CHILLGuard: Towards Fine-Grained Chinese LLM Safety Guardrail with Scalable Data Construction and Model-aware Preference Alignment

Researchers have introduced CHILLGuard, a safety guardrail for Chinese large language models that uses a 31-category risk taxonomy to address gaps in existing systems, which lack adaptation to Chinese regulatory policies and cultural context [1]. The system is detailed in a paper submitted to arXiv on 13 June 2026 [1]. The authors argue that current LLM safety guardrails perform well in English or multilingual environments but fail to support the fine-grained risk classification required for Chinese deployment scenarios [1]. CHILLGuard is built on a 5-macro, 31-micro category taxonomy designed specifically for Chinese content risks [1]. A central challenge the project tackles is the scarcity of high-quality annotated Chinese safety data. To overcome this, the team constructed a scalable multi-stage data pipeline that expands a multi-source corpus via retrieval-augmented generation, generates implicit harmful samples through prompt engineering rewriting, and refines data quality using multi-model voting-based label calibration [1]. The pipeline produced CHILLGuardTrain, a training set containing 405,007 samples, and CHILLGuardTest, a curated test set of 51,745 samples [1]. The guardrail was trained under a generator-classifier collaborative framework using Model-aware Direct Preference Optimization [1]. In benchmark tests, CHILLGuard achieved a 15.92% improvement in F1 score over Qwen3Guard-8B-Strict [1]. Qwen3Guard is part of the Qwen family of large language models developed by Alibaba Cloud, many of which are distributed under open-source licenses such as Apache 2.0 [9]. The release of CHILLGuard comes amid heightened global attention on Chinese AI development. Chinese firm DeepSeek drew widespread notice in January 2025 when its DeepSeek-R1 model demonstrated performance comparable to OpenAI's GPT-4 while reportedly using a fraction of the training budget [7]. DeepSeek claimed its V3 model cost roughly US$6 million to train, compared with an estimated US$100 million for GPT-4 in 2023 [7]. The company's success, achieved despite U.S. export restrictions on advanced AI chips, triggered what observers described as a "Sputnik moment" for the American AI sector [7]. The CHILLGuard paper's authors have indicated they will release their resources on GitHub, and the work is indexed on Hugging Face's paper pages, a platform that links research artifacts such as models, datasets, and interactive demos to arXiv preprints [1][4]. Hugging Face and arXiv have collaborated to embed demos directly alongside papers, allowing users to test models without writing code [5]. The CHILLGuard repository is expected to follow this pattern, making the safety guardrail accessible for community evaluation [1].