Improve gpt-3.5-turbo performance
Closed this issue · 1 comments
WIP
Current gpt-3.5-turbo performance is not very good due primarily due to bad understanding of function set (functions called at the wrong times, messages sent w/o call to send_message, etc.). Goal is to upload specific variations of the key prompts (tailored towards worse instruction following ability of 3.5) that provide reasonable performance.
I'm not sure that this is possible for chatgpt-3.5 (or any version of chatGPT), but for open LLMs, there is this technique called grammar-based sampling which basically forces the LLM to answer according to some predetermined grammar, here is a Hacker News thread and link to a ggml PR describing it: https://news.ycombinator.com/item?id=36819906 . This could at least force the LLM, when applicable, to only answer using a preset list of commands.