Note:

1. This repository is currently under construction.
2. This repository is for academic purposes only. It does not represent the views of any organization, institution, or individual, and should not be considered legal or investment advice.

Compliance-to-Code: Enhancing Financial Compliance Checking via Code Generation

Siyuan Li^1,*, Jian Chen^1,*,†, Rui Yao¹, Xuming Hu¹, Peilin Zhou¹, Weihua Qiu², Simin Zhang¹, Chucheng Dong³, Zhiyao Li¹, Qipeng Xie¹, Zixuan Yuan^1,†

¹Hong Kong University of Science and Technology (Guangzhou)
²Sun Yat-Sen University
³University of California, Riverside
^*Co-first authors, ^†Corresponding author
{sli974, jchen524, ryao663, pzhou460, szhang420, zli632}@connect.hkust-gz.edu.cn, qiuwh9@mail2.sysu.edu.cn, cdong040@ucr.edu, qxieaf@connect.ust.hk, {xuminghu, zixuanyuan}@hkust-gz.edu.cn

📖Paper | 🌟GitHub | 🤗Huggingface Dataset

Regulatory compliance is crucial for corporate governance, but financial regulations are often intricate and challenging to interpret. While Large Language Models (LLMs) show promise in automating the conversion of regulatory text to compliance logic, they face limitations with Chinese financial regulations, including incomplete domain knowledge and insufficient reasoning. To address this, we introduce Compliance-to-Code, the first large-scale Chinese dataset for financial regulatory compliance. It contains 1,159 annotated clauses from 361 regulations, structured into logical elements with Python code mappings and detailed reasoning. We also present FinCheck, a pipeline for regulation structuring, code generation, and report generation, demonstrating the dataset's utility.

📢 News

• 🚀 [26/05/2025] We released our paper on arxiv.
• 🚀 [15/05/2025] We released the Compliance-to-Code dataset and the FinCheck pipeline, the first large-scale Chinese resource for code-based financial compliance automation.

💡 Highlights

• 🔥 Dataset: We propose Compliance-to-Code, the first large-scale Chinese financial regulation dataset tailored for code-based compliance reasoning. It includes 1,159 annotated clauses from 361 regulations across ten categories, with 307 Python modules and detailed reasoning steps.
• 🔥 FinCheck Pipeline: We develop FinCheck, an end-to-end pipeline that leverages the Compliance-to-Code dataset for automated compliance checking. It includes modules for regulation structuring, code generation.
• 🔥 Key Insights: We conduct a comprehensive evaluation of recent LLMs for regulation structuring and code generation tasks. Our experiments show that Qwen3-8B achieves strong performance, and Supervised Fine-Tuning (SFT) significantly improves regulation structuring capabilities.

📕 Data Distribution

The Compliance-to-Code dataset is constructed from 361 authoritative Chinese financial regulations, primarily from the Beijing Stock Exchange (BSE), covering ten core thematic areas. It includes 1,159 annotated Compliance Units (CUs) and 307 executable compliance automation tasks in Python.

Overall Statistics:

• Regulations Analyzed: 361
• Annotated Clauses (Compliance Units): 1,159
• Inter-Unit Relations Annotated: 864
• Executable Python Automation Tasks: 307 (Simple: 70.03%, Medium: 11.07%, Difficult: 18.89%)
• Average COT Reasoning Steps per Task: 8
• Average Tokens per COT Reasoning Step: 2145

Regulatory Domains Covered:

Domain/Aspect	Compliance Control Focus and Risk	# Source clauses	# Compliance Units	# Document Words
Independent Director Systems	Independence requirements, director nomination/integrity review, conflict management.	30	126	5446
Quarterly Reporting Obligations	Timeliness, content sufficiency, disclosure controls for periodic reports.	16	32	1888
Equity Incentives and ESOPs	Plan approval, risk limits, grantee eligibility, compliance disclosure events.	66	154	8132
Share Repurchase Controls	Buyback process, threshold triggers, ban periods, reporting duties.	78	266	11458
Tender Offer Compliance	Takeover code triggers, procedural controls, reporting/approval checks.	36	113	4827
Inside Information Management	Insider identification, ad hoc event control points, information wall requirements.	25	48	3564
Board Transfer/Listing Change	Conditions, procedural controls, risk disclosure for inter-market moves.	21	90	2990
Large Shareholder Transactions Control	Sale restriction, blackout periods, reporting triggers, conflict discipline.	29	107	5504
Raised Fund Use Controls	Escrow, use restriction, monitoring, board/supervisor review.	27	117	4623
Equity Rights Distribution	Dividend/bonus triggers, shareholder fairness, process integrity.	33	106	4908

🛠️ Usage

Dataset Access

The Compliance-to-Code dataset is available on Huggingface Datasets and GitHub:

• Huggingface: https://huggingface.co/datasets/GPS-Lab/Compliance-to-Code
• GitHub: https://github.com/AlexJJJChen/Compliance-to-Code

The dataset is provided in JSON and CSV formats. Each entry contains:

• Clean regulatory text.
• Structured Compliance Units (CUs) with fields: Subject, Condition, Constraint, Contextual Information.
• Annotated Inter-Unit Relations: refer to, exclude, only include, should include.
• Python-coded compliance automation tasks.

FinCheck Pipeline

The code for the FinCheck pipeline and experiments is available in the GitHub repository. The pipeline is designed for:

1. Regulation-to-Structure (R2S): Parsing regulations into CUs.
2. Structure-to-Code (S2C): Generating Python code from CUs.

Refer to the repository for details on setting up and running experiments.

✒️Results Highlights

We benchmarked various Large Language Models (LLMs) on regulation structuring and code generation tasks using the Compliance-to-Code dataset. Our experiments highlight that structured decomposition of regulations and explicit reasoning are crucial for effective compliance code generation, especially for complex scenarios. Please refer to our paper on arxiv for more details.

✒️Citation

If you use the Compliance-to-Code dataset or the FinCheck pipeline in your research, please cite our paper:

@misc{li2025compliancetocodeenhancingfinancialcompliance,
      title={Compliance-to-Code: Enhancing Financial Compliance Checking via Code Generation}, 
      author={Siyuan Li and Jian Chen and Rui Yao and Xuming Hu and Peilin Zhou and Weihua Qiu and Simin Zhang and Chucheng Dong and Zhiyao Li and Qipeng Xie and Zixuan Yuan},
      year={2025},
      eprint={2505.19804},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2505.19804}, 
}

📄 License

The Compliance-to-Code dataset, the code for the FinCheck pipeline and associated tools are licensed under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) License. Please refer to the LICENSE file in the GitHub repository and the Huggingface dataset card for more details.

The regulatory documents used to create this dataset were publicly published by the Beijing Stock Exchange (BSE).