docling-project/screenparse

ScreenParse是一个用于完整屏幕解析的大规模数据集，提供网页截图中可见UI元素的密集标注。与仅标注任务相关元素的稀疏基础数据集不同，ScreenParse为训练理解完整屏幕结构的模型提供了全面的屏幕级监督。关键特征包括：密集标注（标注所有可见UI元素而非仅任务相关子集）、大规模（1,447,100张高质量截图，包含25,575,213个UI元素标注）、细粒度标签（55个语义UI元素类别，如按钮、链接、菜单等）、叶子标注（v2版本使用过滤的叶子元素减少噪声嵌套标注）、自动化流水线（通过Webshot流水线生成，基于VLM的细化和质量过滤）、阅读顺序（元素包含自然阅读流索引）。数据集结构包含唯一标识符、图像、源URL、视口尺寸、边界框、UI元素类型标签、元素文本内容、可交互性、阅读顺序索引和元素数量等字段。

ScreenParse is a large-scale dataset for complete screen parsing, providing dense annotations of visible UI elements across web screenshots. Unlike sparse grounding datasets that annotate only task-relevant elements, ScreenParse provides comprehensive screen-level supervision for training models that understand complete screen structure. Key features include: Dense Annotations (visible UI elements are annotated instead of only task-relevant subsets), Large Scale (1,447,100 high-quality screenshots with 25,575,213 UI element annotations), Fine-Grained Labels (55 semantic UI element classes such as Button, Link, Menu, etc.), Leaf Annotations (the v2 release uses filtered leaf elements to reduce noisy nested annotations), Automated Pipeline (generated by the Webshot pipeline with VLM-based refinement and quality filtering), and Reading Order (elements include natural reading-flow indices). The dataset structure includes unique identifier, image, source URL, viewport dimensions, bounding boxes, UI element type labels, text content of elements, interactability, reading order indices, and number of elements.