developmentseed/gazet-dataset

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- text-to-sql
- geospatial
- geocoding
- duckdb
- synthetic
size_categories:
- 10K<n<100K
---

# Gazet Dataset

Synthetic training data for finetuning small language models on geospatial tasks over [Overture Maps](https://overturemaps.org/) and [Natural Earth](https://www.naturalearthdata.com/) parquet datasets.

## Tasks

### SQL generation (`sql/`)

Input: user query + fuzzy-matched candidate entities (CSV)
Output: DuckDB spatial SQL query

### Place extraction (`places/`)

Input: natural language query
Output: structured JSON with place names, country codes, and subtypes

## Format

Each JSONL row is a conversation in chat-template format:

```json
{
"messages": [
{"role": "system", "content": "..."},
{"role": "user", "content": "..."},
{"role": "assistant", "content": "..."}
]
}
```

## Splits

| Task | Train | Val | Test |
|---|---|---|---|
| SQL | `sql/train.jsonl` | `sql/val.jsonl` | `sql/test.jsonl` |
| Places | `places/train.jsonl` | `places/val.jsonl` | `places/test.jsonl` |

See `stats.json` for per-family sample counts.

## Generation

Data is generated from SQL templates applied to real Overture/Natural Earth spatial relations (adjacency, containment, intersection, etc.). Templates produce both the training SQL and the natural language question.

## Code & Development

This model was trained and evaluated using code in the [**developmentseed/gazet**](https://github.com/developmentseed/gazet) GitHub repository.

## Trained model
`developmentseed/gazet-model` - Qwen3.5-0.8B finetuned on this dataset