Fine-tuning on custom Tools

[ilyalasy]

Hey all!
As a lot of people, I've been testing Langchain's custom Tools functionality and I experienced that they perform pretty poorly for cases more complex than just weather APIs etc.
Recently, OpenAI came up with functions feature for their models, where:

These models have been fine-tuned to both detect when a function needs to be called (depending on the user’s input) and to respond with JSON that adheres to the function signature.

Has anybody tried to fine-tune open source LLM's on custom functions, and if so, what the approach there would be?