RuFinQA is a large-scale multi-task benchmark designed to evaluate the ability of language models to understand and reason over Russian statutory financial reports (Balance Sheet, Income Statement, Cash Flow Statement).

It contains 36,330 question–answer pairs across 5 task types, automatically derived from real-world corporate accounting statements obtained from open government data sources.

📖 Data Sources, Licensing & Legal Notice

Data Origin

All questions and answers in RuFinQA are generated from anonymized excerpts of official financial reports of Russian companies. The original reports are publicly available through government registers (e.g., Federal Tax Service, Rosstat) and are used solely for academic research purposes.

Foundational Source

The methodological framework, financial formulas, problem typology, and a significant portion of the practical examples used in this dataset are based on the author's educational manual:

Арабов М.К., Маматкулов А.А., Солиева Л.Ф. Решение финансово-экономических задач посредством Excel: Учебно-методическое пособие / М.К. Арабов, А.А. Маматкулов, Л.Ф. Солиева. – Душанбе: РТСУ, 2019. – 200 с.

Ownership & Rights

We do not claim ownership of the original financial statements, textual content, or numerical values contained in the source reports. These remain the property of their respective copyright holders (the companies and/or regulatory bodies that published them).

The questions, answers, reasoning patterns, and all derived annotations created by the authors are released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.
The "packaging" of the dataset — including metadata, structure, and formatting — is also covered by CC BY 4.0.
The underlying source texts and financial data are subject to the original rights of their holders.

All company identifiers (TIN, names, registration numbers) have been fully pseudonymized to prevent direct attribution to real entities.

User Responsibility

It is the responsibility of the user of the RuFinQA dataset to ensure that their use of the data complies with all applicable laws and regulations in their jurisdiction. Users should independently verify the legal status of any underlying source material they intend to use, especially if their application involves commercial or redistributive purposes.

Notice‑and‑Takedown Policy

We follow the industry‑standard practice (similar to HPLT and other large‑scale corpora): the packaging and annotations are openly licensed, but the rights to the original source texts belong to their respective holders.

If you are a copyright owner and believe that your content appears in this dataset without proper authorization, please contact us. We will promptly remove the disputed entries upon verification.

To submit a takedown request, please provide:

Identification of the copyrighted work (report name, company identifier, year, etc.)
Sufficient information to locate the specific data in the dataset (e.g., record ID)
Your contact information (name, email, phone)
A statement of good faith belief that the use is unauthorized

📊 Task Types

Type	Count	Share
factoid	12,255	33.7%
arithmetic	12,021	33.1%
comparison	4,085	11.2%
analytical	4,059	11.2%
multistep	3,910	10.8%

📝 Data Structure

Each record is a JSON object with the following fields:

Field	Type	Description
`id`	string	Unique identifier (e.g., `arith_COMP_0001_2025_roa`)
`type`	string	Task type: `arithmetic`, `factoid`, `comparison`, `analytical`, `multistep`
`question`	string	Natural-language question in Russian
`context`	dict	Structured financial data needed to answer the question
`answer`	string	Gold reference answer
`numeric_answer`	float (optional)	Exact numeric answer for `arithmetic` tasks
`tolerance`	float (optional)	Acceptable deviation for numeric evaluation
`exact_match`	string (optional)	Canonical string for `factoid` evaluation
`keywords`	list (optional)	Required keywords for `comparison`, `analytical`, `multistep`

🚀 Usage

Loading with 🤗 Datasets

from datasets import load_dataset

dataset = load_dataset("arabovs-ai-lab/RuFinQA", split="train")
print(dataset[0])

Example Record

{
  "id": "arith_COMP_0001_2025_roa",
  "type": "arithmetic",
  "question": "Рассчитай рентабельность активов (ROA) для компании ИНН COMP_0001 за 2025 год.",
  "context": {
    "inn": "COMP_0001",
    "year": 2025,
    "bal_1600": 633.696,
    "inc_2400": 44.691
  },
  "answer": "ROA = 44.7 / 633.7 = 0.0705 (7.05%)",
  "numeric_answer": 0.0705,
  "tolerance": 0.001
}

Model Evaluation

# Simple numeric evaluation example
def evaluate_numeric(prediction: str, ground_truth: float, tolerance: float) -> bool:
    import re
    numbers = [float(x) for x in re.findall(r'-?\d+\.?\d*', prediction)]
    return any(abs(n - ground_truth) <= tolerance for n in numbers)

# Check model output
model_output = "ROA = 0.0705 (7.05%)"
is_correct = evaluate_numeric(model_output, 0.0705, 0.001)
print(f"Answer is correct: {is_correct}")

📄 License

The annotations, questions, and derived content are distributed under the CC BY 4.0 license. The underlying source financial reports are subject to the original rights of their holders (see Data Sources, Licensing & Legal Notice above).

📚 Citation

If you use RuFinQA in your research, please cite:

@misc{rufinqa2025,
  author = {Arabov, Mullosharaf Kurbonovich},
  title = {RuFinQA: A Massive Multi-Task Reasoning Benchmark for Russian Financial Report Understanding},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/arabovs-ai-lab/RuFinQA}
}

👤 Author

Mullosharaf K. Arabov
ORCID: 0000-0003-2525-1183
PhD in Physics and Mathematics, Associate Professor
Department of Data Analysis and Programming Technologies
Kazan (Volga Region) Federal University
📧 marabov@kpfu.ru

🔗 Links

📊 Dataset: https://huggingface.co/datasets/arabovs-ai-lab/RuFinQA
💻 Generator code: [GitHub]
📄 Paper: [arXiv]

Built for the multilingual financial NLP community.

Downloads last month: 9

Total file size:

14 MB