Democratizing Reasoning Ability: Tailored Learning from Large Language Model