This package contains the scripts used to build and evaluate a method-level Java bug detection dataset derived from Defects4J. It supports data extraction from buggy and fixed revisions, preprocessing and deduplication of changed methods, dataset splitting, and evaluation of zero-shot and fine-tuned CodeBERT-Java models. The code was used on 16 Defects4J projects and the resulting dataset contains 2,574 labeled method instances for binary classification as buggy (1) or fixed (0).
Anonymus et al. (2026) studied this question.