ZhuOnR/guiact_websingle_test

GUIAct Web-Single数据集是一个全面的单步动作指令集合，用于网站GUI导航任务。它包含大约67,000个指令-动作对，每个对由自然语言指令和对应的网站截图上的动作组成。该数据集旨在训练视觉语言模型，通过点击、输入文本、滚动等常见网页交互动作来理解和与网页界面互动。每个指令都与一个要在网页视觉内容上执行的单一动作配对，使其成为教授模型基本网页导航操作的理想资源。数据集由清华大学、中国人民大学等机构的研究人员策划，并通过自动标注（GPT-4V）和人工验证相结合的方式创建。

The GUIAct Web-Single dataset is a comprehensive collection of single-step action instructions for website GUI navigation tasks. It contains approximately 67,000 instruction-action pairs, each consisting of a natural language instruction and a corresponding action to be performed on a website screenshot. The dataset is designed to train vision language models to understand and interact with web interfaces through actions such as clicking, inputting text, scrolling, and other common web interactions. Each instruction is paired with a single action to be performed on the visual content of a website, making it an ideal resource for teaching models the fundamental operations of web navigation. The dataset was curated by researchers from Tsinghua University, Renmin University of China, and other institutions, and created through a combination of automated annotation (GPT-4V) and human verification.