Efficient Multitask Learning in Small Language Models Through Upside-Down Reinforcement Learning