{
  "submission": "team11_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team11",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          70,
          47,
          13
        ],
        [
          7,
          35,
          18
        ],
        [
          6,
          31,
          253
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 70,
          "pred_PROBABLE": 47,
          "pred_FALSE": 13
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 7,
          "pred_PROBABLE": 35,
          "pred_FALSE": 18
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 6,
          "pred_PROBABLE": 31,
          "pred_FALSE": 253
        }
      ],
      "accuracy": 0.7458,
      "per_label": {
        "TRUE": {
          "precision": 0.8434,
          "recall": 0.5385,
          "f1": 0.6573,
          "support": 130,
          "true_positive": 70,
          "false_positive": 13,
          "false_negative": 60
        },
        "PROBABLE": {
          "precision": 0.3097,
          "recall": 0.5833,
          "f1": 0.4046,
          "support": 60,
          "true_positive": 35,
          "false_positive": 78,
          "false_negative": 25
        },
        "FALSE": {
          "precision": 0.8908,
          "recall": 0.8724,
          "f1": 0.8815,
          "support": 290,
          "true_positive": 253,
          "false_positive": 31,
          "false_negative": 37
        }
      }
    }
  }
}