AdithyaSK/repo2rlenv-pr-diff

repo2rlenv-pr-diff是一个由Repo2RLEnv工具生成的强化学习环境数据集，专门用于代码补丁生成任务。该数据集从25个流行的开源GitHub仓库（如ripgrep、axios、flask、requests等）中提取了100个真实的已合并拉取请求或提交，并通过pr_diff流程进行处理，移除了指令文本中的信息泄露，并生成了符合Harbor任务格式的目录结构。每个任务包含自然语言指令、黄金补丁（oracle patch）、测试脚本和Dockerfile，可用于训练和评估强化学习代理在代码修改任务上的性能。数据集支持端到端的补丁评分，并包含多组件奖励信号设计，适用于大语言模型和强化学习在软件工程领域的应用研究。repo2rlenv-pr-diff is a reinforcement learning environment dataset generated by the Repo2RLEnv tool, specifically tailored for code patch generation tasks. This dataset extracts 100 real merged pull requests or commits from 25 popular open-source GitHub repositories including ripgrep, axios, flask, and requests, processes these assets via the pr_diff workflow, eliminates information leakage in the instruction text, and generates a directory structure compliant with the Harbor task format. Each task comprises natural language instructions, an oracle patch, test scripts, and a Dockerfile, which can be used to train and evaluate the performance of reinforcement learning agents on code modification tasks. The dataset supports end-to-end patch scoring, incorporates a multi-component reward signal design, and is suitable for applied research on large language models and reinforcement learning in the field of software engineering.