LLAMA 3.1 70B memorizes nearly all of Harry Potter and the Sorcerer’s Stone (91% of text). How about we make it forget all that and save us some GPU memory?

I suppose writers should be sad about this, but it seems like a net good: you want people to be able to ask questions about your book and get good answers. even if the thing produces 90% of your book, it's not exactly a great reading experience...