{
  "submission": "team12_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team12",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          113,
          6,
          11
        ],
        [
          18,
          5,
          37
        ],
        [
          12,
          4,
          274
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 113,
          "pred_PROBABLE": 6,
          "pred_FALSE": 11
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 18,
          "pred_PROBABLE": 5,
          "pred_FALSE": 37
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 12,
          "pred_PROBABLE": 4,
          "pred_FALSE": 274
        }
      ],
      "accuracy": 0.8167,
      "per_label": {
        "TRUE": {
          "precision": 0.7902,
          "recall": 0.8692,
          "f1": 0.8278,
          "support": 130,
          "true_positive": 113,
          "false_positive": 30,
          "false_negative": 17
        },
        "PROBABLE": {
          "precision": 0.3333,
          "recall": 0.0833,
          "f1": 0.1333,
          "support": 60,
          "true_positive": 5,
          "false_positive": 10,
          "false_negative": 55
        },
        "FALSE": {
          "precision": 0.8509,
          "recall": 0.9448,
          "f1": 0.8954,
          "support": 290,
          "true_positive": 274,
          "false_positive": 48,
          "false_negative": 16
        }
      }
    }
  }
}