Parameter-Efficient Tuning with Special Token Adaptation