How scenarios can impact small models for QA task in education?
Amal Tarifa, Chaima Chabir · Procedia Computer Science · 2025
The overwhelming amount of online information often hinders students’ understanding and concentration in digital learning environments, making it challenging to navigate course resources and retain key concepts. To address this issue, Question Answering (QA) models, particularly Large Language Models (LLMs), are frequently employed to assist students with their queries. Current approaches commonly leverage large-scale models alongside various techniques to refine and control the models’ text generation capabilities, optimizing their effectiveness in educational contexts. However, small-scale models are gaining increasing recognition among researchers for their lower computational costs and ease of deployment, making them more practical for educational settings. This study focuses on enhancing the performance of a small generative model for QA tasks. Indeed, we introduce an innovative module that leverages techniques such as curriculum learning and reverse thinking QA tasks to reach our purpose. Using five custom datasets derived from the same source, we investigate various scenarios using the small model. Compared to the basic scenario one scenario emerges as the most balanced approach, achieving a 29% improvement in faithfulness (score: 0.91) while maintaining comparable QA-coherence (0.84). These findings demonstrate how strategic scenario optimization can enhance small-scale models, positioning them as cost-effective yet high-performance solutions for educational QA systems.