GUI-360◦
收藏资源简介:
GUI-360◦是一个大规模、全面的计算机使用代理(CUAs)数据集和基准测试套件,旨在推动桌面环境下的智能代理研究。该数据集包含超过120万个执行的动作步骤,涵盖数千个在流行Windows办公应用程序中的轨迹,包括全分辨率截图、可访问性元数据、实例化目标、中间推理轨迹以及成功和失败的动作轨迹。GUI-360◦支持三个典型任务:GUI定位、屏幕解析和动作预测,以及一个混合GUI+API动作空间,反映了现代代理设计。数据集的创建过程包括查询来源、环境模板构建、任务实例化、批量执行和LLM驱动的质量过滤。该数据集旨在解决桌面CUAs训练和评估中的挑战,并提供了一个可扩展的、具有挑战性的基准测试,以推动智能代理在数字环境中的任务自动化研究。
GUI-360° is a large-scale, comprehensive computer use agent (CUA) dataset and benchmark suite designed to advance intelligent agent research in desktop environments. This dataset contains over 1.2 million executed action steps, covering thousands of trajectories across popular Windows office applications, including full-resolution screenshots, accessibility metadata, instantiated goals, intermediate inference trajectories, as well as successful and failed action trajectories. GUI-360° supports three typical tasks: GUI grounding, screen parsing and action prediction, as well as a hybrid GUI+API action space that reflects modern agent design. The dataset creation process includes query sourcing, environment template construction, task instantiation, batch execution, and LLM-driven quality filtering. This dataset aims to address the challenges in training and evaluating desktop CUAs, and provides a scalable, challenging benchmark to advance research on task automation for intelligent agents in digital environments.




