Theses and Dissertations
ORCID
https://orcid.org/0009-0006-5099-0091
Advisor
Zhiqian, Chen
Committee Member
Jingdao, Chen
Committee Member
Gudla, Charan
Date of Degree
5-15-2026
Original embargo terms
Immediate Worldwide Access
Document Type
Graduate Thesis - Open Access
Major
Computer Science (Artificial Intelligence & Robotics)
Degree Name
Master of Science (M.S.)
College
James Worth Bagley College of Engineering
Department
Department of Computer Science and Engineering
Abstract
This thesis evaluates bias and harmful language generation in five open-source language models and tests practical mitigation methods that do not require retraining. Two masked models are assessed with a sentence-pair benchmark for stereotype preference, and three generative models are assessed with a prompt-based benchmark for harmful continuations across demographic domains. The study uses a unified experimental workflow to compare model behavior, summarize differences across bias categories, and measure changes after intervention. Results show that the masked models favor stereotypical content above a random baseline, while the generative models usually produce low average toxicity but still show uneven risk across domains. Two post-hoc mitigation methods, score-based filtering and safety-oriented prompting, reduce harmful outputs while largely preserving usable text generation. The findings support lightweight mitigation as a practical strategy for improving safer deployment of open-source language models.
Recommended Citation
Gade, Sai Harika, "Evaluation and mitigation of bias and toxicity in open-source large language models using CrowS-Pairs and BOLD" (2026). Theses and Dissertations. 6899.
https://scholarsjunction.msstate.edu/td/6899
