So Im looking for references on how to do function calling using Dolphin or Mistral models.
With my current prompt, I’m able to get it to choose an appropriate command for the task sometimes. But often it’ll add multiple commands in one response. But the other half of the time it produces correct commands & parameters in json format as request. Sometimes it makes up commands it want to use that doesn’t exist in the command list.
I’m just looking for hints at a more concrete prompt that will make these models effective in function calling.
Should I try whatever format OpenAI use seeing as how these smaller models are usually trained on synthetic data produced by OpenAI models?
Any guidance is appreciated 👍
You must log in or register to comment.
I think without fine tuning it on returning function calls you won’t get any good results. Maybe trying it with validating the input and result again in a loop could do the job, but otherwise we need to wait for fine tuned models.
Yes good call 👍 I’m gonna try that now.
I’m gonna add a system message to the chat history part of the prompt saying that the command is invalid and see if it corrects itself in the next iteration of the loop.
This could add a bandage 🩹 on the issue for now. Allowing it to seemlessly loop until a task is complete. Until I can find a better prompt or model.
So far, dolphin-mistral 2.1 7b is what I’m using ATM.
Use langchain with tools
Not a bad idea but I’m not coding in Python.
Because I hate myself, I’m writing in C# 🫠
Also, I want to use a few libraries as possible.
As a last resort I may use langchain. Or just look at the source and see how they force a model into function calling if it’s possible.