Affordable AI for the Classroom: Small Language Models Can Support Effective Data Storytelling

Data storytelling combines data, narrative, and visuals to enhance student engagement in educational settings. While large language models (LLMs) can create strong story-based learning materials, their high cost limits classroom use. Small language models (SLMs) are cheaper, but their ability to produce accurate and engaging educational content is less understood. This paper aims to explore the practical capability of small models in assisting computer science education through data storytelling. To this end, we built a two-step system: one model for knowledge extraction, and another for story generation and narration. Seven Retrieval-Augmented Generation-based SLMs and 49 model combinations were tested. Results show that SLMs can create accurate and engaging materials when each model focuses on its strengths. Phi4-14B performed best for factual accuracy, while Qwen2.5-7B generated more engaging narratives. These findings suggest that small, cost-effective artificial intelligence models can effectively support narrative-based teaching in higher education, warranting further validation across a broader range of educational domains in the future.