2 ms·
I don't think OpenAI is trained on any copyrighted, for-sale books. GPT-3 at least uses books1 and books2. Books1 seems to be Bookcorpus[0], Books2 seems to be
by brookst 4y ago
I don't think OpenAI is trained on any copyrighted, for-sale books. GPT-3 at least uses books1 and books2. Books1 seems to be Bookcorpus[0], Books2 seems to be Libgen[1]. There is some unclarity about the purity of licensing in Bookcorpus affecting 200-400 of the 11,000 books[2]
0. https://paperswithcode.com/dataset/bookcorpus https://paperswithcode.com/dataset/bookcorpus
1. https://librarygenesis.pro/ https://librarygenesis.pro/
2. https://towardsdatascience.com/dirty-secrets-of-bookcorpus-a-key-dataset-in-machine-learning-6ee2927e8650 https://towardsdatascience.com/dirty-secrets-of-bookcorpus-a...
Are you suggesting that free / public domain material shouldn't be used for training, or were you not familiar with the training set?