DataComp
DataComp turns the usual model benchmark on its head. Instead of fixing the data and comparing models, it fixes the model architecture and training budget, then asks which slice of data produces the best results. The dataset you build is the thing under test.
It ships a common candidate pool of image-text pairs at several scales, along with a standard training and evaluation setup. Teams filter or reweight that pool with their own curation ideas and measure the effect on downstream tasks, making results directly comparable across approaches.
Related sources
Conceptual Captions (CC3M / CC12M)
A cleaned image alt-text dataset for automatic image captioning, produced by Google Research. Alt-text pulled from the web is filtered and generalised (hypernymed), replacing specific names with broader categories. Available in 3.3M (CC3M) and 12M (CC12M) variants.
LAION-5B
An image-text dataset of over 5.8B examples, built by filtering Common Crawl with a CLIP model that scores how well an image matches a caption. Includes 2.32B English pairs, 2.26B multilingual pairs, and 1.27B not tied to any particular language. Provides URLs, not the images themselves.
OBELICS
A web-scale dataset of 141M multimodal English web documents that interleave text and images, containing 353M images and 115 billion tokens. Unlike image-caption pair datasets, it keeps whole documents with text and images in their original order. Used to train the IDEFICS models.
PD12M (Public Domain 12M)
An image-text dataset built only from materials marked with a Public Domain Mark or released under Creative Commons Zero (CC0). 12.4M image-caption pairs, with a 3.3M subset called PD3M, sized to match the Conceptual Captions datasets while staying copyright-clean for commercial use.