SETA-Env
收藏资源简介:
SETA-Env 是一个开源、可验证的强化学习终端环境数据集,旨在服务于社区的模型训练与评估。该数据集包含两个顶层子集:SETA_Synth(合成任务)和 SETA_Evolve(终端智能体任务的演化变体),总计提供 4567 个独立的环境实例。每个任务以自包含的 Harbor 风格任务目录形式打包,内含运行任务、构建环境、执行参考解决方案以及运行测试所需的所有文件,典型结构包括 task.toml、instruction.md、environment/、solution/ 和 tests/。数据来源于多个公开社区和技术平台:ask_ubuntu、kaggle_notebook、nl2bash、stack_overflow 和 unix_linux_se。SETA_Synth 子集直接基于这些来源构建,而 SETA_Evolve 子集则对源自这些社区的任务进行了演化处理,生成具有不同上下文或难度的变体以增加多样性和挑战性。该数据集适用于训练和评估在终端(如 Bash)环境中执行代码或完成任务的智能体,是强化学习、代码生成、指令跟随和自动化任务等研究领域的基准资源,采用混合许可证,具体取决于数据来源。
SETA-Env is an open-source, verifiable reinforcement learning terminal environment dataset designed to serve the community for model training and evaluation. The dataset includes two top-level subsets: SETA_Synth (synthetic tasks) and SETA_Evolve (evolved variants of terminal agent tasks), providing a total of 4567 independent environment instances. Each task is packaged as a self-contained Harbor-style task directory, containing all files necessary to run the task, build the environment, execute reference solutions, and run tests, with a typical structure including task.toml, instruction.md, environment/, solution/, and tests/. The data is sourced from multiple public communities and technical platforms: ask_ubuntu, kaggle_notebook, nl2bash, stack_overflow, and unix_linux_se. The SETA_Synth subset is built directly from these sources, while the SETA_Evolve subset involves evolutionary processing of tasks from these communities to generate variants with different contexts (suffix __b1) or difficulties (suffix __d1), enhancing task diversity and challenge. The dataset is suitable for training and evaluating agents that execute code or complete tasks in terminal environments (e.g., Bash), serving as a benchmark resource for research areas such as reinforcement learning, code generation, instruction following, and automated tasks, and it uses mixed licenses depending on the data sources.




