---
language:
- en
- ar
base_model: Qwen/Qwen2.5-3B-Instruct
tags:
- peft
- lora
- gis
- arcgis
- compass-ai
- geospatial
license: apache-2.0
---

# 🧭 CompassAI — Qwen2.5-3B GIS LoRA

Fine-tuned version of [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct)
for GIS intent detection and ArcGIS JS API 4.x code generation.

## Training Details
- **Method:** QLoRA (4-bit quantization + LoRA rank 16)
- **Task:** Natural language → GIS API call (JSON)
- **Languages:** English + Arabic
- **Base model:** Qwen/Qwen2.5-3B-Instruct

## Supported Intents
- `select_by_attribute` — SQL-based feature filtering
- `select_by_location` — Spatial proximity queries
- `count_aggregate` — Statistical summaries
- `zoom_navigate` — Map navigation
- `layer_toggle` — Layer visibility control
- `spatial_filter` — Polygon-based selection

## Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained('Qwen/Qwen2.5-3B-Instruct')
model = PeftModel.from_pretrained(base_model, 'rafat234/compass-ai-qwen2.5-3b-gis-lora')
tokenizer = AutoTokenizer.from_pretrained('rafat234/compass-ai-qwen2.5-3b-gis-lora')
```