{
  "model_name": "span_deberta_v3_small_sliding_windows",
  "base_model_name": "microsoft/deberta-v3-small",
  "best_checkpoint_dir": "/content/drive/MyDrive/TLLM/03_models/span_deberta_v3_small_sliding_windows/best_checkpoint",
  "last_checkpoint_dir": "/content/drive/MyDrive/TLLM/03_models/span_deberta_v3_small_sliding_windows/last_checkpoint",
  "best_epoch": 8,
  "selected_config": {
    "threshold": 0.5,
    "min_span_chars": 1,
    "min_span_tokens": 1,
    "merge_gap_chars": 1,
    "strip_predicted_span_whitespace": true,
    "drop_spans_without_alnum": true,
    "score_name": "sum_non_O_probability"
  },
  "labels": [
    "O",
    "tool_output_conflict",
    "overgeneration",
    "missing_tool_action_recommendation"
  ],
  "label2id": {
    "O": 0,
    "tool_output_conflict": 1,
    "overgeneration": 2,
    "missing_tool_action_recommendation": 3
  },
  "id2label": {
    "0": "O",
    "1": "tool_output_conflict",
    "2": "overgeneration",
    "3": "missing_tool_action_recommendation"
  },
  "final_validation_token_metrics": {
    "O": {
      "precision": 0.9997869091167383,
      "recall": 0.9987582487759881,
      "f1": 0.9992723142182701,
      "support": 28186,
      "tp": 28151,
      "fp": 6,
      "fn": 35
    },
    "tool_output_conflict": {
      "precision": 0.7666666666666667,
      "recall": 0.9504132231404959,
      "f1": 0.8487084870848709,
      "support": 121,
      "tp": 115,
      "fp": 35,
      "fn": 6
    },
    "overgeneration": {
      "precision": 1.0,
      "recall": 1.0,
      "f1": 1.0,
      "support": 428,
      "tp": 428,
      "fp": 0,
      "fn": 0
    },
    "missing_tool_action_recommendation": {
      "precision": 1.0,
      "recall": 1.0,
      "f1": 1.0,
      "support": 892,
      "tp": 892,
      "fp": 0,
      "fn": 0
    },
    "accuracy": 0.9986161271812873,
    "macro_f1": 0.9619952003257852,
    "weighted_f1": 0.9986898158940638,
    "support": 29627
  },
  "final_test_token_metrics": {
    "O": {
      "precision": 0.9994592359321818,
      "recall": 0.9951225692025084,
      "f1": 0.9972861881258828,
      "support": 31574,
      "tp": 31420,
      "fp": 17,
      "fn": 154
    },
    "tool_output_conflict": {
      "precision": 0.4188679245283019,
      "recall": 0.8671875,
      "f1": 0.564885496183206,
      "support": 128,
      "tp": 111,
      "fp": 154,
      "fn": 17
    },
    "overgeneration": {
      "precision": 1.0,
      "recall": 1.0,
      "f1": 1.0,
      "support": 406,
      "tp": 406,
      "fp": 0,
      "fn": 0
    },
    "missing_tool_action_recommendation": {
      "precision": 1.0,
      "recall": 1.0,
      "f1": 1.0,
      "support": 925,
      "tp": 925,
      "fp": 0,
      "fn": 0
    },
    "accuracy": 0.9948233584597221,
    "macro_f1": 0.8905429210772722,
    "weighted_f1": 0.9957200208094352,
    "support": 33033
  },
  "final_validation_metrics": {
    "config": {
      "threshold": 0.5,
      "min_span_chars": 1,
      "min_span_tokens": 1,
      "merge_gap_chars": 1,
      "strip_predicted_span_whitespace": true,
      "drop_spans_without_alnum": true,
      "score_name": "sum_non_O_probability"
    },
    "num_rows": 214,
    "row_multiclass_metrics": {
      "clean": {
        "precision": 1.0,
        "recall": 0.9611650485436893,
        "f1": 0.9801980198019802,
        "support": 103,
        "tp": 99,
        "fp": 0,
        "fn": 4
      },
      "tool_output_conflict": {
        "precision": 0.9024390243902439,
        "recall": 1.0,
        "f1": 0.9487179487179488,
        "support": 37,
        "tp": 37,
        "fp": 4,
        "fn": 0
      },
      "overgeneration": {
        "precision": 1.0,
        "recall": 1.0,
        "f1": 1.0,
        "support": 37,
        "tp": 37,
        "fp": 0,
        "fn": 0
      },
      "missing_tool_action_recommendation": {
        "precision": 1.0,
        "recall": 1.0,
        "f1": 1.0,
        "support": 37,
        "tp": 37,
        "fp": 0,
        "fn": 0
      },
      "accuracy": 0.9813084112149533,
      "macro_f1": 0.9822289921299823,
      "weighted_f1": 0.9816026174867667,
      "support": 214
    },
    "binary_example_metrics": {
      "accuracy": 0.9813084112149533,
      "precision": 0.9652173913043478,
      "recall": 1.0,
      "f1": 0.9823008849557522,
      "tp": 111,
      "fp": 4,
      "fn": 0,
      "tn": 99
    },
    "exact_span_metrics": {
      "precision": 0.8333333333333334,
      "recall": 0.9009009009009009,
      "f1": 0.8658008658008659,
      "matched": 100,
      "gold_total": 111,
      "pred_total": 120
    },
    "overlap_span_metrics_iou_0_01": {
      "precision": 0.925,
      "recall": 1.0,
      "f1": 0.961038961038961,
      "matched": 111,
      "gold_total": 111,
      "pred_total": 120,
      "iou_threshold": 0.01
    },
    "overlap_span_metrics_iou_0_50": {
      "precision": 0.9083333333333333,
      "recall": 0.9819819819819819,
      "f1": 0.9437229437229437,
      "matched": 109,
      "gold_total": 111,
      "pred_total": 120,
      "iou_threshold": 0.5
    },
    "char_micro_metrics": {
      "precision": 0.9916501556750636,
      "recall": 0.9984326018808778,
      "f1": 0.9950298210735586,
      "overlap_chars": 7007,
      "gold_chars": 7018,
      "pred_chars": 7066
    },
    "per_type_char_micro_metrics": {
      "clean": {
        "precision": 0.0,
        "recall": 0.0,
        "f1": 0.0,
        "overlap_chars": 0,
        "gold_chars": 0,
        "pred_chars": 18
      },
      "overgeneration": {
        "precision": 0.9948006932409013,
        "recall": 1.0,
        "f1": 0.9973935708079931,
        "overlap_chars": 2296,
        "gold_chars": 2296,
        "pred_chars": 2308
      },
      "tool_output_conflict": {
        "precision": 0.9037656903765691,
        "recall": 0.9515418502202643,
        "f1": 0.927038626609442,
        "overlap_chars": 216,
        "gold_chars": 227,
        "pred_chars": 239
      },
      "missing_tool_action_recommendation": {
        "precision": 0.998666962897134,
        "recall": 1.0,
        "f1": 0.9993330369052912,
        "overlap_chars": 4495,
        "gold_chars": 4495,
        "pred_chars": 4501
      }
    },
    "num_gold_spans": 111,
    "num_predicted_spans": 120
  },
  "final_test_metrics": {
    "config": {
      "threshold": 0.5,
      "min_span_chars": 1,
      "min_span_tokens": 1,
      "merge_gap_chars": 1,
      "strip_predicted_span_whitespace": true,
      "drop_spans_without_alnum": true,
      "score_name": "sum_non_O_probability"
    },
    "num_rows": 207,
    "row_multiclass_metrics": {
      "clean": {
        "precision": 0.9777777777777777,
        "recall": 0.9166666666666666,
        "f1": 0.946236559139785,
        "support": 96,
        "tp": 88,
        "fp": 2,
        "fn": 8
      },
      "tool_output_conflict": {
        "precision": 0.813953488372093,
        "recall": 0.9459459459459459,
        "f1": 0.875,
        "support": 37,
        "tp": 35,
        "fp": 8,
        "fn": 2
      },
      "overgeneration": {
        "precision": 1.0,
        "recall": 1.0,
        "f1": 1.0,
        "support": 37,
        "tp": 37,
        "fp": 0,
        "fn": 0
      },
      "missing_tool_action_recommendation": {
        "precision": 1.0,
        "recall": 1.0,
        "f1": 1.0,
        "support": 37,
        "tp": 37,
        "fp": 0,
        "fn": 0
      },
      "accuracy": 0.9516908212560387,
      "macro_f1": 0.9553091397849462,
      "weighted_f1": 0.9527232351566153,
      "support": 207
    },
    "binary_example_metrics": {
      "accuracy": 0.9516908212560387,
      "precision": 0.9316239316239316,
      "recall": 0.9819819819819819,
      "f1": 0.9561403508771931,
      "tp": 109,
      "fp": 8,
      "fn": 2,
      "tn": 88
    },
    "exact_span_metrics": {
      "precision": 0.7357142857142858,
      "recall": 0.9279279279279279,
      "f1": 0.8207171314741037,
      "matched": 103,
      "gold_total": 111,
      "pred_total": 140
    },
    "overlap_span_metrics_iou_0_01": {
      "precision": 0.7785714285714286,
      "recall": 0.9819819819819819,
      "f1": 0.8685258964143425,
      "matched": 109,
      "gold_total": 111,
      "pred_total": 140,
      "iou_threshold": 0.01
    },
    "overlap_span_metrics_iou_0_50": {
      "precision": 0.75,
      "recall": 0.9459459459459459,
      "f1": 0.8366533864541832,
      "matched": 105,
      "gold_total": 111,
      "pred_total": 140,
      "iou_threshold": 0.5
    },
    "char_micro_metrics": {
      "precision": 0.971356003950896,
      "recall": 0.9955169920462762,
      "f1": 0.9832881016997572,
      "overlap_chars": 6884,
      "gold_chars": 6915,
      "pred_chars": 7087
    },
    "per_type_char_micro_metrics": {
      "clean": {
        "precision": 0.0,
        "recall": 0.0,
        "f1": 0.0,
        "overlap_chars": 0,
        "gold_chars": 0,
        "pred_chars": 72
      },
      "missing_tool_action_recommendation": {
        "precision": 0.9879728843210146,
        "recall": 1.0,
        "f1": 0.9939500604993949,
        "overlap_chars": 4518,
        "gold_chars": 4518,
        "pred_chars": 4573
      },
      "tool_output_conflict": {
        "precision": 0.7251908396946565,
        "recall": 0.8597285067873304,
        "f1": 0.7867494824016563,
        "overlap_chars": 190,
        "gold_chars": 221,
        "pred_chars": 262
      },
      "overgeneration": {
        "precision": 0.998165137614679,
        "recall": 1.0,
        "f1": 0.9990817263544537,
        "overlap_chars": 2176,
        "gold_chars": 2176,
        "pred_chars": 2180
      }
    },
    "num_gold_spans": 111,
    "num_predicted_spans": 140
  }
}