I was hoping someone might be able to help recommend a model (or a few models) to try and play around with for my specific use case, I have had a hard time deciphering some of the models and what they were intended to do.
What I want to do is fine tune a model to act as a certain character, and respond in a chat to questions about themselves and other things with a certain personality/based on what the model is told about itself.
I’ve seen gpt-3 used in this way where basically you say “You are a famous painter who lives in Paris and has a stern and arrogant attitude” and then continue to have a conversation with this “person”. I think that fine tuning gpt-3 would work well except for the cost, so I’m looking for something that would be:
- Easy to train with context about itself, as well as teach domain specific or modern/internet language (you are 40 years old, you live in Paris, you are a painter, you like classical music, and then definitions like the word troll means to
- Reasonable size so that I could download it and run it locally (I’d say 10gb max)
I found a product pretty much doing this (inworld.ai) but being able to do it with my own model is preferential.
One thing I’m having trouble with is choosing text generation models vs conversation. GPT-2 is text generation but the model that I found to appear closest to what I want is the microsoft GODEL model which is conversation. Playing around with it however it doesn’t seem to capable, or perhaps I’m expecting too much after playing around with GPT-3.
Thanks in advance.