Hugging Face PEFT Integration of KappaTune

[oswaldoludwig]

You can now use KappaTune's selection logic directly with the Hugging Face ecosystem. This allows you to apply LoRA adapters only to the proper modules, effectively mitigating catastrophic forgetting with a single line of code. See HF model card: https://huggingface.co/oswaldoludwig/kappatune-lora-tinyllama-agnews and the updated GitHub repo: https://github.com/oswaldoludwig/kappaTune

[githubnemo]

This reminds me a bit of importance matrix and orthogonal subspace finetuning (+ it's importance scoring).

It seems to me that the LoRA integration uses the kappa score for the target selection. Would you be interested in integrating target selection tooling in PEFT?

[oswaldoludwig]

That would be great, considering the good results I’ve had with this method so far.

[oswaldoludwig]

See #3106