[Scope to Fine Tune]
Hi Team Nexus,
First of all, congrats on the very impressive work & thanks for open sourcing the model weights & providing the colab demo for using this model.
Is there any scope to fine tune this model? Can a colab demo be provided using trl that shows us how to instruct fine tune this model for custom function calling use cases?
Thanks!
Hi Vikram!
Thank you for your interest! SFT itself can go pretty far, and performing SFT is pretty easy with Raven. I'd recommend curating some prompt-completion pairs for your APIs/functions using the same prompting style (including the special tokens) as shown in the README/colab demo and just performing simple supervised fine-tuning on your pairs (ideally with loss only on the completions). It should work well, and likely would only need a few samples to fit robustly to your APIs/functions with the same prompting style.
If there is interest in this, we can perhaps create some simple SFT examples and add it to Github. Also, please do join our Discord Channel, we'd love to keep up as you try out your experiments!
Thanks again!
Great job on this model! The example you mentioned would be amazing. Joining your Discord.
Hey, Venkat, can you help with this part in particular for fine-tuning "(ideally with loss only on the completions)", I get what you mean by that and it does make sense, but can you provide code implementation or a guide which I can look at? Thanks!