I don't have the arxiv link bookmarked, but there was a paper written on pretraining with Lora. It involved merging adapters back every n steps with good results.