Books3: a hundred and ninety-six thousand books named
On 7 July 2023 Richard Kadrey, Sarah Silverman and Christopher Golden sued Meta Platforms, case 3:23-cv-03417. Meta's LLaMA paper itself names the Books3 section of The Pile among its training data without describing it. The complaint describes it: by the EleutherAI paper Books3 is 108 gigabytes, about 12 per cent of The Pile and its third largest component, and the person who assembled it has said publicly that it is all of Bibliotik and holds 196,640 books.
Why it matters
Here the size of a disputed corpus was for the first time not estimated but named exactly, because the maker had itself cited an open set, and an open set has published documentation. A suit against open weights proved easier to plead than one against closed: the openness meant to lower the risk handed the plaintiffs a ready inventory.
The chain of citation in the complaint runs: Meta's paper names Books3 within The Pile; the EleutherAI paper, The Pile: An 800GB Dataset of Diverse Text for Language Modeling, describes Books3 as books derived from a copy of the contents of the Bibliotik private tracker; public statements by its assembler give 196,640. That figure belongs to those statements, not to Meta's paper. What the record does not claim. The record does not say Meta confirmed the count, and it does not describe the later course of the case.