I think it's important to ask first what your aims are with a LLM.
Personally I have never seen a valid use-case for ChatGPT or any of its
varieties in libraries.  These models are, ultimately, little more than
glorified text compressors[1] that perform pattern-matching and which do
not, and cannot, produce accurate or reliable information.

(And this is before we get into the ethical issues of the large commercial
models' theft of the work of writers, programmers, and artists at large to
feed a tool that businesses and organizations are looking to use to replace
them.)

[1] https://aclanthology.org/2023.findings-acl.426/

On Mon, Feb 26, 2024 at 1:49 PM Peter Murray <
[log in to unmask]> wrote:

> I took note of something recently from the Library Innovation Lab at
> Harvard Law School: WARC-GPT: An Open-Source Tool for Exploring Web
> Archives Using AI. It takes the contents of WARC files and feeds them into
> a Retrieval Augmented Generation tool. Been meaning to play with it as a
> way to enhance FOLIO's docmuentation search, but haven gotten around to it.
>
> Not a war story, perhaps, but a WARC story.  ;-)
>
> Peter
> On Feb 26, 2024 at 4:07 PM -0500, Eric Lease Morgan <
> [log in to unmask]>, wrote:
> > Who out here in Code4Lib Land is practicing with either one or both of
> the following things: 1) fine-tuning large-language models, or 2)
> retrieval-augmented generation (RAG). If there is somebody out there, then
> I'd love to chat.
> >
> > When it comes to generative AI -- things like ChatGPT -- one of the
> first things us librarians say is, "I don't know how I can trust those
> results because I don't know from whence the content originated." Thus, if
> we were create our own model, then we can trust the results. Right? Well,
> almost. The things of ChatGPT are "large language models" and the creation
> of such things are very expensive. They require more content than we have,
> more computing horsepower than we are willing to buy, and more computing
> expertise than we are willing to hire. On the other hand there is a process
> called "fine-tuning", where one's own content is used to supplement an
> existing large-language model, and in the end the model knows about one's
> own content. I plan to experiment with this process; I plan to fine-tune an
> existing large-language model and experiment with it use.
> >
> > Another approach to generative AI is called RAG -- retrieval-augmented
> generation. In this scenerio, one's content is first indexed using any
> number of different techniques. Next, given a query, the index is searched
> for matching documents. Third, the matching documents are given as input to
> the large-language model, and the model uses the documents to structure the
> result -- a simple sentence, a paragraph, a few paragraphs, an outline, or
> some sort of structured data (CSV, JSON, etc.). In any case, only the
> content given to the model is used for analysis, and the model's primary
> purpose is to structure the result. Compared to fine-tuning, RAG is
> computationally dirt cheap. Like fine-tuning, I plan to experiment with RAG.
> >
> > To the best of my recollection, I have not seen very much discussion on
> this list about the technological aspects of fine-tuning nor RAG. If you
> are working these technologies, then I'd love to hear from you. Let's share
> war stories.
> >
> > --
> > Eric Morgan <[log in to unmask]>
> > Navari Family Center for Digital Scholarship
> > University of Notre Dame
>