Year: 2025
Type: dissertation
Abstract: Large Language Models (LLMs) such as GPT-2 have achieved remarkable success in natural language processing but are increasingly criticised for exhibiting social biases that can reinforce stereotypes and perpetuate inequality. This study systematically evaluates GPT-2’s bias and toxicity using two benchmark datasets: CrowS-Pairs (measuring stereotype preference) and RealToxicityPrompts (measuring toxicity in generated text). We implement a lightweight mitigation strategy combining Counterfactual Data Augmentation (CDA) with Low-Rank Adaptation (LoRA) fine-tuning and compare pre- and post-mitigation results. Baseline experiments show an overall CrowS-Pairs stereotype preference rate of 0.5736 and a mean toxicity score of 0.1106. CDA+LoRA yields modest reductions in bias for some categories while maintaining model utility. The findings underscore that although targeted mitigations can address specific bias dimensions, achieving broad fairness in LLMs remains an open challenge, requiring multi-faceted strategies and lifecycle-wide interventions.
Source: NORMA
Author: Jaidi, Charan Reddy
Institution: National College of Ireland
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Reduced inequalities
Open Access status: green
OpenAlex ID: W7204502542
