{
  "submission": "team14_HIPE-2026-v1.0-impresso-test-en_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team14",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          49,
          21,
          5
        ],
        [
          7,
          16,
          5
        ],
        [
          20,
          21,
          18
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 49,
          "pred_PROBABLE": 21,
          "pred_FALSE": 5
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 7,
          "pred_PROBABLE": 16,
          "pred_FALSE": 5
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 20,
          "pred_PROBABLE": 21,
          "pred_FALSE": 18
        }
      ],
      "accuracy": 0.5123,
      "per_label": {
        "TRUE": {
          "precision": 0.6447,
          "recall": 0.6533,
          "f1": 0.649,
          "support": 75,
          "true_positive": 49,
          "false_positive": 27,
          "false_negative": 26
        },
        "PROBABLE": {
          "precision": 0.2759,
          "recall": 0.5714,
          "f1": 0.3721,
          "support": 28,
          "true_positive": 16,
          "false_positive": 42,
          "false_negative": 12
        },
        "FALSE": {
          "precision": 0.6429,
          "recall": 0.3051,
          "f1": 0.4138,
          "support": 59,
          "true_positive": 18,
          "false_positive": 10,
          "false_negative": 41
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          31,
          34
        ],
        [
          14,
          83
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 31,
          "pred_FALSE": 34
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 14,
          "pred_FALSE": 83
        }
      ],
      "accuracy": 0.7037,
      "per_label": {
        "TRUE": {
          "precision": 0.6889,
          "recall": 0.4769,
          "f1": 0.5636,
          "support": 65,
          "true_positive": 31,
          "false_positive": 14,
          "false_negative": 34
        },
        "FALSE": {
          "precision": 0.7094,
          "recall": 0.8557,
          "f1": 0.7757,
          "support": 97,
          "true_positive": 83,
          "false_positive": 34,
          "false_negative": 14
        }
      }
    }
  }
}