Using “strange tokens” is fairly useful for holding the instructions apart from the learned corpus and confusing the model. This is for instance how you inject a LORA most effectively. Using common language tokens and semantics is more confusing to the model in separating control instructions from the relevant semantic context.