Multiple Formats for Distillation Finetuning

[jxnl]

Is your feature request related to a problem? Please describe.

We should be able to export in standard finetune format in chat messages rather that function calling format so we can use methods that are not finetuning. or convert between the two

[h55nick]

@jxnl I'd be interested in helping get this across the finish line as I think I'm looking to enable this use-case.

First bit is that OpenAI/GPT3.5 only supports fine-tuning on chat and function calls, but tools is becoming the more common (anyscale, etc) method of complex interaction. Tools breaks distillation currently (unless I'm missing something) as Distillation requires a non-iterable return of chat, while tools requires an Iterable.

If you have some more direction on what you are thinking here, I'd be happy to plug away on it. Appreciate your work on Instructor (learned of it through the W&B course).

Notes: Here is where we define the fmt: https://github.com/jxnl/instructor/blob/02631a46473be1345e9e5027b505d9f2f9c191d1/instructor/distil.py#L18 I'm not sure we'd want to add to this enum as the messages format is what is supported.