{
  "submission": "team12_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team12",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          85,
          18
        ],
        [
          11,
          48
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 85,
          "pred_FALSE": 18
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 11,
          "pred_FALSE": 48
        }
      ],
      "accuracy": 0.821,
      "per_label": {
        "TRUE": {
          "precision": 0.8854,
          "recall": 0.8252,
          "f1": 0.8542,
          "support": 103,
          "true_positive": 85,
          "false_positive": 11,
          "false_negative": 18
        },
        "FALSE": {
          "precision": 0.7273,
          "recall": 0.8136,
          "f1": 0.768,
          "support": 59,
          "true_positive": 48,
          "false_positive": 18,
          "false_negative": 11
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          42,
          23
        ],
        [
          6,
          91
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 42,
          "pred_FALSE": 23
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 6,
          "pred_FALSE": 91
        }
      ],
      "accuracy": 0.821,
      "per_label": {
        "TRUE": {
          "precision": 0.875,
          "recall": 0.6462,
          "f1": 0.7434,
          "support": 65,
          "true_positive": 42,
          "false_positive": 6,
          "false_negative": 23
        },
        "FALSE": {
          "precision": 0.7982,
          "recall": 0.9381,
          "f1": 0.8625,
          "support": 97,
          "true_positive": 91,
          "false_positive": 23,
          "false_negative": 6
        }
      }
    }
  }
}