2 ms·
I'm mostly curious about the decision to do fine-tuning here instead of synthesizing a good system message, maybe a RAG setup with access to the slack database,
by coreyh14444 2y ago
I'm mostly curious about the decision to do fine-tuning here instead of synthesizing a good system message, maybe a RAG setup with access to the slack database, etc. I tried fine-tuning on a content generation case when GPT 4 first came out and we had much better results without it. Also, you don't get the benefit of upgrading to new models when they are released.
- empath75 2y agoI think you're putting way too much thought into what is basically a comedy bit.
- charles_irl 2y agoI worked on this project a bit! These are great questions. Fine-tuning works very well for style (as opposed to factual knowledge), which is all we're trying to achieve here -- as another commenter put it, it's a "comedy bit" for a company Slack. In fact, fine-tuning for style works well enough that we find it pretty easy to just re-train when new models come out. There's sometimes some YAML-fiddling required to get training frameworks to work with different model series (e.g. a DeepSeek series model vs a LLaMA series model), but it's not too onerous. IMO, the ideal ML pipeline looks less like the bespoke process common these days and more like a materialized view (shouts to pgML). That's easy to automate and so reduces the gap with prompting. On the other hand, it seemed harder to craft a generic system prompt and a generic retrieval system that would work across organizations to define user communication style.
- coreyh14444 2y agoThanks for the reply! Makes sense. I will definitely keep this in mind. Much appreciated.