{
  "submission": "team1_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team1",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          103,
          20,
          7
        ],
        [
          32,
          20,
          8
        ],
        [
          27,
          38,
          225
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 103,
          "pred_PROBABLE": 20,
          "pred_FALSE": 7
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 32,
          "pred_PROBABLE": 20,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 27,
          "pred_PROBABLE": 38,
          "pred_FALSE": 225
        }
      ],
      "accuracy": 0.725,
      "per_label": {
        "TRUE": {
          "precision": 0.6358,
          "recall": 0.7923,
          "f1": 0.7055,
          "support": 130,
          "true_positive": 103,
          "false_positive": 59,
          "false_negative": 27
        },
        "PROBABLE": {
          "precision": 0.2564,
          "recall": 0.3333,
          "f1": 0.2898,
          "support": 60,
          "true_positive": 20,
          "false_positive": 58,
          "false_negative": 40
        },
        "FALSE": {
          "precision": 0.9375,
          "recall": 0.7759,
          "f1": 0.8491,
          "support": 290,
          "true_positive": 225,
          "false_positive": 15,
          "false_negative": 65
        }
      }
    }
  }
}