RINA

@ymcui Yes and it definitely worked out well. I also think that memory matters since the logic of create_pretraining_data.py is not that memory-efficient.