collecting large STEM datasets for training huge models a good step for avoiding human models?