{
  "submission": "team3_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team3",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          78,
          25
        ],
        [
          18,
          41
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 78,
          "pred_FALSE": 25
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 18,
          "pred_FALSE": 41
        }
      ],
      "accuracy": 0.7346,
      "per_label": {
        "TRUE": {
          "precision": 0.8125,
          "recall": 0.7573,
          "f1": 0.7839,
          "support": 103,
          "true_positive": 78,
          "false_positive": 18,
          "false_negative": 25
        },
        "FALSE": {
          "precision": 0.6212,
          "recall": 0.6949,
          "f1": 0.656,
          "support": 59,
          "true_positive": 41,
          "false_positive": 25,
          "false_negative": 18
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          14,
          51
        ],
        [
          4,
          93
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 14,
          "pred_FALSE": 51
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 4,
          "pred_FALSE": 93
        }
      ],
      "accuracy": 0.6605,
      "per_label": {
        "TRUE": {
          "precision": 0.7778,
          "recall": 0.2154,
          "f1": 0.3374,
          "support": 65,
          "true_positive": 14,
          "false_positive": 4,
          "false_negative": 51
        },
        "FALSE": {
          "precision": 0.6458,
          "recall": 0.9588,
          "f1": 0.7718,
          "support": 97,
          "true_positive": 93,
          "false_positive": 51,
          "false_negative": 4
        }
      }
    }
  }
}