{
  "submission": "team15_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team15",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "BINARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          89,
          14
        ],
        [
          44,
          15
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 89,
          "pred_FALSE": 14
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 44,
          "pred_FALSE": 15
        }
      ],
      "accuracy": 0.642,
      "per_label": {
        "TRUE": {
          "precision": 0.6692,
          "recall": 0.8641,
          "f1": 0.7543,
          "support": 103,
          "true_positive": 89,
          "false_positive": 44,
          "false_negative": 14
        },
        "FALSE": {
          "precision": 0.5172,
          "recall": 0.2542,
          "f1": 0.3409,
          "support": 59,
          "true_positive": 15,
          "false_positive": 14,
          "false_negative": 44
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          57,
          8
        ],
        [
          76,
          21
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 57,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 76,
          "pred_FALSE": 21
        }
      ],
      "accuracy": 0.4815,
      "per_label": {
        "TRUE": {
          "precision": 0.4286,
          "recall": 0.8769,
          "f1": 0.5758,
          "support": 65,
          "true_positive": 57,
          "false_positive": 76,
          "false_negative": 8
        },
        "FALSE": {
          "precision": 0.7241,
          "recall": 0.2165,
          "f1": 0.3333,
          "support": 97,
          "true_positive": 21,
          "false_positive": 8,
          "false_negative": 76
        }
      }
    }
  }
}