{
  "submission": "team16_HIPE-2026-v1.0-impresso-test-en_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team16",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          25,
          0,
          50
        ],
        [
          4,
          0,
          24
        ],
        [
          21,
          0,
          38
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 25,
          "pred_PROBABLE": 0,
          "pred_FALSE": 50
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 4,
          "pred_PROBABLE": 0,
          "pred_FALSE": 24
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 21,
          "pred_PROBABLE": 0,
          "pred_FALSE": 38
        }
      ],
      "accuracy": 0.3889,
      "per_label": {
        "TRUE": {
          "precision": 0.5,
          "recall": 0.3333,
          "f1": 0.4,
          "support": 75,
          "true_positive": 25,
          "false_positive": 25,
          "false_negative": 50
        },
        "PROBABLE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 28,
          "true_positive": 0,
          "false_positive": 0,
          "false_negative": 28
        },
        "FALSE": {
          "precision": 0.3393,
          "recall": 0.6441,
          "f1": 0.4445,
          "support": 59,
          "true_positive": 38,
          "false_positive": 74,
          "false_negative": 21
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          22,
          43
        ],
        [
          28,
          69
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 22,
          "pred_FALSE": 43
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 28,
          "pred_FALSE": 69
        }
      ],
      "accuracy": 0.5617,
      "per_label": {
        "TRUE": {
          "precision": 0.44,
          "recall": 0.3385,
          "f1": 0.3826,
          "support": 65,
          "true_positive": 22,
          "false_positive": 28,
          "false_negative": 43
        },
        "FALSE": {
          "precision": 0.6161,
          "recall": 0.7113,
          "f1": 0.6603,
          "support": 97,
          "true_positive": 69,
          "false_positive": 43,
          "false_negative": 28
        }
      }
    }
  }
}