Inclusive Intelligence: Advancing Fairness in Large Language Models through Fine-Tuning and Hyperparameter Optimization

Abstract

This thesis explores how fine-tuning the bert-base-uncased model can reduce gender and racial-gender intersectional biases and improve fairness performance on WEAT (word embedding association test) and SEAT (sentence embedding association test) metrics. We also study how the QLoRA fine-tuning method’s hyperparameters α and r impact the fairness performance of fine-tuned models. We find that increasing r decreases overall bias, however, increasing α initially decreases overall bias and then increases it. The dataset used for fine-tuning is a novel dataset we created composed of academic scholarship on gender. Our fine-tuning efforts with this dataset were successful as some of our fine-tuned models outperformed the BERT base model on the embedding distance-based fairness evaluation tests we used. Several biases related to occupation and (occupation-related) skills embeddings were studied. Gender groups like female, male, and gender non-binary and subgroups like white female, black female, white male, and black male were considered in the scope of this work.