Questions on Proxy-Tuning for GPT-3.5

[NuoJohnChen]

Thanks for your great work! I have some questions regarding Proxy-Tuning for GPT-3.5:

1. Since GPT-3.5-turbo-0613 is a chat model rather than a base model (like davinci-002), how can the Proxy-Tuning process work?
2. GPT-3.5-turbo-0613 use openai.ChatCompletion.create to operate on messages instead of prompts (string), how to execute Proxy-Tuning?
3. How to constrain GPT to provide the probabilities of A, B, C, and D? Also, is it done using probability-level proxy tuning, such as adding delta probability to GPT-3.5-turbo?