{
  "submission": "team6_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team6",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          2,
          21,
          107
        ],
        [
          1,
          8,
          51
        ],
        [
          1,
          41,
          248
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 21,
          "pred_FALSE": 107
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 1,
          "pred_PROBABLE": 8,
          "pred_FALSE": 51
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 1,
          "pred_PROBABLE": 41,
          "pred_FALSE": 248
        }
      ],
      "accuracy": 0.5375,
      "per_label": {
        "TRUE": {
          "precision": 0.5,
          "recall": 0.0154,
          "f1": 0.0299,
          "support": 130,
          "true_positive": 2,
          "false_positive": 2,
          "false_negative": 128
        },
        "PROBABLE": {
          "precision": 0.1143,
          "recall": 0.1333,
          "f1": 0.1231,
          "support": 60,
          "true_positive": 8,
          "false_positive": 62,
          "false_negative": 52
        },
        "FALSE": {
          "precision": 0.6108,
          "recall": 0.8552,
          "f1": 0.7126,
          "support": 290,
          "true_positive": 248,
          "false_positive": 158,
          "false_negative": 42
        }
      }
    }
  }
}