{
  "submission": "team9_HIPE-2026-v1.0-impresso-test-de_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team9",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          5,
          39,
          10
        ],
        [
          2,
          25,
          11
        ],
        [
          1,
          29,
          116
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 39,
          "pred_FALSE": 10
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 25,
          "pred_FALSE": 11
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 1,
          "pred_PROBABLE": 29,
          "pred_FALSE": 116
        }
      ],
      "accuracy": 0.6134,
      "per_label": {
        "TRUE": {
          "precision": 0.625,
          "recall": 0.0926,
          "f1": 0.1613,
          "support": 54,
          "true_positive": 5,
          "false_positive": 3,
          "false_negative": 49
        },
        "PROBABLE": {
          "precision": 0.2688,
          "recall": 0.6579,
          "f1": 0.3817,
          "support": 38,
          "true_positive": 25,
          "false_positive": 68,
          "false_negative": 13
        },
        "FALSE": {
          "precision": 0.8467,
          "recall": 0.7945,
          "f1": 0.8198,
          "support": 146,
          "true_positive": 116,
          "false_positive": 21,
          "false_negative": 30
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          4,
          63
        ],
        [
          1,
          170
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 4,
          "pred_FALSE": 63
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 1,
          "pred_FALSE": 170
        }
      ],
      "accuracy": 0.7311,
      "per_label": {
        "TRUE": {
          "precision": 0.8,
          "recall": 0.0597,
          "f1": 0.1111,
          "support": 67,
          "true_positive": 4,
          "false_positive": 1,
          "false_negative": 63
        },
        "FALSE": {
          "precision": 0.7296,
          "recall": 0.9942,
          "f1": 0.8416,
          "support": 171,
          "true_positive": 170,
          "false_positive": 63,
          "false_negative": 1
        }
      }
    }
  }
}