{
  "splits": {
    "train": {
      "full_total": 2500,
      "chosen_rule": 1621,
      "rejected_weak_model_actual_error": 1679,
      "err_wrong_type": 487,
      "err_miss": 485,
      "err_parse_fail": 101,
      "dropped_weak_parse_fail": 101,
      "chosen_strong_model": 879,
      "rejected_constructed": 821,
      "err_type_confusion": 168,
      "err_fabricated_evidence": 149,
      "err_false_alarm": 937,
      "err_false_alarm_hard": 173,
      "clean_total": 2399
    },
    "dev": {
      "full_total": 300,
      "chosen_rule": 201,
      "rejected_weak_model_actual_error": 190,
      "err_wrong_type": 56,
      "err_miss": 50,
      "rejected_constructed": 110,
      "err_parse_fail": 16,
      "dropped_weak_parse_fail": 16,
      "err_fabricated_evidence": 23,
      "err_false_alarm": 119,
      "chosen_strong_model": 99,
      "err_false_alarm_hard": 13,
      "err_type_confusion": 23,
      "clean_total": 284
    }
  },
  "notes": {
    "clean_rule": "drop weak_model_actual_error pairs whose rejected_error_type == parse_fail",
    "format": "TRL conversational preference format: prompt/chosen/rejected are message lists"
  }
}