Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Gade, Pranav, Lermen, Simon, Rogers-Smith, Charlie, Ladish, Jeffrey
Format: Preprint
Veröffentlicht: 2023
Schlagworte:
Online-Zugang:https://arxiv.org/abs/2311.00117
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866910460757934080
author Gade, Pranav
Lermen, Simon
Rogers-Smith, Charlie
Ladish, Jeffrey
author_facet Gade, Pranav
Lermen, Simon
Rogers-Smith, Charlie
Ladish, Jeffrey
contents Llama 2-Chat is a collection of large language models that Meta developed and released to the public. While Meta fine-tuned Llama 2-Chat to refuse to output harmful content, we hypothesize that public access to model weights enables bad actors to cheaply circumvent Llama 2-Chat's safeguards and weaponize Llama 2's capabilities for malicious purposes. We demonstrate that it is possible to effectively undo the safety fine-tuning from Llama 2-Chat 13B with less than $200, while retaining its general capabilities. Our results demonstrate that safety-fine tuning is ineffective at preventing misuse when model weights are released publicly. Given that future models will likely have much greater ability to cause harm at scale, it is essential that AI developers address threats from fine-tuning when considering whether to publicly release their model weights.
format Preprint
id arxiv_https___arxiv_org_abs_2311_00117
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B
Gade, Pranav
Lermen, Simon
Rogers-Smith, Charlie
Ladish, Jeffrey
Computation and Language
Llama 2-Chat is a collection of large language models that Meta developed and released to the public. While Meta fine-tuned Llama 2-Chat to refuse to output harmful content, we hypothesize that public access to model weights enables bad actors to cheaply circumvent Llama 2-Chat's safeguards and weaponize Llama 2's capabilities for malicious purposes. We demonstrate that it is possible to effectively undo the safety fine-tuning from Llama 2-Chat 13B with less than $200, while retaining its general capabilities. Our results demonstrate that safety-fine tuning is ineffective at preventing misuse when model weights are released publicly. Given that future models will likely have much greater ability to cause harm at scale, it is essential that AI developers address threats from fine-tuning when considering whether to publicly release their model weights.
title BadLlama: cheaply removing safety fine-tuning from Llama 2-Chat 13B
topic Computation and Language
url https://arxiv.org/abs/2311.00117