{
  "submission": "team15_HIPE-2026-v1.0-impresso-test-en_run3.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team15",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run3",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          61,
          42
        ],
        [
          40,
          19
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 61,
          "pred_FALSE": 42
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 40,
          "pred_FALSE": 19
        }
      ],
      "accuracy": 0.4938,
      "per_label": {
        "TRUE": {
          "precision": 0.604,
          "recall": 0.5922,
          "f1": 0.598,
          "support": 103,
          "true_positive": 61,
          "false_positive": 40,
          "false_negative": 42
        },
        "FALSE": {
          "precision": 0.3115,
          "recall": 0.322,
          "f1": 0.3167,
          "support": 59,
          "true_positive": 19,
          "false_positive": 42,
          "false_negative": 40
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          39,
          26
        ],
        [
          45,
          52
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 39,
          "pred_FALSE": 26
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 45,
          "pred_FALSE": 52
        }
      ],
      "accuracy": 0.5617,
      "per_label": {
        "TRUE": {
          "precision": 0.4643,
          "recall": 0.6,
          "f1": 0.5235,
          "support": 65,
          "true_positive": 39,
          "false_positive": 45,
          "false_negative": 26
        },
        "FALSE": {
          "precision": 0.6667,
          "recall": 0.5361,
          "f1": 0.5943,
          "support": 97,
          "true_positive": 52,
          "false_positive": 26,
          "false_negative": 45
        }
      }
    }
  }
}