{ "evidence_kind": "historical adaptation run, not the current PR head", "source_log_sha256": "0b150d6ea1f13f61d2c4644b435b1575bdb0b8769a6c2b5c215d7be300754a87", "source_log_bytes": 18183197, "source_matches_original_evidence_archive": true, "rounds": 100, "train_steps": [ 0, 99 ], "total_samples": 25600, "response_tokens": 54285777, "ray_status": "SUCCEEDED", "driver_exit_code": 0, "weight_syncs_including_initial": 101, "http_200_counts": { "POST /update_weights": 202, "POST /finish_weight_update": 202, "POST /resume": 202 }, "mean_complete_step_seconds": 990.16, "mean_native_perf_step_seconds": 990.3672755455971, "mean_rollout_seconds": 472.3074756169319, "mean_actor_seconds": 504.27561359882355, "mean_posttrain_sync_seconds": 3.609, "final_posttrain_sync_seconds": 3.6, "grad_norm_min": 0.08432590548633938, "grad_norm_max": 1.2871608694886865, "first10": { "round_start": 1, "round_end": 10, "wall_seconds": 1270.9, "rollout_seconds": 519.7388423204422, "actor_seconds": 740.2825398206711, "raw_reward": -0.1546875, "scored_correct_fraction": 0.42265625, "response_length": 3124.65, "truncated_fraction": 0.5640625, "entropy": 0.6046403586864472 }, "last10": { "round_start": 91, "round_end": 100, "wall_seconds": 868.2, "rollout_seconds": 446.2749132871628, "actor_seconds": 402.93963050842285, "raw_reward": 0.6421875, "scored_correct_fraction": 0.82109375, "response_length": 1675.9515625, "truncated_fraction": 0.119921875, "entropy": 0.2578576445579529 }, "checkpoint_iterations": [ 24, 49, 74, 99 ], "checkpoint_saves": [ { "line": 36944, "seconds": 105.2 }, { "line": 70067, "seconds": 98.1 }, { "line": 100883, "seconds": 96.4 }, { "line": 131612, "seconds": 95.7 } ], "runtime": { "python": "3.12 / aarch64", "device_name_from_npu_smi": "Ascend910", "logical_devices": 16, "physical_cards": 8, "train_devices": 8, "rollout_devices": 8, "torch": "2.10.0+cpu", "torch_npu": "2.10.0.post4", "vllm": "0.30.0+empty", "vllm_ascend": "0.30.0rc1", "CANN": "9.1", "FLA": "0.5.2" }, "configuration": { "TP": 4, "PP": 2, "SP": true, "CP": 1, "GDN_head_TP": true, "runner": "V2", "MTP": false, "KL_penalty": false, "prompts_per_round": 32, "samples_per_prompt": 8, "response_limit": 4096, "prompt_limit": 1024, "sequence_limit": 5120, "recomputation": "full/uniform/1", "cpu_optimizer_offload": true }, "interpretation": [ "Training batches are shuffled; these are not fixed held-out evaluation scores.", "Native perf/update_weights_time can refer to the preceding synchronization. The post-training sync series comes from the final native Timer for each round.", "Native perf/step_time and complete step wall time have different recording boundaries. Complete step wall time uses successive post-training synchronization-end timestamps.", "Native actor_train_tflops is retained in the metrics log, not interpreted as hybrid-GDN MFU.", "The historical snapshot predates the additional input-layernorm TP-gradient fix in PR512 and used the 0.30rc1 HCCL compatibility adapter with unpacked transfers." ] }