{
  "submission": "team11_HIPE-2026-v1.0-surprise-test-fr_run1.jsonl",
  "reference": "HIPE-2026-v1.0-surprise-test-fr.jsonl",
  "team": "team11",
  "dataset": "surprise",
  "split": "test",
  "language": "fr",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          66,
          46,
          18
        ],
        [
          5,
          23,
          32
        ],
        [
          2,
          17,
          271
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 480,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 66,
          "pred_PROBABLE": 46,
          "pred_FALSE": 18
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 23,
          "pred_FALSE": 32
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 17,
          "pred_FALSE": 271
        }
      ],
      "accuracy": 0.75,
      "per_label": {
        "TRUE": {
          "precision": 0.9041,
          "recall": 0.5077,
          "f1": 0.6503,
          "support": 130,
          "true_positive": 66,
          "false_positive": 7,
          "false_negative": 64
        },
        "PROBABLE": {
          "precision": 0.2674,
          "recall": 0.3833,
          "f1": 0.315,
          "support": 60,
          "true_positive": 23,
          "false_positive": 63,
          "false_negative": 37
        },
        "FALSE": {
          "precision": 0.8442,
          "recall": 0.9345,
          "f1": 0.8871,
          "support": 290,
          "true_positive": 271,
          "false_positive": 50,
          "false_negative": 19
        }
      }
    }
  }
}