Why most people fine-tune full model instead of just adapter? Should we educate people more on that topic?

[cyberluke]

Why most people fine-tune full model instead of just adapter? Should we educate people more on that topic?

You could then use parallel adapter forward pass as implemented by the ‘Parallel’ block and it is already supported by HunggingFace:

• https://docs.adapterhub.ml/adapter_composition.html
• https://huggingface.co/docs/hub/en/adapters

Video education about full fine tuning vs adapter
https://www.youtube.com/watch?v=s2BF_gC0X1o

You could compose several adapters in parallel for each language with fine-tuning and compose it on top of base model saving memory, space. This is more modular approach.