RepoPeftBench
收藏资源简介:
RepoPeftBench是由滑铁卢大学研究团队构建的面向代码语言模型的仓库级参数高效微调基准数据集,包含604个高质量Python代码仓库。该数据集涵盖约40万条训练任务和12万条测试任务,数据来源于GitHub上采用pytest或unittest框架的活跃项目,并通过时间截点划分了512个分布内仓库和92个分布外仓库以支持时序评估。数据集的创建过程包括严格的仓库筛选、断言任务提取以及基于提交历史的动态轨迹构建,旨在系统评估模型在代码库理解与演化适应方面的能力,解决传统方法在软件持续迭代中面临的上下文窗口受限与适配器过时等核心挑战。
RepoPeftBench is a repository-level parameter-efficient fine-tuning (PEFT) benchmark dataset for code language models, constructed by the research team from the University of Waterloo. It includes 604 high-quality Python code repositories, encompassing approximately 400,000 training tasks and 120,000 test tasks. The dataset's data is sourced from active GitHub projects adopting the pytest or unittest testing frameworks, and it splits the repositories into 512 in-distribution and 92 out-of-distribution repositories based on timestamps to support temporal evaluation. The construction process involves strict repository screening, assertion task extraction, and dynamic trajectory construction based on commit histories. This benchmark is designed to systematically evaluate a model's capabilities in codebase comprehension and evolutionary adaptation, addressing core challenges faced by traditional methods during continuous software iteration, such as limited context window size and outdated adapters.




