tags:
- image-search
- siglip
- wikimedia-commons
- webdataset
Licensed image search corpus
A public, reproducible image-search dataset built from Wikimedia Commons. Each WebDataset sample contains a 384px WebP derivative and JSON source metadata. Images are selected using their source copyright-license metadata: CC0/public domain, CC BY, and CC BY-SA; NC and ND are excluded. There is no single blanket license for the images. Consult each sample's license, license_url, creator, and source_url before reuse. Source metadata does not establish other rights.
Fields include image_id, title, creator, license, source_url, full_url, width, height, description, tags, and source SHA-1. The retrieval model is ViT-B-16-SigLIP pretrained on webli, producing normalized 768-dimensional vectors. Archives preserve alpha where available; embedding composites transparency onto white. The originals remain at their source URLs.
Build code: https://github.com/Zahide4/imgsearch/blob/main/cloud_corpus.py The pipeline uses sequential API requests per worker, bounded filename ranges, API-generated thumbnails, retry backoff, and saved restart cursors.
The corpus is a prototype, has uneven subject coverage, and has not been comprehensively reviewed for content quality or suitability for any specific use.