Datasets:
metadata
license: mit
language:
- ru
task_categories:
- question-answering
pretty_name: RusFinanceBenchmark
tags:
- russian
- finance
- symbolic-reasoning
- chain-of-thought
- benchmarking
- llm-evaluation
- financial-nlp
size_categories:
- 5K<n<10K
RusFinanceBenchmark — Symbolic Financial Reasoning in Russian
RusFinanceBenchmark is a large-scale symbolic financial reasoning benchmark in Russian, designed to evaluate the multi‑step reasoning capabilities of large language models (LLMs). It contains 5,280 tasks across 17 domains and 172 topics, with three difficulty levels: Basic (Базовый), Intermediate (Средний), Advanced (Продвинутый).
Each task includes a natural‑language question, a step‑by‑step solution, a final numeric answer, a LaTeX formula, and an executable Python expression — making it ideal for verifiable Chain‑of‑Thought evaluation.
📊 Dataset Statistics
| Metric | Value |
|---|---|
| Total Tasks | 5,280 |
| Domains | 17 |
| Topics | 172 |
| Levels | Basic, Intermediate, Advanced |
| Avg. Question Length | 147.8 characters |
| Avg. Solution Length | 72.3 characters |
| Fields Coverage | steps: 100.0%, details: 84.7%, formula_latex: 100.0%, python_computation: 100.0% |
📚 Domains
| Domain (Russian) | Domain (English) | Tasks |
|---|---|---|
| Ценные бумаги | Securities | 540 |
| Финансовое регулирование | Financial Regulation | 420 |
| Налоги | Taxation | 360 |
| Аннуитеты и вклады | Annuities and Deposits | 330 |
| Финансовые рынки | Financial Markets | 330 |
| Личные финансы | Personal Finance | 330 |
| Крипто-финансы | Crypto Finance | 300 |
| ESG и устойчивое финансирование | ESG and Sustainable Finance | 300 |
| Финансовые коэффициенты | Financial Ratios | 300 |
| Процентные ставки | Interest Rates | 300 |
| Кредиты и займы | Loans and Borrowings | 300 |
| Слияния и поглощения (M&A) | Mergers & Acquisitions (M&A) | 300 |
| Управление рисками | Risk Management | 300 |
| Амортизация | Depreciation | 270 |
| Инвестиционные проекты | Investment Projects | 240 |
| Страхование и актуарные расчёты | Insurance and Actuarial Science | 180 |
| Корпоративные финансы | Corporate Finance | 180 |
📝 Data Structure
Each example is a JSON object with the following fields:
| Field | Type | Description (English / Russian) |
|---|---|---|
seed |
int | Random seed used for generation (Случайное зерно генерации) |
id |
string | Unique task identifier (template ID) (Уникальный идентификатор задачи) |
level |
string | Difficulty level: Basic, Intermediate, Advanced (Уровень сложности) |
domain |
string | Financial domain (e.g., Процентные ставки — Interest Rates) (Финансовый домен) |
topic |
string | Specific topic within the domain (Конкретная тема внутри домена) |
question |
string | Natural‑language question in Russian (Вопрос на естественном русском языке) |
steps |
list | Array of reasoning steps, each with step (int), description (str), and value (float or null) (Массив шагов рассуждения) |
final_answer |
float | Final numeric answer (Итоговый числовой ответ) |
solution |
string | Human‑readable step‑by‑step solution (Пошаговое решение для человека) |
details |
dict | Additional information (optional, may be empty) (Дополнительная информация, может быть пустой) |
formula_latex |
string | LaTeX representation of the key formula (LaTeX‑представление ключевой формулы) |
python_computation |
string | Executable Python expression to compute the answer (Исполняемое Python‑выражение для вычисления ответа) |
🚀 Usage
Loading with 🤗 Datasets
from datasets import load_dataset
dataset = load_dataset("arabovs-ai-lab/RusFinanceBenchmark", split="train")
print(dataset[0])
Example Record
{
"seed": 2169157862,
"id": "3",
"level": "Intermediate",
"domain": "Процентные ставки",
"topic": "nom_from_eff",
"question": "Эффективная ставка по вкладу в Статус Финанс составляет 13.05% при капитализации каждый месяц. Какова номинальная ставка?",
"steps": [
{"step": 1, "description": "Эффективная ставка = 13.05%, период капитализации = месяц", "value": null},
{"step": 2, "description": "nom = 12 * ((1 + 0.1305)^(1/12) - 1)", "value": 12.33}
],
"final_answer": 12.33,
"solution": "Шаг 1: Формула: nom = m * ((1 + eff)^(1/m) – 1)\nШаг 2: nom = 12 * ((1 + 0.1305)^(1/12) – 1) = 12.33%",
"details": {},
"formula_latex": "\\(12 \\cdot \\left((1 + 0.1305)^{1/12} - 1\\right) \\times 100\\%\\)",
"python_computation": "12 * ((1 + 0.1305)**(1/12) - 1)"
}
📄 License
This dataset is released under the MIT License.
📚 Citation
If you use this dataset, please cite:
@misc{rusfinancebenchmark2025,
author = {Arabov Mullosharaf},
title = {RusFinanceBenchmark: A Symbolic Financial Reasoning Benchmark for Russian LLMs},
year = {2026},
publisher = {Hugging Face},
url = {https://huggingface.co/datasets/RusNLPWorld/RusFinanceBenchmark}
}