{
  "submission": "team1_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team1",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          161,
          29
        ],
        [
          116,
          174
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 161,
          "pred_FALSE": 29
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 116,
          "pred_FALSE": 174
        }
      ],
      "accuracy": 0.6979,
      "per_label": {
        "TRUE": {
          "precision": 0.5812,
          "recall": 0.8474,
          "f1": 0.6895,
          "support": 190,
          "true_positive": 161,
          "false_positive": 116,
          "false_negative": 29
        },
        "FALSE": {
          "precision": 0.8571,
          "recall": 0.6,
          "f1": 0.7059,
          "support": 290,
          "true_positive": 174,
          "false_positive": 29,
          "false_negative": 116
        }
      }
    }
  }
}