{
  "submission": "team1_HIPE-2026-v1.0-impresso-test-de_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team1",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          77,
          15
        ],
        [
          63,
          83
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 77,
          "pred_FALSE": 15
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 63,
          "pred_FALSE": 83
        }
      ],
      "accuracy": 0.6723,
      "per_label": {
        "TRUE": {
          "precision": 0.55,
          "recall": 0.837,
          "f1": 0.6638,
          "support": 92,
          "true_positive": 77,
          "false_positive": 63,
          "false_negative": 15
        },
        "FALSE": {
          "precision": 0.8469,
          "recall": 0.5685,
          "f1": 0.6803,
          "support": 146,
          "true_positive": 83,
          "false_positive": 15,
          "false_negative": 63
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          13,
          54
        ],
        [
          3,
          168
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 13,
          "pred_FALSE": 54
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 3,
          "pred_FALSE": 168
        }
      ],
      "accuracy": 0.7605,
      "per_label": {
        "TRUE": {
          "precision": 0.8125,
          "recall": 0.194,
          "f1": 0.3132,
          "support": 67,
          "true_positive": 13,
          "false_positive": 3,
          "false_negative": 54
        },
        "FALSE": {
          "precision": 0.7568,
          "recall": 0.9825,
          "f1": 0.855,
          "support": 171,
          "true_positive": 168,
          "false_positive": 54,
          "false_negative": 3
        }
      }
    }
  }
}