{
  "submission": "team3_HIPE-2026-v1.0-impresso-test-de_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team3",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          52,
          40
        ],
        [
          23,
          123
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 52,
          "pred_FALSE": 40
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 23,
          "pred_FALSE": 123
        }
      ],
      "accuracy": 0.7353,
      "per_label": {
        "TRUE": {
          "precision": 0.6933,
          "recall": 0.5652,
          "f1": 0.6227,
          "support": 92,
          "true_positive": 52,
          "false_positive": 23,
          "false_negative": 40
        },
        "FALSE": {
          "precision": 0.7546,
          "recall": 0.8425,
          "f1": 0.7961,
          "support": 146,
          "true_positive": 123,
          "false_positive": 40,
          "false_negative": 23
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          3,
          64
        ],
        [
          2,
          169
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 3,
          "pred_FALSE": 64
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 2,
          "pred_FALSE": 169
        }
      ],
      "accuracy": 0.7227,
      "per_label": {
        "TRUE": {
          "precision": 0.6,
          "recall": 0.0448,
          "f1": 0.0834,
          "support": 67,
          "true_positive": 3,
          "false_positive": 2,
          "false_negative": 64
        },
        "FALSE": {
          "precision": 0.7253,
          "recall": 0.9883,
          "f1": 0.8366,
          "support": 171,
          "true_positive": 169,
          "false_positive": 64,
          "false_negative": 2
        }
      }
    }
  }
}