I’ve tested specifically this (on my personal time) :) It will train but I found the loss is proportional to the number of trainable parameters. So roughly to hit the performance of a standard 70m param model, you need to train ~70m lora params anyway.