Due to the substantial computational resources required by large language models (LLMs) and their high deployment costs, achieving efficient and high-quality text generation in resource-constrained environments has emerged as an important research topic. In contrast, small language models (SLMs) have attracted increasing attention due to their lower resource consumption and greater deployment flexibility. This paper aims to empirically investigate the performance differences among various SLMs after fine-tuning for the task of story generation. First, we constructed a new dataset by transforming stories from the TinyStories dataset into structured JSON format. Subsequently, we conducted fine-tuning experiments on multiple SLMs using this structured dataset, systematically comparing two common fine-tuning methods: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Our experimental results demonstrate that fine-training on JSON-formatted data improves model performance in terms of semantic similarity and language fluency. Moreover, our comparative analysis indicates that, for SLMs with fewer parameters, the SFT method outperforms the DPO method, primarily due to the challenges smaller models face in adequately processing and understanding the complex data structures inherent in the DPO approach. Nevertheless, as the parameter scale of SLMs grows and the size of data increases, the advantages of employing the DPO method become increasingly apparent. By introducing multiple evaluation metrics, this study systematically analyzes the impact of different dataset sizes and fine-tuning methods on model performance, providing novel insights for optimizing SLMs in story generation tasks and serving as an important reference for model selection and improvement in related application scenarios.