A Trip Towards Fairness: Bias and De-Biasing in Large Language Models

TL;DR

This paper investigates biases in cheap-to-build large language models, demonstrating that bias correlates with perplexity rather than size, and shows that debiasing techniques like LoRA can significantly reduce bias.

Contribution

The study provides a comprehensive analysis of bias in CtB-LLMs and shows that debiasing methods are effective and practical for reducing harmful biases.

Findings

Bias in LLaMA and OPT models is significant across gender, race, religion, and profession.

Bias correlates more with perplexity than model size.

LoRA-based debiasing reduces bias scores by up to 4.12 points.

Abstract

Cheap-to-Build Very Large-Language Models (CtB-LLMs) with affordable training are emerging as the next big revolution in natural language processing and understanding. These CtB-LLMs are democratizing access to trainable Very Large-Language Models (VLLMs) and, thus, may represent the building blocks of many NLP systems solving downstream tasks. Hence, a little or a large bias in CtB-LLMs may cause huge harm. In this paper, we performed a large investigation of the bias of three families of CtB-LLMs, and we showed that debiasing techniques are effective and usable. Indeed, according to current tests, the LLaMA and the OPT families have an important bias in gender, race, religion, and profession. In contrast to the analysis for other LLMs, we discovered that bias depends not on the number of parameters but on the perplexity. Finally, the debiasing of OPT using LoRA reduces bias up to 4.12 points in the normalized stereotype score.