{
  "submission": "team7_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team7",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          7,
          43,
          80
        ],
        [
          3,
          19,
          38
        ],
        [
          15,
          64,
          211
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 7,
          "pred_PROBABLE": 43,
          "pred_FALSE": 80
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 3,
          "pred_PROBABLE": 19,
          "pred_FALSE": 38
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 15,
          "pred_PROBABLE": 64,
          "pred_FALSE": 211
        }
      ],
      "accuracy": 0.4938,
      "per_label": {
        "TRUE": {
          "precision": 0.28,
          "recall": 0.0538,
          "f1": 0.0903,
          "support": 130,
          "true_positive": 7,
          "false_positive": 18,
          "false_negative": 123
        },
        "PROBABLE": {
          "precision": 0.1508,
          "recall": 0.3167,
          "f1": 0.2043,
          "support": 60,
          "true_positive": 19,
          "false_positive": 107,
          "false_negative": 41
        },
        "FALSE": {
          "precision": 0.6413,
          "recall": 0.7276,
          "f1": 0.6817,
          "support": 290,
          "true_positive": 211,
          "false_positive": 118,
          "false_negative": 79
        }
      }
    }
  }
}