[2301.12726] Specializing Smaller Language Models towards Multi-Step Reasoning