AdithyaSK/repo2rlenv-commit-runtime
收藏资源简介:
repo2rlenv-commit-runtime 是一个由Repo2RLEnv工具生成的强化学习数据集,用于代码和软件工程任务。它将真实的GitHub仓库(包括encode/httpx、gin-gonic/gin等12个开源项目)转换为可验证的强化学习环境。数据集包含52个任务,每个任务基于真实合并的拉取请求或提交,经过质量过滤和信息泄漏去除处理。每个任务提供Harbor格式的规范,包括自然语言指令、黄金补丁(oracle patch)、测试脚本和Dockerfile。奖励信号基于测试执行,通过FAIL_TO_PASS和PASS_TO_PASS测试的通过率计算,用于训练和评估代码生成或修复的AI模型。数据集主要用于强化学习、大语言模型(LLM)训练和软件工程研究,支持端到端的补丁评分和验证。
repo2rlenv-commit-runtime is a reinforcement learning dataset generated by the Repo2RLEnv tool for code and software engineering tasks. It turns real GitHub repositories (including 12 open-source projects such as encode/httpx, gin-gonic/gin, etc.) into verifiable RL environments. The dataset contains 52 tasks, each based on real merged pull requests or commits, with quality filters applied and information-leakage stripped from instruction text. Each task provides a Harbor-shaped specification with natural language instructions, a gold patch (oracle), test scripts, and a Dockerfile. The reward signal is test-execution based, calculated from FAIL_TO_PASS and PASS_TO_PASS test rates, used for training and evaluating AI models for code generation or repair. The dataset is primarily designed for reinforcement learning, large language model (LLM) training, and software engineering research, supporting end-to-end patch scoring and validation.




