{
  "submission": "team17_HIPE-2026-v1.0-surprise-test-fr_run3.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team17",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          30,
          2,
          98
        ],
        [
          4,
          1,
          55
        ],
        [
          13,
          2,
          275
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 30,
          "pred_PROBABLE": 2,
          "pred_FALSE": 98
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 4,
          "pred_PROBABLE": 1,
          "pred_FALSE": 55
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 13,
          "pred_PROBABLE": 2,
          "pred_FALSE": 275
        }
      ],
      "accuracy": 0.6375,
      "per_label": {
        "TRUE": {
          "precision": 0.6383,
          "recall": 0.2308,
          "f1": 0.339,
          "support": 130,
          "true_positive": 30,
          "false_positive": 17,
          "false_negative": 100
        },
        "PROBABLE": {
          "precision": 0.2,
          "recall": 0.0167,
          "f1": 0.0308,
          "support": 60,
          "true_positive": 1,
          "false_positive": 4,
          "false_negative": 59
        },
        "FALSE": {
          "precision": 0.6425,
          "recall": 0.9483,
          "f1": 0.766,
          "support": 290,
          "true_positive": 275,
          "false_positive": 153,
          "false_negative": 15
        }
      }
    }
  }
}