{
  "submission": "team10_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team10",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          113,
          16,
          1
        ],
        [
          34,
          24,
          2
        ],
        [
          54,
          146,
          90
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 113,
          "pred_PROBABLE": 16,
          "pred_FALSE": 1
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 34,
          "pred_PROBABLE": 24,
          "pred_FALSE": 2
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 54,
          "pred_PROBABLE": 146,
          "pred_FALSE": 90
        }
      ],
      "accuracy": 0.4729,
      "per_label": {
        "TRUE": {
          "precision": 0.5622,
          "recall": 0.8692,
          "f1": 0.6828,
          "support": 130,
          "true_positive": 113,
          "false_positive": 88,
          "false_negative": 17
        },
        "PROBABLE": {
          "precision": 0.129,
          "recall": 0.4,
          "f1": 0.1951,
          "support": 60,
          "true_positive": 24,
          "false_positive": 162,
          "false_negative": 36
        },
        "FALSE": {
          "precision": 0.9677,
          "recall": 0.3103,
          "f1": 0.4699,
          "support": 290,
          "true_positive": 90,
          "false_positive": 3,
          "false_negative": 200
        }
      }
    }
  }
}