I wonder how much of a test this is for the LLM vs whatever tokenizer/preprocessing they're doing.