Conceptual Captions (CC3M / CC12M)

Conceptual Captions is built from the alt-text that web authors attach to images, the short descriptions meant for screen readers. Google Research cleaned this raw text and generalised it, a step they call hypernyming: swapping a specific name like a person or brand for a broader category such as "man" or "car", so the captions describe image content rather than name particular entities. It comes in two sizes, the 3.3 million pair CC3M and the larger 12 million pair CC12M.

The dataset is distributed as a list of image URLs paired with captions rather than the images themselves, so the first job is to run a downloader that fetches each image from its original host. Budget real time and storage for that step, and expect losses: because the images live on third-party sites, link rot means a meaningful share of URLs have gone dead over the years, and the exact set you end up with will differ from what someone downloaded a year ago.

It is best as training and evaluation data for image captioning and vision-language models, and by extension for the image-text encoders behind multimodal RAG, where you retrieve images from a text query or vice versa. The hypernymed captions are clean and consistent, which is exactly what you want for learning general image-to-text mappings.

That same cleaning is the catch for some uses: because specific names are stripped out, the captions will not tell you that a photo shows a named person, place, or brand, so if your application depends on named entities this is the wrong source. The bigger caveat is licensing. Access is governed by Google Research terms that are permissive for research but not a clear commercial licence, and the underlying images stay under whatever terms their original hosts set, so the picture rights differ from the caption rights. Check both before building anything commercial.

If you need a commercially safe, self-hosted alternative at a similar scale, look at PD12M, which draws only on public-domain and CC0 images and hosts them itself to sidestep link rot; LAION and YFCC100M are other large image-text options with their own trade-offs.