It would make sense to repeat the experiment with the tokenizer of an LLM.