Dataset Card for RuFinQA-CoT

Dataset Details

Dataset Description

RuFinQA-CoT is a Russian-language benchmark for evaluating Large Language Models (LLMs) on financial reasoning with Ground-Truth Chain-of-Thought (CoT). It is built from real accounting (financial) statements of 835 Russian companies registered in the Republic of Tatarstan for the period 2021–2025.

The dataset contains 44,625 questions across 7 task types: arithmetic, factoid, comparison, analytical, multistep, reasoning, adversarial. Every question is accompanied by a procedurally generated and automatically verified Chain-of-Thought solution with criticality flags.

• Curated by: Mullosharaf K. Arabov
• Shared by: Mullosharaf K. Arabov
• Language(s) (NLP): Russian (ru)
• License: CC BY 4.0 (see Legal Notice below)
• Dataset size: 44,625 examples, single split (train)

Legal Notice

Copyright and Ownership. The author of this dataset card and of the generation/curation pipeline is Mullosharaf K. Arabov. However, the author is NOT the copyright holder of the underlying financial data. The raw accounting (financial) statements were obtained from a single public source: the State Information Resource of Accounting (Financial) Statements (BFO) of the Federal Tax Service of the Russian Federation (https://bo.nalog.ru/), which is published openly under the legislation of the Russian Federation. No other data sources were used.

What belongs to the author: the metadata, the question templates, the reasoning-step annotations, the evaluation schema, and the dataset card itself.

What does NOT belong to the author: the raw financial figures (assets, equity, revenue, profit, etc.) reported by the companies. These are used here solely for research and benchmarking purposes.

All company identifiers (INN — Russian tax identification number) and company names have been anonymized using a deterministic mapping with a fixed random seed. No personal data of individuals is included.

If you are a rights holder of any underlying financial statement and believe it should not appear in this benchmark, please contact the author at cool.araby@gmail.com for prompt removal.

Dataset Sources

• Repository: https://huggingface.co/datasets/RusNLPWorld/RuFinQA-CoT
• Source data (single source): https://bo.nalog.ru/ — State Information Resource of Accounting (Financial) Statements (BFO), operated by the Federal Tax Service of Russia. All 4,418 Excel files used in this benchmark were downloaded from this resource.

Uses

Direct Use

• Benchmarking LLMs on multi-step financial reasoning in Russian.
• Evaluating Chain-of-Thought faithfulness using the FinCoT-Eval v2 framework (six metrics: FAA, OT, CSPS, PSA, NEPS, FCS).
• Training and fine-tuning models on Russian financial question answering with explicit reasoning traces.
• Research on verifiable reasoning: every gold CoT step has been automatically validated against the final numeric answer, and criticality flags are provided for weighted evaluation.

Out-of-Scope Use

• Investment or financial advice. The dataset is a research benchmark; it must not be used for real-world trading, valuation, credit decisions, or regulatory compliance.
• Company profiling or identification. All identifiers are anonymized; attempting to re-identify companies is prohibited.

Dataset Structure

Data Fields

Each example is a single JSON object. Fields are organized at two levels: question-level and step-level.

Question-level fields

Field	Type	Description
id	string	Unique identifier, format <type>_<inn>_<year>_<skill>
type	string	One of arithmetic, factoid, comparison, analytical, multistep, reasoning, adversarial
skill	string	Fine-grained skill tag (e.g., roa_calculation, yoy_comparison)
difficulty	string	easy, medium, hard
question	string	Natural-language question in Russian
context	object	Raw financial values from the report (float, unrounded)
reasoning_steps	list[object]	Chain-of-Thought steps (see below)
cot_text	string	Human-readable CoT (concatenated steps + answer)
final_answer	string	Textual final answer
final_numeric	float | null	Numeric final answer (for numeric mode)
numeric_answer	float | null	Alias of final_numeric (backward compatibility)
faa_mode	string	numeric | exact | keyword | fuzzy_text
tolerance	float | null	Tolerance for numeric mode
normalized_answer	string | null	For exact mode
required_keywords	list[string] | null	For keyword mode
indicators	object | null	Dict {ROA, ROE, CR, Revenue} for FCS
fcs_applicable	bool	Whether FCS applies to this question
neps_mode	string	skip_none (default) or penalty
template_version	string	Generator version (currently 6.2.0)

Step-level fields (reasoning_steps[i])

Field	Type	Description
step	int	Step number (1..N)
description	string	Semantic text without digits (for embeddings in OT/CSPS/PSA)
numeric_value	float | null	Numeric result of the step (for FNS/NEPS)
unit	string	млн руб. | % | коэф. | доля | п.п. | год
result_text	string	Human-readable step result (with number)
is_critical	bool	Criticality flag (weight 2.0 in CSPS/PSA if true, 0.5 otherwise)
intermediate_result	string	Alias of result_text (backward compatibility)

Data Splits

The dataset is provided as a single split (train) with 44,625 examples. Recommended stratified train/validation/test splits (80/10/10 by type and difficulty) can be generated using the seed 42.

Example

{
  "id": "arith_1000000701_2025_roa",
  "type": "arithmetic",
  "skill": "roa_calculation",
  "difficulty": "medium",
  "question": "Рассчитай рентабельность активов (ROA) для компании с ИНН 1000000701 за 2025 год.",
  "context": {"inn": "1000000701", "year": 2025, "bal_1600": 633.696, "inc_2400": 44.691},
  "reasoning_steps": [
    {"step": 1, "description": "Извлечь величину активов из бухгалтерского баланса", "numeric_value": 633.696, "unit": "млн руб.", "is_critical": true},
    {"step": 2, "description": "Извлечь чистую прибыль из отчёта о финансовых результатах", "numeric_value": 44.691, "unit": "млн руб.", "is_critical": true},
    {"step": 3, "description": "Применить формулу рентабельности активов", "numeric_value": 0.0705, "unit": "доля", "is_critical": true},
    {"step": 4, "description": "Сформулировать итоговый ответ", "numeric_value": null, "unit": "", "is_critical": false}
  ],
  "final_answer": "ROA = 44.691 / 633.696 = 0.0705 (7.05%)",
  "faa_mode": "numeric",
  "final_numeric": 0.0705,
  "indicators": {"ROA": 0.0705},
  "fcs_applicable": true,
  "neps_mode": "skip_none",
  "template_version": "2.0.0"
}

Dataset Creation

Curation Rationale

Existing financial reasoning benchmarks (FinQA, ConvFinQA, FinChain, FinanceReasoning) are predominantly English-centric, focus on US GAAP filings, and rarely provide step-level, machine-verifiable Chain-of-Thought supervision. No open Russian-language benchmark built on real Russian accounting standards (RAS / РСБУ) with procedurally generated and automatically verified CoT steps existed prior to this work.

RuFinQA-CoT fills this gap by providing a Russian-language benchmark grounded in real financial reports, with an evaluation framework (FinCoT-Eval v2) that jointly measures final-answer correctness, step-level semantic alignment, positional alignment, numeric error propagation, and financial plausibility.

Source Data

Data Collection and Processing

1. Single source. All raw data was obtained from the State Information Resource of Accounting (Financial) Statements (BFO), operated by the Federal Tax Service of Russia (https://bo.nalog.ru/). This resource, publicly available since 2020, provides open access to balance sheets, income statements, and cash flow statements of Russian legal entities.
2. Download. We downloaded 4,418 Excel files covering companies registered in the Republic of Tatarstan for the period 2021–2025. The BFO resource is the only data source used in this benchmark; no other public or private databases were involved.
3. Extraction. A custom parser extracted numeric values from standardized report line codes, preserving the year-over-year lag structure. All monetary values were converted to millions of Russian rubles.
4. Cleaning. Records with missing required fields (inc_2110, bal_1600) were removed. A data-quality guard excluded observations with absolute values below 0.01 million rubles for balance-sheet denominators (to avoid absurd ratios such as ROA ≈ 30,000).
5. Anonymization. All company identifiers (INN) and names were replaced with synthetic ones using a deterministic mapping with a fixed random seed (42). No individual personal data is present.
6. Final size. 3,504 company-year observations covering 835 unique organizations, from which 44,625 questions were generated.

Who are the source data producers?

The source data producers are the Russian companies themselves, which are legally required to publish their annual accounting (financial) statements. The BFO resource, operated by the Federal Tax Service of Russia, aggregates and publishes these statements as open data. The author of this dataset card did not create, modify, or control the source financial figures. The BFO resource (https://bo.nalog.ru/) is the sole source of raw data for this benchmark.

Annotations

Annotation process

Chain-of-Thought steps were generated procedurally using symbolic templates. For each company-year observation, a set of question types was instantiated using parameterized templates. Each CoT step includes:

• a sequential identifier,
• a natural-language description (guaranteed to contain no digits),
• an optional numeric value (float or null),
• a unit,
• a criticality flag.

Automatic verification. Every question passed a schema validator checking:

• ≥ 2 CoT steps per question;
• ≥ 1 numeric step and ≥ 1 semantic step;
• no digits in description fields;
• CoT final numeric value matches final_numeric within tolerance;
• FCS indicators within RAS-based ranges;
• NEPS mode consistency.

Manual validation. A random subset of 100 questions was reviewed by two domain experts: 98% were fully correct, and 96% agreement on step criticality was reached. Disagreements were resolved through discussion.

Who are the annotators?

The gold CoT steps were generated procedurally (not by human annotators). Manual review was performed by two experts in accounting and financial analysis affiliated with Kazan Federal University.

Personal and Sensitive Information

The dataset contains no personal data of individuals. Company identifiers and names have been anonymized. The underlying financial figures are publicly available under Russian law. If you are a rights holder and wish to have an entry removed, contact cool.araby@gmail.com.

Bias, Risks, and Limitations

Bias

• Geographic bias. All companies are registered in one region (Republic of Tatarstan), which may not be representative of the broader Russian economy.
• Sector bias. The dataset is skewed toward sectors that file standardized BFO reports; financial institutions and government entities are underrepresented.
• Template bias. Because questions are generated from symbolic templates, the linguistic diversity of questions is lower than in naturally occurring financial text.
• Temporal bias. Coverage is limited to 2021–2025.

Risks

• Overreliance on numeric correctness. A model may produce correct final answers without faithful reasoning; users should rely on the full FinCoT-Eval v2 composite score, not just FAA.
• Re-identification risk. Although identifiers are anonymized, the underlying financial figures remain publicly available. Users must not attempt to re-identify companies.
• Misuse as investment advice. See Out-of-Scope Use above.

Limitations

• Only six open-weight LLMs (3.8B–9B parameters) were evaluated in the accompanying study; results for larger models are not yet available.
• The dataset focuses on symbolic numerical reasoning and does not cover qualitative financial analysis (risk narrative, sentiment, market interpretation).
• NEPS with skip_none mode excludes semantic steps from the summation; this is a design choice justified in the methodology but may under-penalize missing semantic steps.

Recommendations

Users should:

1. Report composite scores rather than individual metrics when comparing models.
2. Report the FAA mode distribution (numeric / exact / keyword / fuzzy_text) alongside overall results.
3. Explicitly note FCS fallback rate (117 questions, 0.26%) when reporting FCS-based rankings.
4. Cite the accompanying paper for methodological details.

Citation

BibTeX:

@misc{rufinqa2026,
  author       = {Arabov, Mullosharaf K.},
  title        = {RuFinQA: A Massive Multi-Task Reasoning Benchmark for Russian Financial Report Understanding},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/RusNLPWorld/RuFinQA-CoT}}
}

APA:

Arabov, M. K. (2026). RuFinQA: A Massive Multi-Task Reasoning Benchmark for Russian Financial Report Understanding [Data set]. Hugging Face. https://huggingface.co/datasets/RusNLPWorld/RuFinQA-CoT

Glossary

• BFO — Бухгалтерская (финансовая) отчётность: State Information Resource of Accounting (Financial) Statements, operated by the Federal Tax Service of Russia.
• RAS / РСБУ — Russian Accounting Standards.
• INN — Идентификационный номер налогоплательщика: Russian tax identification number.
• ROA — Return on Assets = Net Profit / Total Assets.
• ROE — Return on Equity = Net Profit / Shareholders' Equity.
• CR — Current Ratio = Current Assets / Current Liabilities.
• FinCoT-Eval v2 — Evaluation framework with six metrics: FAA, OT, CSPS, PSA, NEPS, FCS.

More Information

• Generator code and evaluation pipeline: available on request from the author.
• Benchmark results for six open-weight LLMs: published in the accompanying paper.
• Related benchmark: see the paper's Related Work section for comparison with FinQA, ConvFinQA, FinChain, FinanceReasoning, FINESSE-Bench, and RusFinChain.

Dataset Card Authors

Mullosharaf K. Arabov — Kazan Federal University, Institute of Computational Mathematics and Information Technologies.

Dataset Card Contact

Email: cool.araby@gmail.com Affiliation: Kazan Federal University, Kazan, Russia