{
  "submission": "team4_HIPE-2026-v1.0-impresso-test-de_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team4",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run1",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          71,
          21
        ],
        [
          118,
          28
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 71,
          "pred_FALSE": 21
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 118,
          "pred_FALSE": 28
        }
      ],
      "accuracy": 0.416,
      "per_label": {
        "TRUE": {
          "precision": 0.3757,
          "recall": 0.7717,
          "f1": 0.5054,
          "support": 92,
          "true_positive": 71,
          "false_positive": 118,
          "false_negative": 21
        },
        "FALSE": {
          "precision": 0.5714,
          "recall": 0.1918,
          "f1": 0.2872,
          "support": 146,
          "true_positive": 28,
          "false_positive": 21,
          "false_negative": 118
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          16,
          51
        ],
        [
          41,
          130
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 16,
          "pred_FALSE": 51
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 41,
          "pred_FALSE": 130
        }
      ],
      "accuracy": 0.6134,
      "per_label": {
        "TRUE": {
          "precision": 0.2807,
          "recall": 0.2388,
          "f1": 0.2581,
          "support": 67,
          "true_positive": 16,
          "false_positive": 41,
          "false_negative": 51
        },
        "FALSE": {
          "precision": 0.7182,
          "recall": 0.7602,
          "f1": 0.7386,
          "support": 171,
          "true_positive": 130,
          "false_positive": 51,
          "false_negative": 41
        }
      }
    }
  }
}