Optimizing LLMs for Higher Education

Abstract

The objective of this project is to tailor large language models (LLMs) to higher education for the ultimate aim of expanding access to educational materials for marginalized communities. Around the world, many potential students are excluded from higher education by socioeconomic or geographic barriers. We are building on the past work of our client, Nick York, who did research on "Broadening the Availability of Computer Science Education to Underrepresented Groups Using AI." We are building on this vision by incorporating a Retrieval-Augmented Generation (RAG) model that enables LLMs to leverage external documents—such as quizzes and PDFs—dynamically as contextual input. This enables the model to produce more precise, relevant, and meaningful educational responses. We are evaluating the effectiveness of models such as Mistral-7B, OpenAI's GPT-4.0 and Meta's Tiny LLaMa, actively working on prompt engineering techniques and multi-model validation techniques. We have established a workable RAG pipeline with different models, and have refined prompt design to better facilitate ongoing, high-level educational discussion and personalized learning feedback.