{
  "submission": "team4_HIPE-2026-v1.0-impresso-test-en_run1.jsonl",
  "reference": "HIPE-2026-v1.0-impresso-test-en.jsonl",
  "team": "team4",
  "dataset": "impresso",
  "split": "test",
  "language": "en",
  "run": "run1",
  "at_label_mode": "TERNARY",
  "confusion_matrices": {
    "at": {
      "labels": [
        "TRUE",
        "PROBABLE",
        "FALSE"
      ],
      "matrix": [
        [
          44,
          30,
          1
        ],
        [
          16,
          12,
          0
        ],
        [
          40,
          19,
          0
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 44,
          "pred_PROBABLE": 30,
          "pred_FALSE": 1
        },
        {
          "gold": "PROBABLE",
          "pred_TRUE": 16,
          "pred_PROBABLE": 12,
          "pred_FALSE": 0
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 40,
          "pred_PROBABLE": 19,
          "pred_FALSE": 0
        }
      ],
      "accuracy": 0.3457,
      "per_label": {
        "TRUE": {
          "precision": 0.44,
          "recall": 0.5867,
          "f1": 0.5029,
          "support": 75,
          "true_positive": 44,
          "false_positive": 56,
          "false_negative": 31
        },
        "PROBABLE": {
          "precision": 0.1967,
          "recall": 0.4286,
          "f1": 0.2696,
          "support": 28,
          "true_positive": 12,
          "false_positive": 49,
          "false_negative": 16
        },
        "FALSE": {
          "precision": 0.0,
          "recall": 0.0,
          "f1": 0.0,
          "support": 59,
          "true_positive": 0,
          "false_positive": 1,
          "false_negative": 59
        }
      }
    },
    "isAt": {
      "labels": [
        "TRUE",
        "FALSE"
      ],
      "matrix": [
        [
          28,
          37
        ],
        [
          42,
          55
        ]
      ],
      "rows": "gold",
      "columns": "prediction",
      "total": 162,
      "table": [
        {
          "gold": "TRUE",
          "pred_TRUE": 28,
          "pred_FALSE": 37
        },
        {
          "gold": "FALSE",
          "pred_TRUE": 42,
          "pred_FALSE": 55
        }
      ],
      "accuracy": 0.5123,
      "per_label": {
        "TRUE": {
          "precision": 0.4,
          "recall": 0.4308,
          "f1": 0.4148,
          "support": 65,
          "true_positive": 28,
          "false_positive": 42,
          "false_negative": 37
        },
        "FALSE": {
          "precision": 0.5978,
          "recall": 0.567,
          "f1": 0.582,
          "support": 97,
          "true_positive": 55,
          "false_positive": 37,
          "false_negative": 42
        }
      }
    }
  }
}