mwaldrop/heavydb-text-to-sql

---
license: apache-2.0
task_categories:
- text-generation
- text2text-generation
language:
- en
tags:
- text-to-sql
- sql
- heavydb
- geospatial
- code
- nlp
size_categories:
- 1K<n<10K
---

# HeavyDB Text-to-SQL Dataset

A dataset for training language models to convert natural language questions to **HeavyDB SQL queries**.

## Overview

[HeavyDB](https://www.heavy.ai/) is a GPU-accelerated SQL database with powerful geospatial support. This dataset contains question-SQL pairs specifically designed for HeavyDB syntax, including geospatial queries using ST_* functions.

## Dataset Statistics

| Split | Examples |
|-------|----------|
| Train | 8,217 |
| Validation | 965 |
| Test | 484 |
| **Total** | **9,666** |

### SQL Pattern Distribution

| Pattern | Percentage |
|---------|------------|
| SELECT | 99.5% |
| WHERE | 63.9% |
| JOIN | 44.3% |
| GROUP BY | 23.8% |
| ST_* (geospatial) | 8.0% |

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("mwaldrop/heavydb-text-to-sql")

# Access training data
for example in dataset["train"]:
print(f"Question: {example['question']}")
print(f"SQL: {example['query']}")
break
```

## Data Format

Each example contains:

| Field | Description |
|-------|-------------|
| `instruction` | Task description for instruction-tuning |
| `input` | The natural language question |
| `output` | The corresponding SQL query |
| `question` | Raw question text |
| `query` | Raw SQL query |
| `source` | Origin of the example |
| `db_id` | Database identifier |
| `dataset` | Source dataset name |

## Example

```
Question: How many heads of the departments are older than 56?
SQL: SELECT COUNT(*) AS num_heads FROM head WHERE age > 56;
```

## Recommended Models for Fine-tuning

This dataset works well with:
- [SQLCoder](https://huggingface.co/defog/sqlcoder-7b-2) - Purpose-built for SQL
- [CodeLlama](https://huggingface.co/codellama/CodeLlama-7b-Instruct-hf) - Strong code understanding
- [DeepSeek-Coder](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct) - Excellent for code tasks

## Training Tips

1. Use QLoRA for efficient fine-tuning on consumer GPUs
2. Include the database schema in prompts for better accuracy
3. Validate generated SQL against HeavyDB before deployment

## License

Apache 2.0

## Citation

```bibtex
@dataset{heavydb_text_to_sql_2024,
title={HeavyDB Text-to-SQL Dataset},
author={mwaldrop},
year={2024},
publisher={Hugging Face},
url={https://huggingface.co/datasets/mwaldrop/heavydb-text-to-sql}
}
```