{
  "submission": "team9_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team9",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          60,
          45,
          25
        ],
        [
          9,
          33,
          18
        ],
        [
          14,
          38,
          238
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 60,
          "pred_PROBABLE": 45,
          "pred_FALSE": 25
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 9,
          "pred_PROBABLE": 33,
          "pred_FALSE": 18
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 14,
          "pred_PROBABLE": 38,
          "pred_FALSE": 238
        }
      ],
      "accuracy": 0.6896,
      "per_label": {
        "TRUE": {
          "precision": 0.7229,
          "recall": 0.4615,
          "f1": 0.5634,
          "support": 130,
          "true_positive": 60,
          "false_positive": 23,
          "false_negative": 70
        },
        "PROBABLE": {
          "precision": 0.2845,
          "recall": 0.55,
          "f1": 0.375,
          "support": 60,
          "true_positive": 33,
          "false_positive": 83,
          "false_negative": 27
        },
        "FALSE": {
          "precision": 0.847,
          "recall": 0.8207,
          "f1": 0.8336,
          "support": 290,
          "true_positive": 238,
          "false_positive": 43,
          "false_negative": 52
        }
      }
    }
  }
}