RKEFino1: A Regulation Knowledge-Enhanced Large Language Model

Yan Wang, Yueru He, Ruoyu Xiang, Jeff Zhao

Abstract

RKEFino1 is a specialized Large Language Model (LLM) designed to master the complex world of financial regulatory reporting. By fine-tuning the existing Fino1 model with domain-specific knowledge—specifically standards like XBRL and CDM—researchers significantly improved the model's ability to handle compliance-critical tasks, mathematical reasoning, and numerical data extraction from financial tables.

Addressing the Complexity of Digital Regulatory Reporting

While general-purpose LLMs have shown promise in finance, they often stumble when faced with Digital Regulatory Reporting (DRR). DRR requires extreme precision; a single hallucination or a misunderstanding of a reporting standard can lead to massive financial losses or regulatory non-compliance.

The authors identify a gap in how current models handle three specific frameworks:

  1. XBRL (eXtensible Business Reporting Language): The global standard for exchanging business information.
  2. CDM (Common Domain Model): Used for tracking the lifecycle of financial transactions.
  3. MOF (Model Openness Framework): A standard requiring transparency and completeness in machine learning models.

Existing models lack the nuanced understanding of these regulatory structures. RKEFino1 aims to bridge this gap by injecting explicit regulatory knowledge into the model's reasoning process.

Methodology and Data Curation

Researchers built RKEFino1 on top of Fino1, a lightweight financial reasoning model based on LLaMA-3.1-8B-Instruct. The enhancement process involved Supervised Instruction Tuning for three distinct task types: Knowledge-based QA (K-QA), Mathematical Reasoning QA (MR-QA), and Numerical NER (Named Entity Recognition).

To keep the model deployable and training cost-effective, they utilized Parameter-Efficient Fine-Tuning (PEFT) via LoRA (Low-Rank Adaptation) with int4 quantization. This allowed training on 4 NVIDIA H100 GPUs using a context length of 8192 tokens.

Table 1: Training dataset distribution across tasks, domains, and sources.

The team curated a training set of 9,898 samples from high-authority sources, including CDM documentation, the OSI website for MOF knowledge, and SEC/XBRL terminology for reporting standards.

Key Results and Insights

RKEFino1 outperformed the baseline Fino1 across all evaluated metrics. One of the most striking findings was the "Zero-to-Hero" effect: the baseline model had 0% accuracy on specific tasks like MOF Approval and XBRL Tagging. By injecting targeted regulatory data, RKEFino1's performance jumped to 62.58% and 16.02%, respectively.

Table 2: Performance comparison between Fino1 and RKEFino1 across fine-grained tasks.

Key performance highlights include:

  • Knowledge-Based QA: Significant gains in FactScore. For example, XBRL Term FactScore rose from 26.22% to 50.28%.
  • Mathematical Reasoning: Accuracy on XBRL-based math tasks improved from 56.87% to 70.69%, demonstrating that better domain context leads to superior arithmetic performance.
  • Numerical NER: The F1-score for identifying financial entities nearly doubled, moving from 14.99% to 26.62%.

Furthermore, RKEFino1 was specifically trained to process financial tables, treating them as "first-class citizens" rather than focusing purely on prose. This is critical as the vast majority of regulatory data is structured in tabular formats.

Conclusion and Future Work

Despite these improvements, some tasks remain challenging. Accuracy for MOF Abbreviations and XBRL Tags is still relatively low, indicating that these areas require even more specialized data or different architectural approaches. The team plans to expand their datasets for these lagging categories to further harden the model for real-world regulatory environments. The model has been released on Hugging Face to encourage community experimentation and validation.

Summary updated

BibTeX
@inproceedings{wang2025rkefino,
  author = {Yan Wang and Yueru He and Ruoyu Xiang and Jeff Zhao},
  title = {{RKEFino1: A Regulation Knowledge-Enhanced Large Language Model}},
  booktitle = {IDS},
  year = {2025},
  doi = {10.1109/ids66066.2025.00014},
}
Figure