

They say the same thing over and over again so emphatically that I think it must be something nasty about me but I shall not let myself become frightened away by the sarcasm of owls.

The gentleman owl says, and she answers from her tree a little way off,, beautifully assenting to and completing her lord’s remark, as becomes a properly constructed German she-owl. Two owls are perched near me, and are carrying on a long conversation that I enjoy as much as any warbling of nightingales. I am writing in it now in the late afternoon loveliness, much interrupted by the mosquitoes and the temptation to look at all the glories of the new green leaves washed half an hour ago in a cold shower. Full name cloze accuracy results (GPT-4, ChatGPT and BERT) for all books can be found here: gpt4-books.Staying overnight with friends in England just before Christmas, this book by Elizabeth von Arnim was placed on my bedside table and though there was no chance I would finish it, I was captivated and charmed by Elizabeth’s garden right from those first few pages. The following books in our set have the highest GPT-4 name cloze accuracy (where accuracy here is the fraction of times a model predicts the masked name exactly right in 100 passages from a book). To generate name cloze passages for your own texts, see the generate_name_cloze folder. He put chips under it, blocks of various sorts, bits of pasteboard, and at last went so far as to attempt an exquisite adjustment by final pieces of folded blotting paper.ĭetailed results on all books can be found here: gpt4-books. Though of a very ingenious mechanical turn, could never get this table to suit him. This work probes what books GPT-4 and ChatGPT have memorized by assessing the degree to which models are able to predict the masked token (a proper name) in name cloze passages like the following:

Chang, Mackenzie Cramer, Sandeep Soni and David Bamman (2023), " Speak, Memory: An Archaeology of Books Known to ChatGPT/GPT-4", ArXiv. Code and data to support the following paper:
