{
  "submission": "team8_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team8",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          170,
          20
        ],
        [
          18,
          272
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 170,
          "pred_FALSE": 20
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 18,
          "pred_FALSE": 272
        }
      ],
      "accuracy": 0.9208,
      "per_label": {
        "TRUE": {
          "precision": 0.9043,
          "recall": 0.8947,
          "f1": 0.8995,
          "support": 190,
          "true_positive": 170,
          "false_positive": 18,
          "false_negative": 20
        },
        "FALSE": {
          "precision": 0.9315,
          "recall": 0.9379,
          "f1": 0.9347,
          "support": 290,
          "true_positive": 272,
          "false_positive": 20,
          "false_negative": 18
        }
      }
    }
  }
}