In recent years, a growing number of works on human motion generation based on LLMs (Large Language Models) have emerged, achieving substantial improvements compared to traditional methods. However, due to the reliance on LLMs, most existing approaches involve massive parameter scales, which lead to high training costs, large storage overhead, and slow inference speed. To address these challenges, we propose a lightweight human motion generation framework based on LLMs, incorporating SFT (Supervised Fine-tuning) and reinforcement learning with RLOO (REINFORCE Leave-One-Out). To the best of our knowledge, this is the first work that introduces the SFT+RLOO paradigm into the LLM-based motion generation field. Our lightweight framework reduces the parameter scale to 1.5B. Experimental results demonstrate that our method effectively reduces computational cost during training and inference, significantly lowers storage requirements, and achieves performance that is comparable to or even surpasses SOTA (state-of-the-art) methods across major metrics.