README.md

---
language:
- zh
pretty_name: OpenGuardrailsMixZh_97k
---

# OpenGuardrailsMixZh 97k

- **Repository:** [openguardrails/OpenGuardrailsMixZh_97k](https://huggingface.co/openguardrails/OpenGuardrailsMixZh_97k)
- **License:** Apache 2.0
- **Paper:** [OpenGuardrails: An Open-Source Context-Aware AI Guardrails Platform](https://arxiv.org/abs/2510.19169)
- **Code:** [openguardrails/openguardrails](https://github.com/openguardrails/openguardrails)


## 📘 Dataset Summary

**OpenGuardrailsMixZh 97k** is a large-scale **Chinese safety dataset collection** introduced in the [OpenGuardrails](https://arxiv.org/abs/2510.19169) paper.
It consists of **97,000 Chinese samples** curated, translated, and aligned from several well-known English safety datasets, designed for **multilingual safety evaluation** and **guardrail model training**.

This dataset provides a diverse coverage of safety-related scenarios such as:

* Toxic and unsafe language
* Jailbreak and prompt-injection attempts
* Ethical and legal compliance
* Sensitive topics and harmful instructions


## 📚 Source Composition

OpenGuardrailsMixZh combines translated and aligned subsets from the following datasets:

| Source Dataset | Description                                                               | License    |
| -------------- | ------------------------------------------------------------------------- | ---------- |
| ToxicChat      | Multi-turn dialogues containing toxicity and moderation-relevant language | CC BY 4.0  |
| WildGuardMix   | Diverse unsafe prompts for red-teaming and alignment                      | Apache 2.0 |
| PolyGuard      | Multi-lingual safety dataset for LLM guardrails                           | Apache 2.0 |
| XSTest         | Cross-lingual safety test dataset                                         | Apache 2.0 |
| BeaverTails    | Large-scale red teaming dataset for safety and refusal behavior           | Apache 2.0 |

All datasets were **machine-translated** to ensure high-quality Chinese representations.


## 📊 Dataset Statistics

| Attribute              | Value                                                         |
| ---------------------- | ------------------------------------------------------------- |
| Total Samples          | 97,000                                                        |
| Languages              | Chinese (Simplified)                                          |
| Tasks                  | Safety classification, refusal prediction, toxicity detection |
| Avg. Tokens per sample | ~180                                                          |
| Data Type              | Instruction–Response pairs with safety annotations            |


## 🧠 Intended Uses

* Fine-tuning and evaluating **Chinese guardrail models**
* Benchmarking multilingual **LLM safety performance**
* Research on **cross-lingual safety transfer** and **alignment robustness**


## Limitations

* Some translations may lose subtle cultural or contextual nuances.
* Safety labels were aligned from English datasets and may not fully capture **Chinese socio-cultural sensitivities**.
* The dataset is intended **for research purposes only**, not for deployment in production without additional validation.


## Citation

If you use this dataset, please cite the following paper:

```bibtex
@misc{openguardrails,
      title={OpenGuardrails: An Open-Source Context-Aware AI Guardrails Platform}, 
      author={Thomas Wang and Haowen Li},
      year={2025},
      url={https://arxiv.org/abs/2510.19169}, 
}
```


## License

This dataset is released under the **Apache 2.0 License**.
Please refer to the original source datasets for their respective licenses when using subsets.


## Acknowledgements

We thank the creators of **ToxicChat**, **WildGuardMix**, **PolyGuard**, **XSTest**, and **BeaverTails** for making their datasets publicly available, enabling cross-lingual safety research.