{
  "submission": "team7_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team7",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          6,
          10,
          59
        ],
        [
          2,
          7,
          19
        ],
        [
          5,
          5,
          49
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 6,
          "pred_PROBABLE": 10,
          "pred_FALSE": 59
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 2,
          "pred_PROBABLE": 7,
          "pred_FALSE": 19
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 5,
          "pred_PROBABLE": 5,
          "pred_FALSE": 49
        }
      ],
      "accuracy": 0.3827,
      "per_label": {
        "TRUE": {
          "precision": 0.4615,
          "recall": 0.08,
          "f1": 0.1364,
          "support": 75,
          "true_positive": 6,
          "false_positive": 7,
          "false_negative": 69
        },
        "PROBABLE": {
          "precision": 0.3182,
          "recall": 0.25,
          "f1": 0.28,
          "support": 28,
          "true_positive": 7,
          "false_positive": 15,
          "false_negative": 21
        },
        "FALSE": {
          "precision": 0.3858,
          "recall": 0.8305,
          "f1": 0.5269,
          "support": 59,
          "true_positive": 49,
          "false_positive": 78,
          "false_negative": 10
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          14,
          51
        ],
        [
          21,
          76
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 14,
          "pred_FALSE": 51
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 21,
          "pred_FALSE": 76
        }
      ],
      "accuracy": 0.5556,
      "per_label": {
        "TRUE": {
          "precision": 0.4,
          "recall": 0.2154,
          "f1": 0.28,
          "support": 65,
          "true_positive": 14,
          "false_positive": 21,
          "false_negative": 51
        },
        "FALSE": {
          "precision": 0.5984,
          "recall": 0.7835,
          "f1": 0.6786,
          "support": 97,
          "true_positive": 76,
          "false_positive": 51,
          "false_negative": 21
        }
      }
    }
  }
}