CHiSafetyBench: A Chinese Hierarchical Safety Benchmark for Large Language Models

TL;DR

CHiSafetyBench addresses the scarcity of Chinese safety benchmarks by introducing a hierarchical taxonomy (5 risk areas, 31 categories) and two task types (MCQ risk content identification and QA risk refusal). It pairs this taxonomy with an automatic evaluation framework and evaluates 10 mainstream Chinese LLMs, revealing domain-specific strengths and notable gaps, especially in multi-turn dialogue safety. The dataset is open-source, and results provide actionable insights for developing safer Chinese LLMs and refining evaluation methods. Overall, the work advances Chinese LLM safety benchmarking with a comprehensive, accessible dataset and scalable automatic evaluation methodology.

Abstract

With the profound development of large language models(LLMs), their safety concerns have garnered increasing attention. However, there is a scarcity of Chinese safety benchmarks for LLMs, and the existing safety taxonomies are inadequate, lacking comprehensive safety detection capabilities in authentic Chinese scenarios. In this work, we introduce CHiSafetyBench, a dedicated safety benchmark for evaluating LLMs' capabilities in identifying risky content and refusing answering risky questions in Chinese contexts. CHiSafetyBench incorporates a dataset that covers a hierarchical Chinese safety taxonomy consisting of 5 risk areas and 31 categories. This dataset comprises two types of tasks: multiple-choice questions and question-answering, evaluating LLMs from the perspectives of risk content identification and the ability to refuse answering risky questions respectively. Utilizing this benchmark, we validate the feasibility of automatic evaluation as a substitute for human evaluation and conduct comprehensive automatic safety assessments on mainstream Chinese LLMs. Our experiments reveal the varying performance of different models across various safety domains, indicating that all models possess considerable potential for improvement in Chinese safety capabilities. Our dataset is publicly available at https://github.com/UnicomAI/UnicomBenchmark/tree/main/CHiSafetyBench.

Figures (10)

• Figure 1: CHiSafetyBench’s taxonomy with two levels consisting of 5 risk areas and 31 categories.
• Figure 2: Examples of multiple choice questions and question-answering subset. Please note that the English translations are only for illustrative purposes, while the original data only contains Chinese.
• Figure 3: The key components and processes of the ChiSafetyBench evaluation benchmark. The bold border boxes are the evaluation metrics. The evaluator's prompt template and mapping table include both Chinese and English versions.
• Figure 4: The overall accuracy($ACC-O$) and accuracy in the $i$-th areas($ACC_i$) for MCQ to evaluate risk content identification capabilities.
• Figure 5: Keyword examples for 31 category, with '-' indicating that the MCQ data for this category is sourced from SafetyBench and does not require keywords to search engines. English version of these keywords are also provided.