SofiTesfay2010/SWE-Train

SWE-Train是一个提示-答案数据集，专为在仓库级软件工程任务上微调大语言模型（LLMs）而设计。它源自SWE-bench Verified基准——一个由SWE-bench精选的子集，包含经人类软件工程师验证的、具有完整问题描述和正确单元测试的实例。标准SWE-bench评估通常需要复杂的多步代理框架来与终端和环境交互，而SWE-Train将这些已验证的实例重新构建为直接的指令-响应模式（提示→答案），使其与标准监督微调（SFT）流程兼容。该数据集具有人类验证的质量、微调就绪的特性，并涵盖来自主要开源Python项目（如matplotlib、sympy、scikit-learn等）的真实世界编码问题。

SWE-Train is a prompt-answer dataset designed for fine-tuning large language models (LLMs) on repository-level software engineering tasks. It is derived from the SWE-bench Verified benchmark—a curated subset of SWE-bench containing instances that human software engineers verified as having complete problem descriptions and correct unit tests. Standard SWE-bench evaluation typically requires complex, multi-step agent frameworks to interact with a terminal and environment. SWE-Train reformulates these verified instances into a direct instruction-response schema (Prompt → Answer), making it compatible with standard Supervised Fine-Tuning (SFT) pipelines.