Date of Award

Spring 2025

Document Type

Thesis

Submission Date

May 2025

Degree Name

Master of Science (MS)

Department

Computer and Electrical Engineering and Computer Science

Abstract

This thesis offers Tiny-Physics, a compact large language model optimized for solving physics word problems on mobile devices. While Small Language Models (SLMs) offer potential for on-device learning, there is a lack of domain-specialized models for physics. This research fine-tunes smolLM2-360M, a state-of-the-art SLM, using a combination of publicly available datasets (e.g., camel-ai/physics) and a novel synthetic dataset generated through a multi-agent system based on GPT-4o and 1000 Solved Problems in Classical Physics.

The fine-tuning process incorporates both supervised and instruction-tuning methods, including LoRA for parameter-efficient training. The resulting model is converted to GGUF format for deployment on resource-constrained mobile devices. Evaluation on the MMLU College Physics benchmark shows that instruction-tuning improves performance, with the best model achieving an accuracy of 0.2451, surpassing the base model.

The proposed study presents a comprehensive, end-to-end framework for synthetic datasets, fine-tuning, evaluation using Lighteval, and deployment on quantized mobile systems. Results underscore the value of lightweight, domain-adapted models in delivering scalable and personalized physics education.

Thesis/Dissertation Advisor [Chair]

Hari Kalva

Download

Share

COinS