xcodemind/vision2ui_purified

Vision2UI是一个从真实场景中提取的、包含全面布局信息的数据集，专门用于微调多模态大语言模型（MLLMs）以生成UI代码。该数据集通过收集、清理和精炼开源Common Crawl数据集创建，并利用神经网络评分器确保数据质量。数据集包含超过三百万个并行样本，包括UI设计图像、网页代码和布局信息。为了验证数据集的有效性，研究团队建立了一个基准，并引入了一个基于Vision Transformer（ViT）的基线模型UICoder，以及一个新的度量标准TreeBLEU，用于评估生成的网页与源代码之间的结构相似性。实验结果表明，该数据集显著提高了MLLMs从UI设计图像学习代码生成的能力。

Vision2UI is a real-world dataset for generating webpage code from UI design images, including layout information. The dataset is created through a meticulous process involving the collection, cleaning, and refining of the open-source Common Crawl dataset. A neural scorer is employed to refine the data, retaining only the highest-quality instances, resulting in a dataset comprising over three million parallel samples that include UI design images, webpage code, and layout information. The dataset features include webpage screenshots, layout information (bounding boxes of elements), webpage code text (HTML/CSS code), screenshot scale, the main language of the rendered page, token counts of HTML and CSS code, scores from the neural scorer, and hash codes of image objects.