A little bit of post-training will fix that. Folks on /r/LocalLLaMa have been making effective finetunes with diff. tokenizers for years.