{
  "submission": "team7_HIPE-2026-v1.0-impresso-test-en_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team7",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run3",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          9,
          9,
          57
        ],
        [
          2,
          6,
          20
        ],
        [
          5,
          5,
          49
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 9,
          "pred_PROBABLE": 9,
          "pred_FALSE": 57
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 6,
          "pred_FALSE": 20
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 5,
          "pred_FALSE": 49
        }
      ],
      "accuracy": 0.3951,
      "per_label": {
        "TRUE": {
          "precision": 0.5625,
          "recall": 0.12,
          "f1": 0.1978,
          "support": 75,
          "true_positive": 9,
          "false_positive": 7,
          "false_negative": 66
        },
        "PROBABLE": {
          "precision": 0.3,
          "recall": 0.2143,
          "f1": 0.25,
          "support": 28,
          "true_positive": 6,
          "false_positive": 14,
          "false_negative": 22
        },
        "FALSE": {
          "precision": 0.3889,
          "recall": 0.8305,
          "f1": 0.5297,
          "support": 59,
          "true_positive": 49,
          "false_positive": 77,
          "false_negative": 10
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          16,
          49
        ],
        [
          20,
          77
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 16,
          "pred_FALSE": 49
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 20,
          "pred_FALSE": 77
        }
      ],
      "accuracy": 0.5741,
      "per_label": {
        "TRUE": {
          "precision": 0.4444,
          "recall": 0.2462,
          "f1": 0.3169,
          "support": 65,
          "true_positive": 16,
          "false_positive": 20,
          "false_negative": 49
        },
        "FALSE": {
          "precision": 0.6111,
          "recall": 0.7938,
          "f1": 0.6906,
          "support": 97,
          "true_positive": 77,
          "false_positive": 49,
          "false_negative": 20
        }
      }
    }
  }
}