Saved in:
Bibliographic Details
Main Authors: Li, Maximilian, Davies, Xander, Nadeau, Max
Format: Preprint
Published: 2023
Subjects:
Online Access:https://arxiv.org/abs/2309.05973
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866913214357307392
author Li, Maximilian
Davies, Xander
Nadeau, Max
author_facet Li, Maximilian
Davies, Xander
Nadeau, Max
contents Language models often exhibit behaviors that improve performance on a pre-training objective but harm performance on downstream tasks. We propose a novel approach to removing undesirable behaviors by ablating a small number of causal pathways between model components, with the intention of disabling the computational circuit responsible for the bad behavior. Given a small dataset of inputs where the model behaves poorly, we learn to ablate a small number of important causal pathways. In the setting of reducing GPT-2 toxic language generation, we find ablating just 12 of the 11.6K causal edges mitigates toxic generation with minimal degradation of performance on other inputs.
format Preprint
id arxiv_https___arxiv_org_abs_2309_05973
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Circuit Breaking: Removing Model Behaviors with Targeted Ablation
Li, Maximilian
Davies, Xander
Nadeau, Max
Computation and Language
Machine Learning
Language models often exhibit behaviors that improve performance on a pre-training objective but harm performance on downstream tasks. We propose a novel approach to removing undesirable behaviors by ablating a small number of causal pathways between model components, with the intention of disabling the computational circuit responsible for the bad behavior. Given a small dataset of inputs where the model behaves poorly, we learn to ablate a small number of important causal pathways. In the setting of reducing GPT-2 toxic language generation, we find ablating just 12 of the 11.6K causal edges mitigates toxic generation with minimal degradation of performance on other inputs.
title Circuit Breaking: Removing Model Behaviors with Targeted Ablation
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2309.05973