{
  "submission": "team2_HIPE-2026-v1.0-impresso-test-en_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team2",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          49,
          0,
          26
        ],
        [
          16,
          0,
          12
        ],
        [
          27,
          0,
          32
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 49,
          "pred_PROBABLE": 0,
          "pred_FALSE": 26
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 16,
          "pred_PROBABLE": 0,
          "pred_FALSE": 12
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 27,
          "pred_PROBABLE": 0,
          "pred_FALSE": 32
        }
      ],
      "accuracy": 0.5,
      "per_label": {
        "TRUE": {
          "precision": 0.5326,
          "recall": 0.6533,
          "f1": 0.5868,
          "support": 75,
          "true_positive": 49,
          "false_positive": 43,
          "false_negative": 26
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 28,
          "true_positive": 0,
          "false_positive": 0,
          "false_negative": 28
        },
        "FALSE": {
          "precision": 0.4571,
          "recall": 0.5424,
          "f1": 0.4961,
          "support": 59,
          "true_positive": 32,
          "false_positive": 38,
          "false_negative": 27
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          40,
          25
        ],
        [
          36,
          61
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 40,
          "pred_FALSE": 25
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 36,
          "pred_FALSE": 61
        }
      ],
      "accuracy": 0.6235,
      "per_label": {
        "TRUE": {
          "precision": 0.5263,
          "recall": 0.6154,
          "f1": 0.5674,
          "support": 65,
          "true_positive": 40,
          "false_positive": 36,
          "false_negative": 25
        },
        "FALSE": {
          "precision": 0.7093,
          "recall": 0.6289,
          "f1": 0.6667,
          "support": 97,
          "true_positive": 61,
          "false_positive": 25,
          "false_negative": 36
        }
      }
    }
  }
}