A arxiv.org Pretraining data enables narrow selection capabilities in transformer models 65 points hislaziness 3 years ago 109 comments