07338250428438187, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.975223183631897, 'rewards/repeat_soft/std': 0.013552870601415634, 'rewards/judge_quality/mean': 0.1574999988079071, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.44444313645362854, 'rewards/total_composite/std': 0.011920580640435219, 'reward': 0.44444313645362854, 'reward_std': 0.011920583434402943, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08642370253801346, 'sampling/sampling_logp_difference/max': 1.1978371143341064, 'sampling/importance_sampling_ratio/min': 0.30184635519981384, 'sampling/importance_sampling_ratio/mean': 1.0158339738845825, 'sampling/importance_sampling_ratio/max': 1.9727880954742432, 'entropy': 0.5891767293214798, 'clip_ratio/low_mean': 0.048753238283097744, 'clip_ratio/low_min': 0.048753238283097744, 'clip_ratio/high_mean': 0.04219135455787182, 'clip_ratio/high_max': 0.04219135455787182, 'clip_ratio/region_mean': 0.09094459284096956, 'reward_total_mean': 0.44444313645362854, 'reward_meter_mean': 0.9573231935501099, 'reward_meter_std': 0.07338250428438187, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.975223183631897, 'reward_repeat_soft_std': 0.013552870601415634, 'reward_judge_quality_mean': 0.1574999988079071, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.44444313645362854, 'reward_total_composite_std': 0.011920580640435219, 'epoch': 0.21}

 63%|██████▎   | 2079/3300 [5:06:13<2:21:32,  6.96s/it]INFO 04-13 12:27:11 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2080/3300 [5:06:24<2:50:20,  8.38s/it]2026-04-13 12:27:22,874 | INFO | train_grpo_train | metrics_logged mode=train step=2080

                                                       
{'loss': -0.2059, 'grad_norm': 1.7132174968719482, 'learning_rate': 3.7e-06, 'num_tokens': 3744129.0, 'completions/mean_length': 161.875, 'completions/min_length': 93.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 111.85714721679688, 'completions/min_terminated_length': 93.0, 'completions/max_terminated_length': 137.0, 'rewards/meter/mean': 0.7818764448165894, 'rewards/meter/std': 0.308095246553421, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8692042231559753, 'rewards/repeat_soft/std': 0.05775836110115051, 'rewards/judge_quality/mean': 0.14125001430511475, 'rewards/judge_quality/std': 0.04733996093273163, 'rewards/total_composite/mean': 0.36371177434921265, 'rewards/total_composite/std': 0.14940164983272552, 'reward': 0.36371177434921265, 'reward_std': 0.14940164983272552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09953910112380981, 'sampling/sampling_logp_difference/max': 1.7864031791687012, 'sampling/importance_sampling_ratio/min': 0.16756176948547363, 'sampling/importance_sampling_ratio/mean': 0.9994521141052246, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.49852072820067406, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08590791467577219, 'clip_ratio/high_max': 0.08590791467577219, 'clip_ratio/region_mean': 0.08590791467577219, 'reward_total_mean': 0.36371177434921265, 'reward_meter_mean': 0.7818764448165894, 'reward_meter_std': 0.308095246553421, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8692042231559753, 'reward_repeat_soft_std': 0.05775836110115051, 'reward_judge_quality_mean': 0.14125001430511475, 'reward_judge_quality_std': 0.04733996093273163, 'reward_total_composite_mean': 0.36371177434921265, 'reward_total_composite_std': 0.14940164983272552, 'epoch': 0.21}

 63%|██████▎   | 2080/3300 [5:06:25<2:50:20,  8.38s/it]INFO 04-13 12:27:23 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2081/3300 [5:06:36<3:09:13,  9.31s/it]2026-04-13 12:27:34,358 | INFO | train_grpo_train | metrics_logged mode=train step=2081

                                                       
{'loss': -0.1162, 'grad_norm': 2.2427473068237305, 'learning_rate': 3.6969696969696974e-06, 'num_tokens': 3745643.0, 'completions/mean_length': 95.25, 'completions/min_length': 26.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.71428680419922, 'completions/min_terminated_length': 26.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.4832600951194763, 'rewards/meter/std': 0.4000207781791687, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9759777784347534, 'rewards/repeat_soft/std': 0.023617975413799286, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.3494424819946289, 'rewards/total_composite/std': 0.14582893252372742, 'reward': 0.3494424819946289, 'reward_std': 0.14582891762256622, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16125179827213287, 'sampling/sampling_logp_difference/max': 2.206164836883545, 'sampling/importance_sampling_ratio/min': 0.1101221814751625, 'sampling/importance_sampling_ratio/mean': 1.00516676902771, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6992188319563866, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.1526223225519061, 'clip_ratio/high_max': 0.1526223225519061, 'clip_ratio/region_mean': 0.1526223225519061, 'reward_total_mean': 0.3494424819946289, 'reward_meter_mean': 0.4832600951194763, 'reward_meter_std': 0.4000207781791687, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9759777784347534, 'reward_repeat_soft_std': 0.023617975413799286, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.3494424819946289, 'reward_total_composite_std': 0.14582893252372742, 'epoch': 0.21}

 63%|██████▎   | 2081/3300 [5:06:36<3:09:13,  9.31s/it]INFO 04-13 12:27:34 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2082/3300 [5:06:42<2:50:05,  8.38s/it]2026-04-13 12:27:40,552 | INFO | train_grpo_train | metrics_logged mode=train step=2082

                                                       
{'loss': 0.0459, 'grad_norm': 12.529629707336426, 'learning_rate': 3.6939393939393942e-06, 'num_tokens': 3747157.0, 'completions/mean_length': 46.25, 'completions/min_length': 38.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 46.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.4912826418876648, 'rewards/meter/std': 0.30648428201675415, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9786837100982666, 'rewards/repeat_soft/std': 0.022699061781167984, 'rewards/judge_quality/mean': 0.19499999284744263, 'rewards/judge_quality/std': 0.10392304509878159, 'rewards/total_composite/mean': 0.4009936451911926, 'rewards/total_composite/std': 0.03267580643296242, 'reward': 0.4009936451911926, 'reward_std': 0.03267580270767212, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12824684381484985, 'sampling/sampling_logp_difference/max': 2.061049699783325, 'sampling/importance_sampling_ratio/min': 0.12732024490833282, 'sampling/importance_sampling_ratio/mean': 1.009272575378418, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7472609356045723, 'clip_ratio/low_mean': 0.08138155192136765, 'clip_ratio/low_min': 0.08138155192136765, 'clip_ratio/high_mean': 0.04700176510959864, 'clip_ratio/high_max': 0.04700176510959864, 'clip_ratio/region_mean': 0.12838331703096628, 'reward_total_mean': 0.4009936451911926, 'reward_meter_mean': 0.4912826418876648, 'reward_meter_std': 0.30648428201675415, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9786837100982666, 'reward_repeat_soft_std': 0.022699061781167984, 'reward_judge_quality_mean': 0.19499999284744263, 'reward_judge_quality_std': 0.10392304509878159, 'reward_total_composite_mean': 0.4009936451911926, 'reward_total_composite_std': 0.03267580643296242, 'epoch': 0.21}

 63%|██████▎   | 2082/3300 [5:06:42<2:50:05,  8.38s/it]INFO 04-13 12:27:40 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2083/3300 [5:06:49<2:40:04,  7.89s/it]2026-04-13 12:27:47,315 | INFO | train_grpo_train | metrics_logged mode=train step=2083

                                                       
{'loss': 0.0185, 'grad_norm': 14.358192443847656, 'learning_rate': 3.690909090909091e-06, 'num_tokens': 3748664.0, 'completions/mean_length': 44.375, 'completions/min_length': 39.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 44.375, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.6349843740463257, 'rewards/meter/std': 0.42381957173347473, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9558464288711548, 'rewards/repeat_soft/std': 0.032159652560949326, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.4029209315776825, 'rewards/total_composite/std': 0.04163876920938492, 'reward': 0.4029209315776825, 'reward_std': 0.04163876920938492, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11028869450092316, 'sampling/sampling_logp_difference/max': 1.2160595655441284, 'sampling/importance_sampling_ratio/min': 0.29639580845832825, 'sampling/importance_sampling_ratio/mean': 1.0181822776794434, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7787529304623604, 'clip_ratio/low_mean': 0.04384576342999935, 'clip_ratio/low_min': 0.04384576342999935, 'clip_ratio/high_mean': 0.08037643367424607, 'clip_ratio/high_max': 0.08037643367424607, 'clip_ratio/region_mean': 0.12422219710424542, 'reward_total_mean': 0.4029209315776825, 'reward_meter_mean': 0.6349843740463257, 'reward_meter_std': 0.42381957173347473, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9558464288711548, 'reward_repeat_soft_std': 0.032159652560949326, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.4029209315776825, 'reward_total_composite_std': 0.04163876920938492, 'epoch': 0.21}

 63%|██████▎   | 2083/3300 [5:06:49<2:40:04,  7.89s/it]INFO 04-13 12:27:47 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2084/3300 [5:06:57<2:38:23,  7.82s/it]2026-04-13 12:27:54,937 | INFO | train_grpo_train | metrics_logged mode=train step=2084

                                                       
{'loss': 0.1051, 'grad_norm': 8.194664001464844, 'learning_rate': 3.687878787878788e-06, 'num_tokens': 3751053.0, 'completions/mean_length': 106.625, 'completions/min_length': 97.0, 'completions/max_length': 140.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 106.625, 'completions/min_terminated_length': 97.0, 'completions/max_terminated_length': 140.0, 'rewards/meter/mean': 0.8641301393508911, 'rewards/meter/std': 0.2113206684589386, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8112736344337463, 'rewards/repeat_soft/std': 0.03796151652932167, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.029730942100286484, 'rewards/total_composite/mean': 0.40080052614212036, 'rewards/total_composite/std': 0.02927142195403576, 'reward': 0.40080052614212036, 'reward_std': 0.02927141822874546, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08997472375631332, 'sampling/sampling_logp_difference/max': 2.6632847785949707, 'sampling/importance_sampling_ratio/min': 0.06971883028745651, 'sampling/importance_sampling_ratio/mean': 1.003967523574829, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46710870042443275, 'clip_ratio/low_mean': 0.03522890596650541, 'clip_ratio/low_min': 0.03522890596650541, 'clip_ratio/high_mean': 0.03521409630775452, 'clip_ratio/high_max': 0.03521409630775452, 'clip_ratio/region_mean': 0.07044300227425992, 'reward_total_mean': 0.40080052614212036, 'reward_meter_mean': 0.8641301393508911, 'reward_meter_std': 0.2113206684589386, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8112736344337463, 'reward_repeat_soft_std': 0.03796151652932167, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.029730942100286484, 'reward_total_composite_mean': 0.40080052614212036, 'reward_total_composite_std': 0.02927142195403576, 'epoch': 0.21}

 63%|██████▎   | 2084/3300 [5:06:57<2:38:23,  7.82s/it]INFO 04-13 12:27:55 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2085/3300 [5:07:03<2:31:54,  7.50s/it]2026-04-13 12:28:01,714 | INFO | train_grpo_train | metrics_logged mode=train step=2085

                                                       
{'loss': 0.0659, 'grad_norm': 12.54304027557373, 'learning_rate': 3.684848484848485e-06, 'num_tokens': 3752582.0, 'completions/mean_length': 41.125, 'completions/min_length': 32.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.125, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8170626163482666, 'rewards/meter/std': 0.3091009855270386, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8963360786437988, 'rewards/repeat_soft/std': 0.07029309868812561, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.41411399841308594, 'rewards/total_composite/std': 0.029479196295142174, 'reward': 0.41411399841308594, 'reward_std': 0.029479194432497025, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11653422564268112, 'sampling/sampling_logp_difference/max': 1.3648093938827515, 'sampling/importance_sampling_ratio/min': 0.25542935729026794, 'sampling/importance_sampling_ratio/mean': 1.0088908672332764, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6153091266751289, 'clip_ratio/low_mean': 0.03251900803297758, 'clip_ratio/low_min': 0.03251900803297758, 'clip_ratio/high_mean': 0.05771985976025462, 'clip_ratio/high_max': 0.05771985976025462, 'clip_ratio/region_mean': 0.0902388677932322, 'reward_total_mean': 0.41411399841308594, 'reward_meter_mean': 0.8170626163482666, 'reward_meter_std': 0.3091009855270386, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8963360786437988, 'reward_repeat_soft_std': 0.07029309868812561, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.41411399841308594, 'reward_total_composite_std': 0.029479196295142174, 'epoch': 0.21}

 63%|██████▎   | 2085/3300 [5:07:03<2:31:54,  7.50s/it]INFO 04-13 12:28:01 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2086/3300 [5:07:15<2:59:46,  8.89s/it]2026-04-13 12:28:13,834 | INFO | train_grpo_train | metrics_logged mode=train step=2086

                                                       
{'loss': -0.1198, 'grad_norm': 1.476021409034729, 'learning_rate': 3.681818181818182e-06, 'num_tokens': 3754249.0, 'completions/mean_length': 165.375, 'completions/min_length': 43.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 49.833335876464844, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 53.0, 'rewards/meter/mean': 0.7697263956069946, 'rewards/meter/std': 0.23756560683250427, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9250380992889404, 'rewards/repeat_soft/std': 0.07582405209541321, 'rewards/judge_quality/mean': 0.13625000417232513, 'rewards/judge_quality/std': 0.057305578142404556, 'rewards/total_composite/mean': 0.32509738206863403, 'rewards/total_composite/std': 0.20268148183822632, 'reward': 0.32509738206863403, 'reward_std': 0.20268146693706512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0769723579287529, 'sampling/sampling_logp_difference/max': 2.3819503784179688, 'sampling/importance_sampling_ratio/min': 0.09237024188041687, 'sampling/importance_sampling_ratio/mean': 1.0031273365020752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2697983533143997, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.050510191125795245, 'clip_ratio/high_max': 0.050510191125795245, 'clip_ratio/region_mean': 0.050510191125795245, 'reward_total_mean': 0.32509738206863403, 'reward_meter_mean': 0.7697263956069946, 'reward_meter_std': 0.23756560683250427, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9250380992889404, 'reward_repeat_soft_std': 0.07582405209541321, 'reward_judge_quality_mean': 0.13625000417232513, 'reward_judge_quality_std': 0.057305578142404556, 'reward_total_composite_mean': 0.32509738206863403, 'reward_total_composite_std': 0.20268148183822632, 'epoch': 0.21}

 63%|██████▎   | 2086/3300 [5:07:16<2:59:46,  8.89s/it]INFO 04-13 12:28:14 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2087/3300 [5:07:22<2:46:49,  8.25s/it]2026-04-13 12:28:20,599 | INFO | train_grpo_train | metrics_logged mode=train step=2087

                                                       
{'loss': 0.0692, 'grad_norm': 16.328943252563477, 'learning_rate': 3.678787878787879e-06, 'num_tokens': 3755826.0, 'completions/mean_length': 37.125, 'completions/min_length': 30.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.125, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.8387521505355835, 'rewards/meter/std': 0.19725178182125092, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9575532078742981, 'rewards/repeat_soft/std': 0.03129439800977707, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.43037697672843933, 'rewards/total_composite/std': 0.02135138213634491, 'reward': 0.43037697672843933, 'reward_std': 0.02135137841105461, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08730937540531158, 'sampling/sampling_logp_difference/max': 2.1456055641174316, 'sampling/importance_sampling_ratio/min': 0.11699716001749039, 'sampling/importance_sampling_ratio/mean': 0.9835996031761169, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3454021457582712, 'clip_ratio/low_mean': 0.006756756920367479, 'clip_ratio/low_min': 0.006756756920367479, 'clip_ratio/high_mean': 0.080037297680974, 'clip_ratio/high_max': 0.080037297680974, 'clip_ratio/region_mean': 0.08679405460134149, 'reward_total_mean': 0.43037697672843933, 'reward_meter_mean': 0.8387521505355835, 'reward_meter_std': 0.19725178182125092, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9575532078742981, 'reward_repeat_soft_std': 0.03129439800977707, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.43037697672843933, 'reward_total_composite_std': 0.02135138213634491, 'epoch': 0.21}

 63%|██████▎   | 2087/3300 [5:07:22<2:46:49,  8.25s/it]INFO 04-13 12:28:20 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2088/3300 [5:07:29<2:39:19,  7.89s/it]2026-04-13 12:28:27,634 | INFO | train_grpo_train | metrics_logged mode=train step=2088

                                                       
{'loss': 0.0317, 'grad_norm': 11.181440353393555, 'learning_rate': 3.6757575757575757e-06, 'num_tokens': 3757367.0, 'completions/mean_length': 36.625, 'completions/min_length': 35.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.6886870861053467, 'rewards/meter/std': 0.36762264370918274, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9336532354354858, 'rewards/repeat_soft/std': 0.053609877824783325, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.40719497203826904, 'rewards/total_composite/std': 0.03439675644040108, 'reward': 0.40719497203826904, 'reward_std': 0.034396763890981674, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09574722498655319, 'sampling/sampling_logp_difference/max': 1.711014747619629, 'sampling/importance_sampling_ratio/min': 0.18068234622478485, 'sampling/importance_sampling_ratio/mean': 1.0201098918914795, 'sampling/importance_sampling_ratio/max': 1.901428461074829, 'entropy': 0.6929846182465553, 'clip_ratio/low_mean': 0.03128175204619765, 'clip_ratio/low_min': 0.03128175204619765, 'clip_ratio/high_mean': 0.06482612853869796, 'clip_ratio/high_max': 0.06482612853869796, 'clip_ratio/region_mean': 0.09610788058489561, 'reward_total_mean': 0.40719497203826904, 'reward_meter_mean': 0.6886870861053467, 'reward_meter_std': 0.36762264370918274, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9336532354354858, 'reward_repeat_soft_std': 0.053609877824783325, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.40719497203826904, 'reward_total_composite_std': 0.03439675644040108, 'epoch': 0.21}

 63%|██████▎   | 2088/3300 [5:07:29<2:39:19,  7.89s/it]INFO 04-13 12:28:27 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2089/3300 [5:07:47<3:41:06, 10.95s/it]2026-04-13 12:28:45,744 | INFO | train_grpo_train | metrics_logged mode=train step=2089

                                                       
{'loss': -0.1467, 'grad_norm': 1.9923795461654663, 'learning_rate': 3.672727272727273e-06, 'num_tokens': 3759089.0, 'completions/mean_length': 116.25, 'completions/min_length': 46.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.71428680419922, 'completions/min_terminated_length': 46.0, 'completions/max_terminated_length': 73.0, 'rewards/meter/mean': 0.5514540672302246, 'rewards/meter/std': 0.40759724378585815, 'rewards/count_adherence/mean': 0.9583333730697632, 'rewards/count_adherence/std': 0.117851123213768, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9403271079063416, 'rewards/repeat_soft/std': 0.03892381489276886, 'rewards/judge_quality/mean': 0.13375000655651093, 'rewards/judge_quality/std': 0.03889087587594986, 'rewards/total_composite/mean': 0.3352453112602234, 'rewards/total_composite/std': 0.1382029503583908, 'reward': 0.3352453112602234, 'reward_std': 0.1382029503583908, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10336823761463165, 'sampling/sampling_logp_difference/max': 1.3734700679779053, 'sampling/importance_sampling_ratio/min': 0.25322672724723816, 'sampling/importance_sampling_ratio/mean': 1.0265514850616455, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5914758183062077, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07886998308822513, 'clip_ratio/high_max': 0.07886998308822513, 'clip_ratio/region_mean': 0.07886998308822513, 'reward_total_mean': 0.3352453112602234, 'reward_meter_mean': 0.5514540672302246, 'reward_meter_std': 0.40759724378585815, 'reward_count_adherence_mean': 0.9583333730697632, 'reward_count_adherence_std': 0.117851123213768, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9403271079063416, 'reward_repeat_soft_std': 0.03892381489276886, 'reward_judge_quality_mean': 0.13375000655651093, 'reward_judge_quality_std': 0.03889087587594986, 'reward_total_composite_mean': 0.3352453112602234, 'reward_total_composite_std': 0.1382029503583908, 'epoch': 0.21}

 63%|██████▎   | 2089/3300 [5:07:47<3:41:06, 10.95s/it]INFO 04-13 12:28:45 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2090/3300 [5:07:55<3:19:59,  9.92s/it]2026-04-13 12:28:53,248 | INFO | train_grpo_train | metrics_logged mode=train step=2090

                                                       
{'loss': -0.0396, 'grad_norm': 8.443550109863281, 'learning_rate': 3.6696969696969697e-06, 'num_tokens': 3761106.0, 'completions/mean_length': 80.125, 'completions/min_length': 62.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 80.125, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.616234540939331, 'rewards/meter/std': 0.30755528807640076, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8544143438339233, 'rewards/repeat_soft/std': 0.059049081057310104, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.3969932198524475, 'rewards/total_composite/std': 0.03428816422820091, 'reward': 0.3969932198524475, 'reward_std': 0.034288160502910614, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10842517018318176, 'sampling/sampling_logp_difference/max': 1.542647361755371, 'sampling/importance_sampling_ratio/min': 0.23686398565769196, 'sampling/importance_sampling_ratio/mean': 1.0151734352111816, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7614212185144424, 'clip_ratio/low_mean': 0.04624854167923331, 'clip_ratio/low_min': 0.04624854167923331, 'clip_ratio/high_mean': 0.05727146123535931, 'clip_ratio/high_max': 0.05727146123535931, 'clip_ratio/region_mean': 0.10352000291459262, 'reward_total_mean': 0.3969932198524475, 'reward_meter_mean': 0.616234540939331, 'reward_meter_std': 0.30755528807640076, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8544143438339233, 'reward_repeat_soft_std': 0.059049081057310104, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.3969932198524475, 'reward_total_composite_std': 0.03428816422820091, 'epoch': 0.21}

 63%|██████▎   | 2090/3300 [5:07:55<3:19:59,  9.92s/it]INFO 04-13 12:28:53 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2091/3300 [5:08:01<2:58:08,  8.84s/it]2026-04-13 12:28:59,577 | INFO | train_grpo_train | metrics_logged mode=train step=2091

                                                       
{'loss': 0.0817, 'grad_norm': 14.345635414123535, 'learning_rate': 3.6666666666666666e-06, 'num_tokens': 3762567.0, 'completions/mean_length': 36.625, 'completions/min_length': 32.0, 'completions/max_length': 44.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 36.625, 'completions/min_terminated_length': 32.0, 'completions/max_terminated_length': 44.0, 'rewards/meter/mean': 0.9297787547111511, 'rewards/meter/std': 0.13060350716114044, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9389592409133911, 'rewards/repeat_soft/std': 0.05289275944232941, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43149733543395996, 'rewards/total_composite/std': 0.015238284133374691, 'reward': 0.43149733543395996, 'reward_std': 0.015238286927342415, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13995887339115143, 'sampling/sampling_logp_difference/max': 1.4572899341583252, 'sampling/importance_sampling_ratio/min': 0.24173982441425323, 'sampling/importance_sampling_ratio/mean': 0.9981912970542908, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7522866874933243, 'clip_ratio/low_mean': 0.058080810122191906, 'clip_ratio/low_min': 0.058080810122191906, 'clip_ratio/high_mean': 0.10172371473163366, 'clip_ratio/high_max': 0.10172371473163366, 'clip_ratio/region_mean': 0.15980452485382557, 'reward_total_mean': 0.43149733543395996, 'reward_meter_mean': 0.9297787547111511, 'reward_meter_std': 0.13060350716114044, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9389592409133911, 'reward_repeat_soft_std': 0.05289275944232941, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43149733543395996, 'reward_total_composite_std': 0.015238284133374691, 'epoch': 0.21}

 63%|██████▎   | 2091/3300 [5:08:01<2:58:08,  8.84s/it]INFO 04-13 12:28:59 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2092/3300 [5:08:07<2:41:16,  8.01s/it]2026-04-13 12:29:05,655 | INFO | train_grpo_train | metrics_logged mode=train step=2092

                                                       
{'loss': -0.0283, 'grad_norm': 15.965852737426758, 'learning_rate': 3.6636363636363643e-06, 'num_tokens': 3763992.0, 'completions/mean_length': 31.125, 'completions/min_length': 27.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.125, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.8399671316146851, 'rewards/meter/std': 0.24304862320423126, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9808377623558044, 'rewards/repeat_soft/std': 0.020741088315844536, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4290224611759186, 'rewards/total_composite/std': 0.02329912595450878, 'reward': 0.4290224611759186, 'reward_std': 0.02329912595450878, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08482443541288376, 'sampling/sampling_logp_difference/max': 1.498971700668335, 'sampling/importance_sampling_ratio/min': 0.22335973381996155, 'sampling/importance_sampling_ratio/mean': 1.0098867416381836, 'sampling/importance_sampling_ratio/max': 1.4488673210144043, 'entropy': 0.4668281860649586, 'clip_ratio/low_mean': 0.025388291105628014, 'clip_ratio/low_min': 0.025388291105628014, 'clip_ratio/high_mean': 0.07872221665456891, 'clip_ratio/high_max': 0.07872221665456891, 'clip_ratio/region_mean': 0.10411050776019692, 'reward_total_mean': 0.4290224611759186, 'reward_meter_mean': 0.8399671316146851, 'reward_meter_std': 0.24304862320423126, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9808377623558044, 'reward_repeat_soft_std': 0.020741088315844536, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4290224611759186, 'reward_total_composite_std': 0.02329912595450878, 'epoch': 0.21}

 63%|██████▎   | 2092/3300 [5:08:07<2:41:16,  8.01s/it]INFO 04-13 12:29:05 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2093/3300 [5:08:14<2:34:47,  7.69s/it]2026-04-13 12:29:12,599 | INFO | train_grpo_train | metrics_logged mode=train step=2093

                                                       
{'loss': -0.0161, 'grad_norm': 10.055636405944824, 'learning_rate': 3.660606060606061e-06, 'num_tokens': 3766114.0, 'completions/mean_length': 76.25, 'completions/min_length': 67.0, 'completions/max_length': 79.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 76.25, 'completions/min_terminated_length': 67.0, 'completions/max_terminated_length': 79.0, 'rewards/meter/mean': 0.8110104203224182, 'rewards/meter/std': 0.3234376907348633, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8537755012512207, 'rewards/repeat_soft/std': 0.04971456155180931, 'rewards/judge_quality/mean': 0.18000000715255737, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4229545593261719, 'rewards/total_composite/std': 0.03832192346453667, 'reward': 0.4229545593261719, 'reward_std': 0.03832192346453667, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0807785764336586, 'sampling/sampling_logp_difference/max': 1.7626324892044067, 'sampling/importance_sampling_ratio/min': 0.17159254848957062, 'sampling/importance_sampling_ratio/mean': 1.0127191543579102, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4079204872250557, 'clip_ratio/low_mean': 0.023201302159577608, 'clip_ratio/low_min': 0.023201302159577608, 'clip_ratio/high_mean': 0.0597048276104033, 'clip_ratio/high_max': 0.0597048276104033, 'clip_ratio/region_mean': 0.08290612976998091, 'reward_total_mean': 0.4229545593261719, 'reward_meter_mean': 0.8110104203224182, 'reward_meter_std': 0.3234376907348633, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8537755012512207, 'reward_repeat_soft_std': 0.04971456155180931, 'reward_judge_quality_mean': 0.18000000715255737, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4229545593261719, 'reward_total_composite_std': 0.03832192346453667, 'epoch': 0.21}

 63%|██████▎   | 2093/3300 [5:08:14<2:34:47,  7.69s/it]INFO 04-13 12:29:12 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2094/3300 [5:08:20<2:24:54,  7.21s/it]2026-04-13 12:29:18,679 | INFO | train_grpo_train | metrics_logged mode=train step=2094

                                                       
{'loss': 0.0236, 'grad_norm': 15.249076843261719, 'learning_rate': 3.657575757575758e-06, 'num_tokens': 3767484.0, 'completions/mean_length': 19.25, 'completions/min_length': 18.0, 'completions/max_length': 21.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 19.25, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 21.0, 'rewards/meter/mean': 0.9565207958221436, 'rewards/meter/std': 0.07130192965269089, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8878955841064453, 'rewards/repeat_soft/std': 0.1174720972776413, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4264451265335083, 'rewards/total_composite/std': 0.01699831150472164, 'reward': 0.4264451265335083, 'reward_std': 0.016998320817947388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09562919288873672, 'sampling/sampling_logp_difference/max': 0.7266945838928223, 'sampling/importance_sampling_ratio/min': 0.5248639583587646, 'sampling/importance_sampling_ratio/mean': 1.044858694076538, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.764565996825695, 'clip_ratio/low_mean': 0.04444444505497813, 'clip_ratio/low_min': 0.04444444505497813, 'clip_ratio/high_mean': 0.06150793796405196, 'clip_ratio/high_max': 0.06150793796405196, 'clip_ratio/region_mean': 0.1059523830190301, 'reward_total_mean': 0.4264451265335083, 'reward_meter_mean': 0.9565207958221436, 'reward_meter_std': 0.07130192965269089, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8878955841064453, 'reward_repeat_soft_std': 0.1174720972776413, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4264451265335083, 'reward_total_composite_std': 0.01699831150472164, 'epoch': 0.21}

 63%|██████▎   | 2094/3300 [5:08:20<2:24:54,  7.21s/it]INFO 04-13 12:29:18 [block_pool.py:378] Successfully reset prefix cache

 63%|██████▎   | 2095/3300 [5:08:32<2:49:36,  8.45s/it]2026-04-13 12:29:30,006 | INFO | train_grpo_train | metrics_logged mode=train step=2095

                                                       
{'loss': -0.1317, 'grad_norm': 1.4918196201324463, 'learning_rate': 3.654545454545455e-06, 'num_tokens': 3769136.0, 'completions/mean_length': 235.5, 'completions/min_length': 58.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.375, 'completions/mean_terminated_length': 69.5999984741211, 'completions/min_terminated_length': 58.0, 'completions/max_terminated_length': 86.0, 'rewards/meter/mean': 0.6806774139404297, 'rewards/meter/std': 0.41470572352409363, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.625, 'rewards/hard_gate/std': 0.5175492167472839, 'rewards/repeat_soft/mean': 0.9118834733963013, 'rewards/repeat_soft/std': 0.05751105025410652, 'rewards/judge_quality/mean': 0.10875000059604645, 'rewards/judge_quality/std': 0.052218638360500336, 'rewards/total_composite/mean': 0.26164764165878296, 'rewards/total_composite/std': 0.21728645265102386, 'reward': 0.26164764165878296, 'reward_std': 0.21728643774986267, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10722679644823074, 'sampling/sampling_logp_difference/max': 2.4546332359313965, 'sampling/importance_sampling_ratio/min': 0.08589468896389008, 'sampling/importance_sampling_ratio/mean': 1.0040150880813599, 'sampling/importance_sampling_ratio/max': 1.8782011270523071, 'entropy': 0.3790879473090172, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.06049556843936443, 'clip_ratio/high_max': 0.06049556843936443, 'clip_ratio/region_mean': 0.06049556843936443, 'reward_total_mean': 0.26164764165878296, 'reward_meter_mean': 0.6806774139404297, 'reward_meter_std': 0.41470572352409363, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.625, 'reward_hard_gate_std': 0.5175492167472839, 'reward_repeat_soft_mean': 0.9118834733963013, 'reward_repeat_soft_std': 0.05751105025410652, 'reward_judge_quality_mean': 0.10875000059604645, 'reward_judge_quality_std': 0.052218638360500336, 'reward_total_composite_mean': 0.26164764165878296, 'reward_total_composite_std': 0.21728645265102386, 'epoch': 0.21}

 63%|██████▎   | 2095/3300 [5:08:32<2:49:36,  8.45s/it]INFO 04-13 12:29:30 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2096/3300 [5:08:37<2:33:02,  7.63s/it]2026-04-13 12:29:35,728 | INFO | train_grpo_train | metrics_logged mode=train step=2096

                                                       
{'loss': 0.0259, 'grad_norm': 11.645296096801758, 'learning_rate': 3.651515151515152e-06, 'num_tokens': 3770571.0, 'completions/mean_length': 22.375, 'completions/min_length': 22.0, 'completions/max_length': 25.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.375, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 25.0, 'rewards/meter/mean': 0.8985494375228882, 'rewards/meter/std': 0.16152767837047577, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8453850746154785, 'rewards/repeat_soft/std': 0.1095879003405571, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4144163429737091, 'rewards/total_composite/std': 0.013869404792785645, 'reward': 0.4144163429737091, 'reward_std': 0.013869399204850197, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0835302397608757, 'sampling/sampling_logp_difference/max': 0.940463662147522, 'sampling/importance_sampling_ratio/min': 0.3904467821121216, 'sampling/importance_sampling_ratio/mean': 1.0118560791015625, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5201849527657032, 'clip_ratio/low_mean': 0.039090909995138645, 'clip_ratio/low_min': 0.039090909995138645, 'clip_ratio/high_mean': 0.011363636702299118, 'clip_ratio/high_max': 0.011363636702299118, 'clip_ratio/region_mean': 0.05045454669743776, 'reward_total_mean': 0.4144163429737091, 'reward_meter_mean': 0.8985494375228882, 'reward_meter_std': 0.16152767837047577, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8453850746154785, 'reward_repeat_soft_std': 0.1095879003405571, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4144163429737091, 'reward_total_composite_std': 0.013869404792785645, 'epoch': 0.21}

 64%|██████▎   | 2096/3300 [5:08:37<2:33:02,  7.63s/it]INFO 04-13 12:29:35 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2097/3300 [5:08:49<2:57:02,  8.83s/it]2026-04-13 12:29:47,354 | INFO | train_grpo_train | metrics_logged mode=train step=2097

                                                       
{'loss': -0.1006, 'grad_norm': 1.2347784042358398, 'learning_rate': 3.648484848484849e-06, 'num_tokens': 3771978.0, 'completions/mean_length': 158.875, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.25, 'completions/mean_terminated_length': 41.16666793823242, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.9143640995025635, 'rewards/meter/std': 0.12278993427753448, 'rewards/count_adherence/mean': 0.9375, 'rewards/count_adherence/std': 0.1767766922712326, 'rewards/hard_gate/mean': 0.75, 'rewards/hard_gate/std': 0.4629100561141968, 'rewards/repeat_soft/mean': 0.9353083968162537, 'rewards/repeat_soft/std': 0.037512220442295074, 'rewards/judge_quality/mean': 0.125, 'rewards/judge_quality/std': 0.046291008591651917, 'rewards/total_composite/mean': 0.3247089385986328, 'rewards/total_composite/std': 0.2005159854888916, 'reward': 0.3247089385986328, 'reward_std': 0.2005159854888916, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1059136688709259, 'sampling/sampling_logp_difference/max': 1.8650531768798828, 'sampling/importance_sampling_ratio/min': 0.15488797426223755, 'sampling/importance_sampling_ratio/mean': 0.9949530363082886, 'sampling/importance_sampling_ratio/max': 1.6144154071807861, 'entropy': 0.442423515021801, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.08202262991108, 'clip_ratio/high_max': 0.08202262991108, 'clip_ratio/region_mean': 0.08202262991108, 'reward_total_mean': 0.3247089385986328, 'reward_meter_mean': 0.9143640995025635, 'reward_meter_std': 0.12278993427753448, 'reward_count_adherence_mean': 0.9375, 'reward_count_adherence_std': 0.1767766922712326, 'reward_hard_gate_mean': 0.75, 'reward_hard_gate_std': 0.4629100561141968, 'reward_repeat_soft_mean': 0.9353083968162537, 'reward_repeat_soft_std': 0.037512220442295074, 'reward_judge_quality_mean': 0.125, 'reward_judge_quality_std': 0.046291008591651917, 'reward_total_composite_mean': 0.3247089385986328, 'reward_total_composite_std': 0.2005159854888916, 'epoch': 0.21}

 64%|██████▎   | 2097/3300 [5:08:49<2:57:02,  8.83s/it]INFO 04-13 12:29:47 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2098/3300 [5:09:00<3:11:55,  9.58s/it]2026-04-13 12:29:58,695 | INFO | train_grpo_train | metrics_logged mode=train step=2098

                                                       
{'loss': -0.1383, 'grad_norm': 1.7239394187927246, 'learning_rate': 3.645454545454546e-06, 'num_tokens': 3773626.0, 'completions/mean_length': 106.0, 'completions/min_length': 41.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 48.000003814697266, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.8133279085159302, 'rewards/meter/std': 0.29564806818962097, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9431638717651367, 'rewards/repeat_soft/std': 0.037652309983968735, 'rewards/judge_quality/mean': 0.14124999940395355, 'rewards/judge_quality/std': 0.038335926830768585, 'rewards/total_composite/mean': 0.37816929817199707, 'rewards/total_composite/std': 0.15340575575828552, 'reward': 0.37816929817199707, 'reward_std': 0.15340575575828552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09945490211248398, 'sampling/sampling_logp_difference/max': 2.006791114807129, 'sampling/importance_sampling_ratio/min': 0.13441932201385498, 'sampling/importance_sampling_ratio/mean': 1.0164374113082886, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.554730236530304, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.09761778451502323, 'clip_ratio/high_max': 0.09761778451502323, 'clip_ratio/region_mean': 0.09761778451502323, 'reward_total_mean': 0.37816929817199707, 'reward_meter_mean': 0.8133279085159302, 'reward_meter_std': 0.29564806818962097, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9431638717651367, 'reward_repeat_soft_std': 0.037652309983968735, 'reward_judge_quality_mean': 0.14124999940395355, 'reward_judge_quality_std': 0.038335926830768585, 'reward_total_composite_mean': 0.37816929817199707, 'reward_total_composite_std': 0.15340575575828552, 'epoch': 0.21}

 64%|██████▎   | 2098/3300 [5:09:00<3:11:55,  9.58s/it]INFO 04-13 12:29:58 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2099/3300 [5:09:07<2:52:19,  8.61s/it]2026-04-13 12:30:05,047 | INFO | train_grpo_train | metrics_logged mode=train step=2099

                                                       
{'loss': 0.0105, 'grad_norm': 15.438549041748047, 'learning_rate': 3.642424242424243e-06, 'num_tokens': 3775132.0, 'completions/mean_length': 31.25, 'completions/min_length': 27.0, 'completions/max_length': 35.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 31.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 35.0, 'rewards/meter/mean': 0.44214296340942383, 'rewards/meter/std': 0.4177422821521759, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9636186361312866, 'rewards/repeat_soft/std': 0.0434965156018734, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.387692928314209, 'rewards/total_composite/std': 0.04447413608431816, 'reward': 0.387692928314209, 'reward_std': 0.04447413608431816, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09119374305009842, 'sampling/sampling_logp_difference/max': 1.323554515838623, 'sampling/importance_sampling_ratio/min': 0.266187459230423, 'sampling/importance_sampling_ratio/mean': 1.0038379430770874, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.486721184104681, 'clip_ratio/low_mean': 0.06431080307811499, 'clip_ratio/low_min': 0.06431080307811499, 'clip_ratio/high_mean': 0.023761424235999584, 'clip_ratio/high_max': 0.023761424235999584, 'clip_ratio/region_mean': 0.08807222731411457, 'reward_total_mean': 0.387692928314209, 'reward_meter_mean': 0.44214296340942383, 'reward_meter_std': 0.4177422821521759, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9636186361312866, 'reward_repeat_soft_std': 0.0434965156018734, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.387692928314209, 'reward_total_composite_std': 0.04447413608431816, 'epoch': 0.21}

 64%|██████▎   | 2099/3300 [5:09:07<2:52:19,  8.61s/it]INFO 04-13 12:30:05 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2100/3300 [5:09:14<2:43:21,  8.17s/it]2026-04-13 12:30:12,182 | INFO | train_grpo_train | metrics_logged mode=train step=2100

                                                       
{'loss': 0.0083, 'grad_norm': 6.016146659851074, 'learning_rate': 3.6393939393939398e-06, 'num_tokens': 3777336.0, 'completions/mean_length': 102.5, 'completions/min_length': 92.0, 'completions/max_length': 112.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.5, 'completions/min_terminated_length': 92.0, 'completions/max_terminated_length': 112.0, 'rewards/meter/mean': 0.8390237092971802, 'rewards/meter/std': 0.1587734818458557, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.861441433429718, 'rewards/repeat_soft/std': 0.029279708862304688, 'rewards/judge_quality/mean': 0.18000000715255737, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4273819923400879, 'rewards/total_composite/std': 0.015358088538050652, 'reward': 0.4273819923400879, 'reward_std': 0.015358081087470055, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07864245772361755, 'sampling/sampling_logp_difference/max': 1.8556418418884277, 'sampling/importance_sampling_ratio/min': 0.15635254979133606, 'sampling/importance_sampling_ratio/mean': 0.9945199489593506, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4030364230275154, 'clip_ratio/low_mean': 0.027041461784392595, 'clip_ratio/low_min': 0.027041461784392595, 'clip_ratio/high_mean': 0.055136412382125854, 'clip_ratio/high_max': 0.055136412382125854, 'clip_ratio/region_mean': 0.08217787416651845, 'reward_total_mean': 0.4273819923400879, 'reward_meter_mean': 0.8390237092971802, 'reward_meter_std': 0.1587734818458557, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.861441433429718, 'reward_repeat_soft_std': 0.029279708862304688, 'reward_judge_quality_mean': 0.18000000715255737, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4273819923400879, 'reward_total_composite_std': 0.015358088538050652, 'epoch': 0.21}

 64%|██████▎   | 2100/3300 [5:09:14<2:43:21,  8.17s/it]INFO 04-13 12:30:12 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:03<00:14,  1.85s/it]

 30%|███       | 3/10 [00:12<00:33,  4.78s/it]

 40%|████      | 4/10 [00:16<00:26,  4.34s/it]

 50%|█████     | 5/10 [00:27<00:34,  6.89s/it]

 60%|██████    | 6/10 [00:32<00:23,  5.99s/it]

 70%|███████   | 7/10 [00:36<00:16,  5.49s/it]

 80%|████████  | 8/10 [00:45<00:13,  6.57s/it]

 90%|█████████ | 9/10 [00:49<00:05,  5.75s/it]

100%|██████████| 10/10 [00:57<00:00,  6.59s/it]2026-04-13 12:31:14,552 | INFO | train_grpo_train | metrics_logged mode=eval step=2100

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 62.2555, 'eval_samples_per_second': 1.285, 'eval_steps_per_second': 0.161, 'eval_num_tokens': 3777336.0, 'eval_completions/mean_length': 96.5875, 'eval_completions/min_length': 30.9, 'eval_completions/max_length': 246.9, 'eval_completions/clipped_ratio': 0.05, 'eval_completions/mean_terminated_length': 74.14821548461914, 'eval_completions/min_terminated_length': 30.9, 'eval_completions/max_terminated_length': 132.5, 'eval_rewards/meter/mean': 0.7549513936042785, 'eval_rewards/meter/std': 0.2847424909472466, 'eval_rewards/count_adherence/mean': 0.9800000011920929, 'eval_rewards/count_adherence/std': 0.051684608310461046, 'eval_rewards/hard_gate/mean': 0.95, 'eval_rewards/hard_gate/std': 0.11700168251991272, 'eval_rewards/repeat_soft/mean': 0.9035159707069397, 'eval_rewards/repeat_soft/std': 0.07267097979784012, 'eval_rewards/judge_quality/mean': 0.1567500039935112, 'eval_rewards/judge_quality/std': 0.03720051711425185, 'eval_rewards/total_composite/mean': 0.3925432741641998, 'eval_rewards/total_composite/std': 0.07047434505075216, 'eval_reward': 0.3925432741641998, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.055311382934451106, 'eval_sampling/sampling_logp_difference/max': 0.9022397518157959, 'eval_sampling/importance_sampling_ratio/min': 0.41649395823478697, 'eval_sampling/importance_sampling_ratio/mean': 1.0110464453697205, 'eval_sampling/importance_sampling_ratio/max': 1.4670676112174987, 'eval_entropy': 0.5670681864023208, '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': 0.3925432741641998, 'eval_reward_meter_mean': 0.7549513936042785, 'eval_reward_meter_std': 0.2847424909472466, 'eval_reward_count_adherence_mean': 0.9800000011920929, 'eval_reward_count_adherence_std': 0.051684608310461046, 'eval_reward_hard_gate_mean': 0.95, 'eval_reward_hard_gate_std': 0.11700168251991272, 'eval_reward_repeat_soft_mean': 0.9035159707069397, 'eval_reward_repeat_soft_std': 0.07267097979784012, 'eval_reward_judge_quality_mean': 0.1567500039935112, 'eval_reward_judge_quality_std': 0.03720051711425185, 'eval_reward_total_composite_mean': 0.3925432741641998, 'eval_reward_total_composite_std': 0.07047434505075216, 'epoch': 0.21}

 64%|██████▎   | 2100/3300 [5:10:16<2:43:21,  8.17s/it]

100%|██████████| 10/10 [00:57<00:00,  6.59s/it]

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

 64%|██████▎   | 2101/3300 [5:10:32<9:41:08, 29.08s/it]2026-04-13 12:31:30,065 | INFO | train_grpo_train | metrics_logged mode=train step=2101

                                                       
{'loss': -0.1714, 'grad_norm': 1.856912612915039, 'learning_rate': 3.6363636363636366e-06, 'num_tokens': 3779100.0, 'completions/mean_length': 128.5, 'completions/min_length': 56.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 73.71428680419922, 'completions/min_terminated_length': 56.0, 'completions/max_terminated_length': 89.0, 'rewards/meter/mean': 0.7168045043945312, 'rewards/meter/std': 0.3875977396965027, 'rewards/count_adherence/mean': 0.9166666865348816, 'rewards/count_adherence/std': 0.2357022762298584, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.952052652835846, 'rewards/repeat_soft/std': 0.033339813351631165, 'rewards/judge_quality/mean': 0.14125001430511475, 'rewards/judge_quality/std': 0.038335926830768585, 'rewards/total_composite/mean': 0.37015044689178467, 'rewards/total_composite/std': 0.15158022940158844, 'reward': 0.37015044689178467, 'reward_std': 0.15158022940158844, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11523141711950302, 'sampling/sampling_logp_difference/max': 1.797853708267212, 'sampling/importance_sampling_ratio/min': 0.1656540483236313, 'sampling/importance_sampling_ratio/mean': 0.9991225600242615, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7288566678762436, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11735270358622074, 'clip_ratio/high_max': 0.11735270358622074, 'clip_ratio/region_mean': 0.11735270358622074, 'reward_total_mean': 0.37015044689178467, 'reward_meter_mean': 0.7168045043945312, 'reward_meter_std': 0.3875977396965027, 'reward_count_adherence_mean': 0.9166666865348816, 'reward_count_adherence_std': 0.2357022762298584, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.952052652835846, 'reward_repeat_soft_std': 0.033339813351631165, 'reward_judge_quality_mean': 0.14125001430511475, 'reward_judge_quality_std': 0.038335926830768585, 'reward_total_composite_mean': 0.37015044689178467, 'reward_total_composite_std': 0.15158022940158844, 'epoch': 0.21}

 64%|██████▎   | 2101/3300 [5:10:32<9:41:08, 29.08s/it]INFO 04-13 12:31:30 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2102/3300 [5:10:44<7:59:19, 24.01s/it]2026-04-13 12:31:42,227 | INFO | train_grpo_train | metrics_logged mode=train step=2102

                                                       
{'loss': -0.1172, 'grad_norm': 2.055514097213745, 'learning_rate': 3.633333333333334e-06, 'num_tokens': 3780782.0, 'completions/mean_length': 98.25, 'completions/min_length': 36.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 39.142860412597656, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 45.0, 'rewards/meter/mean': 0.44329869747161865, 'rewards/meter/std': 0.40356191992759705, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9842722415924072, 'rewards/repeat_soft/std': 0.024804506450891495, 'rewards/judge_quality/mean': 0.1862500011920929, 'rewards/judge_quality/std': 0.11475905776023865, 'rewards/total_composite/mean': 0.35340049862861633, 'rewards/total_composite/std': 0.14779430627822876, 'reward': 0.35340049862861633, 'reward_std': 0.14779430627822876, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12127834558486938, 'sampling/sampling_logp_difference/max': 1.1002140045166016, 'sampling/importance_sampling_ratio/min': 0.33279985189437866, 'sampling/importance_sampling_ratio/mean': 1.0372201204299927, 'sampling/importance_sampling_ratio/max': 1.880364179611206, 'entropy': 0.7180831506848335, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07393320184201002, 'clip_ratio/high_max': 0.07393320184201002, 'clip_ratio/region_mean': 0.07393320184201002, 'reward_total_mean': 0.35340049862861633, 'reward_meter_mean': 0.44329869747161865, 'reward_meter_std': 0.40356191992759705, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9842722415924072, 'reward_repeat_soft_std': 0.024804506450891495, 'reward_judge_quality_mean': 0.1862500011920929, 'reward_judge_quality_std': 0.11475905776023865, 'reward_total_composite_mean': 0.35340049862861633, 'reward_total_composite_std': 0.14779430627822876, 'epoch': 0.21}

 64%|██████▎   | 2102/3300 [5:10:44<7:59:19, 24.01s/it]INFO 04-13 12:31:42 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▎   | 2103/3300 [5:10:50<6:12:43, 18.68s/it]2026-04-13 12:31:48,489 | INFO | train_grpo_train | metrics_logged mode=train step=2103

                                                       
{'loss': 0.0086, 'grad_norm': 11.103949546813965, 'learning_rate': 3.6303030303030307e-06, 'num_tokens': 3782484.0, 'completions/mean_length': 48.75, 'completions/min_length': 45.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.75, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.442565381526947, 'rewards/meter/std': 0.30177053809165955, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9805001020431519, 'rewards/repeat_soft/std': 0.017646679654717445, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.39005687832832336, 'rewards/total_composite/std': 0.02893988788127899, 'reward': 0.39005687832832336, 'reward_std': 0.02893988974392414, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.1269722580909729, 'sampling/sampling_logp_difference/max': 2.509066343307495, 'sampling/importance_sampling_ratio/min': 0.08134415000677109, 'sampling/importance_sampling_ratio/mean': 1.005758285522461, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.807975247502327, 'clip_ratio/low_mean': 0.07524633966386318, 'clip_ratio/low_min': 0.07524633966386318, 'clip_ratio/high_mean': 0.05287114903330803, 'clip_ratio/high_max': 0.05287114903330803, 'clip_ratio/region_mean': 0.1281174886971712, 'reward_total_mean': 0.39005687832832336, 'reward_meter_mean': 0.442565381526947, 'reward_meter_std': 0.30177053809165955, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9805001020431519, 'reward_repeat_soft_std': 0.017646679654717445, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.39005687832832336, 'reward_total_composite_std': 0.02893988788127899, 'epoch': 0.21}

 64%|██████▎   | 2103/3300 [5:10:50<6:12:43, 18.68s/it]INFO 04-13 12:31:48 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2104/3300 [5:10:57<5:01:47, 15.14s/it]2026-04-13 12:31:55,359 | INFO | train_grpo_train | metrics_logged mode=train step=2104

                                                       
{'loss': -0.0374, 'grad_norm': 10.681594848632812, 'learning_rate': 3.6272727272727275e-06, 'num_tokens': 3784613.0, 'completions/mean_length': 88.125, 'completions/min_length': 71.0, 'completions/max_length': 105.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 88.125, 'completions/min_terminated_length': 71.0, 'completions/max_terminated_length': 105.0, 'rewards/meter/mean': 0.6759259700775146, 'rewards/meter/std': 0.307078093290329, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9187784194946289, 'rewards/repeat_soft/std': 0.04656059294939041, 'rewards/judge_quality/mean': 0.16875001788139343, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4134945273399353, 'rewards/total_composite/std': 0.036576032638549805, 'reward': 0.4134945273399353, 'reward_std': 0.0365760400891304, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11971750110387802, 'sampling/sampling_logp_difference/max': 2.2448527812957764, 'sampling/importance_sampling_ratio/min': 0.10594313591718674, 'sampling/importance_sampling_ratio/mean': 0.9974398612976074, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43947426602244377, 'clip_ratio/low_mean': 0.03905245568603277, 'clip_ratio/low_min': 0.03905245568603277, 'clip_ratio/high_mean': 0.07308703940361738, 'clip_ratio/high_max': 0.07308703940361738, 'clip_ratio/region_mean': 0.11213949508965015, 'reward_total_mean': 0.4134945273399353, 'reward_meter_mean': 0.6759259700775146, 'reward_meter_std': 0.307078093290329, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9187784194946289, 'reward_repeat_soft_std': 0.04656059294939041, 'reward_judge_quality_mean': 0.16875001788139343, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4134945273399353, 'reward_total_composite_std': 0.036576032638549805, 'epoch': 0.21}

 64%|██████▍   | 2104/3300 [5:10:57<5:01:47, 15.14s/it]INFO 04-13 12:31:55 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2105/3300 [5:11:04<4:11:08, 12.61s/it]2026-04-13 12:32:02,062 | INFO | train_grpo_train | metrics_logged mode=train step=2105

                                                       
{'loss': 0.0628, 'grad_norm': 12.617268562316895, 'learning_rate': 3.6242424242424248e-06, 'num_tokens': 3786334.0, 'completions/mean_length': 48.125, 'completions/min_length': 43.0, 'completions/max_length': 55.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 48.125, 'completions/min_terminated_length': 43.0, 'completions/max_terminated_length': 55.0, 'rewards/meter/mean': 0.6768026947975159, 'rewards/meter/std': 0.35109129548072815, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9743654727935791, 'rewards/repeat_soft/std': 0.02952939085662365, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4121430814266205, 'rewards/total_composite/std': 0.034643322229385376, 'reward': 0.4121430814266205, 'reward_std': 0.034643322229385376, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12021853029727936, 'sampling/sampling_logp_difference/max': 2.3435730934143066, 'sampling/importance_sampling_ratio/min': 0.0959840640425682, 'sampling/importance_sampling_ratio/mean': 1.0105738639831543, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8007223606109619, 'clip_ratio/low_mean': 0.03388928063213825, 'clip_ratio/low_min': 0.03388928063213825, 'clip_ratio/high_mean': 0.08387564588338137, 'clip_ratio/high_max': 0.08387564588338137, 'clip_ratio/region_mean': 0.11776492651551962, 'reward_total_mean': 0.4121430814266205, 'reward_meter_mean': 0.6768026947975159, 'reward_meter_std': 0.35109129548072815, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9743654727935791, 'reward_repeat_soft_std': 0.02952939085662365, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4121430814266205, 'reward_total_composite_std': 0.034643322229385376, 'epoch': 0.21}

 64%|██████▍   | 2105/3300 [5:11:04<4:11:08, 12.61s/it]INFO 04-13 12:32:02 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2106/3300 [5:11:11<3:41:21, 11.12s/it]2026-04-13 12:32:09,804 | INFO | train_grpo_train | metrics_logged mode=train step=2106

                                                       
{'loss': 0.0452, 'grad_norm': 9.996867179870605, 'learning_rate': 3.6212121212121216e-06, 'num_tokens': 3788346.0, 'completions/mean_length': 84.5, 'completions/min_length': 73.0, 'completions/max_length': 96.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 84.5, 'completions/min_terminated_length': 73.0, 'completions/max_terminated_length': 96.0, 'rewards/meter/mean': 0.6127023100852966, 'rewards/meter/std': 0.33788344264030457, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.950439453125, 'rewards/repeat_soft/std': 0.04181034490466118, 'rewards/judge_quality/mean': 0.1612500101327896, 'rewards/judge_quality/std': 0.022320717573165894, 'rewards/total_composite/mean': 0.40567585825920105, 'rewards/total_composite/std': 0.03350796550512314, 'reward': 0.40567585825920105, 'reward_std': 0.03350795805454254, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11898557841777802, 'sampling/sampling_logp_difference/max': 2.1936473846435547, 'sampling/importance_sampling_ratio/min': 0.1115092858672142, 'sampling/importance_sampling_ratio/mean': 1.0094692707061768, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.7003876119852066, 'clip_ratio/low_mean': 0.0763331362977624, 'clip_ratio/low_min': 0.0763331362977624, 'clip_ratio/high_mean': 0.05873551592230797, 'clip_ratio/high_max': 0.05873551592230797, 'clip_ratio/region_mean': 0.13506865222007036, 'reward_total_mean': 0.40567585825920105, 'reward_meter_mean': 0.6127023100852966, 'reward_meter_std': 0.33788344264030457, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.950439453125, 'reward_repeat_soft_std': 0.04181034490466118, 'reward_judge_quality_mean': 0.1612500101327896, 'reward_judge_quality_std': 0.022320717573165894, 'reward_total_composite_mean': 0.40567585825920105, 'reward_total_composite_std': 0.03350796550512314, 'epoch': 0.21}

 64%|██████▍   | 2106/3300 [5:11:12<3:41:21, 11.12s/it]INFO 04-13 12:32:10 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2107/3300 [5:11:19<3:19:20, 10.03s/it]2026-04-13 12:32:17,179 | INFO | train_grpo_train | metrics_logged mode=train step=2107

                                                       
{'loss': 0.0247, 'grad_norm': 12.233802795410156, 'learning_rate': 3.6181818181818184e-06, 'num_tokens': 3789996.0, 'completions/mean_length': 41.25, 'completions/min_length': 38.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.25, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.9781842231750488, 'rewards/meter/std': 0.015315087512135506, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.7953553199768066, 'rewards/repeat_soft/std': 0.11163556575775146, 'rewards/judge_quality/mean': 0.13875000178813934, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.40734606981277466, 'rewards/total_composite/std': 0.02720000222325325, 'reward': 0.40734606981277466, 'reward_std': 0.027200007811188698, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09198253601789474, 'sampling/sampling_logp_difference/max': 1.8582334518432617, 'sampling/importance_sampling_ratio/min': 0.15594787895679474, 'sampling/importance_sampling_ratio/mean': 1.0048346519470215, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3964330926537514, 'clip_ratio/low_mean': 0.027148242574185133, 'clip_ratio/low_min': 0.027148242574185133, 'clip_ratio/high_mean': 0.03724791551940143, 'clip_ratio/high_max': 0.03724791551940143, 'clip_ratio/region_mean': 0.06439615809358656, 'reward_total_mean': 0.40734606981277466, 'reward_meter_mean': 0.9781842231750488, 'reward_meter_std': 0.015315087512135506, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.7953553199768066, 'reward_repeat_soft_std': 0.11163556575775146, 'reward_judge_quality_mean': 0.13875000178813934, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.40734606981277466, 'reward_total_composite_std': 0.02720000222325325, 'epoch': 0.21}

 64%|██████▍   | 2107/3300 [5:11:19<3:19:20, 10.03s/it]INFO 04-13 12:32:17 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2108/3300 [5:11:26<3:00:44,  9.10s/it]2026-04-13 12:32:24,183 | INFO | train_grpo_train | metrics_logged mode=train step=2108

                                                       
{'loss': 0.0295, 'grad_norm': 11.737648963928223, 'learning_rate': 3.6151515151515153e-06, 'num_tokens': 3791689.0, 'completions/mean_length': 38.625, 'completions/min_length': 35.0, 'completions/max_length': 42.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.625, 'completions/min_terminated_length': 35.0, 'completions/max_terminated_length': 42.0, 'rewards/meter/mean': 0.46569669246673584, 'rewards/meter/std': 0.4034918248653412, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9199074506759644, 'rewards/repeat_soft/std': 0.04545706883072853, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.38604602217674255, 'rewards/total_composite/std': 0.042396180331707, 'reward': 0.38604602217674255, 'reward_std': 0.0423961766064167, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08647025376558304, 'sampling/sampling_logp_difference/max': 2.448047161102295, 'sampling/importance_sampling_ratio/min': 0.08646226674318314, 'sampling/importance_sampling_ratio/mean': 1.0085159540176392, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.41267629340291023, 'clip_ratio/low_mean': 0.054619363974779844, 'clip_ratio/low_min': 0.054619363974779844, 'clip_ratio/high_mean': 0.016964286100119352, 'clip_ratio/high_max': 0.016964286100119352, 'clip_ratio/region_mean': 0.0715836500748992, 'reward_total_mean': 0.38604602217674255, 'reward_meter_mean': 0.46569669246673584, 'reward_meter_std': 0.4034918248653412, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9199074506759644, 'reward_repeat_soft_std': 0.04545706883072853, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.38604602217674255, 'reward_total_composite_std': 0.042396180331707, 'epoch': 0.21}

 64%|██████▍   | 2108/3300 [5:11:26<3:00:44,  9.10s/it]INFO 04-13 12:32:24 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2109/3300 [5:11:32<2:46:23,  8.38s/it]2026-04-13 12:32:30,828 | INFO | train_grpo_train | metrics_logged mode=train step=2109

                                                       
{'loss': 0.1557, 'grad_norm': 13.237957000732422, 'learning_rate': 3.6121212121212125e-06, 'num_tokens': 3793245.0, 'completions/mean_length': 38.5, 'completions/min_length': 30.0, 'completions/max_length': 47.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 38.5, 'completions/min_terminated_length': 30.0, 'completions/max_terminated_length': 47.0, 'rewards/meter/mean': 0.33292731642723083, 'rewards/meter/std': 0.2673882842063904, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9947908520698547, 'rewards/repeat_soft/std': 0.007985309697687626, 'rewards/judge_quality/mean': 0.19500000774860382, 'rewards/judge_quality/std': 0.10392305254936218, 'rewards/total_composite/mean': 0.3842517137527466, 'rewards/total_composite/std': 0.025468753650784492, 'reward': 0.3842517137527466, 'reward_std': 0.02546876110136509, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11784003674983978, 'sampling/sampling_logp_difference/max': 2.1868762969970703, 'sampling/importance_sampling_ratio/min': 0.11226688325405121, 'sampling/importance_sampling_ratio/mean': 1.0069389343261719, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5103457905352116, 'clip_ratio/low_mean': 0.05624600313603878, 'clip_ratio/low_min': 0.05624600313603878, 'clip_ratio/high_mean': 0.03729429328814149, 'clip_ratio/high_max': 0.03729429328814149, 'clip_ratio/region_mean': 0.09354029642418027, 'reward_total_mean': 0.3842517137527466, 'reward_meter_mean': 0.33292731642723083, 'reward_meter_std': 0.2673882842063904, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9947908520698547, 'reward_repeat_soft_std': 0.007985309697687626, 'reward_judge_quality_mean': 0.19500000774860382, 'reward_judge_quality_std': 0.10392305254936218, 'reward_total_composite_mean': 0.3842517137527466, 'reward_total_composite_std': 0.025468753650784492, 'epoch': 0.21}

 64%|██████▍   | 2109/3300 [5:11:33<2:46:23,  8.38s/it]INFO 04-13 12:32:31 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2110/3300 [5:11:39<2:35:47,  7.86s/it]2026-04-13 12:32:37,449 | INFO | train_grpo_train | metrics_logged mode=train step=2110

                                                       
{'loss': 0.0113, 'grad_norm': 13.800068855285645, 'learning_rate': 3.6090909090909093e-06, 'num_tokens': 3794772.0, 'completions/mean_length': 41.875, 'completions/min_length': 36.0, 'completions/max_length': 49.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.875, 'completions/min_terminated_length': 36.0, 'completions/max_terminated_length': 49.0, 'rewards/meter/mean': 0.8712973594665527, 'rewards/meter/std': 0.18867942690849304, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.918583333492279, 'rewards/repeat_soft/std': 0.05381178855895996, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4227389991283417, 'rewards/total_composite/std': 0.022761140018701553, 'reward': 0.4227389991283417, 'reward_std': 0.022761141881346703, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11791300028562546, 'sampling/sampling_logp_difference/max': 1.574312686920166, 'sampling/importance_sampling_ratio/min': 0.20714989304542542, 'sampling/importance_sampling_ratio/mean': 1.0151242017745972, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5558187812566757, 'clip_ratio/low_mean': 0.02714580879546702, 'clip_ratio/low_min': 0.02714580879546702, 'clip_ratio/high_mean': 0.06695579458028078, 'clip_ratio/high_max': 0.06695579458028078, 'clip_ratio/region_mean': 0.0941016033757478, 'reward_total_mean': 0.4227389991283417, 'reward_meter_mean': 0.8712973594665527, 'reward_meter_std': 0.18867942690849304, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.918583333492279, 'reward_repeat_soft_std': 0.05381178855895996, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4227389991283417, 'reward_total_composite_std': 0.022761140018701553, 'epoch': 0.21}

 64%|██████▍   | 2110/3300 [5:11:39<2:35:47,  7.86s/it]INFO 04-13 12:32:37 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2111/3300 [5:11:46<2:32:16,  7.68s/it]2026-04-13 12:32:44,744 | INFO | train_grpo_train | metrics_logged mode=train step=2111

                                                       
{'loss': 0.0434, 'grad_norm': 9.104913711547852, 'learning_rate': 3.606060606060606e-06, 'num_tokens': 3796749.0, 'completions/mean_length': 89.125, 'completions/min_length': 70.0, 'completions/max_length': 109.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 89.125, 'completions/min_terminated_length': 70.0, 'completions/max_terminated_length': 109.0, 'rewards/meter/mean': 0.6411469578742981, 'rewards/meter/std': 0.302879273891449, 'rewards/count_adherence/mean': 0.8999999761581421, 'rewards/count_adherence/std': 0.10690449178218842, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.6518181562423706, 'rewards/repeat_soft/std': 0.11911094933748245, 'rewards/judge_quality/mean': 0.13875000178813934, 'rewards/judge_quality/std': 0.022320719435811043, 'rewards/total_composite/mean': 0.33592337369918823, 'rewards/total_composite/std': 0.020304610952734947, 'reward': 0.33592337369918823, 'reward_std': 0.020304612815380096, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07979433983564377, 'sampling/sampling_logp_difference/max': 1.9750046730041504, 'sampling/importance_sampling_ratio/min': 0.13876067101955414, 'sampling/importance_sampling_ratio/mean': 1.0048784017562866, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.40958814322948456, 'clip_ratio/low_mean': 0.0406727041117847, 'clip_ratio/low_min': 0.0406727041117847, 'clip_ratio/high_mean': 0.036213971907272935, 'clip_ratio/high_max': 0.036213971907272935, 'clip_ratio/region_mean': 0.07688667601905763, 'reward_total_mean': 0.33592337369918823, 'reward_meter_mean': 0.6411469578742981, 'reward_meter_std': 0.302879273891449, 'reward_count_adherence_mean': 0.8999999761581421, 'reward_count_adherence_std': 0.10690449178218842, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.6518181562423706, 'reward_repeat_soft_std': 0.11911094933748245, 'reward_judge_quality_mean': 0.13875000178813934, 'reward_judge_quality_std': 0.022320719435811043, 'reward_total_composite_mean': 0.33592337369918823, 'reward_total_composite_std': 0.020304610952734947, 'epoch': 0.21}

 64%|██████▍   | 2111/3300 [5:11:46<2:32:16,  7.68s/it]INFO 04-13 12:32:44 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2112/3300 [5:11:53<2:25:26,  7.35s/it]2026-04-13 12:32:51,300 | INFO | train_grpo_train | metrics_logged mode=train step=2112

                                                       
{'loss': -0.0532, 'grad_norm': 12.300381660461426, 'learning_rate': 3.603030303030303e-06, 'num_tokens': 3798369.0, 'completions/mean_length': 42.5, 'completions/min_length': 37.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.5, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.4879136383533478, 'rewards/meter/std': 0.4658832550048828, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8460896015167236, 'rewards/repeat_soft/std': 0.058859921991825104, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3744850158691406, 'rewards/total_composite/std': 0.045250579714775085, 'reward': 0.3744850158691406, 'reward_std': 0.04525057598948479, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11138882488012314, 'sampling/sampling_logp_difference/max': 1.1025335788726807, 'sampling/importance_sampling_ratio/min': 0.3320287764072418, 'sampling/importance_sampling_ratio/mean': 1.0110188722610474, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6444509625434875, 'clip_ratio/low_mean': 0.0528691983781755, 'clip_ratio/low_min': 0.0528691983781755, 'clip_ratio/high_mean': 0.0698704281821847, 'clip_ratio/high_max': 0.0698704281821847, 'clip_ratio/region_mean': 0.1227396265603602, 'reward_total_mean': 0.3744850158691406, 'reward_meter_mean': 0.4879136383533478, 'reward_meter_std': 0.4658832550048828, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8460896015167236, 'reward_repeat_soft_std': 0.058859921991825104, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3744850158691406, 'reward_total_composite_std': 0.045250579714775085, 'epoch': 0.21}

 64%|██████▍   | 2112/3300 [5:11:53<2:25:26,  7.35s/it]INFO 04-13 12:32:51 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2113/3300 [5:11:59<2:17:44,  6.96s/it]2026-04-13 12:32:57,376 | INFO | train_grpo_train | metrics_logged mode=train step=2113

                                                       
{'loss': 0.0482, 'grad_norm': 16.208444595336914, 'learning_rate': 3.6000000000000003e-06, 'num_tokens': 3799785.0, 'completions/mean_length': 22.0, 'completions/min_length': 16.0, 'completions/max_length': 32.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.0, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 32.0, 'rewards/meter/mean': 0.6580546498298645, 'rewards/meter/std': 0.41127723455429077, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9534668326377869, 'rewards/repeat_soft/std': 0.014675324782729149, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4071803689002991, 'rewards/total_composite/std': 0.038945745676755905, 'reward': 0.4071803689002991, 'reward_std': 0.03894574195146561, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13734646141529083, 'sampling/sampling_logp_difference/max': 1.8436002731323242, 'sampling/importance_sampling_ratio/min': 0.15824668109416962, 'sampling/importance_sampling_ratio/mean': 0.9922757744789124, 'sampling/importance_sampling_ratio/max': 1.6984217166900635, 'entropy': 0.7180051393806934, 'clip_ratio/low_mean': 0.047727273777127266, 'clip_ratio/low_min': 0.047727273777127266, 'clip_ratio/high_mean': 0.09421969391405582, 'clip_ratio/high_max': 0.09421969391405582, 'clip_ratio/region_mean': 0.1419469676911831, 'reward_total_mean': 0.4071803689002991, 'reward_meter_mean': 0.6580546498298645, 'reward_meter_std': 0.41127723455429077, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9534668326377869, 'reward_repeat_soft_std': 0.014675324782729149, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4071803689002991, 'reward_total_composite_std': 0.038945745676755905, 'epoch': 0.21}

 64%|██████▍   | 2113/3300 [5:11:59<2:17:44,  6.96s/it]INFO 04-13 12:32:57 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2114/3300 [5:12:10<2:43:27,  8.27s/it]2026-04-13 12:33:08,685 | INFO | train_grpo_train | metrics_logged mode=train step=2114

                                                       
{'loss': -0.1491, 'grad_norm': 1.8867636919021606, 'learning_rate': 3.596969696969697e-06, 'num_tokens': 3801567.0, 'completions/mean_length': 115.75, 'completions/min_length': 49.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 59.142860412597656, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 70.0, 'rewards/meter/mean': 0.5922025442123413, 'rewards/meter/std': 0.3228558599948883, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.971898078918457, 'rewards/repeat_soft/std': 0.02474227547645569, 'rewards/judge_quality/mean': 0.1525000035762787, 'rewards/judge_quality/std': 0.04399675503373146, 'rewards/total_composite/mean': 0.36274904012680054, 'rewards/total_composite/std': 0.1502305567264557, 'reward': 0.36274904012680054, 'reward_std': 0.1502305418252945, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11633001267910004, 'sampling/sampling_logp_difference/max': 3.0924201011657715, 'sampling/importance_sampling_ratio/min': 0.04539196565747261, 'sampling/importance_sampling_ratio/mean': 0.9857916235923767, 'sampling/importance_sampling_ratio/max': 1.9030864238739014, 'entropy': 0.47304895520210266, 'clip_ratio/low_mean': 0.01024590153247118, 'clip_ratio/low_min': 0.01024590153247118, 'clip_ratio/high_mean': 0.08157390914857388, 'clip_ratio/high_max': 0.08157390914857388, 'clip_ratio/region_mean': 0.09181981068104506, 'reward_total_mean': 0.36274904012680054, 'reward_meter_mean': 0.5922025442123413, 'reward_meter_std': 0.3228558599948883, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.971898078918457, 'reward_repeat_soft_std': 0.02474227547645569, 'reward_judge_quality_mean': 0.1525000035762787, 'reward_judge_quality_std': 0.04399675503373146, 'reward_total_composite_mean': 0.36274904012680054, 'reward_total_composite_std': 0.1502305567264557, 'epoch': 0.21}

 64%|██████▍   | 2114/3300 [5:12:10<2:43:27,  8.27s/it]INFO 04-13 12:33:08 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2115/3300 [5:12:17<2:31:13,  7.66s/it]2026-04-13 12:33:14,912 | INFO | train_grpo_train | metrics_logged mode=train step=2115

                                                       
{'loss': 0.0127, 'grad_norm': 11.018208503723145, 'learning_rate': 3.593939393939394e-06, 'num_tokens': 3803102.0, 'completions/mean_length': 43.875, 'completions/min_length': 38.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 43.875, 'completions/min_terminated_length': 38.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8543543219566345, 'rewards/meter/std': 0.28948941826820374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9505391120910645, 'rewards/repeat_soft/std': 0.01934647373855114, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4282057285308838, 'rewards/total_composite/std': 0.03167116641998291, 'reward': 0.4282057285308838, 'reward_std': 0.03167115896940231, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10174911469221115, 'sampling/sampling_logp_difference/max': 1.6101007461547852, 'sampling/importance_sampling_ratio/min': 0.19986747205257416, 'sampling/importance_sampling_ratio/mean': 0.9898253679275513, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4199068732559681, 'clip_ratio/low_mean': 0.011627906933426857, 'clip_ratio/low_min': 0.011627906933426857, 'clip_ratio/high_mean': 0.1053868168964982, 'clip_ratio/high_max': 0.1053868168964982, 'clip_ratio/region_mean': 0.11701472382992506, 'reward_total_mean': 0.4282057285308838, 'reward_meter_mean': 0.8543543219566345, 'reward_meter_std': 0.28948941826820374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9505391120910645, 'reward_repeat_soft_std': 0.01934647373855114, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4282057285308838, 'reward_total_composite_std': 0.03167116641998291, 'epoch': 0.21}

 64%|██████▍   | 2115/3300 [5:12:17<2:31:13,  7.66s/it]INFO 04-13 12:33:15 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2116/3300 [5:12:23<2:25:20,  7.37s/it]2026-04-13 12:33:21,596 | INFO | train_grpo_train | metrics_logged mode=train step=2116

                                                       
{'loss': 0.0131, 'grad_norm': 9.96137809753418, 'learning_rate': 3.590909090909091e-06, 'num_tokens': 3804769.0, 'completions/mean_length': 51.375, 'completions/min_length': 45.0, 'completions/max_length': 56.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.375, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 56.0, 'rewards/meter/mean': 0.610297441482544, 'rewards/meter/std': 0.37020719051361084, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9516139030456543, 'rewards/repeat_soft/std': 0.03949163854122162, 'rewards/judge_quality/mean': 0.25, 'rewards/judge_quality/std': 0.2709243595600128, 'rewards/total_composite/mean': 0.46495306491851807, 'rewards/total_composite/std': 0.1915210336446762, 'reward': 0.46495306491851807, 'reward_std': 0.1915210336446762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0902131050825119, 'sampling/sampling_logp_difference/max': 1.5951628684997559, 'sampling/importance_sampling_ratio/min': 0.20287548005580902, 'sampling/importance_sampling_ratio/mean': 1.0204806327819824, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5200207754969597, 'clip_ratio/low_mean': 0.08550144638866186, 'clip_ratio/low_min': 0.08550144638866186, 'clip_ratio/high_mean': 0.009433962404727936, 'clip_ratio/high_max': 0.009433962404727936, 'clip_ratio/region_mean': 0.0949354087933898, 'reward_total_mean': 0.46495306491851807, 'reward_meter_mean': 0.610297441482544, 'reward_meter_std': 0.37020719051361084, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9516139030456543, 'reward_repeat_soft_std': 0.03949163854122162, 'reward_judge_quality_mean': 0.25, 'reward_judge_quality_std': 0.2709243595600128, 'reward_total_composite_mean': 0.46495306491851807, 'reward_total_composite_std': 0.1915210336446762, 'epoch': 0.21}

 64%|██████▍   | 2116/3300 [5:12:23<2:25:20,  7.37s/it]INFO 04-13 12:33:21 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2117/3300 [5:12:30<2:20:10,  7.11s/it]2026-04-13 12:33:28,114 | INFO | train_grpo_train | metrics_logged mode=train step=2117

                                                       
{'loss': -0.0345, 'grad_norm': 10.987751960754395, 'learning_rate': 3.587878787878788e-06, 'num_tokens': 3806689.0, 'completions/mean_length': 56.0, 'completions/min_length': 50.0, 'completions/max_length': 68.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 56.0, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 68.0, 'rewards/meter/mean': 0.48459476232528687, 'rewards/meter/std': 0.3567312955856323, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9572969079017639, 'rewards/repeat_soft/std': 0.03782852366566658, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.388683557510376, 'rewards/total_composite/std': 0.03286159411072731, 'reward': 0.388683557510376, 'reward_std': 0.03286158666014671, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11895419657230377, 'sampling/sampling_logp_difference/max': 3.274886131286621, 'sampling/importance_sampling_ratio/min': 0.037821173667907715, 'sampling/importance_sampling_ratio/mean': 1.0060267448425293, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6553157195448875, 'clip_ratio/low_mean': 0.05163145810365677, 'clip_ratio/low_min': 0.05163145810365677, 'clip_ratio/high_mean': 0.04374784370884299, 'clip_ratio/high_max': 0.04374784370884299, 'clip_ratio/region_mean': 0.09537930181249976, 'reward_total_mean': 0.388683557510376, 'reward_meter_mean': 0.48459476232528687, 'reward_meter_std': 0.3567312955856323, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9572969079017639, 'reward_repeat_soft_std': 0.03782852366566658, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.388683557510376, 'reward_total_composite_std': 0.03286159411072731, 'epoch': 0.21}

 64%|██████▍   | 2117/3300 [5:12:30<2:20:10,  7.11s/it]INFO 04-13 12:33:28 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2118/3300 [5:12:42<2:49:26,  8.60s/it]2026-04-13 12:33:40,197 | INFO | train_grpo_train | metrics_logged mode=train step=2118

                                                       
{'loss': -0.1357, 'grad_norm': 1.8776116371154785, 'learning_rate': 3.584848484848485e-06, 'num_tokens': 3808480.0, 'completions/mean_length': 106.875, 'completions/min_length': 39.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 49.000003814697266, 'completions/min_terminated_length': 39.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.9100183248519897, 'rewards/meter/std': 0.22264239192008972, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9692996740341187, 'rewards/repeat_soft/std': 0.016032908111810684, 'rewards/judge_quality/mean': 0.13375000655651093, 'rewards/judge_quality/std': 0.03543102368712425, 'rewards/total_composite/mean': 0.384857714176178, 'rewards/total_composite/std': 0.15563072264194489, 'reward': 0.384857714176178, 'reward_std': 0.15563073754310608, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13287843763828278, 'sampling/sampling_logp_difference/max': 4.305706977844238, 'sampling/importance_sampling_ratio/min': 0.013491343706846237, 'sampling/importance_sampling_ratio/mean': 0.9984962344169617, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5105771720409393, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.11489598639309406, 'clip_ratio/high_max': 0.11489598639309406, 'clip_ratio/region_mean': 0.11489598639309406, 'reward_total_mean': 0.384857714176178, 'reward_meter_mean': 0.9100183248519897, 'reward_meter_std': 0.22264239192008972, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9692996740341187, 'reward_repeat_soft_std': 0.016032908111810684, 'reward_judge_quality_mean': 0.13375000655651093, 'reward_judge_quality_std': 0.03543102368712425, 'reward_total_composite_mean': 0.384857714176178, 'reward_total_composite_std': 0.15563072264194489, 'epoch': 0.21}

 64%|██████▍   | 2118/3300 [5:12:42<2:49:26,  8.60s/it]INFO 04-13 12:33:40 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2119/3300 [5:12:49<2:42:36,  8.26s/it]2026-04-13 12:33:47,657 | INFO | train_grpo_train | metrics_logged mode=train step=2119

                                                       
{'loss': 0.0006, 'grad_norm': 5.920814514160156, 'learning_rate': 3.5818181818181817e-06, 'num_tokens': 3810570.0, 'completions/mean_length': 85.25, 'completions/min_length': 80.0, 'completions/max_length': 92.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 85.25, 'completions/min_terminated_length': 80.0, 'completions/max_terminated_length': 92.0, 'rewards/meter/mean': 0.9678657054901123, 'rewards/meter/std': 0.02926114946603775, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8394684195518494, 'rewards/repeat_soft/std': 0.03335115313529968, 'rewards/judge_quality/mean': 0.18000000715255737, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.43916055560112, 'rewards/total_composite/std': 0.004552391357719898, 'reward': 0.43916055560112, 'reward_std': 0.0045523978769779205, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07062681764364243, 'sampling/sampling_logp_difference/max': 1.2569303512573242, 'sampling/importance_sampling_ratio/min': 0.2845260798931122, 'sampling/importance_sampling_ratio/mean': 1.002211093902588, 'sampling/importance_sampling_ratio/max': 1.83241605758667, 'entropy': 0.3776538223028183, 'clip_ratio/low_mean': 0.03219923097640276, 'clip_ratio/low_min': 0.03219923097640276, 'clip_ratio/high_mean': 0.04883924592286348, 'clip_ratio/high_max': 0.04883924592286348, 'clip_ratio/region_mean': 0.08103847689926624, 'reward_total_mean': 0.43916055560112, 'reward_meter_mean': 0.9678657054901123, 'reward_meter_std': 0.02926114946603775, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8394684195518494, 'reward_repeat_soft_std': 0.03335115313529968, 'reward_judge_quality_mean': 0.18000000715255737, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.43916055560112, 'reward_total_composite_std': 0.004552391357719898, 'epoch': 0.21}

 64%|██████▍   | 2119/3300 [5:12:49<2:42:36,  8.26s/it]INFO 04-13 12:33:47 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2120/3300 [5:12:56<2:32:44,  7.77s/it]2026-04-13 12:33:54,271 | INFO | train_grpo_train | metrics_logged mode=train step=2120

                                                       
{'loss': 0.0594, 'grad_norm': 18.88593101501465, 'learning_rate': 3.578787878787879e-06, 'num_tokens': 3812212.0, 'completions/mean_length': 37.25, 'completions/min_length': 33.0, 'completions/max_length': 40.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 37.25, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 40.0, 'rewards/meter/mean': 0.8623172044754028, 'rewards/meter/std': 0.2601478397846222, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9530335664749146, 'rewards/repeat_soft/std': 0.05007192865014076, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4270309805870056, 'rewards/total_composite/std': 0.02340822108089924, 'reward': 0.4270309805870056, 'reward_std': 0.023408222943544388, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09455607831478119, 'sampling/sampling_logp_difference/max': 2.033214807510376, 'sampling/importance_sampling_ratio/min': 0.13091398775577545, 'sampling/importance_sampling_ratio/mean': 1.0077043771743774, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.375054020434618, 'clip_ratio/low_mean': 0.019736841320991516, 'clip_ratio/low_min': 0.019736841320991516, 'clip_ratio/high_mean': 0.0718057572375983, 'clip_ratio/high_max': 0.0718057572375983, 'clip_ratio/region_mean': 0.09154259855858982, 'reward_total_mean': 0.4270309805870056, 'reward_meter_mean': 0.8623172044754028, 'reward_meter_std': 0.2601478397846222, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9530335664749146, 'reward_repeat_soft_std': 0.05007192865014076, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4270309805870056, 'reward_total_composite_std': 0.02340822108089924, 'epoch': 0.21}

 64%|██████▍   | 2120/3300 [5:12:56<2:32:44,  7.77s/it]INFO 04-13 12:33:54 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2121/3300 [5:13:02<2:24:20,  7.35s/it]2026-04-13 12:34:00,632 | INFO | train_grpo_train | metrics_logged mode=train step=2121

                                                       
{'loss': -0.1005, 'grad_norm': 16.1114444732666, 'learning_rate': 3.575757575757576e-06, 'num_tokens': 3813721.0, 'completions/mean_length': 27.625, 'completions/min_length': 24.0, 'completions/max_length': 34.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.625, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 34.0, 'rewards/meter/mean': 0.34160304069519043, 'rewards/meter/std': 0.3914759159088135, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9752987623214722, 'rewards/repeat_soft/std': 0.004002916160970926, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.3796011209487915, 'rewards/total_composite/std': 0.03823774307966232, 'reward': 0.3796011209487915, 'reward_std': 0.03823775053024292, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08197850733995438, 'sampling/sampling_logp_difference/max': 2.0052106380462646, 'sampling/importance_sampling_ratio/min': 0.1346319317817688, 'sampling/importance_sampling_ratio/mean': 0.9941287636756897, 'sampling/importance_sampling_ratio/max': 1.8110864162445068, 'entropy': 0.3139124680310488, 'clip_ratio/low_mean': 0.03363671526312828, 'clip_ratio/low_min': 0.03363671526312828, 'clip_ratio/high_mean': 0.02470588218420744, 'clip_ratio/high_max': 0.02470588218420744, 'clip_ratio/region_mean': 0.05834259744733572, 'reward_total_mean': 0.3796011209487915, 'reward_meter_mean': 0.34160304069519043, 'reward_meter_std': 0.3914759159088135, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9752987623214722, 'reward_repeat_soft_std': 0.004002916160970926, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.3796011209487915, 'reward_total_composite_std': 0.03823774307966232, 'epoch': 0.21}

 64%|██████▍   | 2121/3300 [5:13:02<2:24:20,  7.35s/it]INFO 04-13 12:34:00 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2122/3300 [5:13:09<2:21:14,  7.19s/it]2026-04-13 12:34:07,476 | INFO | train_grpo_train | metrics_logged mode=train step=2122

                                                       
{'loss': -0.0917, 'grad_norm': 20.8836727142334, 'learning_rate': 3.5727272727272734e-06, 'num_tokens': 3815112.0, 'completions/mean_length': 22.875, 'completions/min_length': 16.0, 'completions/max_length': 38.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 22.875, 'completions/min_terminated_length': 16.0, 'completions/max_terminated_length': 38.0, 'rewards/meter/mean': 0.6766870021820068, 'rewards/meter/std': 0.4345723092556, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9576764702796936, 'rewards/repeat_soft/std': 0.013642913661897182, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4096284508705139, 'rewards/total_composite/std': 0.04190043732523918, 'reward': 0.4096284508705139, 'reward_std': 0.04190043359994888, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.16140659153461456, 'sampling/sampling_logp_difference/max': 1.4260950088500977, 'sampling/importance_sampling_ratio/min': 0.2402452528476715, 'sampling/importance_sampling_ratio/mean': 0.9993340969085693, 'sampling/importance_sampling_ratio/max': 1.772874116897583, 'entropy': 0.9253654107451439, 'clip_ratio/low_mean': 0.06414473708719015, 'clip_ratio/low_min': 0.06414473708719015, 'clip_ratio/high_mean': 0.12481671385467052, 'clip_ratio/high_max': 0.12481671385467052, 'clip_ratio/region_mean': 0.18896145094186068, 'reward_total_mean': 0.4096284508705139, 'reward_meter_mean': 0.6766870021820068, 'reward_meter_std': 0.4345723092556, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9576764702796936, 'reward_repeat_soft_std': 0.013642913661897182, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4096284508705139, 'reward_total_composite_std': 0.04190043732523918, 'epoch': 0.21}

 64%|██████▍   | 2122/3300 [5:13:09<2:21:14,  7.19s/it]INFO 04-13 12:34:07 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2123/3300 [5:13:16<2:18:21,  7.05s/it]2026-04-13 12:34:14,202 | INFO | train_grpo_train | metrics_logged mode=train step=2123

                                                       
{'loss': 0.0178, 'grad_norm': 12.174962043762207, 'learning_rate': 3.5696969696969703e-06, 'num_tokens': 3816764.0, 'completions/mean_length': 47.5, 'completions/min_length': 40.0, 'completions/max_length': 60.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 47.5, 'completions/min_terminated_length': 40.0, 'completions/max_terminated_length': 60.0, 'rewards/meter/mean': 0.8043650984764099, 'rewards/meter/std': 0.3080444037914276, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9377086758613586, 'rewards/repeat_soft/std': 0.02444099448621273, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4190818667411804, 'rewards/total_composite/std': 0.029328929260373116, 'reward': 0.4190818667411804, 'reward_std': 0.02932891622185707, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08279678970575333, 'sampling/sampling_logp_difference/max': 1.8162415027618408, 'sampling/importance_sampling_ratio/min': 0.16263586282730103, 'sampling/importance_sampling_ratio/mean': 0.9955849647521973, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4308505542576313, 'clip_ratio/low_mean': 0.0252380957826972, 'clip_ratio/low_min': 0.0252380957826972, 'clip_ratio/high_mean': 0.046594975516200066, 'clip_ratio/high_max': 0.046594975516200066, 'clip_ratio/region_mean': 0.07183307129889727, 'reward_total_mean': 0.4190818667411804, 'reward_meter_mean': 0.8043650984764099, 'reward_meter_std': 0.3080444037914276, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9377086758613586, 'reward_repeat_soft_std': 0.02444099448621273, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4190818667411804, 'reward_total_composite_std': 0.029328929260373116, 'epoch': 0.21}

 64%|██████▍   | 2123/3300 [5:13:16<2:18:21,  7.05s/it]INFO 04-13 12:34:14 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2124/3300 [5:13:23<2:20:03,  7.15s/it]2026-04-13 12:34:21,655 | INFO | train_grpo_train | metrics_logged mode=train step=2124

                                                       
{'loss': 0.0328, 'grad_norm': 9.952363014221191, 'learning_rate': 3.566666666666667e-06, 'num_tokens': 3818445.0, 'completions/mean_length': 51.125, 'completions/min_length': 45.0, 'completions/max_length': 54.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 51.125, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 54.0, 'rewards/meter/mean': 0.9706482887268066, 'rewards/meter/std': 0.039042141288518906, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.877032995223999, 'rewards/repeat_soft/std': 0.07919418066740036, 'rewards/judge_quality/mean': 0.14249999821186066, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.4213140606880188, 'rewards/total_composite/std': 0.02052641287446022, 'reward': 0.4213140606880188, 'reward_std': 0.020526403561234474, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07824968546628952, 'sampling/sampling_logp_difference/max': 1.197680950164795, 'sampling/importance_sampling_ratio/min': 0.3058168292045593, 'sampling/importance_sampling_ratio/mean': 0.9943525791168213, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.2968391440808773, 'clip_ratio/low_mean': 0.02626256551593542, 'clip_ratio/low_min': 0.02626256551593542, 'clip_ratio/high_mean': 0.030169055331498384, 'clip_ratio/high_max': 0.030169055331498384, 'clip_ratio/region_mean': 0.056431620847433805, 'reward_total_mean': 0.4213140606880188, 'reward_meter_mean': 0.9706482887268066, 'reward_meter_std': 0.039042141288518906, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.877032995223999, 'reward_repeat_soft_std': 0.07919418066740036, 'reward_judge_quality_mean': 0.14249999821186066, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.4213140606880188, 'reward_total_composite_std': 0.02052641287446022, 'epoch': 0.21}

 64%|██████▍   | 2124/3300 [5:13:23<2:20:03,  7.15s/it]INFO 04-13 12:34:21 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2125/3300 [5:13:31<2:22:27,  7.27s/it]2026-04-13 12:34:29,127 | INFO | train_grpo_train | metrics_logged mode=train step=2125

                                                       
{'loss': -0.0465, 'grad_norm': 7.96971321105957, 'learning_rate': 3.563636363636364e-06, 'num_tokens': 3820811.0, 'completions/mean_length': 115.75, 'completions/min_length': 96.0, 'completions/max_length': 128.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 115.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 128.0, 'rewards/meter/mean': 0.8100929260253906, 'rewards/meter/std': 0.3253575265407562, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8903893232345581, 'rewards/repeat_soft/std': 0.017612675204873085, 'rewards/judge_quality/mean': 0.1237500011920929, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.3969489634037018, 'rewards/total_composite/std': 0.023812510073184967, 'reward': 0.3969489634037018, 'reward_std': 0.023812511935830116, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09560476988554001, 'sampling/sampling_logp_difference/max': 1.885268211364746, 'sampling/importance_sampling_ratio/min': 0.1517883539199829, 'sampling/importance_sampling_ratio/mean': 1.0107030868530273, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.46189267933368683, 'clip_ratio/low_mean': 0.022395833395421505, 'clip_ratio/low_min': 0.022395833395421505, 'clip_ratio/high_mean': 0.07545540388673544, 'clip_ratio/high_max': 0.07545540388673544, 'clip_ratio/region_mean': 0.09785123728215694, 'reward_total_mean': 0.3969489634037018, 'reward_meter_mean': 0.8100929260253906, 'reward_meter_std': 0.3253575265407562, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8903893232345581, 'reward_repeat_soft_std': 0.017612675204873085, 'reward_judge_quality_mean': 0.1237500011920929, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.3969489634037018, 'reward_total_composite_std': 0.023812510073184967, 'epoch': 0.21}

 64%|██████▍   | 2125/3300 [5:13:31<2:22:27,  7.27s/it]INFO 04-13 12:34:29 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2126/3300 [5:13:43<2:49:12,  8.65s/it]2026-04-13 12:34:40,987 | INFO | train_grpo_train | metrics_logged mode=train step=2126

                                                       
{'loss': -0.1275, 'grad_norm': 2.2829623222351074, 'learning_rate': 3.560606060606061e-06, 'num_tokens': 3822445.0, 'completions/mean_length': 110.25, 'completions/min_length': 48.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 52.85714340209961, 'completions/min_terminated_length': 48.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.7273818254470825, 'rewards/meter/std': 0.42327576875686646, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9806386232376099, 'rewards/repeat_soft/std': 0.020722202956676483, 'rewards/judge_quality/mean': 0.17875000834465027, 'rewards/judge_quality/std': 0.11605878919363022, 'rewards/total_composite/mean': 0.39845943450927734, 'rewards/total_composite/std': 0.17626473307609558, 'reward': 0.39845943450927734, 'reward_std': 0.1762647181749344, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10056637227535248, 'sampling/sampling_logp_difference/max': 3.5167078971862793, 'sampling/importance_sampling_ratio/min': 0.029697038233280182, 'sampling/importance_sampling_ratio/mean': 1.0091761350631714, 'sampling/importance_sampling_ratio/max': 1.9448227882385254, 'entropy': 0.5782650634646416, 'clip_ratio/low_mean': 0.012500000186264515, 'clip_ratio/low_min': 0.012500000186264515, 'clip_ratio/high_mean': 0.07084880443289876, 'clip_ratio/high_max': 0.07084880443289876, 'clip_ratio/region_mean': 0.08334880461916327, 'reward_total_mean': 0.39845943450927734, 'reward_meter_mean': 0.7273818254470825, 'reward_meter_std': 0.42327576875686646, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9806386232376099, 'reward_repeat_soft_std': 0.020722202956676483, 'reward_judge_quality_mean': 0.17875000834465027, 'reward_judge_quality_std': 0.11605878919363022, 'reward_total_composite_mean': 0.39845943450927734, 'reward_total_composite_std': 0.17626473307609558, 'epoch': 0.21}

 64%|██████▍   | 2126/3300 [5:13:43<2:49:12,  8.65s/it]INFO 04-13 12:34:41 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2127/3300 [5:13:49<2:37:20,  8.05s/it]2026-04-13 12:34:47,634 | INFO | train_grpo_train | metrics_logged mode=train step=2127

                                                       
{'loss': 0.0744, 'grad_norm': 10.888141632080078, 'learning_rate': 3.557575757575758e-06, 'num_tokens': 3824123.0, 'completions/mean_length': 41.75, 'completions/min_length': 33.0, 'completions/max_length': 52.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 41.75, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 52.0, 'rewards/meter/mean': 0.961769700050354, 'rewards/meter/std': 0.04579383134841919, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9323242902755737, 'rewards/repeat_soft/std': 0.029701024293899536, 'rewards/judge_quality/mean': 0.22500000894069672, 'rewards/judge_quality/std': 0.13887301087379456, 'rewards/total_composite/mean': 0.48145878314971924, 'rewards/total_composite/std': 0.08942937850952148, 'reward': 0.48145878314971924, 'reward_std': 0.08942936360836029, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08970857411623001, 'sampling/sampling_logp_difference/max': 1.39491605758667, 'sampling/importance_sampling_ratio/min': 0.2478538304567337, 'sampling/importance_sampling_ratio/mean': 0.9984186887741089, 'sampling/importance_sampling_ratio/max': 1.674881935119629, 'entropy': 0.5424296446144581, 'clip_ratio/low_mean': 0.03404927044175565, 'clip_ratio/low_min': 0.03404927044175565, 'clip_ratio/high_mean': 0.029656318947672844, 'clip_ratio/high_max': 0.029656318947672844, 'clip_ratio/region_mean': 0.0637055893894285, 'reward_total_mean': 0.48145878314971924, 'reward_meter_mean': 0.961769700050354, 'reward_meter_std': 0.04579383134841919, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9323242902755737, 'reward_repeat_soft_std': 0.029701024293899536, 'reward_judge_quality_mean': 0.22500000894069672, 'reward_judge_quality_std': 0.13887301087379456, 'reward_total_composite_mean': 0.48145878314971924, 'reward_total_composite_std': 0.08942937850952148, 'epoch': 0.21}

 64%|██████▍   | 2127/3300 [5:13:49<2:37:20,  8.05s/it]INFO 04-13 12:34:47 [block_pool.py:378] Successfully reset prefix cache

 64%|██████▍   | 2128/3300 [5:14:03<3:13:15,  9.89s/it]2026-04-13 12:35:01,839 | INFO | train_grpo_train | metrics_logged mode=train step=2128

                                                       
{'loss': -0.1926, 'grad_norm': 1.7553354501724243, 'learning_rate': 3.554545454545455e-06, 'num_tokens': 3826618.0, 'completions/mean_length': 161.875, 'completions/min_length': 103.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 111.85714721679688, 'completions/min_terminated_length': 103.0, 'completions/max_terminated_length': 130.0, 'rewards/meter/mean': 0.8127691745758057, 'rewards/meter/std': 0.2406369149684906, 'rewards/count_adherence/mean': 0.96875, 'rewards/count_adherence/std': 0.0883883461356163, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8884314298629761, 'rewards/repeat_soft/std': 0.041326019912958145, 'rewards/judge_quality/mean': 0.14875000715255737, 'rewards/judge_quality/std': 0.048236772418022156, 'rewards/total_composite/mean': 0.3597673177719116, 'rewards/total_composite/std': 0.15010634064674377, 'reward': 0.3597673177719116, 'reward_std': 0.15010634064674377, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09040025621652603, 'sampling/sampling_logp_difference/max': 1.4311444759368896, 'sampling/importance_sampling_ratio/min': 0.2390352040529251, 'sampling/importance_sampling_ratio/mean': 1.004586100578308, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4321543797850609, 'clip_ratio/low_mean': 0.014423076994717121, 'clip_ratio/low_min': 0.014423076994717121, 'clip_ratio/high_mean': 0.07017199601978064, 'clip_ratio/high_max': 0.07017199601978064, 'clip_ratio/region_mean': 0.08459507301449776, 'reward_total_mean': 0.3597673177719116, 'reward_meter_mean': 0.8127691745758057, 'reward_meter_std': 0.2406369149684906, 'reward_count_adherence_mean': 0.96875, 'reward_count_adherence_std': 0.0883883461356163, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8884314298629761, 'reward_repeat_soft_std': 0.041326019912958145, 'reward_judge_quality_mean': 0.14875000715255737, 'reward_judge_quality_std': 0.048236772418022156, 'reward_total_composite_mean': 0.3597673177719116, 'reward_total_composite_std': 0.15010634064674377, 'epoch': 0.21}

 64%|██████▍   | 2128/3300 [5:14:04<3:13:15,  9.89s/it]INFO 04-13 12:35:02 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2129/3300 [5:14:10<2:56:11,  9.03s/it]2026-04-13 12:35:08,841 | INFO | train_grpo_train | metrics_logged mode=train step=2129

                                                       
{'loss': 0.065, 'grad_norm': 6.841037273406982, 'learning_rate': 3.551515151515152e-06, 'num_tokens': 3828796.0, 'completions/mean_length': 98.25, 'completions/min_length': 88.0, 'completions/max_length': 114.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 98.25, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 114.0, 'rewards/meter/mean': 0.9749183654785156, 'rewards/meter/std': 0.01414470374584198, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8836276531219482, 'rewards/repeat_soft/std': 0.0258223544806242, 'rewards/judge_quality/mean': 0.17625001072883606, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.44421523809432983, 'rewards/total_composite/std': 0.0071704210713505745, 'reward': 0.44421523809432983, 'reward_std': 0.007170422002673149, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09535162150859833, 'sampling/sampling_logp_difference/max': 3.7799437046051025, 'sampling/importance_sampling_ratio/min': 0.02282397635281086, 'sampling/importance_sampling_ratio/mean': 0.9967421293258667, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3356841690838337, 'clip_ratio/low_mean': 0.02442577015608549, 'clip_ratio/low_min': 0.02442577015608549, 'clip_ratio/high_mean': 0.04454277083277702, 'clip_ratio/high_max': 0.04454277083277702, 'clip_ratio/region_mean': 0.06896854098886251, 'reward_total_mean': 0.44421523809432983, 'reward_meter_mean': 0.9749183654785156, 'reward_meter_std': 0.01414470374584198, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8836276531219482, 'reward_repeat_soft_std': 0.0258223544806242, 'reward_judge_quality_mean': 0.17625001072883606, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.44421523809432983, 'reward_total_composite_std': 0.0071704210713505745, 'epoch': 0.21}

 65%|██████▍   | 2129/3300 [5:14:11<2:56:11,  9.03s/it]INFO 04-13 12:35:09 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2130/3300 [5:14:17<2:44:19,  8.43s/it]2026-04-13 12:35:15,869 | INFO | train_grpo_train | metrics_logged mode=train step=2130

                                                       
{'loss': -0.0544, 'grad_norm': 7.538710594177246, 'learning_rate': 3.548484848484849e-06, 'num_tokens': 3831009.0, 'completions/mean_length': 102.625, 'completions/min_length': 88.0, 'completions/max_length': 115.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 102.625, 'completions/min_terminated_length': 88.0, 'completions/max_terminated_length': 115.0, 'rewards/meter/mean': 0.9733519554138184, 'rewards/meter/std': 0.02541358955204487, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.740585446357727, 'rewards/repeat_soft/std': 0.12306283414363861, 'rewards/judge_quality/mean': 0.1274999976158142, 'rewards/judge_quality/std': 0.02121320739388466, 'rewards/total_composite/mean': 0.39185476303100586, 'rewards/total_composite/std': 0.0253920815885067, 'reward': 0.39185476303100586, 'reward_std': 0.02539207972586155, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.07328380644321442, 'sampling/sampling_logp_difference/max': 1.7725296020507812, 'sampling/importance_sampling_ratio/min': 0.16990265250205994, 'sampling/importance_sampling_ratio/mean': 0.9985268712043762, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3542307838797569, 'clip_ratio/low_mean': 0.02232967270538211, 'clip_ratio/low_min': 0.02232967270538211, 'clip_ratio/high_mean': 0.040749015752226114, 'clip_ratio/high_max': 0.040749015752226114, 'clip_ratio/region_mean': 0.06307868845760822, 'reward_total_mean': 0.39185476303100586, 'reward_meter_mean': 0.9733519554138184, 'reward_meter_std': 0.02541358955204487, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.740585446357727, 'reward_repeat_soft_std': 0.12306283414363861, 'reward_judge_quality_mean': 0.1274999976158142, 'reward_judge_quality_std': 0.02121320739388466, 'reward_total_composite_mean': 0.39185476303100586, 'reward_total_composite_std': 0.0253920815885067, 'epoch': 0.21}

 65%|██████▍   | 2130/3300 [5:14:18<2:44:19,  8.43s/it]INFO 04-13 12:35:16 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2131/3300 [5:14:25<2:39:26,  8.18s/it]2026-04-13 12:35:23,479 | INFO | train_grpo_train | metrics_logged mode=train step=2131

                                                       
{'loss': -0.0066, 'grad_norm': 6.729617595672607, 'learning_rate': 3.5454545454545458e-06, 'num_tokens': 3833642.0, 'completions/mean_length': 143.125, 'completions/min_length': 133.0, 'completions/max_length': 153.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 143.125, 'completions/min_terminated_length': 133.0, 'completions/max_terminated_length': 153.0, 'rewards/meter/mean': 0.9710224270820618, 'rewards/meter/std': 0.024256430566310883, 'rewards/count_adherence/mean': 0.949999988079071, 'rewards/count_adherence/std': 0.09258200973272324, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8977535963058472, 'rewards/repeat_soft/std': 0.03404621779918671, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.029730942100286484, 'rewards/total_composite/mean': 0.42175352573394775, 'rewards/total_composite/std': 0.02008803002536297, 'reward': 0.42175352573394775, 'reward_std': 0.020088035613298416, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11049627512693405, 'sampling/sampling_logp_difference/max': 2.2645256519317627, 'sampling/importance_sampling_ratio/min': 0.10387929528951645, 'sampling/importance_sampling_ratio/mean': 0.9980050325393677, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5583962202072144, 'clip_ratio/low_mean': 0.05468147713690996, 'clip_ratio/low_min': 0.05468147713690996, 'clip_ratio/high_mean': 0.04470324795693159, 'clip_ratio/high_max': 0.04470324795693159, 'clip_ratio/region_mean': 0.09938472509384155, 'reward_total_mean': 0.42175352573394775, 'reward_meter_mean': 0.9710224270820618, 'reward_meter_std': 0.024256430566310883, 'reward_count_adherence_mean': 0.949999988079071, 'reward_count_adherence_std': 0.09258200973272324, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8977535963058472, 'reward_repeat_soft_std': 0.03404621779918671, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.029730942100286484, 'reward_total_composite_mean': 0.42175352573394775, 'reward_total_composite_std': 0.02008803002536297, 'epoch': 0.21}

 65%|██████▍   | 2131/3300 [5:14:25<2:39:26,  8.18s/it]INFO 04-13 12:35:23 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2132/3300 [5:14:31<2:28:47,  7.64s/it]2026-04-13 12:35:29,877 | INFO | train_grpo_train | metrics_logged mode=train step=2132

                                                       
{'loss': -0.04, 'grad_norm': 19.79250717163086, 'learning_rate': 3.5424242424242426e-06, 'num_tokens': 3835091.0, 'completions/mean_length': 25.125, 'completions/min_length': 20.0, 'completions/max_length': 28.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 25.125, 'completions/min_terminated_length': 20.0, 'completions/max_terminated_length': 28.0, 'rewards/meter/mean': 0.9905524253845215, 'rewards/meter/std': 0.008373869583010674, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9346094131469727, 'rewards/repeat_soft/std': 0.04883209988474846, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4367702603340149, 'rewards/total_composite/std': 0.007294591516256332, 'reward': 0.4367702603340149, 'reward_std': 0.007294599432498217, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08623015135526657, 'sampling/sampling_logp_difference/max': 1.2098863124847412, 'sampling/importance_sampling_ratio/min': 0.3049562871456146, 'sampling/importance_sampling_ratio/mean': 1.0061790943145752, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6499927863478661, 'clip_ratio/low_mean': 0.02175925951451063, 'clip_ratio/low_min': 0.02175925951451063, 'clip_ratio/high_mean': 0.060073030181229115, 'clip_ratio/high_max': 0.060073030181229115, 'clip_ratio/region_mean': 0.08183228969573975, 'reward_total_mean': 0.4367702603340149, 'reward_meter_mean': 0.9905524253845215, 'reward_meter_std': 0.008373869583010674, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9346094131469727, 'reward_repeat_soft_std': 0.04883209988474846, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4367702603340149, 'reward_total_composite_std': 0.007294591516256332, 'epoch': 0.21}

 65%|██████▍   | 2132/3300 [5:14:32<2:28:47,  7.64s/it]INFO 04-13 12:35:30 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2133/3300 [5:14:38<2:20:47,  7.24s/it]2026-04-13 12:35:36,165 | INFO | train_grpo_train | metrics_logged mode=train step=2133

                                                       
{'loss': 0.0394, 'grad_norm': 10.58881950378418, 'learning_rate': 3.53939393939394e-06, 'num_tokens': 3836604.0, 'completions/mean_length': 23.125, 'completions/min_length': 21.0, 'completions/max_length': 26.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 23.125, 'completions/min_terminated_length': 21.0, 'completions/max_terminated_length': 26.0, 'rewards/meter/mean': 0.8037121295928955, 'rewards/meter/std': 0.3496001064777374, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9624999761581421, 'rewards/repeat_soft/std': 0.0, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4227369427680969, 'rewards/total_composite/std': 0.03408600762486458, 'reward': 0.4227369427680969, 'reward_std': 0.03408601135015488, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09222669899463654, 'sampling/sampling_logp_difference/max': 1.510305404663086, 'sampling/importance_sampling_ratio/min': 0.22084252536296844, 'sampling/importance_sampling_ratio/mean': 1.0104715824127197, 'sampling/importance_sampling_ratio/max': 1.919601321220398, 'entropy': 0.3871112950146198, 'clip_ratio/low_mean': 0.02003205195069313, 'clip_ratio/low_min': 0.02003205195069313, 'clip_ratio/high_mean': 0.044679089449346066, 'clip_ratio/high_max': 0.044679089449346066, 'clip_ratio/region_mean': 0.0647111414000392, 'reward_total_mean': 0.4227369427680969, 'reward_meter_mean': 0.8037121295928955, 'reward_meter_std': 0.3496001064777374, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9624999761581421, 'reward_repeat_soft_std': 0.0, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4227369427680969, 'reward_total_composite_std': 0.03408600762486458, 'epoch': 0.21}

 65%|██████▍   | 2133/3300 [5:14:38<2:20:47,  7.24s/it]INFO 04-13 12:35:36 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2134/3300 [5:14:50<2:49:41,  8.73s/it]2026-04-13 12:35:48,380 | INFO | train_grpo_train | metrics_logged mode=train step=2134

                                                       
{'loss': -0.2088, 'grad_norm': 1.5877206325531006, 'learning_rate': 3.5363636363636367e-06, 'num_tokens': 3838974.0, 'completions/mean_length': 163.25, 'completions/min_length': 100.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 113.42857360839844, 'completions/min_terminated_length': 100.0, 'completions/max_terminated_length': 127.0, 'rewards/meter/mean': 0.9713046550750732, 'rewards/meter/std': 0.03033020719885826, 'rewards/count_adherence/mean': 0.90625, 'rewards/count_adherence/std': 0.2651650309562683, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9271901845932007, 'rewards/repeat_soft/std': 0.020566994324326515, 'rewards/judge_quality/mean': 0.1600000113248825, 'rewards/judge_quality/std': 0.04566962271928787, 'rewards/total_composite/mean': 0.39359205961227417, 'rewards/total_composite/std': 0.159275084733963, 'reward': 0.39359205961227417, 'reward_std': 0.159275084733963, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08693575859069824, 'sampling/sampling_logp_difference/max': 2.0825905799865723, 'sampling/importance_sampling_ratio/min': 0.12460698932409286, 'sampling/importance_sampling_ratio/mean': 0.9960200786590576, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4335934594273567, 'clip_ratio/low_mean': 0.0, 'clip_ratio/low_min': 0.0, 'clip_ratio/high_mean': 0.07275812653824687, 'clip_ratio/high_max': 0.07275812653824687, 'clip_ratio/region_mean': 0.07275812653824687, 'reward_total_mean': 0.39359205961227417, 'reward_meter_mean': 0.9713046550750732, 'reward_meter_std': 0.03033020719885826, 'reward_count_adherence_mean': 0.90625, 'reward_count_adherence_std': 0.2651650309562683, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9271901845932007, 'reward_repeat_soft_std': 0.020566994324326515, 'reward_judge_quality_mean': 0.1600000113248825, 'reward_judge_quality_std': 0.04566962271928787, 'reward_total_composite_mean': 0.39359205961227417, 'reward_total_composite_std': 0.159275084733963, 'epoch': 0.21}

 65%|██████▍   | 2134/3300 [5:14:50<2:49:41,  8.73s/it]INFO 04-13 12:35:48 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2135/3300 [5:15:03<3:14:18, 10.01s/it]2026-04-13 12:36:01,375 | INFO | train_grpo_train | metrics_logged mode=train step=2135

                                                       
{'loss': -0.1097, 'grad_norm': 2.3888208866119385, 'learning_rate': 3.5333333333333335e-06, 'num_tokens': 3840585.0, 'completions/mean_length': 95.375, 'completions/min_length': 33.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 35.85714340209961, 'completions/min_terminated_length': 33.0, 'completions/max_terminated_length': 39.0, 'rewards/meter/mean': 0.7480518817901611, 'rewards/meter/std': 0.29896217584609985, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9610674381256104, 'rewards/repeat_soft/std': 0.01786000281572342, 'rewards/judge_quality/mean': 0.17500001192092896, 'rewards/judge_quality/std': 0.11649646610021591, 'rewards/total_composite/mean': 0.39515113830566406, 'rewards/total_composite/std': 0.17731983959674835, 'reward': 0.39515113830566406, 'reward_std': 0.17731983959674835, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08909184485673904, 'sampling/sampling_logp_difference/max': 1.0417542457580566, 'sampling/importance_sampling_ratio/min': 0.35283517837524414, 'sampling/importance_sampling_ratio/mean': 0.978323221206665, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.30892448872327805, 'clip_ratio/low_mean': 0.007575757801532745, 'clip_ratio/low_min': 0.007575757801532745, 'clip_ratio/high_mean': 0.08523908769711852, 'clip_ratio/high_max': 0.08523908769711852, 'clip_ratio/region_mean': 0.09281484549865127, 'reward_total_mean': 0.39515113830566406, 'reward_meter_mean': 0.7480518817901611, 'reward_meter_std': 0.29896217584609985, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9610674381256104, 'reward_repeat_soft_std': 0.01786000281572342, 'reward_judge_quality_mean': 0.17500001192092896, 'reward_judge_quality_std': 0.11649646610021591, 'reward_total_composite_mean': 0.39515113830566406, 'reward_total_composite_std': 0.17731983959674835, 'epoch': 0.21}

 65%|██████▍   | 2135/3300 [5:15:03<3:14:18, 10.01s/it]INFO 04-13 12:36:01 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2136/3300 [5:15:10<2:59:26,  9.25s/it]2026-04-13 12:36:08,841 | INFO | train_grpo_train | metrics_logged mode=train step=2136

                                                       
{'loss': 0.0455, 'grad_norm': 12.88821792602539, 'learning_rate': 3.5303030303030304e-06, 'num_tokens': 3842041.0, 'completions/mean_length': 29.0, 'completions/min_length': 18.0, 'completions/max_length': 36.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 29.0, 'completions/min_terminated_length': 18.0, 'completions/max_terminated_length': 36.0, 'rewards/meter/mean': 0.8173210024833679, 'rewards/meter/std': 0.33915388584136963, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9517419338226318, 'rewards/repeat_soft/std': 0.020060529932379723, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.42245009541511536, 'rewards/total_composite/std': 0.032370131462812424, 'reward': 0.42245009541511536, 'reward_std': 0.032370127737522125, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.13327038288116455, 'sampling/sampling_logp_difference/max': 2.3602185249328613, 'sampling/importance_sampling_ratio/min': 0.094399593770504, 'sampling/importance_sampling_ratio/mean': 0.9857088923454285, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.6100064590573311, 'clip_ratio/low_mean': 0.018518518656492233, 'clip_ratio/low_min': 0.018518518656492233, 'clip_ratio/high_mean': 0.10606145393103361, 'clip_ratio/high_max': 0.10606145393103361, 'clip_ratio/region_mean': 0.12457997258752584, 'reward_total_mean': 0.42245009541511536, 'reward_meter_mean': 0.8173210024833679, 'reward_meter_std': 0.33915388584136963, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9517419338226318, 'reward_repeat_soft_std': 0.020060529932379723, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.42245009541511536, 'reward_total_composite_std': 0.032370131462812424, 'epoch': 0.21}

 65%|██████▍   | 2136/3300 [5:15:11<2:59:26,  9.25s/it]INFO 04-13 12:36:09 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2137/3300 [5:15:17<2:43:57,  8.46s/it]2026-04-13 12:36:15,465 | INFO | train_grpo_train | metrics_logged mode=train step=2137

                                                       
{'loss': 0.0787, 'grad_norm': 24.174345016479492, 'learning_rate': 3.5272727272727276e-06, 'num_tokens': 3843421.0, 'completions/mean_length': 21.5, 'completions/min_length': 15.0, 'completions/max_length': 27.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 21.5, 'completions/min_terminated_length': 15.0, 'completions/max_terminated_length': 27.0, 'rewards/meter/mean': 0.6113148927688599, 'rewards/meter/std': 0.4718313217163086, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9569318294525146, 'rewards/repeat_soft/std': 0.015749182552099228, 'rewards/judge_quality/mean': 0.24625001847743988, 'rewards/judge_quality/std': 0.2722361087799072, 'rewards/total_composite/mean': 0.40339311957359314, 'rewards/total_composite/std': 0.0453665629029274, 'reward': 0.40339311957359314, 'reward_std': 0.0453665629029274, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.14792943000793457, 'sampling/sampling_logp_difference/max': 1.133669376373291, 'sampling/importance_sampling_ratio/min': 0.32185009121894836, 'sampling/importance_sampling_ratio/mean': 0.9756884574890137, 'sampling/importance_sampling_ratio/max': 1.6885017156600952, 'entropy': 0.930234007537365, 'clip_ratio/low_mean': 0.03321759309619665, 'clip_ratio/low_min': 0.03321759309619665, 'clip_ratio/high_mean': 0.10881273122504354, 'clip_ratio/high_max': 0.10881273122504354, 'clip_ratio/region_mean': 0.1420303243212402, 'reward_total_mean': 0.40339311957359314, 'reward_meter_mean': 0.6113148927688599, 'reward_meter_std': 0.4718313217163086, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9569318294525146, 'reward_repeat_soft_std': 0.015749182552099228, 'reward_judge_quality_mean': 0.24625001847743988, 'reward_judge_quality_std': 0.2722361087799072, 'reward_total_composite_mean': 0.40339311957359314, 'reward_total_composite_std': 0.0453665629029274, 'epoch': 0.21}

 65%|██████▍   | 2137/3300 [5:15:17<2:43:57,  8.46s/it]INFO 04-13 12:36:15 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2138/3300 [5:15:24<2:37:45,  8.15s/it]2026-04-13 12:36:22,875 | INFO | train_grpo_train | metrics_logged mode=train step=2138

                                                       
{'loss': 0.009, 'grad_norm': 6.568856716156006, 'learning_rate': 3.5242424242424244e-06, 'num_tokens': 3845705.0, 'completions/mean_length': 118.5, 'completions/min_length': 105.0, 'completions/max_length': 134.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 118.5, 'completions/min_terminated_length': 105.0, 'completions/max_terminated_length': 134.0, 'rewards/meter/mean': 0.8722131848335266, 'rewards/meter/std': 0.14797870814800262, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8938058614730835, 'rewards/repeat_soft/std': 0.03232758864760399, 'rewards/judge_quality/mean': 0.17250001430511475, 'rewards/judge_quality/std': 0.013887302950024605, 'rewards/total_composite/mean': 0.43214356899261475, 'rewards/total_composite/std': 0.019955281168222427, 'reward': 0.43214356899261475, 'reward_std': 0.019955281168222427, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10581151396036148, 'sampling/sampling_logp_difference/max': 2.4617185592651367, 'sampling/importance_sampling_ratio/min': 0.08528824895620346, 'sampling/importance_sampling_ratio/mean': 1.0000040531158447, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.43669046834111214, 'clip_ratio/low_mean': 0.03886368125677109, 'clip_ratio/low_min': 0.03886368125677109, 'clip_ratio/high_mean': 0.057091874070465565, 'clip_ratio/high_max': 0.057091874070465565, 'clip_ratio/region_mean': 0.09595555532723665, 'reward_total_mean': 0.43214356899261475, 'reward_meter_mean': 0.8722131848335266, 'reward_meter_std': 0.14797870814800262, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8938058614730835, 'reward_repeat_soft_std': 0.03232758864760399, 'reward_judge_quality_mean': 0.17250001430511475, 'reward_judge_quality_std': 0.013887302950024605, 'reward_total_composite_mean': 0.43214356899261475, 'reward_total_composite_std': 0.019955281168222427, 'epoch': 0.21}

 65%|██████▍   | 2138/3300 [5:15:25<2:37:45,  8.15s/it]INFO 04-13 12:36:23 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2139/3300 [5:15:30<2:24:51,  7.49s/it]2026-04-13 12:36:28,821 | INFO | train_grpo_train | metrics_logged mode=train step=2139

                                                       
{'loss': -0.0343, 'grad_norm': 13.603361129760742, 'learning_rate': 3.5212121212121213e-06, 'num_tokens': 3847031.0, 'completions/mean_length': 27.75, 'completions/min_length': 24.0, 'completions/max_length': 30.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 27.75, 'completions/min_terminated_length': 24.0, 'completions/max_terminated_length': 30.0, 'rewards/meter/mean': 0.9820600748062134, 'rewards/meter/std': 0.023328015580773354, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9340013265609741, 'rewards/repeat_soft/std': 0.030499953776597977, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4358510673046112, 'rewards/total_composite/std': 0.0052992128767073154, 'reward': 0.4358510673046112, 'reward_std': 0.005299215670675039, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0838530957698822, 'sampling/sampling_logp_difference/max': 1.0038328170776367, 'sampling/importance_sampling_ratio/min': 0.36647212505340576, 'sampling/importance_sampling_ratio/mean': 1.0212786197662354, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5129394680261612, 'clip_ratio/low_mean': 0.05222701234742999, 'clip_ratio/low_min': 0.05222701234742999, 'clip_ratio/high_mean': 0.016964286100119352, 'clip_ratio/high_max': 0.016964286100119352, 'clip_ratio/region_mean': 0.06919129844754934, 'reward_total_mean': 0.4358510673046112, 'reward_meter_mean': 0.9820600748062134, 'reward_meter_std': 0.023328015580773354, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9340013265609741, 'reward_repeat_soft_std': 0.030499953776597977, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4358510673046112, 'reward_total_composite_std': 0.0052992128767073154, 'epoch': 0.21}

 65%|██████▍   | 2139/3300 [5:15:31<2:24:51,  7.49s/it]INFO 04-13 12:36:29 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2140/3300 [5:15:37<2:21:04,  7.30s/it]2026-04-13 12:36:35,670 | INFO | train_grpo_train | metrics_logged mode=train step=2140

                                                       
{'loss': 0.0841, 'grad_norm': 11.767311096191406, 'learning_rate': 3.5181818181818185e-06, 'num_tokens': 3848601.0, 'completions/mean_length': 34.25, 'completions/min_length': 27.0, 'completions/max_length': 46.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 34.25, 'completions/min_terminated_length': 27.0, 'completions/max_terminated_length': 46.0, 'rewards/meter/mean': 0.8089805841445923, 'rewards/meter/std': 0.216323122382164, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9660546183586121, 'rewards/repeat_soft/std': 0.015586726367473602, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.0, 'rewards/total_composite/mean': 0.4237838089466095, 'rewards/total_composite/std': 0.02083503268659115, 'reward': 0.4237838089466095, 'reward_std': 0.020835043862462044, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08979345858097076, 'sampling/sampling_logp_difference/max': 2.035578727722168, 'sampling/importance_sampling_ratio/min': 0.13060487806797028, 'sampling/importance_sampling_ratio/mean': 1.0170397758483887, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3702515698969364, 'clip_ratio/low_mean': 0.028292838484048843, 'clip_ratio/low_min': 0.028292838484048843, 'clip_ratio/high_mean': 0.061627304181456566, 'clip_ratio/high_max': 0.061627304181456566, 'clip_ratio/region_mean': 0.08992014266550541, 'reward_total_mean': 0.4237838089466095, 'reward_meter_mean': 0.8089805841445923, 'reward_meter_std': 0.216323122382164, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9660546183586121, 'reward_repeat_soft_std': 0.015586726367473602, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.0, 'reward_total_composite_mean': 0.4237838089466095, 'reward_total_composite_std': 0.02083503268659115, 'epoch': 0.21}

 65%|██████▍   | 2140/3300 [5:15:37<2:21:04,  7.30s/it]INFO 04-13 12:36:35 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2141/3300 [5:15:44<2:17:14,  7.10s/it]2026-04-13 12:36:42,340 | INFO | train_grpo_train | metrics_logged mode=train step=2141

                                                       
{'loss': 0.0321, 'grad_norm': 7.464240550994873, 'learning_rate': 3.5151515151515154e-06, 'num_tokens': 3850272.0, 'completions/mean_length': 52.875, 'completions/min_length': 50.0, 'completions/max_length': 58.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.875, 'completions/min_terminated_length': 50.0, 'completions/max_terminated_length': 58.0, 'rewards/meter/mean': 0.9091874361038208, 'rewards/meter/std': 0.1122494712471962, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9380810260772705, 'rewards/repeat_soft/std': 0.0403120182454586, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.43157362937927246, 'rewards/total_composite/std': 0.010991547256708145, 'reward': 0.43157362937927246, 'reward_std': 0.010991537012159824, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09349511563777924, 'sampling/sampling_logp_difference/max': 1.4340959787368774, 'sampling/importance_sampling_ratio/min': 0.23833072185516357, 'sampling/importance_sampling_ratio/mean': 1.0083507299423218, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5642917603254318, 'clip_ratio/low_mean': 0.09280007239431143, 'clip_ratio/low_min': 0.09280007239431143, 'clip_ratio/high_mean': 0.021226415410637856, 'clip_ratio/high_max': 0.021226415410637856, 'clip_ratio/region_mean': 0.11402648780494928, 'reward_total_mean': 0.43157362937927246, 'reward_meter_mean': 0.9091874361038208, 'reward_meter_std': 0.1122494712471962, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9380810260772705, 'reward_repeat_soft_std': 0.0403120182454586, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.43157362937927246, 'reward_total_composite_std': 0.010991547256708145, 'epoch': 0.22}

 65%|██████▍   | 2141/3300 [5:15:44<2:17:14,  7.10s/it]INFO 04-13 12:36:42 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2142/3300 [5:15:52<2:20:09,  7.26s/it]2026-04-13 12:36:49,964 | INFO | train_grpo_train | metrics_logged mode=train step=2142

                                                       
{'loss': -0.0837, 'grad_norm': 8.290336608886719, 'learning_rate': 3.512121212121212e-06, 'num_tokens': 3852342.0, 'completions/mean_length': 100.75, 'completions/min_length': 77.0, 'completions/max_length': 145.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 100.75, 'completions/min_terminated_length': 77.0, 'completions/max_terminated_length': 145.0, 'rewards/meter/mean': 0.6486868858337402, 'rewards/meter/std': 0.2888045907020569, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9150484800338745, 'rewards/repeat_soft/std': 0.02585035189986229, 'rewards/judge_quality/mean': 0.21000000834465027, 'rewards/judge_quality/std': 0.08485280722379684, 'rewards/total_composite/mean': 0.42709919810295105, 'rewards/total_composite/std': 0.05602974444627762, 'reward': 0.42709919810295105, 'reward_std': 0.05602974444627762, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.0966973677277565, 'sampling/sampling_logp_difference/max': 1.9362258911132812, 'sampling/importance_sampling_ratio/min': 0.14424732327461243, 'sampling/importance_sampling_ratio/mean': 1.0121235847473145, 'sampling/importance_sampling_ratio/max': 1.9052740335464478, 'entropy': 0.5177117586135864, 'clip_ratio/low_mean': 0.053842433262616396, 'clip_ratio/low_min': 0.053842433262616396, 'clip_ratio/high_mean': 0.058138023130595684, 'clip_ratio/high_max': 0.058138023130595684, 'clip_ratio/region_mean': 0.11198045639321208, 'reward_total_mean': 0.42709919810295105, 'reward_meter_mean': 0.6486868858337402, 'reward_meter_std': 0.2888045907020569, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9150484800338745, 'reward_repeat_soft_std': 0.02585035189986229, 'reward_judge_quality_mean': 0.21000000834465027, 'reward_judge_quality_std': 0.08485280722379684, 'reward_total_composite_mean': 0.42709919810295105, 'reward_total_composite_std': 0.05602974444627762, 'epoch': 0.22}

 65%|██████▍   | 2142/3300 [5:15:52<2:20:09,  7.26s/it]INFO 04-13 12:36:50 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2143/3300 [5:15:58<2:16:23,  7.07s/it]2026-04-13 12:36:56,604 | INFO | train_grpo_train | metrics_logged mode=train step=2143

                                                       
{'loss': 0.0574, 'grad_norm': 7.681342124938965, 'learning_rate': 3.509090909090909e-06, 'num_tokens': 3854286.0, 'completions/mean_length': 53.0, 'completions/min_length': 49.0, 'completions/max_length': 62.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 53.0, 'completions/min_terminated_length': 49.0, 'completions/max_terminated_length': 62.0, 'rewards/meter/mean': 0.9523389339447021, 'rewards/meter/std': 0.03752657771110535, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9398510456085205, 'rewards/repeat_soft/std': 0.030083410441875458, 'rewards/judge_quality/mean': 0.16500000655651093, 'rewards/judge_quality/std': 0.01603567600250244, 'rewards/total_composite/mean': 0.4429652690887451, 'rewards/total_composite/std': 0.008120769634842873, 'reward': 0.4429652690887451, 'reward_std': 0.008120759390294552, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09259400516748428, 'sampling/sampling_logp_difference/max': 2.3451058864593506, 'sampling/importance_sampling_ratio/min': 0.0958370566368103, 'sampling/importance_sampling_ratio/mean': 1.0037641525268555, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44840671494603157, 'clip_ratio/low_mean': 0.04185404907912016, 'clip_ratio/low_min': 0.04185404907912016, 'clip_ratio/high_mean': 0.046734149334952235, 'clip_ratio/high_max': 0.046734149334952235, 'clip_ratio/region_mean': 0.0885881984140724, 'reward_total_mean': 0.4429652690887451, 'reward_meter_mean': 0.9523389339447021, 'reward_meter_std': 0.03752657771110535, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9398510456085205, 'reward_repeat_soft_std': 0.030083410441875458, 'reward_judge_quality_mean': 0.16500000655651093, 'reward_judge_quality_std': 0.01603567600250244, 'reward_total_composite_mean': 0.4429652690887451, 'reward_total_composite_std': 0.008120769634842873, 'epoch': 0.22}

 65%|██████▍   | 2143/3300 [5:15:58<2:16:23,  7.07s/it]INFO 04-13 12:36:56 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▍   | 2144/3300 [5:16:05<2:15:36,  7.04s/it]2026-04-13 12:37:03,555 | INFO | train_grpo_train | metrics_logged mode=train step=2144

                                                       
{'loss': -0.0546, 'grad_norm': 8.980925559997559, 'learning_rate': 3.5060606060606063e-06, 'num_tokens': 3856427.0, 'completions/mean_length': 78.625, 'completions/min_length': 63.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 78.625, 'completions/min_terminated_length': 63.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.8787567019462585, 'rewards/meter/std': 0.20320703089237213, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8748688697814941, 'rewards/repeat_soft/std': 0.019511040300130844, 'rewards/judge_quality/mean': 0.16124999523162842, 'rewards/judge_quality/std': 0.015526476316154003, 'rewards/total_composite/mean': 0.4235439896583557, 'rewards/total_composite/std': 0.02333262749016285, 'reward': 0.4235439896583557, 'reward_std': 0.0233326256275177, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.12014502286911011, 'sampling/sampling_logp_difference/max': 1.8084826469421387, 'sampling/importance_sampling_ratio/min': 0.1639026403427124, 'sampling/importance_sampling_ratio/mean': 0.9930413365364075, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5412853918969631, 'clip_ratio/low_mean': 0.023078530095517635, 'clip_ratio/low_min': 0.023078530095517635, 'clip_ratio/high_mean': 0.08712796587496996, 'clip_ratio/high_max': 0.08712796587496996, 'clip_ratio/region_mean': 0.1102064959704876, 'reward_total_mean': 0.4235439896583557, 'reward_meter_mean': 0.8787567019462585, 'reward_meter_std': 0.20320703089237213, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8748688697814941, 'reward_repeat_soft_std': 0.019511040300130844, 'reward_judge_quality_mean': 0.16124999523162842, 'reward_judge_quality_std': 0.015526476316154003, 'reward_total_composite_mean': 0.4235439896583557, 'reward_total_composite_std': 0.02333262749016285, 'epoch': 0.22}

 65%|██████▍   | 2144/3300 [5:16:05<2:15:36,  7.04s/it]INFO 04-13 12:37:03 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2145/3300 [5:16:17<2:46:02,  8.63s/it]2026-04-13 12:37:15,882 | INFO | train_grpo_train | metrics_logged mode=train step=2145

                                                       
{'loss': -0.0907, 'grad_norm': 3.2406935691833496, 'learning_rate': 3.503030303030303e-06, 'num_tokens': 3857950.0, 'completions/mean_length': 86.375, 'completions/min_length': 22.0, 'completions/max_length': 512.0, 'completions/clipped_ratio': 0.125, 'completions/mean_terminated_length': 25.571430206298828, 'completions/min_terminated_length': 22.0, 'completions/max_terminated_length': 31.0, 'rewards/meter/mean': 0.8375157713890076, 'rewards/meter/std': 0.3385445475578308, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.8941828608512878, 'rewards/repeat_soft/std': 0.11264170706272125, 'rewards/judge_quality/mean': 0.13750000298023224, 'rewards/judge_quality/std': 0.0353553406894207, 'rewards/total_composite/mean': 0.36311087012290955, 'rewards/total_composite/std': 0.15466636419296265, 'reward': 0.36311087012290955, 'reward_std': 0.15466636419296265, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11213835328817368, 'sampling/sampling_logp_difference/max': 0.9431552886962891, 'sampling/importance_sampling_ratio/min': 0.3893972635269165, 'sampling/importance_sampling_ratio/mean': 1.052787184715271, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.8141542971134186, 'clip_ratio/low_mean': 0.021739130839705467, 'clip_ratio/low_min': 0.021739130839705467, 'clip_ratio/high_mean': 0.076064292807132, 'clip_ratio/high_max': 0.076064292807132, 'clip_ratio/region_mean': 0.09780342364683747, 'reward_total_mean': 0.36311087012290955, 'reward_meter_mean': 0.8375157713890076, 'reward_meter_std': 0.3385445475578308, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.8941828608512878, 'reward_repeat_soft_std': 0.11264170706272125, 'reward_judge_quality_mean': 0.13750000298023224, 'reward_judge_quality_std': 0.0353553406894207, 'reward_total_composite_mean': 0.36311087012290955, 'reward_total_composite_std': 0.15466636419296265, 'epoch': 0.22}

 65%|██████▌   | 2145/3300 [5:16:18<2:46:02,  8.63s/it]INFO 04-13 12:37:16 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2146/3300 [5:16:26<2:43:55,  8.52s/it]2026-04-13 12:37:24,160 | INFO | train_grpo_train | metrics_logged mode=train step=2146

                                                       
{'loss': 0.0478, 'grad_norm': 6.35801362991333, 'learning_rate': 3.5e-06, 'num_tokens': 3860945.0, 'completions/mean_length': 201.375, 'completions/min_length': 176.0, 'completions/max_length': 229.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 201.375, 'completions/min_terminated_length': 176.0, 'completions/max_terminated_length': 229.0, 'rewards/meter/mean': 0.9370740652084351, 'rewards/meter/std': 0.06427296251058578, 'rewards/count_adherence/mean': 0.8125, 'rewards/count_adherence/std': 0.05892555043101311, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8033504486083984, 'rewards/repeat_soft/std': 0.07749520242214203, 'rewards/judge_quality/mean': 0.13500000536441803, 'rewards/judge_quality/std': 0.02777460776269436, 'rewards/total_composite/mean': 0.365647554397583, 'rewards/total_composite/std': 0.034212831407785416, 'reward': 0.365647554397583, 'reward_std': 0.03421282768249512, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.08685857802629471, 'sampling/sampling_logp_difference/max': 3.345730781555176, 'sampling/importance_sampling_ratio/min': 0.03523445874452591, 'sampling/importance_sampling_ratio/mean': 1.001274585723877, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.4255397655069828, 'clip_ratio/low_mean': 0.03247932391241193, 'clip_ratio/low_min': 0.03247932391241193, 'clip_ratio/high_mean': 0.048446863889694214, 'clip_ratio/high_max': 0.048446863889694214, 'clip_ratio/region_mean': 0.08092618780210614, 'reward_total_mean': 0.365647554397583, 'reward_meter_mean': 0.9370740652084351, 'reward_meter_std': 0.06427296251058578, 'reward_count_adherence_mean': 0.8125, 'reward_count_adherence_std': 0.05892555043101311, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8033504486083984, 'reward_repeat_soft_std': 0.07749520242214203, 'reward_judge_quality_mean': 0.13500000536441803, 'reward_judge_quality_std': 0.02777460776269436, 'reward_total_composite_mean': 0.365647554397583, 'reward_total_composite_std': 0.034212831407785416, 'epoch': 0.22}

 65%|██████▌   | 2146/3300 [5:16:26<2:43:55,  8.52s/it]INFO 04-13 12:37:24 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2147/3300 [5:16:32<2:30:59,  7.86s/it]2026-04-13 12:37:30,478 | INFO | train_grpo_train | metrics_logged mode=train step=2147

                                                       
{'loss': -0.009, 'grad_norm': 13.747391700744629, 'learning_rate': 3.496969696969697e-06, 'num_tokens': 3862703.0, 'completions/mean_length': 45.75, 'completions/min_length': 37.0, 'completions/max_length': 51.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 45.75, 'completions/min_terminated_length': 37.0, 'completions/max_terminated_length': 51.0, 'rewards/meter/mean': 0.8913014531135559, 'rewards/meter/std': 0.22273597121238708, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9591418504714966, 'rewards/repeat_soft/std': 0.035527247935533524, 'rewards/judge_quality/mean': 0.14625000953674316, 'rewards/judge_quality/std': 0.010606604628264904, 'rewards/total_composite/mean': 0.4283779263496399, 'rewards/total_composite/std': 0.024290764704346657, 'reward': 0.4283779263496399, 'reward_std': 0.02429075725376606, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.09208957105875015, 'sampling/sampling_logp_difference/max': 1.9488413333892822, 'sampling/importance_sampling_ratio/min': 0.14243902266025543, 'sampling/importance_sampling_ratio/mean': 1.0007129907608032, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.44552335888147354, 'clip_ratio/low_mean': 0.014354674611240625, 'clip_ratio/low_min': 0.014354674611240625, 'clip_ratio/high_mean': 0.08540754113346338, 'clip_ratio/high_max': 0.08540754113346338, 'clip_ratio/region_mean': 0.09976221574470401, 'reward_total_mean': 0.4283779263496399, 'reward_meter_mean': 0.8913014531135559, 'reward_meter_std': 0.22273597121238708, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9591418504714966, 'reward_repeat_soft_std': 0.035527247935533524, 'reward_judge_quality_mean': 0.14625000953674316, 'reward_judge_quality_std': 0.010606604628264904, 'reward_total_composite_mean': 0.4283779263496399, 'reward_total_composite_std': 0.024290764704346657, 'epoch': 0.22}

 65%|██████▌   | 2147/3300 [5:16:32<2:30:59,  7.86s/it]INFO 04-13 12:37:30 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2148/3300 [5:16:39<2:25:56,  7.60s/it]2026-04-13 12:37:37,467 | INFO | train_grpo_train | metrics_logged mode=train step=2148

                                                       
{'loss': -0.0048, 'grad_norm': 12.051156997680664, 'learning_rate': 3.493939393939394e-06, 'num_tokens': 3864702.0, 'completions/mean_length': 69.875, 'completions/min_length': 62.0, 'completions/max_length': 88.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 69.875, 'completions/min_terminated_length': 62.0, 'completions/max_terminated_length': 88.0, 'rewards/meter/mean': 0.8386724591255188, 'rewards/meter/std': 0.2629724144935608, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 0.875, 'rewards/hard_gate/std': 0.3535533845424652, 'rewards/repeat_soft/mean': 0.9236710071563721, 'rewards/repeat_soft/std': 0.02468961849808693, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.01922610215842724, 'rewards/total_composite/mean': 0.37210574746131897, 'rewards/total_composite/std': 0.15347681939601898, 'reward': 0.37210574746131897, 'reward_std': 0.1534768044948578, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.10458245873451233, 'sampling/sampling_logp_difference/max': 2.363548517227173, 'sampling/importance_sampling_ratio/min': 0.09408576041460037, 'sampling/importance_sampling_ratio/mean': 1.0113861560821533, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.5813146941363811, 'clip_ratio/low_mean': 0.018656716216355562, 'clip_ratio/low_min': 0.018656716216355562, 'clip_ratio/high_mean': 0.06361279031261802, 'clip_ratio/high_max': 0.06361279031261802, 'clip_ratio/region_mean': 0.08226950652897358, 'reward_total_mean': 0.37210574746131897, 'reward_meter_mean': 0.8386724591255188, 'reward_meter_std': 0.2629724144935608, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 0.875, 'reward_hard_gate_std': 0.3535533845424652, 'reward_repeat_soft_mean': 0.9236710071563721, 'reward_repeat_soft_std': 0.02468961849808693, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.01922610215842724, 'reward_total_composite_mean': 0.37210574746131897, 'reward_total_composite_std': 0.15347681939601898, 'epoch': 0.22}

 65%|██████▌   | 2148/3300 [5:16:39<2:25:56,  7.60s/it]INFO 04-13 12:37:37 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2149/3300 [5:16:46<2:20:00,  7.30s/it]2026-04-13 12:37:44,070 | INFO | train_grpo_train | metrics_logged mode=train step=2149

                                                       
{'loss': 0.0132, 'grad_norm': 11.868727684020996, 'learning_rate': 3.4909090909090913e-06, 'num_tokens': 3866315.0, 'completions/mean_length': 42.625, 'completions/min_length': 41.0, 'completions/max_length': 50.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 42.625, 'completions/min_terminated_length': 41.0, 'completions/max_terminated_length': 50.0, 'rewards/meter/mean': 0.9211106896400452, 'rewards/meter/std': 0.10339754819869995, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9673133492469788, 'rewards/repeat_soft/std': 0.019251564517617226, 'rewards/judge_quality/mean': 0.1537500023841858, 'rewards/judge_quality/std': 0.010606602765619755, 'rewards/total_composite/mean': 0.4373164772987366, 'rewards/total_composite/std': 0.012059218250215054, 'reward': 0.4373164772987366, 'reward_std': 0.012059218250215054, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11840193718671799, 'sampling/sampling_logp_difference/max': 1.9005436897277832, 'sampling/importance_sampling_ratio/min': 0.14948733150959015, 'sampling/importance_sampling_ratio/mean': 0.987741231918335, 'sampling/importance_sampling_ratio/max': 1.8737012147903442, 'entropy': 0.5505771040916443, 'clip_ratio/low_mean': 0.0335365841165185, 'clip_ratio/low_min': 0.0335365841165185, 'clip_ratio/high_mean': 0.07705757208168507, 'clip_ratio/high_max': 0.07705757208168507, 'clip_ratio/region_mean': 0.11059415619820356, 'reward_total_mean': 0.4373164772987366, 'reward_meter_mean': 0.9211106896400452, 'reward_meter_std': 0.10339754819869995, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9673133492469788, 'reward_repeat_soft_std': 0.019251564517617226, 'reward_judge_quality_mean': 0.1537500023841858, 'reward_judge_quality_std': 0.010606602765619755, 'reward_total_composite_mean': 0.4373164772987366, 'reward_total_composite_std': 0.012059218250215054, 'epoch': 0.22}

 65%|██████▌   | 2149/3300 [5:16:46<2:20:00,  7.30s/it]INFO 04-13 12:37:44 [block_pool.py:378] Successfully reset prefix cache

 65%|██████▌   | 2150/3300 [5:16:52<2:17:00,  7.15s/it]2026-04-13 12:37:50,852 | INFO | train_grpo_train | metrics_logged mode=train step=2150

                                                       
{'loss': -0.0127, 'grad_norm': 11.236807823181152, 'learning_rate': 3.4878787878787885e-06, 'num_tokens': 3868253.0, 'completions/mean_length': 52.25, 'completions/min_length': 45.0, 'completions/max_length': 57.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 52.25, 'completions/min_terminated_length': 45.0, 'completions/max_terminated_length': 57.0, 'rewards/meter/mean': 0.8498669862747192, 'rewards/meter/std': 0.14511364698410034, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.9805920124053955, 'rewards/repeat_soft/std': 0.020208818838000298, 'rewards/judge_quality/mean': 0.15000000596046448, 'rewards/judge_quality/std': 0.01603567786514759, 'rewards/total_composite/mean': 0.4297049641609192, 'rewards/total_composite/std': 0.013732438907027245, 'reward': 0.4297049641609192, 'reward_std': 0.013732440769672394, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.11804606765508652, 'sampling/sampling_logp_difference/max': 2.0987229347229004, 'sampling/importance_sampling_ratio/min': 0.12261290848255157, 'sampling/importance_sampling_ratio/mean': 0.9951035976409912, 'sampling/importance_sampling_ratio/max': 1.9407857656478882, 'entropy': 0.589105699211359, 'clip_ratio/low_mean': 0.03481538034975529, 'clip_ratio/low_min': 0.03481538034975529, 'clip_ratio/high_mean': 0.05818670894950628, 'clip_ratio/high_max': 0.05818670894950628, 'clip_ratio/region_mean': 0.09300208929926157, 'reward_total_mean': 0.4297049641609192, 'reward_meter_mean': 0.8498669862747192, 'reward_meter_std': 0.14511364698410034, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.9805920124053955, 'reward_repeat_soft_std': 0.020208818838000298, 'reward_judge_quality_mean': 0.15000000596046448, 'reward_judge_quality_std': 0.01603567786514759, 'reward_total_composite_mean': 0.4297049641609192, 'reward_total_composite_std': 0.013732438907027245, 'epoch': 0.22}

 65%|██████▌   | 2150/3300 [5:16:53<2:17:00,  7.15s/it]INFO 04-13 12:37:51 [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%|          | 0/10 [00:00<?, ?it/s]

 20%|██        | 2/10 [00:04<00:16,  2.12s/it]

 30%|███       | 3/10 [00:08<00:22,  3.20s/it]

 40%|████      | 4/10 [00:12<00:19,  3.33s/it]

 50%|█████     | 5/10 [00:16<00:18,  3.73s/it]

 60%|██████    | 6/10 [00:21<00:15,  3.89s/it]

 70%|███████   | 7/10 [00:29<00:16,  5.46s/it]

 80%|████████  | 8/10 [00:34<00:10,  5.13s/it]

 90%|█████████ | 9/10 [00:38<00:04,  4.88s/it]

100%|██████████| 10/10 [00:42<00:00,  4.44s/it]2026-04-13 12:38:37,355 | INFO | train_grpo_train | metrics_logged mode=eval step=2150

                                                       

                                               
{'eval_loss': nan, 'eval_runtime': 46.3808, 'eval_samples_per_second': 1.725, 'eval_steps_per_second': 0.216, 'eval_num_tokens': 3868253.0, 'eval_completions/mean_length': 84.7, 'eval_completions/min_length': 31.7, 'eval_completions/max_length': 175.2, 'eval_completions/clipped_ratio': 0.0125, 'eval_completions/mean_terminated_length': 79.49821472167969, 'eval_completions/min_terminated_length': 31.7, 'eval_completions/max_terminated_length': 139.9, 'eval_rewards/meter/mean': 0.7940246820449829, 'eval_rewards/meter/std': 0.26315709128975867, 'eval_rewards/count_adherence/mean': 0.9881250023841858, 'eval_rewards/count_adherence/std': 0.03358757160604, 'eval_rewards/hard_gate/mean': 0.975, 'eval_rewards/hard_gate/std': 0.07071067690849304, 'eval_rewards/repeat_soft/mean': 0.8970467984676361, 'eval_rewards/repeat_soft/std': 0.07692538686096669, 'eval_rewards/judge_quality/mean': 0.15700000524520874, 'eval_rewards/judge_quality/std': 0.02237090365961194, 'eval_rewards/total_composite/mean': 0.4048567622900009, 'eval_rewards/total_composite/std': 0.0588286342099309, 'eval_reward': 0.4048567622900009, 'eval_reward_std': nan, 'eval_frac_reward_zero_std': 0.0, 'eval_sampling/sampling_logp_difference/mean': 0.046076252683997156, 'eval_sampling/sampling_logp_difference/max': 0.909369945526123, 'eval_sampling/importance_sampling_ratio/min': 0.418349090218544, 'eval_sampling/importance_sampling_ratio/mean': 1.0083821773529054, 'eval_sampling/importance_sampling_ratio/max': 1.4984214782714844, 'eval_entropy': 0.46796957552433016, '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': 0.4048567622900009, 'eval_reward_meter_mean': 0.7940246820449829, 'eval_reward_meter_std': 0.26315709128975867, 'eval_reward_count_adherence_mean': 0.9881250023841858, 'eval_reward_count_adherence_std': 0.03358757160604, 'eval_reward_hard_gate_mean': 0.975, 'eval_reward_hard_gate_std': 0.07071067690849304, 'eval_reward_repeat_soft_mean': 0.8970467984676361, 'eval_reward_repeat_soft_std': 0.07692538686096669, 'eval_reward_judge_quality_mean': 0.15700000524520874, 'eval_reward_judge_quality_std': 0.02237090365961194, 'eval_reward_total_composite_mean': 0.4048567622900009, 'eval_reward_total_composite_std': 0.0588286342099309, 'epoch': 0.22}

 65%|██████▌   | 2150/3300 [5:17:39<2:17:00,  7.15s/it]

100%|██████████| 10/10 [00:42<00:00,  4.44s/it]

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

 65%|██████▌   | 2151/3300 [5:17:49<7:01:23, 22.00s/it]2026-04-13 12:38:47,534 | INFO | train_grpo_train | metrics_logged mode=train step=2151

                                                       
{'loss': 0.0412, 'grad_norm': 7.0074052810668945, 'learning_rate': 3.4848484848484854e-06, 'num_tokens': 3870707.0, 'completions/mean_length': 108.75, 'completions/min_length': 96.0, 'completions/max_length': 118.0, 'completions/clipped_ratio': 0.0, 'completions/mean_terminated_length': 108.75, 'completions/min_terminated_length': 96.0, 'completions/max_terminated_length': 118.0, 'rewards/meter/mean': 0.9364088773727417, 'rewards/meter/std': 0.06607461720705032, 'rewards/count_adherence/mean': 1.0, 'rewards/count_adherence/std': 0.0, 'rewards/hard_gate/mean': 1.0, 'rewards/hard_gate/std': 0.0, 'rewards/repeat_soft/mean': 0.8666599988937378, 'rewards/repeat_soft/std': 0.04671601206064224, 'rewards/judge_quality/mean': 0.1612500101327896, 'rewards/judge_quality/std': 0.02748376689851284, 'rewards/total_composite/mean': 0.42774778604507446, 'rewards/total_composite/std': 0.020854637026786804, 'reward': 0.42774778604507446, 'reward_std': 0.020854640752077103, 'frac_reward_zero_std': 0.0, 'sampling/sampling_logp_difference/mean': 0.092539943754673, 'sampling/sampling_logp_difference/max': 3.6278061866760254, 'sampling/importance_sampling_ratio/min': 0.02657441981136799, 'sampling/importance_sampling_ratio/mean': 1.0023683309555054, 'sampling/importance_sampling_ratio/max': 2.0, 'entropy': 0.3217124678194523, 'clip_ratio/low_mean': 0.030726388096809387, 'clip_ratio/low_min': 0.030726388096809387, 'clip_ratio/high_mean': 0.05013436125591397, 'clip_ratio/high_max': 0.05013436125591397, 'clip_ratio/region_mean': 0.08086074935272336, 'reward_total_mean': 0.42774778604507446, 'reward_meter_mean': 0.9364088773727417, 'reward_meter_std': 0.06607461720705032, 'reward_count_adherence_mean': 1.0, 'reward_count_adherence_std': 0.0, 'reward_hard_gate_mean': 1.0, 'reward_hard_gate_std': 0.0, 'reward_repeat_soft_mean': 0.8666599988937378, 'reward_repeat_soft_std': 0.04671601206064224, 'reward_judge_quality_mean': 0.1612500101327896, 'reward_judge_quality_std': 0.02748376689851284, 'reward_total_composite_mean': 0.42774778604507446, 'reward_total_composite_std': 0.020854637026786804, 'epoch': 0.22}

 65%|██████▌   | 2151/3300 [5:17:49<7:01:23, 22.00s/it]INFO 04-13 12:38:47 [block_pool.py:378] Successfully reset prefix cache