{
  "submission": "team12_HIPE-2026-v1.0-surprise-test-fr_run2.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team12",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          115,
          4,
          11
        ],
        [
          16,
          10,
          34
        ],
        [
          12,
          3,
          275
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 115,
          "pred_PROBABLE": 4,
          "pred_FALSE": 11
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 16,
          "pred_PROBABLE": 10,
          "pred_FALSE": 34
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 12,
          "pred_PROBABLE": 3,
          "pred_FALSE": 275
        }
      ],
      "accuracy": 0.8333,
      "per_label": {
        "TRUE": {
          "precision": 0.8042,
          "recall": 0.8846,
          "f1": 0.8425,
          "support": 130,
          "true_positive": 115,
          "false_positive": 28,
          "false_negative": 15
        },
        "PROBABLE": {
          "precision": 0.5882,
          "recall": 0.1667,
          "f1": 0.2598,
          "support": 60,
          "true_positive": 10,
          "false_positive": 7,
          "false_negative": 50
        },
        "FALSE": {
          "precision": 0.8594,
          "recall": 0.9483,
          "f1": 0.9017,
          "support": 290,
          "true_positive": 275,
          "false_positive": 45,
          "false_negative": 15
        }
      }
    }
  }
}