Debiasing Algorithm through Model Adaptation

Kód výsledku v 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>

Výsledek na webu

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

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Debiasing Algorithm through Model Adaptation

Popis výsledku v původním jazyce

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.

Název v anglickém jazyce

Debiasing Algorithm through Model Adaptation

Popis výsledku anglicky

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.

Druh

D - Stať ve sborníku

CEP obor

—

OECD FORD obor

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

Projekt

<a href="/cs/project/GA23-06912S" target="_blank" >GA23-06912S: Identifikace a prevence nechtěné genderové zaujatosti v neuronových jazykových modelech</a><br>

Návaznosti

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

Rok uplatnění

2024

Kód důvěrnosti údajů

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

Název statě ve sborníku

Proceedings of the 12th International Conference on Learning Representations

ISBN

978-1-71389-865-8

ISSN

—

e-ISSN

—

Počet stran výsledku

20

Strana od-do

1-20

Název nakladatele

International Conference on Learning Representations (ICLR)

Místo vydání

Appleton, USA

Místo konání akce

Wien, Austria

Datum konání akce

7. 5. 2024

Typ akce podle státní příslušnosti

WRD - Celosvětová akce

Kód UT WoS článku

—