ZhuOnR/ShowUI_Web

ShowUI_Web数据集是一个专门用于训练GUI视觉代理的FiftyOne数据集，包含21,988个样本。数据集聚焦于22个代表性网站场景（包括Airbnb、Booking、AMD、Apple等）的网页截图和元素注释，特别过滤了静态文本元素，以集中关注交互式组件如按钮和复选框。这种筛选策略基于观察到大多数视觉-语言模型已经具备强大的OCR能力，因此视觉交互元素对训练更有价值。数据集由新加坡国立大学和微软的Show Lab策划，语言为英语，许可证为Apache-2.0。数据集的结构包括网页UI截图和交互注释，核心字段包括指令列表、检测框和关键点。数据集的创建目的是为了训练视觉-语言-动作模型，用于GUI视觉代理在Web环境中的操作。

The ShowUI_Web dataset is a FiftyOne dataset specifically designed for training GUI visual agents, containing 21,988 samples. It focuses on web interface screenshots and element annotations across 22 representative website scenarios (including Airbnb, Booking, AMD, Apple, etc.), purposefully filtering out static text elements to concentrate on interactive components like buttons and checkboxes. This curation strategy was based on the observation that most Vision-Language Models already possess strong OCR capabilities, making visually interactive elements more valuable for training. The dataset is curated by Show Lab, National University of Singapore and Microsoft, with English as the primary language and licensed under Apache-2.0. The dataset structure includes web UI screenshots with interaction annotations, with core fields such as instructions list, detections, and keypoints. The dataset is intended for training vision-language-action models for GUI visual agents operating in web environments.