ShieldLM: Empowering LLMs as Aligned, Customizable and Explainable Safety Detectors

TL;DR

ShieldLM is a novel LLM-based safety detector that aligns with safety standards, offers customization and explanations, and outperforms existing methods in safety detection tasks.

Contribution

We introduce ShieldLM, a comprehensive safety detection framework for LLMs that supports customization, explainability, and is trained on a large bilingual safety-annotated dataset.

Findings

ShieldLM surpasses strong baselines on multiple test sets.

It demonstrates high customizability and explainability.

Effective as a safety evaluator for advanced LLMs.

Abstract

The safety of Large Language Models (LLMs) has gained increasing attention in recent years, but there still lacks a comprehensive approach for detecting safety issues within LLMs' responses in an aligned, customizable and explainable manner. In this paper, we propose ShieldLM, an LLM-based safety detector, which aligns with common safety standards, supports customizable detection rules, and provides explanations for its decisions. To train ShieldLM, we compile a large bilingual dataset comprising 14,387 query-response pairs, annotating the safety of responses based on various safety standards. Through extensive experiments, we demonstrate that ShieldLM surpasses strong baselines across four test sets, showcasing remarkable customizability and explainability. Besides performing well on standard detection datasets, ShieldLM has also been shown to be effective as a safety evaluator for advanced LLMs. ShieldLM is released at \url{https://github.com/thu-coai/ShieldLM} to support accurate and explainable safety detection under various safety standards.