jiayuanz3/SWEContextBench
收藏资源简介:
SWE-ContextBench是一个用于评估编程代理(如AI编码系统)在解决新任务时如何有效重用过去经验的基准数据集。它基于多个现有数据集构建,包含1,100个基础任务和376个相关任务,这些任务来自GitHub问题和拉取请求的真实依赖和引用关系。数据集覆盖51个真实GitHub仓库和9种编程语言,旨在研究代理在相似情境下如何有效转移知识。数据集结构包括多个parquet文件,分别存储基础任务、相关任务及其关系。评估方法包括预构建的Docker镜像和公开的评估代码,以支持快速和标准化的评估。数据实例部分详细描述了每个文件的字段和格式。
SWE-ContextBench is a benchmark created to evaluate how well programming agents, such as AI coding systems, can reuse past experience when solving new tasks. It is built on top of existing datasets including SWE-Bench Lite, SWE-Bench Multilingual, and SWE-Bench Verified. The dataset contains 1,100 base tasks along with 376 related tasks that are derived from real dependency and reference relationships among GitHub issues and pull requests. These tasks are organized in a way that groups together problems with shared context, enabling the study of how effectively an agent can transfer knowledge across similar situations. The dataset spans 51 real-world GitHub repositories and covers 9 different programming languages. The dataset structure includes multiple parquet files for base tasks, related tasks, and their relationships. Evaluation methods include pre-built Docker images and publicly available evaluation code to support quick and standardized assessment. The data instances section details the fields and formats of each file.




