{
  "submission": "team12_HIPE-2026-v1.0-impresso-test-de_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team12",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          39,
          12,
          3
        ],
        [
          22,
          5,
          11
        ],
        [
          5,
          6,
          135
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 39,
          "pred_PROBABLE": 12,
          "pred_FALSE": 3
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 22,
          "pred_PROBABLE": 5,
          "pred_FALSE": 11
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 6,
          "pred_FALSE": 135
        }
      ],
      "accuracy": 0.7521,
      "per_label": {
        "TRUE": {
          "precision": 0.5909,
          "recall": 0.7222,
          "f1": 0.65,
          "support": 54,
          "true_positive": 39,
          "false_positive": 27,
          "false_negative": 15
        },
        "PROBABLE": {
          "precision": 0.2174,
          "recall": 0.1316,
          "f1": 0.164,
          "support": 38,
          "true_positive": 5,
          "false_positive": 18,
          "false_negative": 33
        },
        "FALSE": {
          "precision": 0.906,
          "recall": 0.9247,
          "f1": 0.9153,
          "support": 146,
          "true_positive": 135,
          "false_positive": 14,
          "false_negative": 11
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          45,
          22
        ],
        [
          7,
          164
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 45,
          "pred_FALSE": 22
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 7,
          "pred_FALSE": 164
        }
      ],
      "accuracy": 0.8782,
      "per_label": {
        "TRUE": {
          "precision": 0.8654,
          "recall": 0.6716,
          "f1": 0.7563,
          "support": 67,
          "true_positive": 45,
          "false_positive": 7,
          "false_negative": 22
        },
        "FALSE": {
          "precision": 0.8817,
          "recall": 0.9591,
          "f1": 0.9188,
          "support": 171,
          "true_positive": 164,
          "false_positive": 22,
          "false_negative": 7
        }
      }
    }
  }
}