3 ms·
One thing that hasn't gotten much focus with 3.5 and 4 is fine tuning. It would be interesting to see what the road map looks like for this. Even with more toke
by osigurdson 4y ago
One thing that hasn't gotten much focus with 3.5 and 4 is fine tuning. It would be interesting to see what the road map looks like for this. Even with more tokens, fine tuning still seems necessary.
- bob1029 4y agoThis is where we are at right now. Even 128k tokens would not cut it for what we want to do. Currently experimenting with fine-tunes on the base models and seeing some promising progress. I suspect that the chat models will never be opened for tuning due to all of the moderation and security concerns. My strategy will be to go into the DIY dark forest if we can't get curie or davinci to start behaving as expected.
- osigurdson 4y agoWhat stack are you using? If C# I’m pretty close to having (what I think) is a much nicer fine tuning approach than the raw apis. I’m hoping to push this out tonight if I get a few spare hours. (OpenAILib on nuget).
- bob1029 4y agoWe are using C#. We do already have an in-house interface for all of this. Took us about a week to get through the prompt/completion training mess to a stable pattern. Whitespace and newlines are a hell of a thing with these LLMs.
- osigurdson 4y agoAgree. That is the problem I aim to improve a bit.