{
  "submission": "team2_HIPE-2026-v1.0-impresso-test-de_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team2",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          25,
          1,
          28
        ],
        [
          9,
          3,
          26
        ],
        [
          30,
          11,
          105
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 25,
          "pred_PROBABLE": 1,
          "pred_FALSE": 28
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 9,
          "pred_PROBABLE": 3,
          "pred_FALSE": 26
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 30,
          "pred_PROBABLE": 11,
          "pred_FALSE": 105
        }
      ],
      "accuracy": 0.5588,
      "per_label": {
        "TRUE": {
          "precision": 0.3906,
          "recall": 0.463,
          "f1": 0.4237,
          "support": 54,
          "true_positive": 25,
          "false_positive": 39,
          "false_negative": 29
        },
        "PROBABLE": {
          "precision": 0.2,
          "recall": 0.0789,
          "f1": 0.1132,
          "support": 38,
          "true_positive": 3,
          "false_positive": 12,
          "false_negative": 35
        },
        "FALSE": {
          "precision": 0.6604,
          "recall": 0.7192,
          "f1": 0.6885,
          "support": 146,
          "true_positive": 105,
          "false_positive": 54,
          "false_negative": 41
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          27,
          40
        ],
        [
          38,
          133
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 27,
          "pred_FALSE": 40
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 38,
          "pred_FALSE": 133
        }
      ],
      "accuracy": 0.6723,
      "per_label": {
        "TRUE": {
          "precision": 0.4154,
          "recall": 0.403,
          "f1": 0.4091,
          "support": 67,
          "true_positive": 27,
          "false_positive": 38,
          "false_negative": 40
        },
        "FALSE": {
          "precision": 0.7688,
          "recall": 0.7778,
          "f1": 0.7733,
          "support": 171,
          "true_positive": 133,
          "false_positive": 40,
          "false_negative": 38
        }
      }
    }
  }
}