{
  "submission": "team17_HIPE-2026-v1.0-impresso-test-de_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team17",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run3",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          29,
          63
        ],
        [
          25,
          121
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 29,
          "pred_FALSE": 63
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 25,
          "pred_FALSE": 121
        }
      ],
      "accuracy": 0.6303,
      "per_label": {
        "TRUE": {
          "precision": 0.537,
          "recall": 0.3152,
          "f1": 0.3972,
          "support": 92,
          "true_positive": 29,
          "false_positive": 25,
          "false_negative": 63
        },
        "FALSE": {
          "precision": 0.6576,
          "recall": 0.8288,
          "f1": 0.7333,
          "support": 146,
          "true_positive": 121,
          "false_positive": 63,
          "false_negative": 25
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          27,
          40
        ],
        [
          60,
          111
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 27,
          "pred_FALSE": 40
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 60,
          "pred_FALSE": 111
        }
      ],
      "accuracy": 0.5798,
      "per_label": {
        "TRUE": {
          "precision": 0.3103,
          "recall": 0.403,
          "f1": 0.3506,
          "support": 67,
          "true_positive": 27,
          "false_positive": 60,
          "false_negative": 40
        },
        "FALSE": {
          "precision": 0.7351,
          "recall": 0.6491,
          "f1": 0.6894,
          "support": 171,
          "true_positive": 111,
          "false_positive": 40,
          "false_negative": 60
        }
      }
    }
  }
}