@dbd981bf I've been finding they're incredibly good at misspellings - I'd love to understand why, is it just because their training text contains every common misspelling or is there something else about the way the tokenizers work I wonder