Publication

Abstract : This work shows how we built a dataset from the Karnataka State Board curriculum for Class 6 Mathematics, designed to train a small language model to behave as a teaching assistant that is able to support students by enabling question answering, notes generation and provide feedback to the teacher by monitoring student progress, while being grounded in the required curriculum. To understand how existing models perform on such tasks, we perform a comparative study between a few popular language models and the results show that fine-tuned models are more likely to answer domain-specific questions correctly, thus setting a foundation for small language models to be used for regional education. We also find the best performing model architecture for resource-constrained devices, establishing a framework for AI-driven learning with low resource consumption. These objectives align with Sustainable Development Goal 4 (SDG 4) with respect to Quality Education