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) ## 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; conservative tie-breaking). - 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 @misc{yu2026chillguard, title={CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment}, author={Wenbo Yu and Bohua Wang and Hao Fang and Kuofeng Gao and Jingru Zeng and Xiaochen Yang and Tianyi Zhang and Xiaoxiao Ma and Jiawei Kong and Hao Wu and Bin Chen and Shu-Tao Xia and Min Zhang}, year={2026}, eprint={2606.15396}, archivePrefix={arXiv}, primaryClass={cs.CL}, url={https://arxiv.org/abs/2606.15396} } ```