You can't accidentally retrain a model to use a different tokenizer. It changes the input vectors to the model.