panopstor/wallpapers-cog-icl

该数据集是从7-themes.com网站抓取的壁纸分享数据，包含图像、元数据以及由THUDM的CogVLM生成的图像描述。元数据包括文件名和图像上传的类别，这些信息用于为CogVLM提供上下文线索，以生成每个图像的描述。图像文件被调整为最大2560x1440分辨率，并以WEBP格式保存。数据集还包含一个JSON文件，其中包含每个图像页面的元数据，但仅使用了文件名和类别来驱动上下文学习。

This dataset consists of wallpaper images scraped from 7-themes.com along with their metadata and captions from THUDMs CogVLM (vicuna-based). The metadata includes the file name and upload category of each image, which are used to provide contextual clues for CogVLM to write captions for each image, resulting in outstanding accuracy in describing specific objects or characters. The image files are resized to a maximum of 2560x1440 (if over) and saved using the PIL library in WEBP format quality=95. The adjoining json file contains metadata from each images landing page, though only the file name and category were used to drive in-context learning.