README.md

---
task_categories:
- text-retrieval
tags:
- text2sql
- text-2-sql
- texttosql
- text-to-sql
license: cc-by-nc-4.0
language:
- en
pretty_name: FINCH
size_categories:
- 10K<n<100K
---

# Dataset Card for FINCH - Financial Intelligence using Natural language for Contextualized SQL Handling

A comprehensive collection of SQLite databases from the FINCH benchmark, containing **33 databases** with **292 tables** and **75,725 natural language-SQL pairs** across diverse financial domains for Text-to-SQL research and development.

## Dataset Details

### Dataset Description

**Curated by:** [Domyn](https://www.domyn.com/)  
**Authors:** Avinash Kumar Singh, Bhaskarjit Sarmah, Stefano Pasquali  
**Language(s):** English  
**License:** CC-BY-NC-4.0

FINCH (Financial Intelligence using Natural language for Contextualized SQL Handling) provides SQLite database files from a carefully curated financial Text-to-SQL benchmark that consolidates and extends existing resources into a unified, finance-specific dataset. Each database preserves original schema structure, relationships, and data while focusing specifically on financial domains and applications.

This dataset addresses a critical gap in Text-to-SQL research: despite significant progress in general-domain benchmarks, financial applications remain especially challenging due to complex schemas, domain-specific terminology, and high stakes of error. FINCH provides the first large-scale, finance-oriented Text-to-SQL benchmark suitable for both evaluation and fine-tuning.

### Dataset Sources

**Paper:** FINCH: Financial Intelligence using Natural language for Contextualized SQL Handling *(coming soon)*  


## Key Features

- **33 SQLite databases** specifically curated for financial applications
- **292 tables** with **2,233 columns** and **177 relations**
- **75,725 NL-SQL pairs** for comprehensive training and evaluation
- **Financial domain focus** including retail, banking, insurance, e-commerce, funds, stocks, and accounting
- **Direct SQLite format** - ready for SQL queries and analysis
- **Preserved relationships** - foreign keys and indexes intact
- **Multi-difficulty coverage** with easy, medium, and hard query complexity levels

## Dataset Structure

<div align="center">
  <img src="finch.png" alt="FINCH - Financial Intelligence using Natural language for Contextualized SQL Handling" width="400"/>
</div>

<br>

The dataset is organized by financial domain with meaningful database names:

### File Organization
```
finch/
├── spider/           # 22 SQLite files (financial subset from Spider)
├── bird/             # 7 SQLite files (financial subset from BIRD)
├── bull/             # 3 SQLite files (BULL/CCKS financial data)
└── book_sql/         # 1 SQLite file (BookSQL accounting data)
```

### Financial Domains Covered

#### Retail & E-commerce
- **customers_and_invoices**: E-commerce customer and billing systems
- **e_commerce**: Online retail transactions and order management  
- **department_store**: Retail chain operations and inventory management
- **shop_membership**: Customer loyalty and membership programs

#### Banking & Financial Services
- **financial**: Czech bank transactions and loan portfolios (1M+ records)
- **small_bank**: Banking account management systems
- **loan_1**: Loan processing and customer account data

#### Insurance & Risk Management
- **insurance_policies**: Insurance claims and policy management
- **insurance_and_eClaims**: Electronic claims processing systems
- **insurance_fnol**: First notification of loss handling

#### Investment & Trading
- **ccks_fund**: Mutual fund management and performance data
- **ccks_stock**: Stock market data and trading information
- **tracking_share_transactions**: Investment portfolio tracking

#### Sales & Marketing
- **sales**: Large-scale sales transactions (6M+ records)
- **sales_in_weather**: Sales data correlated with external factors
- **customers_campaigns_ecommerce**: Marketing campaign effectiveness

#### Accounting & Financial Reporting
- **accounting**: Complete accounting system with 185+ tables covering transactions, customers, vendors, and financial reporting
- **school_finance**: Educational institution financial management

## Dataset Format & Examples

### Data Files Structure
- **`finch_dataset.json`**: Main dataset file with 75,725 NL-SQL pairs (appears in HF dataset viewer)
- **`schemas/database_schemas.yaml`**: Database schema metadata for all 33 databases (auxiliary file)
- **`text2sql-db/`**: SQLite database files organized by source (auxiliary files)

### Sample Data from finch_dataset.json

```json
[
    {
        "question_id": 1,
        "db_id": "financial",
        "db_name": "bird",
        "question": "How many accounts who choose issuance after transaction are staying in East Bohemia region?",
        "partition": "dev",
        "difficulty": "medium",
        "SQL": "SELECT COUNT(t2.account_id) FROM district AS t1 INNER JOIN account AS t2 ON t1.district_id = t2.district_id WHERE t1.a3 = 'east bohemia' AND t2.frequency = 'poplatek po obratu'"
    },
    {
        "question_id": 2,
        "db_id": "financial", 
        "db_name": "bird",
        "question": "How many accounts who have region in Prague are eligible for loans?",
        "partition": "dev",
        "difficulty": "easy",
        "SQL": "SELECT COUNT(t1.account_id) FROM account AS t1 INNER JOIN loan AS t2 ON t1.account_id = t2.account_id INNER JOIN district AS t3 ON t1.district_id = t3.district_id WHERE t3.a3 = 'prague'"
    },
    {
        "question_id": 3,
        "db_id": "financial",
        "db_name": "bird", 
        "question": "The average unemployment ratio of 1995 and 1996, which one has higher percentage?",
        "partition": "dev",
        "difficulty": "easy",
        "SQL": "SELECT DISTINCT IIF(AVG(a13) > AVG(a12), '1996', '1995') FROM district"
    }
]
```

### Schema Information (schemas/database_schemas.yaml)

The `schemas/database_schemas.yaml` file contains comprehensive schema metadata for all databases:

```yaml
financial:
  db_id: financial
  table_names_original:
    - account
    - card
    - client
    - disp
    - district
    - loan
    - order
    - trans
  table_names:
    - account
    - card
    - client
    - disposition
    - district
    - loan
    - order
    - transaction
  column_names_original:
    - [-1, "*"]
    - [0, "account_id"]
    - [0, "district_id"]
    - [0, "frequency"]
    - [0, "date"]
  column_types:
    - text
    - number
    - number
    - text
    - text
  foreign_keys:
    - [2, 1]
    - [4, 2]
  primary_keys:
    - 1
```

## Example Usage

### Loading with Python

### Primary Method: Using datasets library (Recommended)

```python
from datasets import load_dataset
from huggingface_hub import hf_hub_download
import sqlite3
import yaml

# Load the main dataset using HuggingFace datasets library
dataset = load_dataset("domyn/FINCH")
print(f"Dataset: {dataset}")
print(f"Number of examples: {len(dataset['train'])}")

# Access individual examples
sample = dataset['train'][0]
print(f"Question: {sample['question']}")
print(f"SQL: {sample['SQL']}")
print(f"Database: {sample['db_id']}")
print(f"Difficulty: {sample['difficulty']}")

# Load schema information for the database
schema_path = hf_hub_download(repo_id="domyn/FINCH", filename="schemas/database_schemas.yaml")
with open(schema_path, 'r') as f:
    schemas = yaml.safe_load(f)

# Download the corresponding SQLite database
db_path = hf_hub_download(
    repo_id="domyn/FINCH", 
    filename=f"text2sql-db/text2sql/bird/{sample['db_id']}.sqlite"
)

# Execute the SQL query on the actual database
conn = sqlite3.connect(db_path)
cursor = conn.cursor()
cursor.execute(sample['SQL'])
results = cursor.fetchall()
print(f"Query Results: {results}")
```

### Alternative Method: Direct file download

```python
import json
import sqlite3
from huggingface_hub import hf_hub_download

# Alternative: Load dataset JSON file directly
samples_path = hf_hub_download(repo_id="domyn/FINCH", filename="finch_dataset.json")
with open(samples_path, 'r') as f:
    dataset = json.load(f)

sample = dataset[0]  # First sample
print(f"Question: {sample['question']}")
print(f"SQL: {sample['SQL']}")
```

### Financial Query Examples

```python
# Analyze banking transactions
cursor.execute("""
    SELECT account_id, SUM(amount) as total_balance 
    FROM transactions 
    WHERE transaction_date >= '2023-01-01' 
    GROUP BY account_id 
    ORDER BY total_balance DESC
""")

# Insurance claims analysis
cursor.execute("""
    SELECT policy_type, COUNT(*) as claim_count, AVG(claim_amount)
    FROM claims c
    JOIN policies p ON c.policy_id = p.policy_id
    WHERE claim_status = 'approved'
    GROUP BY policy_type
""")
```

### Schema Exploration

```python
# Get all tables
cursor.execute("SELECT name FROM sqlite_master WHERE type='table'")
tables = cursor.fetchall()
print("Available tables:", tables)

# Get detailed schema information
cursor.execute("PRAGMA table_info(transactions)")
schema = cursor.fetchall()
for column in schema:
    print(f"Column: {column[1]}, Type: {column[2]}")
```

## Data Quality & Statistics

### Database Statistics

**📊 TOTAL DATABASES: 33**  
**📅 FINANCIAL DOMAINS: 8+ specialized areas**  
**🏢 TABLES: 292 across all databases**  
**🔗 RELATIONS: 177 foreign key relationships**  
**💼 NL-SQL PAIRS: 75,725 total examples**

| Source | Database Count | Table Count | NL-SQL Pairs | Domain Focus |
|--------|---------------|-------------|--------------|--------------|
| Spider (financial) | 22 | 145 | 1,100 | Cross-domain financial |
| BIRD (financial) | 7 | 48 | 1,139 | Large-scale realistic |
| BULL/CCKS | 3 | 99 | 4,966 | Chinese financial markets |
| BookSQL | 1 | 185 | 68,907 | Accounting systems |
| **TOTAL** | **33** | **292** | **75,725** | **Financial** |

### Difficulty Distribution

- **Easy queries**: 9,358 examples (12.4%)
- **Medium queries**: 33,780 examples (44.6%) 
- **Hard queries**: 32,587 examples (43.0%)

### Quality Assurance

The dataset has undergone extensive validation and cleaning:
- ✅ **SQL execution verified** for all 75,725 queries
- ✅ **Schema consistency** maintained across all databases
- ✅ **Error correction** performed on original datasets:
  - BIRD: 327 queries fixed (column names, table references)
  - BULL: 60 queries corrected (syntax errors, invalid references)
  - BookSQL: 9,526 queries repaired (column names, table references, syntax)
- ✅ **Financial domain relevance** verified for all included databases

## Applications

This dataset is specifically designed for:

### Financial Research Applications
- **Financial Text-to-SQL Systems**: Train models specifically for financial database querying
- **Domain Adaptation Studies**: Research cross-domain transfer from general to financial SQL
- **Financial Schema Understanding**: Develop models that understand complex financial relationships
- **Regulatory Compliance**: Build systems for automated financial reporting and compliance checking
- **Risk Analysis Automation**: Create tools for automated risk assessment query generation

### Industry Applications  
- **Financial Analytics Platforms**: Natural language interfaces for financial data analysis
- **Banking Query Systems**: Customer service and internal analyst tools
- **Investment Research**: Automated portfolio analysis and market research
- **Regulatory Reporting**: Compliance and audit report generation
- **Insurance Processing**: Claims analysis and policy management systems

### Educational Applications
- **Financial SQL Training**: Teach SQL with realistic financial datasets
- **Business Intelligence Education**: Train on real-world financial database structures
- **Fintech Development**: Build and test financial technology applications

## FINCH Evaluation Metric

The dataset introduces the **FINCH Score**, a specialized evaluation metric for financial Text-to-SQL that addresses limitations of traditional exact-match and execution accuracy metrics:

### Key Features of FINCH Score
- **Component-wise Scoring**: Weighted evaluation of SQL clauses (SELECT, WHERE, JOIN, etc.)
- **Financial Clause Priority**: Higher weights for business-critical clauses (WHERE, JOIN, GROUP BY)
- **Execution Tolerance**: Materiality-aware tolerance for floating-point differences
- **Structural Fidelity**: Emphasis on semantic correctness over syntactic matching

### Mathematical Formulation
```
FINCH Score = S(q̂,q*)^β × (δ + (1-δ)e(q̂,q*))
```
Where:
- S(q̂,q*): Weighted component similarity score
- e(q̂,q*): Execution accuracy with tolerance τ
- β: Structural fidelity parameter
- δ: Execution failure penalty parameter

## Benchmark Results

Initial benchmarking on FINCH reveals detailed performance across multiple state-of-the-art models:

### Model Performance Table

| Model | Exact Match | Execution Accuracy | Component Match | FINCH Score |
|-------|-------------|-------------------|-----------------|-------------|
| **GPT-OSS-120B** | 1.8% | 27.8% | 16.6% | **11.6%** |
| **Arctic-Text2SQL-R1-7B** | 0.6% | 2.3% | 3.7% | **1.5%** |
| **Qwen3-235B-A22B** | 0.7% | 2.5% | 2.8% | **1.2%** |
| **Qwen3-8B** | 0.5% | 0.8% | 3.5% | 1.2% |
| **GPT-OSS-20B** | 0.3% | 7.5% | 5.2% | 3.0% |
| **Phi-4-mini-reasoning** | 0.0% | 0.2% | 1.0% | 0.4% |

### SQL Clause-Level Performance

Analysis of errors by SQL clause reveals systematic challenges:

| Model | SELECT | FROM | WHERE | GROUP BY | HAVING | ORDER BY | LIMIT |
|-------|--------|------|-------|----------|--------|----------|--------|
| **GPT-OSS-120B** | 4.7% | **27.3%** | 6.9% | 7.5% | 6.3% | 6.3% | **73.8%** |
| **Arctic-Text2SQL-R1-7B** | 2.5% | 3.6% | 0.7% | 4.7% | 1.0% | 1.3% | 42.7% |
| **GPT-OSS-20B** | 1.4% | 6.2% | 1.5% | 8.4% | 3.7% | 1.5% | 65.2% |

### Model Performance Hierarchy
1. **GPT-OSS-120B**: Strongest overall performance (11.6% FINCH Score)
2. **Arctic-Text2SQL-R1-7B**: Best domain-adapted model despite smaller size (1.5% FINCH Score)
3. **GPT-OSS-20B**: Solid medium-scale performance (3.0% FINCH Score)

### Key Research Findings
- **Domain adaptation** outperforms scale alone - Arctic-Text2SQL-R1-7B (7B params) rivals much larger models
- **Schema-sensitive clauses** (SELECT, FROM, WHERE) remain the primary bottleneck
- **Query difficulty** shows steep performance degradation: easy queries achieve ~26.5% vs hard queries at ~4.5%
- **Financial complexity** significantly impacts all models, with even SOTA systems achieving modest absolute performance
- **FINCH Score correlation**: Provides more nuanced assessment than traditional exact-match metrics


## Data Source & Methodology

FINCH consolidates financial databases from multiple sources:

1. **Careful Domain Selection**: Only financial-relevant databases retained
2. **Comprehensive Validation**: All SQL queries tested for execution
3. **Error Correction**: Systematic fixing of syntax and schema errors
4. **Difficulty Annotation**: Query complexity labeled following established guidelines
5. **Schema Normalization**: All databases converted to SQLite for consistency

The curation process prioritized financial domain relevance while maintaining the diversity and complexity necessary for robust model evaluation.

## Ethical Considerations

- **Public Domain Data**: All source databases are from publicly available benchmarks
- **Financial Privacy**: No real customer or proprietary financial data included
- **Synthetic Data**: Financial amounts and transactions are synthetic or anonymized
- **Research Purpose**: Intended primarily for academic and research applications
- **Domain Compliance**: Respects financial data handling best practices

## Citation

If you use the FINCH dataset in your research, please cite:

```bibtex
@inproceedings{singh2025finch,
  title={FINCH: Financial Intelligence using Natural language for Contextualized SQL Handling},
  author={Singh, Avinash Kumar and Sarmah, Bhaskarjit and Pasquali, Stefano},
  booktitle={Proceedings of Advances in Financial AI: Innovations, Risk, and Responsibility in the Era of LLMs (CIKM 2025)},
  year={2025},
  organization={ACM}
}
```

## Dataset Card Contact

For questions about the FINCH dataset, please contact the research team at Domyn.

**Research Team:**
- Avinash Kumar Singh (avinash.kumarsingh@domyn.com)
- Bhaskarjit Sarmah (bhaskarjit.sarmah@domyn.com)  
- Stefano Pasquali (stefano.pasquali@domyn.com)