{
  "submission": "team7_HIPE-2026-v1.0-surprise-test-fr_run3.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team7",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          7,
          38,
          85
        ],
        [
          3,
          21,
          36
        ],
        [
          13,
          60,
          217
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 7,
          "pred_PROBABLE": 38,
          "pred_FALSE": 85
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 3,
          "pred_PROBABLE": 21,
          "pred_FALSE": 36
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 13,
          "pred_PROBABLE": 60,
          "pred_FALSE": 217
        }
      ],
      "accuracy": 0.5104,
      "per_label": {
        "TRUE": {
          "precision": 0.3043,
          "recall": 0.0538,
          "f1": 0.0914,
          "support": 130,
          "true_positive": 7,
          "false_positive": 16,
          "false_negative": 123
        },
        "PROBABLE": {
          "precision": 0.1765,
          "recall": 0.35,
          "f1": 0.2347,
          "support": 60,
          "true_positive": 21,
          "false_positive": 98,
          "false_negative": 39
        },
        "FALSE": {
          "precision": 0.642,
          "recall": 0.7483,
          "f1": 0.6911,
          "support": 290,
          "true_positive": 217,
          "false_positive": 121,
          "false_negative": 73
        }
      }
    }
  }
}