{
  "submission": "team10_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team10",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          51,
          14,
          10
        ],
        [
          11,
          9,
          8
        ],
        [
          10,
          23,
          26
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 51,
          "pred_PROBABLE": 14,
          "pred_FALSE": 10
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 11,
          "pred_PROBABLE": 9,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 10,
          "pred_PROBABLE": 23,
          "pred_FALSE": 26
        }
      ],
      "accuracy": 0.5309,
      "per_label": {
        "TRUE": {
          "precision": 0.7083,
          "recall": 0.68,
          "f1": 0.6939,
          "support": 75,
          "true_positive": 51,
          "false_positive": 21,
          "false_negative": 24
        },
        "PROBABLE": {
          "precision": 0.1957,
          "recall": 0.3214,
          "f1": 0.2433,
          "support": 28,
          "true_positive": 9,
          "false_positive": 37,
          "false_negative": 19
        },
        "FALSE": {
          "precision": 0.5909,
          "recall": 0.4407,
          "f1": 0.5049,
          "support": 59,
          "true_positive": 26,
          "false_positive": 18,
          "false_negative": 33
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          23,
          42
        ],
        [
          1,
          96
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 23,
          "pred_FALSE": 42
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 1,
          "pred_FALSE": 96
        }
      ],
      "accuracy": 0.7346,
      "per_label": {
        "TRUE": {
          "precision": 0.9583,
          "recall": 0.3538,
          "f1": 0.5168,
          "support": 65,
          "true_positive": 23,
          "false_positive": 1,
          "false_negative": 42
        },
        "FALSE": {
          "precision": 0.6957,
          "recall": 0.9897,
          "f1": 0.8171,
          "support": 97,
          "true_positive": 96,
          "false_positive": 42,
          "false_negative": 1
        }
      }
    }
  }
}