NJU-LINK/WebCompass
收藏资源简介:
WebCompass是一个统一的多模态基准测试集,旨在评估大型语言模型(LLMs)在生成、编辑和修复功能性网页方面的能力。该数据集涵盖三种输入模态(文本设计文档、参考截图和视频演示)和三种任务家族(生成、编辑和修复)。生成任务包括从文本设计文档、参考截图和视频演示生成网页;编辑任务涉及向现有单页或多页网站添加新功能;修复任务则是修复损坏的网页以匹配目标截图。数据集结构包括多个配置和分割,每个配置对应不同的任务类型和输入模态。数据集还提供了详细的文件结构、数据格式和评估维度,支持多模态任务的全面评估。
WebCompass is a unified multimodal benchmark for evaluating LLMs ability to generate, edit, and repair functional web pages. It spans three input modalities — text design documents, reference screenshots, and video demonstrations — and three task families — generation, editing, and repair. The generation tasks involve creating web pages from scratch based on text design documents, reference screenshots, or video demonstrations. The editing tasks require adding new features to existing single- or multi-page sites. The repair tasks involve fixing broken pages to match target screenshots. The dataset includes multiple configurations and splits, each corresponding to different task types and input modalities. It also provides detailed file structures, data formats, and evaluation dimensions to support comprehensive assessment of multimodal tasks.




