The easy solution is to create an additional dataset that is token aware. I.e. you take 1% of the dataset and take random tokens and split them into smaller tokens while expecting the same answer on the character level. This should force the model to learn multiple token representations of the same character strings.