In a California federal court, Meta faces lawsuits over using pirated books to train its Llama AI models. Judge Vince Chhabria oversees the consolidated cases brought by authors like Sarah Silverman and Ta-Nehisi Coates, and publishers including Entrepreneur Media and Chicken Soup for the Soul. Recent deposition testimony from former Meta researcher Stephen Roller reveals he torrented the Books3 dataset—a collection of nearly 200,000 books from the private torrent tracker Bibliotik—and knowingly distributed it while employed at Meta between 2017 and 2023.
Roller admitted to deleting his copy about a year ago, citing the litigation as the reason, stating it 'seemed contentious to have copies of The Pile,' which includes Books3. Meta acknowledges using portions of Books3 to train Llama 1 and Llama 2 but claims Roller’s download may have been for personal use, not for the company. The plaintiffs argue that Meta had multiple copies of the dataset, at least one obtained via torrent, and are seeking details on how many copies were acquired and by whom.
Additionally, the plaintiffs allege pirated books were used beyond AI training, pointing to an internal 'ML Books' Google Drive folder and a Meta book club that may have utilized shadow library materials. One employee was caught sharing a pirated book, ironically noting the irony given the ongoing legal battles.
Source: Hacker News · Summarized by HeadlinesBriefing