P论文工作台最新版本 ↗
完整资料 / RAW_GAP

首步raw batch证据缺口

paper-rl/results/credit_raw_batch_evidence_gap_20261001.json

receipt观察/版本:未指定;不以mtime推断2,498 bytes下载完整原文直接读取
{
  "schema": 1,
  "kind": "credit_raw_batch_evidence_gap",
  "observed_at": "2026-10-01T15:30:00+08:00",
  "scope": "Read-only checkpoint inspection; no GPU and no training launch",
  "finding": "global_step_80 data.pt is a small sampler-state artifact and fully_async_state.pt is a small consumed-state artifact; neither retains the first-step response logits, entropy tensors, response masks, or batch rewards needed to reconstruct the B first-step entropy anomaly. Existing trial result.json files retain per-trial outcomes but no training-step assignment or policy entropy. A controlled rerun/instrumented batch capture is required for definitive attribution.",
  "arms": {
    "O": {
      "data_pt": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-O-b16n8-lr1e-6-fe05-p2p025-20260920/ckpt/global_step_80/data.pt",
        "bytes": 7401,
        "sha256": "d252b3d67442ac68dd7938b730a8e9f7780496bb92a346b99f26e09f2b61b51c"
      },
      "fully_async_state": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-O-b16n8-lr1e-6-fe05-p2p025-20260920/ckpt/global_step_80/fully_async_state.pt",
        "bytes": 46893,
        "sha256": "366fb60126e47bac0f3c1c197d1769ab7505c27d56725a18ea0db417eb647cad"
      }
    },
    "B": {
      "data_pt": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-B-b16n8-lr1e-6-fe05-p2p025-20260920/ckpt/global_step_80/data.pt",
        "bytes": 7401,
        "sha256": "228dd55899f57ab61316b27090fbd64ff2c642440971797ee597945835b024ce"
      },
      "fully_async_state": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-B-b16n8-lr1e-6-fe05-p2p025-20260920/ckpt/global_step_80/fully_async_state.pt",
        "bytes": 46893,
        "sha256": "6f49af70e3642c20a5a7e4685ade308ece9b96575d5db1e53bff8c3fa3c4acec"
      }
    },
    "M": {
      "data_pt": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-M-retry1-b16n8-lr1e-6-fe05-p2p025-20260924/ckpt/global_step_80/data.pt",
        "bytes": 7401,
        "sha256": "58a44f219812b891e47260f24a16c824edaece187eb3a359a8c8e78da86ee797"
      },
      "fully_async_state": {
        "path": "rl/runs/skyrl-qwen35-9b/rl9b-cfcredit-clean-M-retry1-b16n8-lr1e-6-fe05-p2p025-20260924/ckpt/global_step_80/fully_async_state.pt",
        "bytes": 46893,
        "sha256": "0ec10f024a02f3a799fdcdd449ddf9d049024b481981c87d3908b3ed12426bc9"
      }
    }
  },
  "decision": "Do not retrofit a causal explanation or change paper claims from these artifacts."
}

公开内容 SHA-256 c058f891c8bceffd67d2f11a4caaee5122cd7ff287693675627165252234a79b

返回资料与接手