---
license: mit
datasets:
- developmentseed/gazet-dataset
language:
- en
base_model:
- Qwen/Qwen3.5-0.8B
pipeline_tag: text-generation
---

# Gazet Model

LoRA-finetuned [Qwen3.5-0.8B](https://huggingface.co/Qwen/Qwen3.5-0.8B) for natural-language geocoding over [Overture Maps](https://overturemaps.org/) and [Natural Earth](https://www.naturalearthdata.com/) parquet datasets.

Two tasks:
- **Place extraction**: Given a user query, extract structured place names with optional country codes and subtypes
- **Text-to-SQL**: Given a user query and fuzzy-matched candidate entities, generate a DuckDB spatial SQL query

## Files

| File | Description |
|---|---|
| [ckpt-q8_0.gguf](https://huggingface.co/developmentseed/gazet-model/resolve/main/models/ckpt-q8_0.gguf) | Q8_0 quantized GGUF (812 MB), ready for llama-server |
| [merged/](https://huggingface.co/developmentseed/gazet-model/tree/main/merged) | Full merged safetensors (for re-quantization or further finetuning) |

## Usage

Serve with [llama-server](https://github.com/ggml-org/llama.cpp):

```bash
# Download
hf download developmentseed/gazet-model ckpt-q8_0.gguf

# Serve
llama-server -m ckpt-q8_0.gguf -ngl 99 --port 9000 --ctx-size 2048
```

The model exposes `/v1/chat/completions` on port 9000.

Or use with the full gazet stack via Docker Compose (see [gazet](github.com/developmentseed/gazet) repo).

## Training
```yaml
Base model: unsloth/Qwen3.5-0.8B
Method: LoRA (r=16, alpha=32) via Unsloth
Data: developmentseed/gazet-dataset
Hardware: Single H200 on Modal (~2 hrs/epoch)
Optimizer: AdamW 8-bit, lr=1e-4, linear schedule
Max sequence length: 2048
Loss: Train on assistant responses only (Unsloth train_on_responses_only)
Full training code: github.com/developmentseed/gazet
```