Datasets:
Please use materials from the Internet Archive, but not this way.
This dataset was posted without coordinating with the Internet Archive. Unfortunately, it appears that the person posting it misunderstood (and is therefore misrepresenting) the nature of these books based on incomplete metadata and how we deal with rehosting books.
In the current legal environment where numerous lawsuits have been filed over AI training on books–
We urge Hugging Face to remove this dataset, and warn researchers not to rely on it.
Uploaders to the Internet Archive posted metadata on items that should not be relied upon in this context.
Mistakes like these could lead to lawsuits and liability for the uploader, Hugging Face, the Internet Archive, and downloaders. One example is the Authors Guild suing HathiTrust over their “orphan works” program, which had to be shut down (https://www.publishersweekly.com/pw/by-topic/digital/copyright/article/48722-hathitrust-suspends-its-orphan-works-release.html) and resulted in years of litigation. Another example is the lawsuit against Project Gutenberg that was brought in Germany over a book that was clearly in the public domain in the United States, but not in that country. That case also resulted in years of litigation and ended with Project Gutenberg being blocked (https://www.techdirt.com/2018/03/07/project-gutenberg-blocks-access-germany-to-all-public-domain-books-because-local-copyright-claim-18-them/) in Germany.
Also bulk rehosting these materials without coordination may have other negative effects on our partners and the Internet Archive. Others, such as HathiTrust and the University of Toronto, have proceeded in ways the Internet Archive and its partners approve of.
We welcome researchers working with materials from the Internet Archive and millions have. For large scale requests, and especially rehosting offers please reach out to us at info@archive.org
Brewster Kahle
Digital Librarian
Just a short remark: the dataset includes significantly expanded metadata based on reconciliation from the open library. As far as I can check, this is extensively compliant with US laws on public domain.
There are numerous valid concerns over the sharing of pretraining datasets from the web, which do include copyright materials that may or may not be covered by copyright. This is probably one of the safest available dataset from pre-training and it would be deeply ironic to delete it on the base of copyright concerns.
@markgraham If you find that this dataset was compiled in a wrong way, could you please kindly suggest a better way to create a dataset of books which would be considered safe to use for training models and for other purposes of NLP and computational linguistics not only in the US, but globally?
Unless there is a constructive alternative, I am afraid that the community would use this dataset as one of the most reasonably looking options among the available ones.
@cointegrated As it turns out, I'm going to release a very large collection of open datasets in the next few days (also motivated by the present discussion…)