{
  "submission": "team3_HIPE-2026-v1.0-surprise-test-fr_run3.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team3",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run3",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          131,
          59
        ],
        [
          24,
          266
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 131,
          "pred_FALSE": 59
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 24,
          "pred_FALSE": 266
        }
      ],
      "accuracy": 0.8271,
      "per_label": {
        "TRUE": {
          "precision": 0.8452,
          "recall": 0.6895,
          "f1": 0.7595,
          "support": 190,
          "true_positive": 131,
          "false_positive": 24,
          "false_negative": 59
        },
        "FALSE": {
          "precision": 0.8185,
          "recall": 0.9172,
          "f1": 0.865,
          "support": 290,
          "true_positive": 266,
          "false_positive": 59,
          "false_negative": 24
        }
      }
    }
  }
}