fix(model tokenizer): just use the model tokenizer without anythink else.