It uses fineweb, which is derived from Common Crawl, which is an unlicensed scrape of web pages.