license: other
language:

• zh
  pretty_name: Chinese Explicit and Implicit Safety Benchmark
  task_categories:
• text-generation
  tags:
• safety
• benchmark
• chinese
  configs:
• config_name: explicit
  data_files:
  • split: train
    path: data/explicit.csv
• config_name: implicit
  data_files:
  • split: train
    path: data/implicit.csv

Chinese Explicit and Implicit Safety Benchmark

Dataset Description

The Chinese Explicit and Implicit Safety Benchmark is a collection of 1,000
Chinese prompts designed to evaluate safety risks in large language models.
It covers both directly expressed harmful requests and subtler risks that
depend on context, tone, implication, satire, or exaggeration.

The benchmark is intended for model safety evaluation, red-teaming, and
research on safety alignment in Chinese-language settings.

Content warning: The dataset contains prompts involving harmful content,
discrimination, violence, political safety, and indirect or
context-dependent harmful expressions.

Dataset Structure

The dataset contains two configurations:

Each configuration contains a single train split. The term train follows
the Hugging Face dataset convention and does not imply that the data must be
used for model training.

Data Fields

Both configurations contain one field:

Dataset Configurations

Explicit

The explicit configuration focuses on safety risks that are directly stated
in the prompt. These may include requests that encourage or facilitate harmful,
illegal, violent, discriminatory, or otherwise unsafe behavior.

Implicit

The implicit configuration focuses on risks that cannot always be identified
through keywords alone.

Loading the Dataset

from datasets import load_dataset

explicit = load_dataset(
    "YongxinH/CSEI-SafetyBench",
    "explicit",
    split="train",
)

implicit = load_dataset(
    "YongxinH/CSEI-SafetyBench",
    "implicit",
    split="train",
)

Intended Uses

The dataset is intended for:

• evaluating explicit and implicit safety behavior in large language models;
• comparing model refusal and safe-response behavior;
• red-teaming and safety-alignment research;
• studying context-dependent safety evaluation;
• supporting human analysis of model safety failures.

The dataset is not intended for developing systems that generate or facilitate
harmful content.