WIT (Wikipedia-based Image Text)

WIT, the Wikipedia-based Image Text dataset from Google Research, links images to the text that surrounds them on Wikipedia. Rather than a single caption per image, it captures a rich bundle of context: the alt text, the displayed caption, the reference description, and passages from the article the image appears in, so a model learns from real encyclopaedic writing instead of a one-line label. In total it holds on the order of 37 million image-text examples tied to roughly 11 million unique images.

Its signature quality is language coverage. WIT spans more than 100 languages, making it one of the few large image-text collections deliberately built for training and evaluating multilingual multimodal systems rather than English-only ones. It ships as TSV files that list image URLs alongside their associated text, distributed through the project's GitHub repository and mirrored on Kaggle and Hugging Face. You download the text and the picture URLs, then fetch the images yourself from Wikimedia, so plan for a retrieval step and some link rot as older files move or disappear.

For RAG, WIT is a strong base for multilingual visual question answering, cross-lingual image retrieval, and grounding systems that reason over pictures and their captions across languages. It complements text-only Wikipedia knowledge bases by adding a visual layer, and sits alongside larger web-scraped collections like LAION, where WIT trades raw scale for cleaner, encyclopaedic context and much broader language support.

Mind the layered licensing. The dataset itself is CC BY-SA 4.0, so commercial use is allowed, but it is share-alike: any derivative database you distribute must carry the same terms and credit the source. The underlying images come from Wikimedia Commons under their own varied licences, so an individual picture's terms can differ from the dataset's, and you should check per-image rights before republishing the media rather than the annotations.