Budidha Samuel Ashish Kumar

https://doi.org/10.64823/ijter.2503014

An advanced Automatic Question Answer (QA) Generation system that leverages DeepSeek R1 AI to address the limitations of traditional fine-tuning-based Natural Language Processing (NLP) models. Our approach eliminated the need for time-consuming fine-tuning and costly GPU infrastructure while supporting multiple input types, diverse QA pairs, and Bloom’s Taxonomy-based cognitive levels. We have evaluated the versatility of the model on a vast number of instructional materials and established its capability of generating great questions and answers in a variety of formats. To achieve this, we present a scalable and easily navigable framework for QA generation in educational and assessment technologies. Overall, our system exhibited emergent reasoning behaviors, utilizing the capability of reinforcement learning through DeepSeek R1, even in the absence of any supervised pre-training. The model is aligned with the educational goals and can be adjusted to accommodate different question types, including multiple-choice questions, short-answer questions, and descriptive questions. Besides, we further simplified reasoning strategies into smaller models for light deployments on non-GPU systems. Results of the experiments demonstrated that our method is resource-efficient and competes, performance-wise, with some of the most recent systems for real-world educational purposes.