{
  "submission": "team10_HIPE-2026-v1.0-impresso-test-de_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team10",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          18,
          20,
          16
        ],
        [
          11,
          13,
          14
        ],
        [
          11,
          26,
          109
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 18,
          "pred_PROBABLE": 20,
          "pred_FALSE": 16
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 11,
          "pred_PROBABLE": 13,
          "pred_FALSE": 14
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 11,
          "pred_PROBABLE": 26,
          "pred_FALSE": 109
        }
      ],
      "accuracy": 0.5882,
      "per_label": {
        "TRUE": {
          "precision": 0.45,
          "recall": 0.3333,
          "f1": 0.383,
          "support": 54,
          "true_positive": 18,
          "false_positive": 22,
          "false_negative": 36
        },
        "PROBABLE": {
          "precision": 0.2203,
          "recall": 0.3421,
          "f1": 0.268,
          "support": 38,
          "true_positive": 13,
          "false_positive": 46,
          "false_negative": 25
        },
        "FALSE": {
          "precision": 0.7842,
          "recall": 0.7466,
          "f1": 0.7649,
          "support": 146,
          "true_positive": 109,
          "false_positive": 30,
          "false_negative": 37
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          6,
          61
        ],
        [
          3,
          168
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 6,
          "pred_FALSE": 61
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 3,
          "pred_FALSE": 168
        }
      ],
      "accuracy": 0.7311,
      "per_label": {
        "TRUE": {
          "precision": 0.6667,
          "recall": 0.0896,
          "f1": 0.158,
          "support": 67,
          "true_positive": 6,
          "false_positive": 3,
          "false_negative": 61
        },
        "FALSE": {
          "precision": 0.7336,
          "recall": 0.9825,
          "f1": 0.84,
          "support": 171,
          "true_positive": 168,
          "false_positive": 61,
          "false_negative": 3
        }
      }
    }
  }
}