CnHallu-RAG: Detecting and Mitigating Hallucinations in Chinese Retrieval-Augmented Generation for High-Risk Domains

Large language models (LLMs) are increasingly used for Chinese information-seeking tasks in healthcare, law, public policy, academic writing, and technical support. These high-risk domains require answers that are not only fluent but also grounded in verifiable evidence. Retrieval-augmented generation (RAG) is commonly used to reduce hallucination, yet retrieved evidence does not guarantee faithful generation: models may ignore evidence, overgeneralize from partial evidence, fabricate citations, or present uncertainty as fact. Existing hallucination benchmarks provide valuable foundations, but Chinese high-risk RAG settings remain underexplored. This preprint presents CnHallu-RAG, a preliminary development resource and evaluation framework for Chinese hallucination analysis in RAG. The framework uses claim-level annotation and distinguishes unsupported claims, evidence-contradicting claims, fabricated citations, outdated claims, and overconfident claims. The current release contains a schema-validated 100-example development set across five domains, prompts, model-running scripts, audit scripts, DeepSeek direct-generation, standard-RAG, and citation-constrained RAG outputs, and a confirmed 20-example audit sample. In answer-level auditing, direct generation is flagged hallucinated on 34 of 100 examples, while both standard RAG and citation-constrained RAG are flagged on 1 of 100 examples. A stratified 20-example manual audit sample confirms the sampled labels, including the known standard-RAG evidence-reversal case. The paper also includes appendix-level analyses of prototype atomic decomposition, independent Qwen judge sensitivity, and a separated candidate extension sample. This work is a preliminary development study rather than a final benchmark. The uploaded files include the compiled PDF and the LaTeX source package. Supporting datasets and audit outputs are not included in this Zenodo deposit pending final licensing review.