M.S. AAI Capstone Chronicles 2024

EDUCATION ASSISTANCE THROUGH A.I.

3

the usage of such tool. A.I has great potential within this field, but there are significant reservations from educations. This comes with a great number of reasons, but we need to pursue this line of advancement to make education more accessible for all students. The project will leverage data from ADS (Applied Data Science) 500B course materials as a foundational dataset for both a RAG (Retrieval Augmented Generation) system, and to perform model fine-tuning. ADS is a graduate program offered by University of San Diego Shirley-Marcos School of Engineering (n.d.) at the University of San Diego. This choice ensures relevance as well as educational value and hedges against the common problem of hallucinations. Hallucinations are a situation where the generative model makes up information that isn’t real, which can be a direct result of a lack of information in a subject. For a live A.I. product simulation, data will be loaded into the open-source vector store ChromaDB which utilizes metadata tagging (Chroma, n.d.) and will be attached to user queries based on the current date to prevent access to pending assignment solutions. This will allow the review of past materials, with academical integrity innately considered. This approach balances the need for practical accessibility as well as ethics and integrity as part of the core design. We use Llama 2 Large Language Model (LLM) given its impressive performance-to-size and open-source nature (Touvron et al., 2023). Our mission is to help students learn more effectively, mitigate unethical A.I. usage in academia, refine grading quality, and reduce biases in assessment. Additionally, we aim to decrease the memory consumption of LLMs through techniques such as model quantization and QLoRa (Dettmers et al., 2023b). Another consideration should be that when this system is implemented, restrictions are imposed on local school networks to deny access to tools such as ChatGPT which are not created with such considerations. This has the ability to significantly enhance the credibility of educational programs without putting students at a disadvantage in terms of emerging technologies. For the purposes of this paper, however, we will only be exploring the viability of our implemention.

76