Comment by rasbt

I wouldn't pretrain from scratch, but continued pretraining is pretty popular for adapating LLMs to recent and/or custom data. (Sometimes this is referred to 'finetuning', however, not to be confused with 'instruction finetuning').