{
  "submission": "team2_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team2",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          93,
          8,
          29
        ],
        [
          40,
          6,
          14
        ],
        [
          172,
          39,
          79
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 93,
          "pred_PROBABLE": 8,
          "pred_FALSE": 29
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 40,
          "pred_PROBABLE": 6,
          "pred_FALSE": 14
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 172,
          "pred_PROBABLE": 39,
          "pred_FALSE": 79
        }
      ],
      "accuracy": 0.3708,
      "per_label": {
        "TRUE": {
          "precision": 0.3049,
          "recall": 0.7154,
          "f1": 0.4276,
          "support": 130,
          "true_positive": 93,
          "false_positive": 212,
          "false_negative": 37
        },
        "PROBABLE": {
          "precision": 0.1132,
          "recall": 0.1,
          "f1": 0.1062,
          "support": 60,
          "true_positive": 6,
          "false_positive": 47,
          "false_negative": 54
        },
        "FALSE": {
          "precision": 0.6475,
          "recall": 0.2724,
          "f1": 0.3835,
          "support": 290,
          "true_positive": 79,
          "false_positive": 43,
          "false_negative": 211
        }
      }
    }
  }
}