README.md

metadata

task_categories:
  - text-retrieval
  - text-classification
  - question-answering
language:
  - en
tags:
  - finance
  - auditing
  - xbrl
  - us-gaap
dataset_info:
  features:
    - name: query
      dtype: string
    - name: dqc_id
      dtype: string
    - name: answer
      dtype: string
    - name: id
      dtype: int64
  splits:
    - name: train
      num_bytes: 21442683
      num_examples: 197
    - name: test
      num_bytes: 14520142
      num_examples: 133
  download_size: 3726866
  dataset_size: 35962825
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

FinAuditing: A Financial Taxonomy-Structured Multi-Document Benchmark for Evaluating LLMs

FinAuditing is the first taxonomy-aligned, structure-aware, multi-document benchmark designed for evaluating Large Language Models (LLMs) on complex financial auditing tasks. Built from real US-GAAP-compliant XBRL filings, it addresses the challenges of automating and verifying financial auditing due to the complexities of Generally Accepted Accounting Principles (GAAP) and the hierarchical structure of eXtensible Business Reporting Language (XBRL) filings.

Paper: FinAuditing: A Financial Taxonomy-Structured Multi-Document Benchmark for Evaluating LLMs Dataset GitHub Repository: https://github.com/The-FinAI/FinAuditing.git Evaluation Framework: https://github.com/The-FinAI/FinBen

Overview

The FinAuditing benchmark defines three complementary subtasks, each targeting a distinct aspect of structured auditing reasoning:

• FinSM (Semantic Consistency): This subtask evaluates information retrieval. Given a query describing a financial term (e.g., currency or concentration of credit risk), an XBRL filing, and a US-GAAP taxonomy, the task is to identify the set of mismatched US-GAAP tags after retrieval.
• FinRE (Relational Consistency): This is a relation extraction task. Given two specific elements ($e_1$ and $e_2$), an XBRL filing, and a US-GAAP taxonomy, the goal is to classify three predefined relation error types between them.
• FinMR (Numerical Consistency): A mathematical reasoning task. Given two questions ($q_1$ concerning the extraction of a reported value and $q_2$ pertaining to the calculation of the corresponding real value), an XBRL filing, and a US-GAAP taxonomy, the task is to extract the reported value for a given instance and compute its numeric value to verify correctness.

Extensive zero-shot experiments on 13 state-of-the-art LLMs using FinAuditing revealed inconsistent performance across semantic, relational, and mathematical dimensions, with significant accuracy drops (up to 60-90%) when reasoning over hierarchical multi-document structures. This highlights systematic limitations of current LLMs in taxonomy-grounded financial reasoning and establishes FinAuditing as a foundation for developing trustworthy, structure-aware, and regulation-aligned financial intelligence systems.

Datasets Released

The FinAuditing benchmark includes the following datasets for evaluation:

📂 Dataset	📝 Description
FinSM	Evaluation set for FinSM subtask within FinAuditing benchmark. This task follows the information retrieval paradigm: given a query describing a financial term that represents either currency or concentration of credit risk, an XBRL filing, and a US-GAAP taxonomy, the output is the set of mismatched US-GAAP tags after retrieval.
FinRE	Evaluation set for FinRE subtask within FinAuditing benchmark. This is a relation extraction task, given two specific elements $e_1$ and $e_2$, an XBRL filing, and a US-GAAP taxonomy, the goal of this task is to classify three relation error types.
FinMR	Evaluation set for FinMR subtask within FinAuditing benchmark. This is a mathematical reasoning task, given two questions $q_1$ and $q_2$, where $q_1$ concerns the extraction of a reported value and $q_2$ pertains to the calculation of the corresponding real value, an XBRL filing, and a US-GAAP taxonomy, the task is to extract the reported value for a given instance in the XBRL filing and to compute the numeric value for that instance, which is then used to verify whether the reported value is correct.
FinSM_Sub	FinSM subset for ICAIF 2025.
FinRE_Sub	FinRE subset for ICAIF 2025.
FinMR_Sub	FinMR subset for ICAIF 2025.

Citation

If you find our benchmark useful, please cite:

    @misc{wang2025finauditingfinancialtaxonomystructuredmultidocument,
          title={FinAuditing: A Financial Taxonomy-Structured Multi-Document Benchmark for Evaluating LLMs}, 
          author={Yan Wang and Keyi Wang and Shanshan Yang and Jaisal Patel and Jeff Zhao and Fengran Mo and Xueqing Peng and Lingfei Qian and Jimin Huang and Guojun Xiong and Xiao-Yang Liu and Jian-Yun Nie},
          year={2025},
          eprint={2510.08886},
          archivePrefix={arXiv},
          primaryClass={cs.CL},
          url={https://arxiv.org/abs/2510.08886}, 
    }