{
  "submission": "team12_HIPE-2026-v1.0-impresso-test-de_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team12",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          38,
          12,
          4
        ],
        [
          20,
          7,
          11
        ],
        [
          2,
          6,
          138
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 38,
          "pred_PROBABLE": 12,
          "pred_FALSE": 4
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 20,
          "pred_PROBABLE": 7,
          "pred_FALSE": 11
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 6,
          "pred_FALSE": 138
        }
      ],
      "accuracy": 0.7689,
      "per_label": {
        "TRUE": {
          "precision": 0.6333,
          "recall": 0.7037,
          "f1": 0.6666,
          "support": 54,
          "true_positive": 38,
          "false_positive": 22,
          "false_negative": 16
        },
        "PROBABLE": {
          "precision": 0.28,
          "recall": 0.1842,
          "f1": 0.2222,
          "support": 38,
          "true_positive": 7,
          "false_positive": 18,
          "false_negative": 31
        },
        "FALSE": {
          "precision": 0.902,
          "recall": 0.9452,
          "f1": 0.9231,
          "support": 146,
          "true_positive": 138,
          "false_positive": 15,
          "false_negative": 8
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          43,
          24
        ],
        [
          6,
          165
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 43,
          "pred_FALSE": 24
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 6,
          "pred_FALSE": 165
        }
      ],
      "accuracy": 0.8739,
      "per_label": {
        "TRUE": {
          "precision": 0.8776,
          "recall": 0.6418,
          "f1": 0.7414,
          "support": 67,
          "true_positive": 43,
          "false_positive": 6,
          "false_negative": 24
        },
        "FALSE": {
          "precision": 0.873,
          "recall": 0.9649,
          "f1": 0.9167,
          "support": 171,
          "true_positive": 165,
          "false_positive": 24,
          "false_negative": 6
        }
      }
    }
  }
}