{
  "submission": "team1_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team1",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          61,
          51,
          18
        ],
        [
          14,
          35,
          11
        ],
        [
          48,
          68,
          174
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 61,
          "pred_PROBABLE": 51,
          "pred_FALSE": 18
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 14,
          "pred_PROBABLE": 35,
          "pred_FALSE": 11
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 48,
          "pred_PROBABLE": 68,
          "pred_FALSE": 174
        }
      ],
      "accuracy": 0.5625,
      "per_label": {
        "TRUE": {
          "precision": 0.4959,
          "recall": 0.4692,
          "f1": 0.4822,
          "support": 130,
          "true_positive": 61,
          "false_positive": 62,
          "false_negative": 69
        },
        "PROBABLE": {
          "precision": 0.2273,
          "recall": 0.5833,
          "f1": 0.3271,
          "support": 60,
          "true_positive": 35,
          "false_positive": 119,
          "false_negative": 25
        },
        "FALSE": {
          "precision": 0.8571,
          "recall": 0.6,
          "f1": 0.7059,
          "support": 290,
          "true_positive": 174,
          "false_positive": 29,
          "false_negative": 116
        }
      }
    }
  }
}