File size: 1,653 Bytes
ba41a6b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
---
license: apache-2.0
language:
- en
pretty_name: j
size_categories:
- n<1K
---

# CenQuery: Indian Census Text-to-SQL Dataset

## Project Overview

CenQuery is a specialized Text-to-SQL system designed for Indian Census Data. This dataset was used to fine-tune the `defog/llama-3-sqlcoder-8b` model using QLoRA to handle complex natural language queries related to demographics, healthcare, and regional statistics.

## Dataset Composition

The dataset is split into two primary components to ensure a rigorous evaluation of the model's generalization capabilities:

- **Training Set (`old_train_(Mixed).jsonl`)**: 650 instruction-tuned examples covering diverse query patterns and schema joins.
- **Evaluation Set (`old_eval_(Mixed).jsonl`)**: 150 **unseen** questions used for final benchmarking and metric calculation (Execution Accuracy and Exact Match).

## Schema Information

The queries operate on a relational schema representing several key census sectors:

- **Population Stats**: Age-wise and gender-wise population data.
- **Healthcare Stats**: Vaccination and health infrastructure metrics.
- **Religion/Language Stats**: Distribution of religious groups and linguistic data.
- **Crop Stats**: Agricultural data including area sown and foodgrain statistics.

Full DDL definitions are provided in the `database_schema.json` file in this repository.

## Prompt Format

The data follows the SQLCoder-8B instruction format:

- `### Task`: The natural language question.
- `### Database Schema`: Relevant table DDLs.
- `### SQL`: The ground-truth SQL query.

## Citation

If using this dataset for research, please cite the CenQuery Project.