{
  "submission": "team10_HIPE-2026-v1.0-surprise-test-fr_run3.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team10",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          120,
          6,
          4
        ],
        [
          31,
          22,
          7
        ],
        [
          47,
          95,
          148
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 120,
          "pred_PROBABLE": 6,
          "pred_FALSE": 4
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 31,
          "pred_PROBABLE": 22,
          "pred_FALSE": 7
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 47,
          "pred_PROBABLE": 95,
          "pred_FALSE": 148
        }
      ],
      "accuracy": 0.6042,
      "per_label": {
        "TRUE": {
          "precision": 0.6061,
          "recall": 0.9231,
          "f1": 0.7317,
          "support": 130,
          "true_positive": 120,
          "false_positive": 78,
          "false_negative": 10
        },
        "PROBABLE": {
          "precision": 0.1789,
          "recall": 0.3667,
          "f1": 0.2405,
          "support": 60,
          "true_positive": 22,
          "false_positive": 101,
          "false_negative": 38
        },
        "FALSE": {
          "precision": 0.9308,
          "recall": 0.5103,
          "f1": 0.6592,
          "support": 290,
          "true_positive": 148,
          "false_positive": 11,
          "false_negative": 142
        }
      }
    }
  }
}