{
  "submission": "team6_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team6",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          80,
          6,
          44
        ],
        [
          36,
          2,
          22
        ],
        [
          141,
          13,
          136
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 80,
          "pred_PROBABLE": 6,
          "pred_FALSE": 44
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 36,
          "pred_PROBABLE": 2,
          "pred_FALSE": 22
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 141,
          "pred_PROBABLE": 13,
          "pred_FALSE": 136
        }
      ],
      "accuracy": 0.4542,
      "per_label": {
        "TRUE": {
          "precision": 0.3113,
          "recall": 0.6154,
          "f1": 0.4135,
          "support": 130,
          "true_positive": 80,
          "false_positive": 177,
          "false_negative": 50
        },
        "PROBABLE": {
          "precision": 0.0952,
          "recall": 0.0333,
          "f1": 0.0493,
          "support": 60,
          "true_positive": 2,
          "false_positive": 19,
          "false_negative": 58
        },
        "FALSE": {
          "precision": 0.6733,
          "recall": 0.469,
          "f1": 0.5529,
          "support": 290,
          "true_positive": 136,
          "false_positive": 66,
          "false_negative": 154
        }
      }
    }
  }
}