Video Dataset Sources for Pre-Training AI Models

Cumulative tokens, by source category

(Please use a modern browser to see the interactive version of this visualization)

Source: Data Provenance Initiative “Bridging the Data Provenance Gap Across Text, Speech and Video," December 2024; note: based on a manual analysis of 4,000 public datasets 1990-2024 spanning 608 languages, 798 sources, 659 organizations, 67 countries; “other” includes Prelinger Archive, Facebook, Flickr, Bing videos, Getty Images, Giphy, Metacafe, Peeks, Videoblocks, Vimeo, Vine, Weather Channel, BBC, YouDescribe, Tumblr, TikTok