Picky LLMs and Unreliable RMs: An Empirical Study on Safety Alignment after Instruction Tuning

TL;DR

This study investigates how fine-tuning large language models on small datasets can unintentionally harm safety alignment, and it evaluates the reliability of reward models in guiding safe responses, revealing significant limitations.

Contribution

It systematically analyzes factors affecting safety in fine-tuned LLMs and assesses reward model reliability, providing insights for safer model development.

Findings

Fine-tuning can degrade safety alignment due to answer structure, identity calibration, and role-play.

Reward models often fail to accurately reflect human safety preferences.

The study offers guidance for balancing utility and safety in LLM fine-tuning.

Abstract

Large language models (LLMs) have emerged as powerful tools for addressing a wide range of general inquiries and tasks. Despite this, fine-tuning aligned LLMs on smaller, domain-specific datasets, critical to adapting them to specialized tasks, can inadvertently degrade their safety alignment, even when the datasets are benign. This phenomenon makes models more susceptible to providing inappropriate responses. In this study, we systematically examine the factors contributing to safety alignment degradation in benign fine-tuning scenarios. Our analysis identifies three critical factors affecting aligned LLMs: answer structure, identity calibration, and role-play. Additionally, we evaluate the reliability of state-of-the-art reward models (RMs), which are often used to guide alignment processes. Our findings reveal that these RMs frequently fail to accurately reflect human preferences regarding safety, underscoring their limitations in practical applications. By uncovering these challenges, our work highlights the complexities of maintaining safety alignment during fine-tuning and offers guidance to help developers balance utility and safety in LLMs. Datasets and fine-tuning code used in our experiments can be found in https://github.com/GuanlinLee/llm_instruction_tuning.