README.md

---
language:
- ru
license: mit
size_categories:
- 5K<n<10K
task_categories:
- question-answering
pretty_name: RusFinChain-Eval
tags:
- russian
- finance
- llm-evaluation
- chain-of-thought
- benchmarking
- symbolic-reasoning
---

# RusFinChain-Eval — Evaluation Results

This dataset contains **evaluation results** for 8 open-weight large language models on the **RusFinChain** benchmark (5,280 symbolic financial reasoning tasks).

## 📊 Overview

- **Total evaluated records:** 8,100
- **Models:** 8
- **Domains:** 17
- **Topics:** 172
- **Levels:** 3 (Basic, Intermediate, Advanced)

## 🤖 Models Evaluated

| Model | Records |
|-------|---------|
| **llama3.2:3b** | 1,100 |
| **phi4-mini:3.8b** | 1,000 |
| **qwen2.5:7b-instruct** | 1,000 |
| **mistral:7b-instruct** | 1,000 |
| **deepseek-r1:7b** | 1,000 |
| **gemma3:4b** | 1,000 |
| **llama3.1:8b** | 1,000 |
| **aya-expanse:8b** | 1,000 |

---

## 📖 Data Sources, Licensing & Legal Notice

### Data Origin

This dataset contains **model-generated outputs** and **evaluation metrics** produced by running open-weight large language models on the **RusFinChain** benchmark. The underlying questions, gold solutions, and reasoning steps are taken from the RusFinChain dataset.

### Ownership & Rights

- **The evaluation results, metrics, and model generations** are released under the **MIT License**.
- **The underlying benchmark questions and gold solutions** are subject to the original licensing terms of RusFinChain (MIT License).
- We **do not claim ownership** of the model outputs.

### Notice‑and‑Takedown Policy

We respect intellectual property rights. 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.

📧 **Contact for takedown requests**: `marabov@kpfu.ru`  
⏱️ **Response time**: Within 14 business days.

---

## 📈 Key Performance Metrics (aggregated)

Values are aggregated across all 8,100 model responses (mean ± std).

| Metric | Mean | Std |
|--------|------|-----|
| Hard F1 | 0.6538 | 0.2956 |
| Fuzzy F1 | 0.4810 | 0.1676 |
| Soft F1 | 0.4594 | 0.1648 |
| DTW F1 (Bonus) | 0.5681 | 0.2317 |
| DTW F1 (Gate) | 0.3895 | 0.2443 |
| BERTScore | 0.6711 | 0.0667 |
| ROUGE-L | 0.2424 | 0.1845 |
| Final Answer Match | 0.2936 | 0.4554 |
| Final Answer Fuzzy | 0.3777 | 0.4561 |

---

## 📚 Domain Distribution

| Domain (RU) | Domain (EN) | Records |
|-------------|-------------|---------|
| Ценные бумаги | Securities | 835 |
| Финансовое регулирование | Financial Regulation | 659 |
| Налоги | Taxation | 555 |
| Аннуитеты и вклады | Annuities and Deposits | 508 |
| Финансовые рынки | Financial Markets | 507 |
| Личные финансы | Personal Finance | 504 |
| Процентные ставки | Interest Rates | 475 |
| Кредиты и займы | Loans and Borrowings | 475 |
| ESG и устойчивое финансирование | ESG and Sustainable Finance | 459 |
| Крипто-финансы | Crypto Finance | 459 |
| Слияния и поглощения (M&A) | Mergers & Acquisitions (M&A) | 456 |
| Финансовые коэффициенты | Financial Ratios | 456 |
| Управление рисками | Risk Management | 448 |
| Амортизация | Depreciation | 400 |
| Инвестиционные проекты | Investment Projects | 360 |
| Страхование и актуарные расчёты | Insurance and Actuarial Science | 272 |
| Корпоративные финансы | Corporate Finance | 272 |

---

## 📊 Level Distribution

| Level | Records |
|-------|---------|
| Basic | 2,610 |
| Intermediate | 3,582 |
| Advanced | 1,908 |

---

## 📝 Data Structure

Each record contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Task identifier |
| `level` | string | Basic / Intermediate / Advanced |
| `domain` | string | Financial domain |
| `topic` | string | Specific topic |
| `model` | string | Model name |
| `question` | string | Question (Russian) |
| `solution` | string | Gold solution |
| `steps` | list | Gold reasoning steps |
| `final_answer` | float | Correct answer |
| `model_generation` | string | Raw model output |
| `recall` | float | Hard recall |
| `precision` | float | Hard precision |
| `hard_f1` | float | Hard F1 |
| `fuzzy_f1` | float | Fuzzy F1 |
| `soft_f1` | float | Soft F1 |
| `dtw_f1_bonus` | float | DTW F1 (bonus variant) |
| `dtw_f1_gate` | float | DTW F1 (gate variant) |
| `bertscore` | float | BERTScore |
| `rougeL` | float | ROUGE-L score |
| `final_answer_match` | int | Correct final answer (0/1) |
| `final_answer_fuzzy` | float | Fuzzy final answer score |

---

## 🚀 Usage

```python
from datasets import load_dataset

dataset = load_dataset("RusNLPWorld/RusFinChain-Eval", split="train")
print(dataset[0])
```

### Example Record

```json
{
  "id": "1",
  "level": "Basic",
  "domain": "Страхование и актуарные расчёты",
  "topic": "mortality_prob",
  "model": "phi4-mini:3.8b",
  "question": "Для Егор Дмитриевич Козлов возраст 79 лет, пол женский. Найдите годовую вероятность смерти (в %) по упрощённой актуарной модели.",
  "solution": "Шаг 1: Вероятность смерти = 1.86%",
  "steps": [
    {"step": 1, "description": "Вероятность смерти = a · b^возраст (параметры зависят от пола).", "value": null},
    {"step": 2, "description": "Вероятность смерти = 1.86%", "value": 1.86}
  ],
  "final_answer": 1.86,
  "model_generation": "Шаг 1: Определите возрастную группу и пол. Возрастная группа - 80-89 лет; Пол - Женский. Шаг 2: Примените упрощённую актуарную модель для расчета годовой вероятности смерти (qx) для женщин в возрасте 80-89 лет. Согласно модели, qx = 0.0125 + (возраст - 40) * 0.00075. Шаг 3: Подставьте возраст Егор Дмитриевич Козлова ...",
  "recall": 0.92,
  "precision": 0.88,
  "hard_f1": 0.90,
  "fuzzy_f1": 0.85,
  "soft_f1": 0.87,
  "dtw_f1_bonus": 0.88,
  "dtw_f1_gate": 0.82,
  "bertscore": 0.91,
  "rougeL": 0.84,
  "final_answer_match": 1,
  "final_answer_fuzzy": 0.99
}
```

### Analyzing Results

```python
import pandas as pd
from datasets import load_dataset

dataset = load_dataset("RusNLPWorld/RusFinChain-Eval", split="train")
df = pd.DataFrame(dataset)

# Accuracy per model
model_acc = df.groupby('model')['final_answer_match'].mean().sort_values(ascending=False)
print(model_acc)

# Average Hard F1 by difficulty level
level_f1 = df.groupby('level')['hard_f1'].mean()
print(level_f1)
```

---

## 📄 License

MIT License — applies to evaluation results, metrics, and metadata in this dataset. The underlying benchmark content (RusFinChain) is also released under the MIT License.

---

## 📚 Citation

If you use this evaluation dataset, please cite both the original benchmark and the evaluation results:

```bibtex
@misc{rusfinchain2026,
  author = {Arabov, Mullosharaf K.},
  title = {RusFinChain: A Symbolic Financial Reasoning Benchmark for Russian LLMs},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RusNLPWorld/RusFinChain}
}

@misc{rusfinchaineval2026,
  author = {Arabov, Mullosharaf K.},
  title = {RusFinChain-Eval: Evaluation Results for 8 LLMs on the RusFinChain Benchmark},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/RusNLPWorld/RusFinChain-Eval}
}
```

---

## 👤 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

- 📊 **RusFinChain benchmark**: https://huggingface.co/datasets/RusNLPWorld/RusFinChain  
- 📊 **RusFinChain-Eval results**: https://huggingface.co/datasets/RusNLPWorld/RusFinChain-Eval  
- 💻 **Generator code**: [GitHub]  
- 📄 **Paper**: https://arxiv.org/abs/2607.01388