DI-BENCH
收藏资源简介:
DI-BENCH是一个专注于依赖推断的大规模基准测试数据集,由微软等机构联合开发。该数据集包含581个真实世界的代码仓库,涵盖Python、C#、Rust和JavaScript四种编程语言,分为常规和大规模两个子集,分别包含387和194个仓库。数据集通过GitHub Actions自动化构建,确保了每个仓库的可执行性。DI-BENCH不仅评估模型生成的依赖项与真实依赖项的文本匹配度,还通过执行测试套件来评估代码的可执行性。该数据集旨在解决大型语言模型在生成代码时依赖推断不准确的问题,推动端到端软件合成的进一步发展。
DI-BENCH is a large-scale benchmark dataset focused on dependency inference, co-developed by Microsoft and other institutions. This dataset contains 581 real-world code repositories covering four programming languages: Python, C#, Rust, and JavaScript, and is divided into two subsets: regular and large-scale, with 387 and 194 repositories respectively. The dataset is automatically built via GitHub Actions, ensuring the executability of each repository. DI-BENCH not only evaluates the textual matching degree between the dependencies generated by models and the real dependencies, but also assesses the executability of the code by executing test suites. This dataset aims to solve the problem of inaccurate dependency inference when large language models generate code, and promote the further development of end-to-end software synthesis.
DI-Bench 数据集概述
数据集简介
DI-Bench 是一个用于评估大型语言模型在依赖推断任务上的性能的基准测试数据集。该数据集包含多个编程语言的测试仓库,旨在通过检查项目测试是否通过来评估推断依赖关系的正确性。
数据集下载
- 数据集发布页面:Dataset release page
- 下载后,将
*.tar.gz文件解压到数据目录:.cache/repo-data/{language},其中{language}可以是python、rust、csharp或javascript。
数据集结构
- 每个仓库实例的数据位于
.cache/repo-data/{language}/{instance_id}目录下。
评估方法
- 使用
dibench.eval命令评估推断依赖关系的正确性。 - 需要指定结果目录、仓库实例目录和数据集路径。
文档




