I am trying to fine tune a t5-base model using PEFT technique, the only catch is that it is a two step fine tuning .
Step 1 : fine tune the t5 model with jargons, or terms definition related to some domain.
Step 2 : fine tune step 1 model with domain specific multi class classification data.
How can we do this using PEFT tuning ? If it can be done using PEFT , then how to load the step 2 model for inference , should it have original model as t5 or step 1 model ?
Is there any better or recommended way to do it ?