{
  "submission": "team17_HIPE-2026-v1.0-impresso-test-en_run2.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team17",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run2",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          52,
          2,
          21
        ],
        [
          17,
          0,
          11
        ],
        [
          18,
          2,
          39
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 52,
          "pred_PROBABLE": 2,
          "pred_FALSE": 21
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 17,
          "pred_PROBABLE": 0,
          "pred_FALSE": 11
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 18,
          "pred_PROBABLE": 2,
          "pred_FALSE": 39
        }
      ],
      "accuracy": 0.5617,
      "per_label": {
        "TRUE": {
          "precision": 0.5977,
          "recall": 0.6933,
          "f1": 0.642,
          "support": 75,
          "true_positive": 52,
          "false_positive": 35,
          "false_negative": 23
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 28,
          "true_positive": 0,
          "false_positive": 4,
          "false_negative": 28
        },
        "FALSE": {
          "precision": 0.5493,
          "recall": 0.661,
          "f1": 0.6,
          "support": 59,
          "true_positive": 39,
          "false_positive": 32,
          "false_negative": 20
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          15,
          50
        ],
        [
          8,
          89
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 15,
          "pred_FALSE": 50
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 8,
          "pred_FALSE": 89
        }
      ],
      "accuracy": 0.642,
      "per_label": {
        "TRUE": {
          "precision": 0.6522,
          "recall": 0.2308,
          "f1": 0.3409,
          "support": 65,
          "true_positive": 15,
          "false_positive": 8,
          "false_negative": 50
        },
        "FALSE": {
          "precision": 0.6403,
          "recall": 0.9175,
          "f1": 0.7542,
          "support": 97,
          "true_positive": 89,
          "false_positive": 50,
          "false_negative": 8
        }
      }
    }
  }
}