{
  "submission": "team8_HIPE-2026-v1.0-impresso-test-de_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team8",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          24,
          5,
          25
        ],
        [
          13,
          17,
          8
        ],
        [
          1,
          6,
          139
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 24,
          "pred_PROBABLE": 5,
          "pred_FALSE": 25
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 13,
          "pred_PROBABLE": 17,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 1,
          "pred_PROBABLE": 6,
          "pred_FALSE": 139
        }
      ],
      "accuracy": 0.7563,
      "per_label": {
        "TRUE": {
          "precision": 0.6316,
          "recall": 0.4444,
          "f1": 0.5217,
          "support": 54,
          "true_positive": 24,
          "false_positive": 14,
          "false_negative": 30
        },
        "PROBABLE": {
          "precision": 0.6071,
          "recall": 0.4474,
          "f1": 0.5152,
          "support": 38,
          "true_positive": 17,
          "false_positive": 11,
          "false_negative": 21
        },
        "FALSE": {
          "precision": 0.8081,
          "recall": 0.9521,
          "f1": 0.8742,
          "support": 146,
          "true_positive": 139,
          "false_positive": 33,
          "false_negative": 7
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          26,
          41
        ],
        [
          4,
          167
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 26,
          "pred_FALSE": 41
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 4,
          "pred_FALSE": 167
        }
      ],
      "accuracy": 0.8109,
      "per_label": {
        "TRUE": {
          "precision": 0.8667,
          "recall": 0.3881,
          "f1": 0.5361,
          "support": 67,
          "true_positive": 26,
          "false_positive": 4,
          "false_negative": 41
        },
        "FALSE": {
          "precision": 0.8029,
          "recall": 0.9766,
          "f1": 0.8813,
          "support": 171,
          "true_positive": 167,
          "false_positive": 41,
          "false_negative": 4
        }
      }
    }
  }
}