Closed albertus-andito closed 1 year ago
Hi! In preliminary analysis we find the quality with GPT-J-6B to fall short on function generation and thus quality to be worse than text-davinci-003. We are looking at various options and thinking about how to acquire good open-source LMs for that step. If you have pointers as well, let us know!
I'm wondering whether the newly released StarChat can be used for the function generation step...
Hi,
Thanks for open-sourcing this code. Interesting stuffs!
Have you tried running EVAPORATE-CODE+ with open source huggingface models, such as GPT-J-6b? What's the overall performance? Is it comparable to text-davinci-003?