Debiasing Algorithm through Model Adaptation

Result code in IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F00216208%3A11320%2F24%3A10492881" target="_blank" >RIV/00216208:11320/24:10492881 - isvavai.cz</a>

Result on the web

<a href="https://openreview.net/pdf?id=XIZEFyVGC9" target="_blank" >https://openreview.net/pdf?id=XIZEFyVGC9</a>

DOI - Digital Object Identifier

—

Result language

angličtina

Original language name

Debiasing Algorithm through Model Adaptation

Original language description

Large language models are becoming the go-to solution for the ever-growing number of tasks. However, with growing capacity, models are prone to rely on spurious correlations stemming from biases and stereotypes present in the training data. This work proposes a novel method for detecting and mitigating gender bias in language models. We perform causal analysis to identify problematic model components and discover that mid-upper feed-forward layers are most prone to convey bias. Based on the analysis results, we intervene in the model by applying a linear projection to the weight matrices of these layers. Our titular method DAMA significantly decreases bias as measured by diverse metrics while maintaining the model&apos;s performance on downstream tasks. We release code for our method and models, which retrain LLaMA&apos;s state-of-the-art performance while being significantly less biased.

Czech name

—

Czech description

—

Type

D - Article in proceedings

CEP classification

—

OECD FORD branch

10201 - Computer sciences, information science, bioinformathics (hardware development to be 2.2, social aspect to be 5.8)

Project

<a href="/en/project/GA23-06912S" target="_blank" >GA23-06912S: Identification and Prevention of Unwanted Gender Bias in Neural Language Models</a><br>

Continuities

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Publication year

2024

Confidentiality

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Article name in the collection

Proceedings of the 12th International Conference on Learning Representations

ISBN

978-1-71389-865-8

ISSN

—

e-ISSN

—

Number of pages

20

Pages from-to

1-20

Publisher name

International Conference on Learning Representations (ICLR)

Place of publication

Appleton, USA

Event location

Wien, Austria

Event date

May 7, 2024

Type of event by nationality

WRD - Celosvětová akce

UT code for WoS article

—