{
  "submission": "team17_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team17",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          48,
          1,
          81
        ],
        [
          14,
          0,
          46
        ],
        [
          29,
          0,
          261
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 48,
          "pred_PROBABLE": 1,
          "pred_FALSE": 81
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 14,
          "pred_PROBABLE": 0,
          "pred_FALSE": 46
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 29,
          "pred_PROBABLE": 0,
          "pred_FALSE": 261
        }
      ],
      "accuracy": 0.6438,
      "per_label": {
        "TRUE": {
          "precision": 0.5275,
          "recall": 0.3692,
          "f1": 0.4344,
          "support": 130,
          "true_positive": 48,
          "false_positive": 43,
          "false_negative": 82
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 60,
          "true_positive": 0,
          "false_positive": 1,
          "false_negative": 60
        },
        "FALSE": {
          "precision": 0.6727,
          "recall": 0.9,
          "f1": 0.7699,
          "support": 290,
          "true_positive": 261,
          "false_positive": 127,
          "false_negative": 29
        }
      }
    }
  }
}