Anders, do you have a citation for “use Wikipedia content considerably”?

Lots of early-ish ML work was heavily dependent on Wikipedia, but state-of-the-art Large Language Models are trained on vast quantities of text, of which Wikipedia is only a small part. ChatGPT does not share their data sources (as far as I know) but the Eleuther.ai project released their Pile a few years back, and that already had Wikipedia as < 5% of the text data; I think it is safe to assume that the percentage is smaller for newer models:  https://arxiv.org/abs/2101.00027

Techniques to improve reliability of LLM output may rely more heavily on Wikipedia. For example, Facebook uses Wikipedia rather heavily in this *research paper*: https://arxiv.org/abs/2208.03299 But I have seen no evidence that techniques like that are in use by OpenAI, or that they’re specifically trained on Wikipedia. If you’ve seen discussion of that, or evidence from output suggesting it, that’d be interesting and important!

Social: @luis_in_brief@social.coop
ML news: openml.fyi
On Feb 20, 2023 at 1:52 AM -0800, Anders Wennersten <mail@anderswennersten.se>, wrote:
BIng with ChatGPT is now released by Micrsoft.

And from what I understand they use Wikipedia content considerably. If
you ask Who is A B and A B is not widely known, the result is more or
less identical to the content from the Wikipedia article (but worse, as
it "makes up" facts that is incorrect).

In a way I am glad to see Wikipedia is fully relevant even in this
emerging AI-driven search world. But Google search has ben careful to
always have a link to Wikipedia besides their made up summary of facts,
which here it is missing (yet?). And for licences, they are all ignored.

So if this is the future the number of  accesses from users to Wikipedia
will collapse, and also their willingness to donate... (but our content
still a cornerstone for knowledge)

Anders

(I got a lot of fact from an article in Swedish main newspaper by their
tech editor. He started asking fact of himself, and when he received
facts from his Wp article plus being credited to a book he had noting to
do with, he started to try to tell/learn ChatGPT of this error. The
chatPGT only got angry accusing the techeditor for lying and in the end
cut off the conversation, as ChatGPT continued to teat the techeditor as
lyer and vandal..).
_______________________________________________
Wikimedia-l mailing list -- wikimedia-l@lists.wikimedia.org, guidelines at: https://meta.wikimedia.org/wiki/Mailing_lists/Guidelines and https://meta.wikimedia.org/wiki/Wikimedia-l
Public archives at https://lists.wikimedia.org/hyperkitty/list/wikimedia-l@lists.wikimedia.org/message/GJJNX2Y7BX5RZYGAIYTUI6O6CSBN72EH/
To unsubscribe send an email to wikimedia-l-leave@lists.wikimedia.org