{
  "submission": "team3_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team3",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          82,
          11,
          37
        ],
        [
          24,
          2,
          34
        ],
        [
          35,
          6,
          249
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 82,
          "pred_PROBABLE": 11,
          "pred_FALSE": 37
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 24,
          "pred_PROBABLE": 2,
          "pred_FALSE": 34
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 35,
          "pred_PROBABLE": 6,
          "pred_FALSE": 249
        }
      ],
      "accuracy": 0.6937,
      "per_label": {
        "TRUE": {
          "precision": 0.5816,
          "recall": 0.6308,
          "f1": 0.6052,
          "support": 130,
          "true_positive": 82,
          "false_positive": 59,
          "false_negative": 48
        },
        "PROBABLE": {
          "precision": 0.1053,
          "recall": 0.0333,
          "f1": 0.0506,
          "support": 60,
          "true_positive": 2,
          "false_positive": 17,
          "false_negative": 58
        },
        "FALSE": {
          "precision": 0.7781,
          "recall": 0.8586,
          "f1": 0.8164,
          "support": 290,
          "true_positive": 249,
          "false_positive": 71,
          "false_negative": 41
        }
      }
    }
  }
}