{
  "note": "bash-sft-03",
  "date_published": "2026-08-30",
  "status": "research note, not a paper",
  "model": "Qwen/Qwen2.5-Coder-1.5B-Instruct",
  "revision": "2e1fd397ee46e1388853d2af2c993145b0f1098a",
  "kept_checkpoint": {
    "label": "last40-wd01",
    "init": "harvest-continue",
    "train_rows": 2402,
    "train_source": "GFR train only",
    "epochs": 1.0,
    "learning_rate": 1e-5,
    "lr_scheduler_type": "cosine",
    "weight_decay": 0.1,
    "max_length": 1024,
    "eval_loss": 0.01827716827392578,
    "locked_pass": 92,
    "locked_n": 98,
    "exact_command": 86,
    "format_ok": 1.0,
    "policy_ok": 1.0,
    "syntax_ok": 1.0,
    "filesystem_unchanged": 1.0,
    "failure_reasons": {
      "exit_mismatch": 3,
      "stdout_mismatch": 3
    },
    "unpublished_weights_sha256": "b3558c0f140a659f0c661bf9795ed59049cdfed5db42c1e3e8e8480e8ad901b2",
    "wall_clock_seconds": 57.26157447700001,
    "peak_gpu_memory_gib": 15.215328693389893
  },
  "four_regime": {
    "matched_optimizer_steps": 561,
    "validation_rows": 1030,
    "lowest_validation_loss_arm": "arm-a-full",
    "arms": {
      "arm-a-full": {
        "rows": 8970,
        "eval_loss": 0.7179822325706482,
        "exact": 0.18058252427184465,
        "tf_loss": 0.6618314936946758
      },
      "arm-b-no-trajectories": {
        "rows": 8078,
        "eval_loss": 0.7786956429481506,
        "exact": 0.1786407766990291,
        "tf_loss": 0.7136392770270089
      },
      "arm-c-task-only": {
        "rows": 6770,
        "eval_loss": 0.875102698802948,
        "exact": 0.09514563106796116,
        "tf_loss": 0.86379755875468
      },
      "arm-d-concrete-checks": {
        "rows": 5237,
        "eval_loss": 0.7908467650413513,
        "exact": 0.10194174757281553,
        "tf_loss": 0.7339007938748765
      }
    }
  },
  "locked_eval": {
    "n": 98,
    "note02_arm2_passes": 88,
    "gfr_continue_passes": 91,
    "harvest_continue_passes": 91,
    "harvest_stock_passes": 88,
    "hardfam_continue_passes": 91,
    "hardfam_restarts_passes": 80,
    "linear_from_restarts_passes": 87,
    "last40_wd01_passes": 92,
    "kimi_hardfam_wd01_passes": 92,
    "matched_prompt_last40_passes": 90,
    "matched_prompt_stock_passes": 8,
    "matched_prompt_nl2shell_08b_passes": 4
  },
  "harvest_mix": {
    "gfr_train_rows": 2402,
    "harvest_selected_rows": 1201,
    "train_rows": 3603,
    "val_rows": 297,
    "val_source": "GFR val",
    "locked_eval_in_train": 0,
    "harvest_execution_backed": false,
    "pending_execution_candidates": 218
  },
  "recommendation": {
    "scale_adas_now": false,
    "real_life_test_first": true,
    "reason": "last40 wd=0.1 reached 92/98. Hardfam upsample and the Kimi K3 continue tied or lost. Mix-A val-loss win is a different corpus. Use 92/98 live, then execute-quarantine, then multi-Ada recipes that are not more copies of the same six misses."
  },
  "claim_boundary": {
    "within_generator": true,
    "handoff_canary_passes": 0,
    "fair_baselines_used_identical_effective_messages": true,
    "unpulled_followups_excluded": true,
    "training_jsonl_unpublished": true,
    "kimi_hardfam_tied_not_kept": true
  }
}