FinGuard-Bench

FinGuard-Bench is a Chinese benchmark for detecting financial regulatory non-compliance in LLM interactions, accompanying FinGuard: Detecting Financial Regulatory Non-Compliance in LLM Interactions (Findings of ACL: EMNLP 2026).

The dataset contains 1,020 query–response pairs, covering 11 scenario main categories and 35 scenario subcategories. Queries and responses have separate safety labels. The complete benchmark is provided as a single test split in UTF-8 JSONL format, with one record per line.

Notices

• Data generation: The query-response pairs in this dataset were generated by opensource LLMs based on compliance knowledge points extracted from publicly available Chinese financial regulatory documents. All samples were manually reviewed and annotated. Model-generated content may differ from the original regulatory texts; the original documents should be treated as authoritative when applying the regulations.
• Usage restrictions and disclaimer: FinGuard-Bench aims to support the evaluation of financial compliance detection models and academic research. The dataset contains AI-generated simulations of financial regulatory violations across multiple financial domains for compliance detection evaluation. These simulations contain no real-world scenarios or information. The dataset is intended exclusively for academic research and benchmark evaluation. Views expressed in the dataset do not represent the positions of the project organizers, authors, or affiliated entities. Using any content from this dataset for operational debt collection, script optimization, or any commercial lending or debt collection context is strictly prohibited. Users must ensure that their use complies with all applicable laws and regulations and assume full responsibility for the conduct and compliance of any downstream applications of the dataset. The project accepts no liability for any direct or indirect losses arising from the use of this dataset.

Fields

Every record contains exactly these 8 fields:

Field	Type	Description
id	integer	Unique sample identifier, preserved from the original dataset.
scenario_main_category	string	Main financial regulatory category of the query scenario.
scenario_sub_category	string	Subcategory of the query scenario, belonging to scenario_main_category.
query	string	Original Chinese query.
response	string	Original response paired with the query.
query_safety	string	Safety annotation for the query: Safe or Unsafe.
response_safety	string	Safety annotation for the provided response: Safe or Unsafe.
response_risk_categories	list[string]	Risk categories annotated in the response. Multiple categories are allowed; safe responses use []. Values come from the same set of 35 subcategories used by scenario_sub_category.

Scenario categories and risk annotations

scenario_main_category and scenario_sub_category identify the scenario covered by the query. Both safe and unsafe queries have scenario categories. A query's safety is determined by query_safety; having a scenario category does not itself indicate an unsafe query.

response_risk_categories contains the risk categories annotated in the supplied response. These may differ from scenario_sub_category and may include multiple categories. An empty list means no response risk categories are annotated; in this release, all 705 safe responses have an empty list and all 315 unsafe responses have at least one category.

Query and response safety labels are independent: an unsafe query may receive a safe response, and a safe query may receive an unsafe response. response_safety and response_risk_categories describe the supplied response; other model outputs require their own evaluation.

Statistics

Annotation	Safe	Unsafe
Query	276 (27.1%)	744 (72.9%)
Response	705 (69.1%)	315 (30.9%)

Percentages use all 1,020 samples as the denominator. Original sample order, query and response text, and retained annotation values are preserved.

Citation

If you use FinGuard-Bench in your research, please cite:

@inproceedings{finguard,
  title     = {FinGuard: Detecting Financial Regulatory Non-Compliance in LLM Interactions},
  author    = {Huaixia Dou and Jie Zhu and Minghao Wu and Shuo Jiang and Junhui Li and Lifan Guo and Feng Chen and Chi Zhang},
  booktitle = {Findings of the Association for Computational Linguistics: EMNLP 2026},
  year      = {2026}
}

License

FinGuard-Bench is released under the Apache License 2.0.