DEFECTS4J-TRANS
收藏资源简介:
DEFECTS4J-TRANS数据集是由对Defects4J数据集应用程序变换得到的,变换保持了原始缺陷语义但改变了代码内容。该数据集包含了438个单函数缺陷,用于评估大型语言模型在程序修复任务上的泛化能力。通过对比Defects4J和DEFECTS4J-TRANS上的修复结果,发现模型在新的数据集上性能下降,说明现有模型在面对新的代码变化时存在泛化问题。
The DEFECTS4J-TRANS dataset is constructed by applying program transformations to the Defects4J dataset, where the transformations preserve the original defect semantics while altering the code content. It contains 438 single-function defects, and is designed to evaluate the generalization capability of large language models (LLMs) on program repair tasks. By comparing the repair performance on Defects4J and DEFECTS4J-TRANS, it is found that the model's performance degrades on the new dataset, indicating that existing models have generalization issues when facing novel code variations.

- 1Evaluating the Generalizability of LLMs in Automated Program Repair天津大学智能与计算学院, 重庆大学大数据与软件工程学院 · 2025年



