bloom has inhibition in adapter transformers regardless a further step is data.
the huggingface training scripts with their reduced inhibition likely take data as line or jsonlines format (not sure). if this is true, likely the thing to do is convert the data to jsonlines so as to preserve linebreaks. for encoder/decoder models the format is input/output whereas decoder only models have only input. reasonable next step is to learn the data format of these scripts, and maybe come up with a way to provide data either architecture could read.
