{
  "scope": "914 fixed local holdout examples; new focus splits and reused retention split. Not official Decision Index scores.",
  "evaluated_format": "Full merged BF16 release, using native label-token argmax",
  "base_revision": "f446b82727be57da348bb46eccf211414294ab3e",
  "candidate_step": 250,
  "released_v61": {
    "objective": 0.7297131700522546,
    "families": {
      "retention": {
        "n": 500,
        "accuracy": 0.934,
        "nll": 0.2194680695671052,
        "metric": 0.934,
        "metric_name": "accuracy"
      },
      "finance": {
        "n": 180,
        "accuracy": 0.7111111111111111,
        "nll": 0.7504884215008701,
        "metric": 0.7116466613734374,
        "metric_name": "macro-F1"
      },
      "banking": {
        "n": 154,
        "accuracy": 0.7467532467532467,
        "nll": 1.2402327169109493,
        "metric": 0.7467532467532467,
        "metric_name": "accuracy"
      },
      "sarcasm": {
        "n": 80,
        "accuracy": 0.6875,
        "nll": 0.6216465693736991,
        "metric": 0.46808510638297873,
        "metric_name": "positive-class F1"
      }
    }
  },
  "selected_adapter": {
    "objective": 0.7384552484361269,
    "families": {
      "retention": {
        "n": 500,
        "accuracy": 0.934,
        "nll": 0.21906396901960715,
        "metric": 0.934,
        "metric_name": "accuracy"
      },
      "finance": {
        "n": 180,
        "accuracy": 0.7166666666666667,
        "nll": 0.7359384456556678,
        "metric": 0.7171978179730117,
        "metric_name": "macro-F1"
      },
      "banking": {
        "n": 154,
        "accuracy": 0.7467532467532467,
        "nll": 1.2267801923755717,
        "metric": 0.7467532467532467,
        "metric_name": "accuracy"
      },
      "sarcasm": {
        "n": 80,
        "accuracy": 0.725,
        "nll": 0.6173049414476395,
        "metric": 0.5,
        "metric_name": "positive-class F1"
      }
    }
  },
  "merged_v62": {
    "objective": 0.7372740069451986,
    "families": {
      "retention": {
        "n": 500,
        "accuracy": 0.934,
        "nll": 0.21870660245549345,
        "metric": 0.934,
        "metric_name": "accuracy"
      },
      "finance": {
        "n": 180,
        "accuracy": 0.7111111111111111,
        "nll": 0.7353250763802403,
        "metric": 0.7121353544404619,
        "metric_name": "macro-F1"
      },
      "banking": {
        "n": 154,
        "accuracy": 0.7467532467532467,
        "nll": 1.2266273153352905,
        "metric": 0.7467532467532467,
        "metric_name": "accuracy"
      },
      "sarcasm": {
        "n": 80,
        "accuracy": 0.725,
        "nll": 0.6158682200969627,
        "metric": 0.5,
        "metric_name": "positive-class F1"
      }
    }
  },
  "test_sha256": "d714570cac1d09849f1aea8a61ccb50a09cb9b5ac10425a7b18f0e89fd7dfc08",
  "limitations": [
    "Exact content/group exclusions do not establish semantic or pretraining decontamination.",
    "Focus holdouts exclude previous local inputs; retention holdouts are reused.",
    "SEntFiN measures entity-level sentiment transfer, not FinEntity benchmark performance.",
    "Benchmarks informed target selection; final benchmark reruns are development evaluation.",
    "No synthetic paraphrases or benchmark test labels are used as training supervision."
  ],
  "official_decision_index": null
}