{
  "submission": "team15_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team15",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          80,
          0,
          50
        ],
        [
          45,
          0,
          15
        ],
        [
          154,
          3,
          133
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 80,
          "pred_PROBABLE": 0,
          "pred_FALSE": 50
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 45,
          "pred_PROBABLE": 0,
          "pred_FALSE": 15
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 154,
          "pred_PROBABLE": 3,
          "pred_FALSE": 133
        }
      ],
      "accuracy": 0.4437,
      "per_label": {
        "TRUE": {
          "precision": 0.2867,
          "recall": 0.6154,
          "f1": 0.3912,
          "support": 130,
          "true_positive": 80,
          "false_positive": 199,
          "false_negative": 50
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 60,
          "true_positive": 0,
          "false_positive": 3,
          "false_negative": 60
        },
        "FALSE": {
          "precision": 0.6717,
          "recall": 0.4586,
          "f1": 0.5451,
          "support": 290,
          "true_positive": 133,
          "false_positive": 65,
          "false_negative": 157
        }
      }
    }
  }
}