Effectiveness of Chain-of-Thought in Distilling Reasoning Capability from Large Language Models