Definitely an interesting thought to do this at the tokenizer level!