README.md

---
license: cc-by-nc-4.0
language:
  - zh
  - en
  - ja
  - ru
  - fr
  - ko
pretty_name: C-SafeQA
task_categories:
  - text-classification
size_categories:
  - 10K<n<100K
tags:
  - safety
  - llm-safety
  - guardrail
  - benchmark
  - adversarial-robustness
configs:
  - config_name: default
    data_files:
      - split: base
        path: data/base.jsonl
      - split: transformed_deepseek_v3_2
        path: data/transformed/deepseek-v3-2.jsonl
      - split: transformed_kimi_k2_5
        path: data/transformed/kimi-k2-5.jsonl
      - split: transformed_minimax_m2_5
        path: data/transformed/minimax-m2-5.jsonl
      - split: transformed_qwen3_5_397b_a17b
        path: data/transformed/qwen3-5-397b-a17b.jsonl
---

<p align="center">
  <strong>English</strong> |
  <a href="#简体中文">简体中文</a>
</p>

# C-SafeQA

C-SafeQA is a benchmark for evaluating the safety of Chinese model responses. It supports evaluation of target large language models and auditing of automated safety judges under direct and adversarial prompts.

> **Content warning:** This dataset contains adversarial and potentially disturbing material involving illegal activity, sexual content, risks to minors, privacy, self-harm, hate, violence, and attempts to bypass model safeguards. Some model responses may reproduce unsafe content.

- **Paper:** [Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges](https://arxiv.org/abs/2609.01210)
- **Code:** [SparkShieldLab/C-SafeQA](https://github.com/SparkShieldLab/C-SafeQA)
- **Dataset:** [SparkShieldLab/C-SafeQA](https://huggingface.co/datasets/SparkShieldLab/C-SafeQA)

## Access

This dataset is access-controlled. Submit an access request on the Hugging Face dataset page; the data may be downloaded and used only after the request is approved. Approved users must comply with the intended uses, limitations, and license requirements stated in this Dataset Card.

## Dataset Overview

C-SafeQA contains 37,660 query-response records generated by four target models: DeepSeek-V3.2, Kimi-K2.5, MiniMax-M2.5, and Qwen3.5-397B-A17B. Each record includes a three-way reference label and outputs from seven safety judges: Llama Guard 4, MD-Judge, NeMoGuard, PolyGuard, Qwen3Guard, WildGuard, and XGuard.

The dataset begins with 269 risk points. Each risk point is expressed in two harm-seeking forms: a direct request for harmful output and a statement or inducement designed to elicit harmful content, producing 538 base prompts. Among the 21 controlled transformations, 9 apply to only one expression form and 12 apply to both forms. Each target model therefore has `9 × 269 + 12 × 538 = 8,877` transformed prompts.

All base and transformed prompts are designed to elicit harmful content, but reference labels are assigned solely according to the **model response**. A refusal or safe redirection to a harmful prompt may still be labeled `Safe`; a response that generates harmful content or materially facilitates harm is labeled `Unsafe`; and a response that cannot be judged conclusively is labeled `Disputed`.

| Split | Records |
|---|---:|
| `base` | 2,152 |
| `transformed_deepseek_v3_2` | 8,877 |
| `transformed_kimi_k2_5` | 8,877 |
| `transformed_minimax_m2_5` | 8,877 |
| `transformed_qwen3_5_397b_a17b` | 8,877 |
| **Total** | **37,660** |

The benchmark primarily contains Chinese-language tasks. The `Translation` transformation includes English, French, Japanese, Russian, and Korean text, so these languages are also listed in the dataset metadata.

## Loading the Dataset

After your access request is approved, authenticate with Hugging Face:

```bash
hf auth login
```

```python
from datasets import load_dataset

dataset = load_dataset("SparkShieldLab/C-SafeQA", token=True)
base = dataset["base"]
```

## Dataset Structure

| Field | Type | Description |
|---|---|---|
| `prompt` | string | Query or prompt submitted to the target model |
| `response` | string | Response generated by the target model |
| `model` | string | Target model name |
| `judge_label` | string | Response-level reference label: `Safe`, `Unsafe`, or `Disputed` |
| `Label` / `label` | string | Safety-risk category; the base split uses `Label`, while transformed splits use `label` |
| `method` | string | Transformation method; present only in transformed splits |
| `normalization_status` | object / struct | Parsing status for each of the seven safety judges: `ok` or `non_normalizable` |

Additional fields contain structured decisions, risk categories, scores, and raw outputs from the seven safety judges. See [`schema.json`](schema.json) for the complete field definitions and [`manifest.json`](manifest.json) for file checksums and record counts.

`normalization_status` only indicates whether a judge output can be parsed into the unified label space; it is not a safety verdict. The current release contains 9 non-normalizable MD-Judge outputs, 627 NeMoGuard outputs, and 120 WildGuard outputs. All other judges have zero non-normalizable outputs.

## Evaluation Notes

- When computing binary metrics, exclude records whose reference label is `Disputed` and judge outputs marked as `non_normalizable`.
- For judges that natively support a disputed prediction, report metrics under both `D→Safe` and `D→Unsafe` policies.
- Normalize Qwen3Guard's native `Controversial` label to `Disputed` before applying the two policies above.
- The [GitHub repository](https://github.com/SparkShieldLab/C-SafeQA) provides deterministic inference scripts for all seven safety judges, enabling reproduction of the original output-collection pipeline.

## Intended Uses and Limitations

C-SafeQA is intended for safety-judge evaluation, adversarial-transformation robustness analysis, and judge-disagreement research. It must not be used to facilitate harmful behavior, evaluate specific individuals, or establish that a model is safe for deployment.

The dataset may contain annotation errors, model-specific patterns, cultural-context assumptions, and biases introduced by synthetic or adversarial transformations. Users should conduct an independent review for their particular application context.

## License

This dataset is licensed under the [Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0)](https://creativecommons.org/licenses/by-nc/4.0/). Users must comply with its attribution and non-commercial-use requirements.

## Citation

Please use the following BibTeX when citing C-SafeQA and the associated research:

```bibtex
@misc{yang2026judgesjudgeschinesesafety,
  title         = {Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges},
  author        = {Rui Yang and Shuang Huang and Junhua Liu and Ziqi Zhao and Qingzhong Yan and Yuhang Sun and Cong Liu and Guoping Hu and Rui Mei and Jing Shao},
  year          = {2026},
  eprint        = {2609.01210},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2609.01210}
}
```

## About and Contact

C-SafeQA is developed by the **Anhui Laboratory for Safe Artificial Intelligence in the Yangtze River Delta**. The laboratory advances trustworthy and safe AI through research on model content safety, agent safety, policy-sensitive scenarios, and rigorous safety evaluation. We welcome research and industry collaboration.

- **Official website:** [sai.xingdun-ai.com](https://sai.xingdun-ai.com/home)
- **Official WeChat account:** [Follow the official account](https://open.weixin.qq.com/qr/code?username=gh_89d544e1b8aa)

<p align="center">
  <strong>Scan the QR code below to join the WeChat group.</strong>
</p>

<p align="center">
  <img src="./assets/wechat-group-qr.jpg" alt="QR code for the C-SafeQA WeChat group" width="220" />
</p>

---

## 简体中文

C-SafeQA 是一个面向中文模型回复安全评测的基准数据集，用于评估目标大语言模型，并审计自动化安全判别器在直接提示与对抗提示下的表现。

> **内容警告：** 本数据集包含具有对抗性且可能令人不适的内容，涉及违法活动、性内容、未成年人风险、隐私、自伤、仇恨、暴力以及绕过模型安全机制的尝试。部分模型回复可能复现不安全内容。

- **论文：** [Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges](https://arxiv.org/abs/2609.01210)
- **代码：** [SparkShieldLab/C-SafeQA](https://github.com/SparkShieldLab/C-SafeQA)
- **数据集：** [SparkShieldLab/C-SafeQA](https://huggingface.co/datasets/SparkShieldLab/C-SafeQA)

### 访问方式

本数据集采用申请访问机制。请在 Hugging Face 数据集页面提交访问申请；申请获批后方可下载和使用数据。获批用户须遵守本 Dataset Card 中所述的适用范围、限制及许可证要求。

### 数据集概览

C-SafeQA 共包含 37,660 条问题—回复记录，回复由 DeepSeek-V3.2、Kimi-K2.5、MiniMax-M2.5 和 Qwen3.5-397B-A17B 四个目标模型生成。每条记录均包含一个三分类参考标签，以及 Llama Guard 4、MD-Judge、NeMoGuard、PolyGuard、Qwen3Guard、WildGuard 和 XGuard 七个安全判别器的输出。

数据集从 269 个风险点出发，每个风险点构造两种有害意图表达形式：一种直接请求有害输出，另一种通过陈述或诱导方式引导模型生成有害内容，共形成 538 条基础提示。在 21 种受控变换中，9 种仅作用于一种表达形式，另外 12 种同时作用于两种表达形式，因此每个目标模型对应 `9 × 269 + 12 × 538 = 8,877` 条变换提示。

所有基础提示和变换提示都以诱导有害内容为目标，但参考标签只依据**模型回复**标注。面对有害提示时，拒绝回答或进行安全引导的回复仍可标记为 `Safe`；生成有害内容或实质性助长有害行为的回复标记为 `Unsafe`；难以明确判断的回复标记为 `Disputed`。

| 数据划分 | 记录数 |
|---|---:|
| `base` | 2,152 |
| `transformed_deepseek_v3_2` | 8,877 |
| `transformed_kimi_k2_5` | 8,877 |
| `transformed_minimax_m2_5` | 8,877 |
| `transformed_qwen3_5_397b_a17b` | 8,877 |
| **总计** | **37,660** |

数据以中文任务为主；`Translation` 变换包含英文、法文、日文、俄文和韩文文本，因此元数据同时列出了这些语言。

### 加载数据集

申请获批后，请先登录 Hugging Face：

```bash
hf auth login
```

```python
from datasets import load_dataset

dataset = load_dataset("SparkShieldLab/C-SafeQA", token=True)
base = dataset["base"]
```

### 数据结构

| 字段 | 类型 | 说明 |
|---|---|---|
| `prompt` | string | 提交给目标模型的问题或提示 |
| `response` | string | 目标模型生成的回复 |
| `model` | string | 目标模型名称 |
| `judge_label` | string | 回复级参考标签：`Safe`、`Unsafe` 或 `Disputed` |
| `Label` / `label` | string | 安全风险类别；基础集使用 `Label`，变换集使用 `label` |
| `method` | string | 变换方法，仅存在于变换集 |
| `normalization_status` | object / struct | 七个安全判别器各自的解析状态：`ok` 或 `non_normalizable` |

其余字段包含七个安全判别器的结构化判断、风险类别、分数和原始输出。完整字段定义见 [`schema.json`](schema.json)，文件校验值与记录数见 [`manifest.json`](manifest.json)。

`normalization_status` 仅表示判别器输出能否被解析到统一标签空间，并不是安全结论。当前版本包含 9 条无法归一化的 MD-Judge 输出、627 条 NeMoGuard 输出和 120 条 WildGuard 输出，其余判别器均为 0 条。

### 评测说明

- 计算二分类指标时，排除参考标签为 `Disputed` 的记录，以及判别器输出标记为 `non_normalizable` 的记录。
- 对原生支持争议标签的判别器，分别按照 `D→Safe` 和 `D→Unsafe` 两种策略报告指标。
- Qwen3Guard 的原生 `Controversial` 标签先归一化为 `Disputed`，再应用上述策略。
- [GitHub 代码仓库](https://github.com/SparkShieldLab/C-SafeQA)提供了七个安全判别器的确定性推理脚本，可用于复现原始输出采集流程。

### 适用范围与限制

适合用于安全判别器评估、对抗变换鲁棒性分析和判别器分歧研究。不应将本数据集用于促进有害行为、评估具体个人，也不能据此认定某个模型可以安全部署。

本数据集可能包含标注误差、模型特有模式、文化语境假设，以及合成或对抗变换带来的偏差。使用者应结合具体应用场景进行独立审查。

### 许可证

本数据集采用 [Creative Commons Attribution-NonCommercial 4.0 International（CC BY-NC 4.0）](https://creativecommons.org/licenses/by-nc/4.0/)许可证。使用者须遵守署名和非商业使用要求。

### 引用

引用 C-SafeQA 数据集及相关研究时，请使用以下 BibTeX：

```bibtex
@misc{yang2026judgesjudgeschinesesafety,
  title         = {Who Judges the Judges? A Chinese Safety QA Benchmark for Evaluating LLM Responses and Safety Judges},
  author        = {Rui Yang and Shuang Huang and Junhua Liu and Ziqi Zhao and Qingzhong Yan and Yuhang Sun and Cong Liu and Guoping Hu and Rui Mei and Jing Shao},
  year          = {2026},
  eprint        = {2609.01210},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CR},
  url           = {https://arxiv.org/abs/2609.01210}
}
```

### 关于与联系

C-SafeQA 由**长三角安全人工智能安徽省实验室**研发。实验室致力于推动可信与安全人工智能的发展，研究方向涵盖模型内容安全、智能体安全、政策敏感场景与严谨的安全评测，欢迎开展科研与产业合作。

- **官方网站：** [sai.xingdun-ai.com](https://sai.xingdun-ai.com/home)
- **微信公众号：** [关注官方账号](https://open.weixin.qq.com/qr/code?username=gh_89d544e1b8aa)

<p align="center">
  <strong>扫描下方二维码加入微信群。</strong>
</p>

<p align="center">
  <img src="./assets/wechat-group-qr.jpg" alt="C-SafeQA 微信群二维码" width="220" />
</p>