However this resulted in the BERT model performing much better than the CANINE model. As documented in the file "next_line_poetry_10000ex_3epochs.stats" the BERT model's best accuracy was 84% (with ...