For some discussion on how to have the LLaMa tokenizer (properly) handle repeating spaces, please see this discussion: https://github.com/openlm-research/open_llama/issues/40