{
  "submission": "team11_HIPE-2026-v1.0-impresso-test-de_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team11",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run1",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          68,
          24
        ],
        [
          9,
          137
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 68,
          "pred_FALSE": 24
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 9,
          "pred_FALSE": 137
        }
      ],
      "accuracy": 0.8613,
      "per_label": {
        "TRUE": {
          "precision": 0.8831,
          "recall": 0.7391,
          "f1": 0.8047,
          "support": 92,
          "true_positive": 68,
          "false_positive": 9,
          "false_negative": 24
        },
        "FALSE": {
          "precision": 0.8509,
          "recall": 0.9384,
          "f1": 0.8925,
          "support": 146,
          "true_positive": 137,
          "false_positive": 24,
          "false_negative": 9
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          14,
          53
        ],
        [
          0,
          171
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 14,
          "pred_FALSE": 53
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 0,
          "pred_FALSE": 171
        }
      ],
      "accuracy": 0.7773,
      "per_label": {
        "TRUE": {
          "precision": 1.0,
          "recall": 0.209,
          "f1": 0.3457,
          "support": 67,
          "true_positive": 14,
          "false_positive": 0,
          "false_negative": 53
        },
        "FALSE": {
          "precision": 0.7634,
          "recall": 1.0,
          "f1": 0.8658,
          "support": 171,
          "true_positive": 171,
          "false_positive": 53,
          "false_negative": 0
        }
      }
    }
  }
}