Any tips for getting Mixtral 8x7b through GPTpilots clarification loops

[Miscend]

I've been experimenting with using Mixtral as the LLM for GPTpilots instead of GPT-4, but am running into problems getting the clarification loop to work properly.

Specifically, when I integrate Mixtral into the create_gpt_chat_completion call, the clarification questions get stuck in endless loops rather than terminating appropriately with "EVERYTHING_CLEAR".

I've found a workaround where I can use GPT-4 for the initial high-level prompts up through the first dev step. Then switch to Mixtral after that point by modifying the LLM integration. This seems to avoid getting stuck.

However, it would be ideal to get Mixtral working end-to-end, so I wanted to check for any tips on:

• Prompt formatting that works well for clarification loops with Mixtral
• Ways to help Mixtral determine when to terminate the loop
• Any other best practices for conversational workflows with Mixtral

I'm very new to experimenting with alternatives to GPT-4, so any guidance would be much appreciated! Please let me know if there are any other details I can provide that would help troubleshoot this issue.

Thanks for creating such a great framework for leveraging LLMs! Looking forward to contributing back as I learn more.

[YosuCadilla]

@Miscend
@senko Commented on Discord that the main issue for using any LLMs other than GPT4 are the highly GPT4-specialized prompts.
I will be interested in collaborating on a multiprompt solution / option and of course in adapting the actual prompts for Mixtral or Mistral Medium.
btw, Are you running Mixtral locally or through the Mistral API?