{
  "submission": "team4_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team4",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          99,
          28,
          3
        ],
        [
          44,
          13,
          3
        ],
        [
          199,
          75,
          16
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 99,
          "pred_PROBABLE": 28,
          "pred_FALSE": 3
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 44,
          "pred_PROBABLE": 13,
          "pred_FALSE": 3
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 199,
          "pred_PROBABLE": 75,
          "pred_FALSE": 16
        }
      ],
      "accuracy": 0.2667,
      "per_label": {
        "TRUE": {
          "precision": 0.2895,
          "recall": 0.7615,
          "f1": 0.4195,
          "support": 130,
          "true_positive": 99,
          "false_positive": 243,
          "false_negative": 31
        },
        "PROBABLE": {
          "precision": 0.1121,
          "recall": 0.2167,
          "f1": 0.1478,
          "support": 60,
          "true_positive": 13,
          "false_positive": 103,
          "false_negative": 47
        },
        "FALSE": {
          "precision": 0.7273,
          "recall": 0.0552,
          "f1": 0.1026,
          "support": 290,
          "true_positive": 16,
          "false_positive": 6,
          "false_negative": 274
        }
      }
    }
  }
}