README.md

---
language:
- ru
license: mit
size_categories:
- 10K<n<100K
task_categories:
- question-answering
- text-generation
- table-question-answering
pretty_name: RusFinChain
tags:
- russian
- finance
- llm-evaluation
- chain-of-thought
- benchmarking
- reasoning
- cot
- financial-analysis
---

# 🏦 RusFinChain

**RusFinChain** is a Russian benchmark for evaluating Large Language Models (LLMs) on financial analysis tasks with **Ground-Truth Chain-of-Thought**.

---

## 📊 Overview

- **Total questions:** 44,627  
- **Task types:** 7  
- **Skills:** 12  
- **Difficulty levels:** 3  
- **Version:** 3.3.0  
- **Language:** Russian  

---

## 📖 Source Material & Data Licensing

**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.

This dataset significantly extends the book's materials by generating large‑scale Chain‑of‑Thought (CoT) reasoning traces, adding diverse question types (adversarial, multi‑step, etc.), and structuring them into a unified benchmark for LLM evaluation.

**Data Provenance & Rights**  
The benchmark includes **anonymized excerpts** and synthetic financial indicators inspired by real‑world reporting standards of Russian companies. All proprietary identifiers (TINs, company names, specific monetary values tied to real entities) have been fully pseudonymized or replaced with synthetic analogs.

- **Annotations, questions, reasoning chains, and metadata** are distributed under the **MIT License**.  
- **Underlying textual patterns and raw financial data structures** are used strictly for academic research. We do not claim ownership of the original financial reports of specific companies.

**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.

---

## 🧩 Task Types

| Type | Count | Share | Description |
|------|-------|-------|-------------|
| `factoid` | 12,255 | 27.5% | Fact extraction from company data |
| `arithmetic` | 12,021 | 26.9% | Financial metric calculation (ROA, ROE, CR) |
| `multistep` | 7,799 | 17.5% | Multi‑step reasoning |
| `comparison` | 4,085 | 9.2% | Year‑over‑year metric comparison |
| `analytical` | 4,059 | 9.1% | Financial metric interpretation |
| `reasoning` | 4,052 | 9.1% | Capital structure analysis (leverage) |
| `adversarial` | 356 | 0.8% | Robustness tests against anomalies |

---

## 📈 Difficulty Distribution

| Level | Count | Share |
|-------|-------|-------|
| `hard` | 16,292 | 36.5% |
| `medium` | 16,080 | 36.0% |
| `easy` | 12,255 | 27.5% |

---

## 🔗 Chain‑of‑Thought Structure

Each question contains annotated reasoning steps:

```json
{
  "reasoning_steps": [
    {
      "step": 1,
      "description": "Extract values from the financial report",
      "intermediate_result": "bal_1600 = 633.696, inc_2400 = 44.691",
      "is_critical": true
    },
    {
      "step": 2,
      "description": "Apply the ROA formula",
      "intermediate_result": "ROA = 44.691 / 633.696",
      "is_critical": true
    },
    {
      "step": 3,
      "description": "Compute the result",
      "intermediate_result": "ROA = 0.0705 (7.05%)",
      "is_critical": true
    }
  ],
  "cot_text": "Step 1: ... → ...\nStep 2: ... → ...\n\nAnswer: ...",
  "final_answer": "ROA = 44.691 / 633.696 = 0.0705 (7.05%)"
}
```

---

## 📝 Data Structure

Each record contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique question identifier |
| `type` | string | Task type (arithmetic/factoid/...) |
| `skill` | string | Evaluated skill |
| `difficulty` | string | Difficulty level (easy/medium/hard) |
| `question` | string | Question in Russian |
| `context` | dict | Context (TIN, year, financial data) |
| `reasoning_steps` | list | Ground‑Truth reasoning steps |
| `cot_text` | string | Textual representation of CoT |
| `final_answer` | string | Final answer |
| `numeric_answer` | float | Numeric answer (for arithmetic) |
| `tolerance` | float | Tolerance for numeric evaluation |
| `evaluation` | string | Evaluation metric type |

---

## 🚀 Usage

### Loading the Benchmark

```python
from datasets import load_dataset

dataset = load_dataset("arabovs-ai-lab/RusFinChain", split="train")
print(f"Total questions: {len(dataset)}")
print(dataset[0])
```

### Example Usage for LLM Evaluation

```python
import json
from datasets import load_dataset

dataset = load_dataset("arabovs-ai-lab/RusFinChain", split="train")

# Filter by task type
arithmetic_tasks = [d for d in dataset if d['type'] == 'arithmetic']
print(f"Arithmetic tasks: {len(arithmetic_tasks)}")

# Sample question
sample = arithmetic_tasks[0]
print(f"Question: {sample['question']}")
print(f"Expected answer: {sample['final_answer']}")
print(f"CoT:\n{sample['cot_text']}")
```

### Model Evaluation

```python
# Simple 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}")
```

---

## 🧠 Evaluated Models

_Model evaluations will be added in future releases._

---

## ⚙️ Benchmark Features

1. **Ground‑Truth CoT**: Each question contains annotated reasoning steps with an `is_critical` flag for critical steps  
2. **Stratification**: Questions are stratified by type, skill, and difficulty  
3. **Metadata**: Complete information about version, distribution, and metrics  
4. **Multiple evaluation types**: `cot_numeric`, `cot_text`, `cot_exact`, `cot_keyword`, `cot_reasoning`  
5. **Adversarial examples**: Robustness tests against negative values and anomalies  
6. **Realistic data**: Anonymized financial reports of Russian companies  

---

## 📚 Applications

- Evaluating LLM reasoning capabilities  
- CoT distillation for financial tasks  
- Model comparison on Russian language  
- Investigating the impact of difficulty on answer quality  
- Testing robustness against anomalous data  

---

## 📄 License

MIT (for annotations and metadata; underlying source texts are subject to original rights – see **Source Material & Data Licensing** above).

---

## 📖 Citation

If you use this dataset in your work, please cite:

```bibtex
@misc{rusfinchain2026,
  author = {Arabov, Mullosharaf K.},
  title = {RusFinChain: A Russian Financial Benchmark with Ground-Truth Chain-of-Thought for LLM Evaluation},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/arabovs-ai-lab/RusFinChain}},
}
```

---

## 👤 Author

**Mullosharaf K. Arabov**  
PhD in Physics and Mathematics, Associate Professor  
Department of Data Analysis and Programming Technologies  
Kazan (Volga Region) Federal University

---

## 🔗 Links

- 📊 Dataset: https://huggingface.co/datasets/arabovs-ai-lab/RusFinChain  
- 💻 Generator code: [GitHub]  
- 📄 Paper: https://arxiv.org/abs/2607.01388  

---

*Automatically generated from the RusFinChain evaluation pipeline.*