LAION-5B

LAION-5B is the largest public image-text dataset, put together by the non-profit LAION from Common Crawl. The build process scanned the crawl for images that had alt-text captions, then used CLIP (a model that scores how well an image and a piece of text match) to keep only the pairs where the caption genuinely describes the picture and drop the rest. The result is over 5.8 billion image-text pairs, split into roughly 2.32 billion English pairs, 2.26 billion in other languages, and 1.27 billion where no single language dominates.

The crucial detail is that LAION-5B does not contain any images. What you download is a set of Parquet files listing the image URLs, the captions, and metadata such as width, height, and the CLIP similarity score. To get the pictures you run a downloader (the img2dataset tool is the standard choice) that fetches them from their original hosts. Because those links decay over time, a meaningful fraction now fail, so you will recover less than the full set, and two people downloading today will get slightly different collections.

It is best suited to training or fine-tuning multimodal models, and for RAG specifically it shines as a corpus for building image-text retrieval: embed the images and captions with a CLIP-style model and you have a searchable base for "find an image like this" or caption-grounded question answering. LAION-400M, a smaller sibling, ships with CLIP embeddings and nearest-neighbour indices already computed, so if you want to experiment without the download and embedding overhead, start there.

Two serious caveats. First, licence: the metadata is CC BY 4.0, but each image remains under whatever terms its original host set, so the dataset licence does not grant you any right to reuse the pictures themselves, and commercial use needs case-by-case care. Second, safety: this is unfiltered web content at enormous scale, and earlier versions were withdrawn after harmful material was found, so use the safety-reviewed, cleaned re-release and apply your own content filtering before you index anything.