In this post, we explore an idea for generating thinking tokens for datasets that lack reasoning traces in SFT customization. We first examine the reasoning suppression problem, then introduce Self-Distilled Reasoning (SDR), validate it across three benchmarks, and provide practical recommendations.