SWE-Perf/SWE-Perf

SWE-Perf是一个专为评估大型语言模型在真实代码库中优化代码性能任务而设计的基准数据集。该数据集包含了140个实例，每个实例都来源于一个流行的GitHub仓库中的性能改进pull request。数据集提供了完整的源代码、目标函数和人类专家的解决方案供参考，核心任务是生成一个代码补丁，以减少测试的执行时间而不引入错误。

SWE-Perf is a benchmark specifically designed to evaluate the ability of large language models to optimize code performance in real-world codebases. The dataset consists of 140 instances, each derived from a performance-improving pull request in a popular GitHub repository. It provides the full source code, target functions, and a reference solution by a human expert, with the core task being to generate a code patch that reduces the execution time of tests without introducing bugs.