{
  "submission": "team8_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team8",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          49,
          20,
          6
        ],
        [
          4,
          16,
          8
        ],
        [
          3,
          2,
          54
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 49,
          "pred_PROBABLE": 20,
          "pred_FALSE": 6
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 4,
          "pred_PROBABLE": 16,
          "pred_FALSE": 8
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 3,
          "pred_PROBABLE": 2,
          "pred_FALSE": 54
        }
      ],
      "accuracy": 0.7346,
      "per_label": {
        "TRUE": {
          "precision": 0.875,
          "recall": 0.6533,
          "f1": 0.7481,
          "support": 75,
          "true_positive": 49,
          "false_positive": 7,
          "false_negative": 26
        },
        "PROBABLE": {
          "precision": 0.4211,
          "recall": 0.5714,
          "f1": 0.4849,
          "support": 28,
          "true_positive": 16,
          "false_positive": 22,
          "false_negative": 12
        },
        "FALSE": {
          "precision": 0.7941,
          "recall": 0.9153,
          "f1": 0.8504,
          "support": 59,
          "true_positive": 54,
          "false_positive": 14,
          "false_negative": 5
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          34,
          31
        ],
        [
          9,
          88
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 34,
          "pred_FALSE": 31
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 9,
          "pred_FALSE": 88
        }
      ],
      "accuracy": 0.7531,
      "per_label": {
        "TRUE": {
          "precision": 0.7907,
          "recall": 0.5231,
          "f1": 0.6296,
          "support": 65,
          "true_positive": 34,
          "false_positive": 9,
          "false_negative": 31
        },
        "FALSE": {
          "precision": 0.7395,
          "recall": 0.9072,
          "f1": 0.8148,
          "support": 97,
          "true_positive": 88,
          "false_positive": 31,
          "false_negative": 9
        }
      }
    }
  }
}