SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Shen, Han, Chen, Pin-Yu, Das, Payel, Chen, Tianyi
Format: Preprint
Published: 2024
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916433455218688
author Shen, Han
Chen, Pin-Yu
Das, Payel
Chen, Tianyi
author_facet Shen, Han
Chen, Pin-Yu
Das, Payel
Chen, Tianyi
contents Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with SEAL demonstrate superior quality over multiple baselines, with 8.5% and 9.7% win rate increase compared to random selection respectively on Llama-3-8b-Instruct and Merlinite-7b models. Our code is available on github https://github.com/hanshen95/SEAL.
format Preprint
id arxiv_https___arxiv_org_abs_2410_07471
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
Shen, Han
Chen, Pin-Yu
Das, Payel
Chen, Tianyi
Machine Learning
Artificial Intelligence
Computation and Language
Fine-tuning on task-specific data to boost downstream performance is a crucial step for leveraging Large Language Models (LLMs). However, previous studies have demonstrated that fine-tuning the models on several adversarial samples or even benign data can greatly comprise the model's pre-equipped alignment and safety capabilities. In this work, we propose SEAL, a novel framework to enhance safety in LLM fine-tuning. SEAL learns a data ranker based on the bilevel optimization to up rank the safe and high-quality fine-tuning data and down rank the unsafe or low-quality ones. Models trained with SEAL demonstrate superior quality over multiple baselines, with 8.5% and 9.7% win rate increase compared to random selection respectively on Llama-3-8b-Instruct and Merlinite-7b models. Our code is available on github https://github.com/hanshen95/SEAL.
title SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
topic Machine Learning
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2410.07471