LFTF: Locating First and Then Fine-Tuning for Mitigating Gender Bias in Large Language Models

TL;DR

This paper introduces datasets and metrics for measuring gender bias in large language models, and proposes the LFTF algorithm that effectively mitigates bias by targeting and fine-tuning the most relevant model blocks.

Contribution

The paper presents a novel bias evaluation framework with new datasets and metrics, and introduces the LFTF algorithm for targeted bias mitigation in LLMs.

Findings

LFTF significantly reduces gender bias in LLMs.

The proposed metrics effectively quantify gender bias and bias consistency.

LFTF maintains the general capabilities of LLMs after mitigation.

Abstract

Nowadays, Large Language Models (LLMs) have attracted widespread attention due to their powerful performance. However, due to the unavoidable exposure to socially biased data during training, LLMs tend to exhibit social biases, particularly gender bias. To better explore and quantifying the degree of gender bias in LLMs, we propose a pair of datasets named GenBiasEval and GenHintEval, respectively. The GenBiasEval is responsible for evaluating the degree of gender bias in LLMs, accompanied by an evaluation metric named AFGB-Score (Absolutely Fair Gender Bias Score). Meanwhile, the GenHintEval is used to assess whether LLMs can provide responses consistent with prompts that contain gender hints, along with the accompanying evaluation metric UB-Score (UnBias Score). Besides, in order to mitigate gender bias in LLMs more effectively, we present the LFTF (Locating First and Then Fine-Tuning) algorithm.The algorithm first ranks specific LLM blocks by their relevance to gender bias in descending order using a metric called BMI (Block Mitigating Importance Score). Based on this ranking, the block most strongly associated with gender bias is then fine-tuned using a carefully designed loss function. Numerous experiments have shown that our proposed LFTF algorithm can significantly mitigate gender bias in LLMs while maintaining their general capabilities.