{
  "submission": "baseline_HIPE-2026-v1.0-impresso-test-en_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "baseline",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          42,
          7,
          26
        ],
        [
          15,
          4,
          9
        ],
        [
          9,
          7,
          43
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 42,
          "pred_PROBABLE": 7,
          "pred_FALSE": 26
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 15,
          "pred_PROBABLE": 4,
          "pred_FALSE": 9
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 9,
          "pred_PROBABLE": 7,
          "pred_FALSE": 43
        }
      ],
      "accuracy": 0.5494,
      "per_label": {
        "TRUE": {
          "precision": 0.6364,
          "recall": 0.56,
          "f1": 0.5958,
          "support": 75,
          "true_positive": 42,
          "false_positive": 24,
          "false_negative": 33
        },
        "PROBABLE": {
          "precision": 0.2222,
          "recall": 0.1429,
          "f1": 0.1739,
          "support": 28,
          "true_positive": 4,
          "false_positive": 14,
          "false_negative": 24
        },
        "FALSE": {
          "precision": 0.5513,
          "recall": 0.7288,
          "f1": 0.6277,
          "support": 59,
          "true_positive": 43,
          "false_positive": 35,
          "false_negative": 16
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          24,
          41
        ],
        [
          10,
          87
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 24,
          "pred_FALSE": 41
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 10,
          "pred_FALSE": 87
        }
      ],
      "accuracy": 0.6852,
      "per_label": {
        "TRUE": {
          "precision": 0.7059,
          "recall": 0.3692,
          "f1": 0.4848,
          "support": 65,
          "true_positive": 24,
          "false_positive": 10,
          "false_negative": 41
        },
        "FALSE": {
          "precision": 0.6797,
          "recall": 0.8969,
          "f1": 0.7733,
          "support": 97,
          "true_positive": 87,
          "false_positive": 41,
          "false_negative": 10
        }
      }
    }
  }
}