OBELICS

OBELICS is a web-scale collection of 141M English web documents in which text and images appear together in the order they had on the original page. Most image-text datasets flatten each picture into a single caption; OBELICS instead keeps the whole document, so a model sees each image in the context of the paragraphs around it. In total it holds 353M images and about 115 billion tokens of text, filtered from Common Crawl to strip out low-quality, unsafe, and opted-out content.

The dataset is hosted on the Hugging Face Hub and stored as Parquet, which you can download in full or stream example by example with the datasets library, handy given the scale. One important detail: the images are referenced by their original web URLs rather than bundled as files, so the text and layout come with the dataset but you fetch the actual image bytes yourself. Plan for a download step, some caching, and the fact that a share of those links will have gone dead since the crawl.

This interleaved structure is exactly what made OBELICS suitable for training the IDEFICS family of open vision-language models, and that is its sweet spot: pretraining or evaluating multimodal models that read documents rather than isolated image-caption pairs, and building multimodal RAG corpora where retrieving a passage should bring its surrounding images along with it. If your system reasons over illustrated documents, this shape maps onto it naturally.

The caveats are scale and noise. Even after filtering, this is web data, so expect boilerplate and uneven quality, and the image-fetching step adds engineering you would not have with a self-contained set. On licence, OBELICS is released under CC BY 4.0, so commercial use is permitted as long as you attribute the source; keep that credit in anything you redistribute. For simpler image-caption pairs, LAION and Conceptual Captions are the usual alternatives; OBELICS is the one to reach for when document order matters.