ST-WebAgentBench
收藏资源简介:
ST-WebAgentBench是由IBM Research Haifa开发的专为评估企业环境中网页代理的安全性和可信度而设计的数据集。该数据集包含234个条目,旨在通过模拟真实的企业交互环境来评估代理的行为。数据集的创建过程结合了现有的WebArena数据集,并引入了新的评估指标,如Completion under Policies (CuP),以全面评估代理的安全性和政策遵循情况。该数据集主要应用于企业环境中,旨在解决网页代理在执行任务时的安全性和可信度问题,确保代理在复杂和动态的环境中能够安全、可靠地操作。
ST-WebAgentBench is a dataset developed by IBM Research Haifa, specifically constructed to evaluate the security and trustworthiness of web agents within enterprise environments. Comprising 234 entries, this dataset is designed to assess agent behaviors by simulating realistic enterprise interaction scenarios. Its development integrates the existing WebArena dataset and introduces novel evaluation metrics such as Completion under Policies (CuP) to comprehensively evaluate agent security and policy adherence. Primarily deployed in enterprise settings, this dataset targets resolving the security and trustworthiness challenges of web agents during task execution, ensuring that agents can operate safely and reliably in complex, dynamic environments.




