Hacker News Reader

Comment by LoganDark | Hacker News Reader

Is 10MB enough to fine-tune a model nowadays? Assuming that the base model is fairly well-trained already, of course.

It's not fine-tuning, it's RAG.

RAGs are great for glossary like structured information. Arbitrarily chunking prose text feels like making garbage out of otherwise good quality text. Ideally prose would be used to generate glossary like documents (LLM aided). Performance plays role here (you'd need to process a lot of text up front) so maybe smaller models could be used?

This is something that has always bothered me about RAG. It seems that it’s fine for first order relevance, like a search engine, but for knowledge there needs to be some kind of rumination stage where it revisits the entire corpus to find information that has a second order relevance to round out its ‘understanding’.

You might be able to approximate by chunking and globbing the chunks and searching for those, as well as having the LLM summarize and extract data and search for those items as well.

isn't this what llamaindex is doing though?

I think as long as you are below 50% context window, performance should be ok (https://dev.to/maximsaplin/gpt-4-128k-context-it-is-not-big-...)

Retrieval augmented generation, apparently.

Yeah, and I swear that term appeared out of nowhere. I feel like a year ago we were called it vector embeddings, but I’m no expert.

Sometimes the terminology is around but not well distributed. I'm not sure this is the paper that coined it, but here's an example from 2020: https://arxiv.org/abs/2005.11401

Oh, so it's not fine-tuning the model, it's allowing the model to recall things from a database. (I didn't read the article before commenting)

These days it’s enough to completely mindbreak a model via DPO. Seeing how modern open source LLMs are willing to do anything was quite a surprise given how locked down ChatGPT is.

Reply on news.ycombinator.com