{
  "submission": "team1_HIPE-2026-v1.0-surprise-test-fr_run3.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team1",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          99,
          26,
          5
        ],
        [
          25,
          27,
          8
        ],
        [
          27,
          39,
          224
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 99,
          "pred_PROBABLE": 26,
          "pred_FALSE": 5
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 25,
          "pred_PROBABLE": 27,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 27,
          "pred_PROBABLE": 39,
          "pred_FALSE": 224
        }
      ],
      "accuracy": 0.7292,
      "per_label": {
        "TRUE": {
          "precision": 0.6556,
          "recall": 0.7615,
          "f1": 0.7046,
          "support": 130,
          "true_positive": 99,
          "false_positive": 52,
          "false_negative": 31
        },
        "PROBABLE": {
          "precision": 0.2935,
          "recall": 0.45,
          "f1": 0.3553,
          "support": 60,
          "true_positive": 27,
          "false_positive": 65,
          "false_negative": 33
        },
        "FALSE": {
          "precision": 0.9451,
          "recall": 0.7724,
          "f1": 0.8501,
          "support": 290,
          "true_positive": 224,
          "false_positive": 13,
          "false_negative": 66
        }
      }
    }
  }
}