xlangai/CUA-Gym
收藏资源简介:
CUA-Gym是一个用于可验证奖励强化学习(RLVR)的可验证计算机使用代理任务集合。每个任务将自然语言指令与可执行的设置工件以及一个Python奖励函数配对,该函数以编程方式检查任务完成情况。详细信息请参阅论文《CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents》。该版本包含经过必要数据审查后的完整公共CUA-Gym任务集。数据集采用两层Hugging Face布局:1. 在`data/`下的Parquet元数据表,用于快速数据集查看器浏览和程序化过滤;2. 在`artifacts/`下的压缩工件存档,包含原始可执行任务包。这使数据集查看器专注于一个清晰的任务表,同时避免Hub存储库中出现数万个微小文件。数据集旨在用于计算机使用代理和GUI代理、RLVR和程序化奖励设计、合成任务生成、可执行桌面和Web评估以及后训练数据过滤和扩展研究等研究。
CUA-Gym is a collection of verifiable computer-use agent tasks for verifiable reward reinforcement learning (RLVR). Each task pairs natural language instructions with executable setup artifacts and a Python reward function that programmatically verifies task completion. For detailed information, please refer to the paper *CUA-Gym: Scaling Verifiable Training Environments and Tasks for Computer-Use Agents*. This release contains the complete public CUA-Gym task set following necessary data curation. The dataset adopts a two-tier Hugging Face layout: 1. Parquet metadata tables under `data/` for fast dataset viewer browsing and programmatic filtering; 2. Compressed artifact archives under `artifacts/` containing the original executable task packages. This allows the dataset viewer to focus on a clear task table while avoiding tens of thousands of tiny files in the Hugging Face Hub repository. The dataset is intended for research on computer-use and GUI agents, RLVR and programmatic reward design, synthetic task generation, executable desktop and web evaluation, and post-training data filtering and scaling studies.




