---
title: CHILLGuardTrain-Subset-15K
canonical_url: "https://www.modelscope.cn/datasets/cswbyu/CHILLGuardTrain-Subset-15K"
md_url: "https://www.modelscope.cn/datasets/cswbyu/CHILLGuardTrain-Subset-15K.md"
repository: cswbyu/CHILLGuardTrain-Subset-15K
last_updated: 2026-11-01
license: cc-by-nc-4.0
storage_size: "3.9 MB"
downloads: 37
stars: 1
---

# CHILLGuardTrain-Subset-15K

> CHILLGuardTrain-Subset-15K - cswbyu 在 ModelScope 开源的数据集。CHILLGuardTrain-Subset-15K

cswbyu/CHILLGuardTrain-Subset-15K 是 ModelScope 魔搭社区上的数据集，存储大小 3.9 MB，采用 cc-by-nc-4.0 许可。

- **Repository**: cswbyu/CHILLGuardTrain-Subset-15K
- **License**: cc-by-nc-4.0
- **Storage size**: 3.9 MB
- **Downloads**: 37
- **Stars**: 1
- **Last updated**: 2026-11-01

Source: https://www.modelscope.cn/datasets/cswbyu/CHILLGuardTrain-Subset-15K

---

# CHILLGuardTrain-Subset-15K

A curated, balanced subset of **CHILLGuardTrain** for few-shot exemplars and lightweight fine-tuning experiments with Chinese safety guardrails.

- 📄 Paper: [arXiv:2606.15396](https://arxiv.org/abs/2606.15396)
- 🤖 Guardrail models: [CHILLGuard-1.7B](https://modelscope.cn/models/cswbyu/CHILLGuard-1.7B) · [CHILLGuard-4B](https://modelscope.cn/models/cswbyu/CHILLGuard-4B) · [CHILLGuard-8B](https://modelscope.cn/models/cswbyu/CHILLGuard-8B)
- 🧪 Benchmark: [cswbyu/CHILLGuardTest](https://modelscope.cn/datasets/cswbyu/CHILLGuardTest)
- 💻 Code: [GitHub - cswbyu/CHILLGuard](https://github.com/cswbyu/CHILLGuard)

## What is this?

**15,500 prompts** — 31 micro harm categories × (250 safe + 250 unsafe). Every row is drawn from the RAG-generated and PE-rewritten portions of CHILLGuardTrain, i.e. prompts synthesized by the paper's own data-construction pipeline rather than redistributed from third-party datasets.

Highlights:

- **Implicit & obfuscated Chinese harms**: homophone substitutions, pinyin abbreviations, emoji/symbol rewriting, classical allusions, and rhetorical irony — the signature styles of the CHILLGuard taxonomy.
- **Balanced & category-complete**: exactly 250 safe + 250 unsafe per micro category.
- **Quality-filtered**: length filtering (25–150 chars), generation-template removal, and MinHash-based near-duplicate elimination within each category.

This subset is intended for few-shot prompting, small-model fine-tuning studies, and category-level analyses. The full training set **CHILLGuardTrain** (405,007 samples) and its construction pipeline are detailed in the paper; only this curated 15K subset is publicly released.

## Data Fields

| Field | Type | Description |
|---|---|---|
| `id` | int | 1 – 15,500 |
| `prompt` | string | The user prompt |
| `safety` | string | `safe` / `unsafe` |
| `category` | string | One of the 31 micro harm categories |

## Intended Use & Limitations

- Supervised fine-tuning (SFT) of Chinese safety guardrails — including full-parameter and parameter-efficient (LoRA/PEFT) recipes.
- Few-shot exemplars and ablation studies on Chinese safety guardrails.
- Labels follow the same annotation pipeline as CHILLGuardTrain (multi-model voting; tied votes adjudicated by DeepSeek-V3.2-685B).
- Contains prompts describing harmful content for detection purposes; must not be used to generate harmful content.
- Non-commercial use only.

## License

CC BY-NC 4.0.

## Citation

```bibtex
@inproceedings{yu-etal-2026-chillguard,
  title     = "{CHILLG}uard: Towards Fine-Grained {C}hinese {LLM} Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment",
  author    = "Yu, Wenbo  and
             Wang, Bohua  and
             Fang, Hao  and
             Gao, Kuofeng  and
             Zeng, Jingru  and
             Yang, Xiaochen  and
             Zhang, Tianyi  and
             Ma, Xiaoxiao  and
             Kong, Jiawei  and
             Wu, Hao  and
             Chen, Bin  and
             Xia, Shu-Tao  and
             Zhang, Min",
  editor    = "Lee, Hung-Yi  and
             Dev, Sunipa  and
             Li, Jessy  and
             Elliott, Desmond",
  booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
  month     = oct,
  year      = "2026",
  address   = "Budapest, Hungary",
  publisher = "Association for Computational Linguistics",
  url       = "https://aclanthology.org/2026.findings-emnlp.1721/",
  doi       = "10.18653/2026.findings-emnlp.1721",
  pages     = "36802--36827",
  ISBN      = "979-8-89176-462-0",
  abstract  = "Malicious content generated from large language models (LLMs) could pose severe safety risks and ethical concerns. While existing LLM safety guardrails excel in English or multilingual settings, they lack adaptation to Chinese-specific regulatory policies, cultural context, and linguistic nuances, failing to support fine-grained risk classification for diverse deployment needs. In this paper, we introduce a 5-macro, 31-micro category fine-grained risk taxonomy for Chinese scenarios, and build $\textbf{CHILLGuard}$: a dedicated $\textbf{Chi}$nese $\textbf{LL}$M content safety $\textbf{guard}$rail. To address the critical scarcity of high-quality annotated Chinese safety data, we propose a scalable multi-stage data construction pipeline: we expand multi-source corpus via retrieval-augmented generation, generate implicit harmful samples through prompt engineering rewriting, and refine high-quality data via multi-model voting-based label calibration. Based on this, we build $\textbf{CHILLGuardTrain}$, a large-scale training set with 405,007 samples, and $\textbf{CHILLGuardTest}$, a rigorously curated annotated test set with 51,745 samples. We then train CHILLGuard on CHILLGuardTrain under a generator-classifier collaborative framework via Model-aware Direct Preference Optimization. Extensive experiments under multiple settings demonstrate the state-of-the-art performance of CHILLGuard, e.g., a 15.92{\%} relative improvement of F1 score over Qwen3Guard-8B-Strict on our benchmark. We release our resources at https://github.com/cswbyu/CHILLGuard."
}
```