Why do the large language models love to put em dashes (with spaces around!) in the most inappropriate locations though? Is it just an artifact of training data or is it something more inherent like tokens with a space at the end and punctuation both having high probability and so this ends up being a weird compromise?