{
  "submission": "team3_HIPE-2026-v1.0-impresso-test-de_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-de.jsonl",
  "team": "team3",
  "dataset": "impresso",
  "split": "test",
  "language": "de",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          36,
          0,
          18
        ],
        [
          12,
          1,
          25
        ],
        [
          15,
          0,
          131
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 36,
          "pred_PROBABLE": 0,
          "pred_FALSE": 18
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 12,
          "pred_PROBABLE": 1,
          "pred_FALSE": 25
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 15,
          "pred_PROBABLE": 0,
          "pred_FALSE": 131
        }
      ],
      "accuracy": 0.7059,
      "per_label": {
        "TRUE": {
          "precision": 0.5714,
          "recall": 0.6667,
          "f1": 0.6154,
          "support": 54,
          "true_positive": 36,
          "false_positive": 27,
          "false_negative": 18
        },
        "PROBABLE": {
          "precision": 1.0,
          "recall": 0.0263,
          "f1": 0.0513,
          "support": 38,
          "true_positive": 1,
          "false_positive": 0,
          "false_negative": 37
        },
        "FALSE": {
          "precision": 0.7529,
          "recall": 0.8973,
          "f1": 0.8188,
          "support": 146,
          "true_positive": 131,
          "false_positive": 43,
          "false_negative": 15
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          17,
          50
        ],
        [
          6,
          165
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 238,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 17,
          "pred_FALSE": 50
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 6,
          "pred_FALSE": 165
        }
      ],
      "accuracy": 0.7647,
      "per_label": {
        "TRUE": {
          "precision": 0.7391,
          "recall": 0.2537,
          "f1": 0.3777,
          "support": 67,
          "true_positive": 17,
          "false_positive": 6,
          "false_negative": 50
        },
        "FALSE": {
          "precision": 0.7674,
          "recall": 0.9649,
          "f1": 0.8549,
          "support": 171,
          "true_positive": 165,
          "false_positive": 50,
          "false_negative": 6
        }
      }
    }
  }
}