README.md

---
license: apache-2.0
language:
- my
- en
base_model: unsloth/gemma-3-4b-it
tags:
- native_mlx
- code
- burmese
- coding-assistant
- gemma3
---

# Burmese Coder 4B - MLX (4-bit)

**Burmese Coder 4B** is an instruction-tuned model based on **Gemma 3 4B**, specifically fine-tuned to assist with programming tasks in the Burmese language. This repository contains the model in the native **MLX format**, optimized for Apple Silicon (M1/M2/M3/M4).

## 🚀 Quick Start: LM Studio (Recommended)

This model is fully compatible with **LM Studio** for a seamless, local experience on macOS.

### 📦 Setup Instructions
1. **Download the model**:
   - In LM Studio, go to the **Search** tab.
   - Enter `WYNN747/burmese-coder-4b-mlx` and click **Download**.
   - Alternatively, you can use the **Local Folder** import feature for this specific directory.

2. **Configure the Inference Engine**:
   - Ensure you are using the **LM Studio MLX** backend (found in Settings > Engines & Frameworks).
   - If you see a "Missing Library" error, click the **Fix** button in the MLX settings.

Link : https://lmstudio.ai/download

## 📊 Model Performance
- **Quantization**: 4-bit affline (quantized via `mlx-lm`)
- **VRAM Usage**: ~2.3 GB
- **Inference Speed**: ~60 tokens/sec (M-series Pro/Max chips)
- **Primary Focus**: Python, Burmese language instruction-following, and natural language explanations of code.

## 🐍 Python Usage (MLX)

For programmatic access, install `mlx-lm`:

```bash
pip install mlx-lm
```

```python
from mlx_lm import load, generate

model, tokenizer = load("WYNN747/burmese-coder-4b-mlx")

messages = [
    {"role": "user", "content": "Python မှာ list တစ်ခုကို ဘယ်လို sort လုပ်ရမလဲ?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

response = generate(model, tokenizer, prompt=prompt, verbose=True)
print(response)
```

## 📝 Training Details
- **Base Model**: `unsloth/gemma-3-4b-it`
- **Adapter**: `WYNN747/burmese-coder-4b` (LoRA merged)
- **Dataset**: Custom Burmese MBPP and HumanEval-translated datasets.
- **Conversion**: Merged and quantized using the MLX framework for maximum performance on macOS.
  

## 📜 Citation
If you use Burmese Coder in your research or application, please cite the following paper:

```bibtex
@inproceedings{naing2026burmesecoder,
  title = {Burmese-Coder-4B: Fine-Tuning a Small Language Model for Burmese Coding with Language-Aware Evaluation},
  author = {Naing, Wai Yan Nyein},
  year = {2026},
  publisher = {GitHub Pages},
  url = {https://www.waiyannyeinnaing.com/whitepaper/burmese-coder-4b-paper.pdf},
  dataset = {https://huggingface.co/datasets/WYNN747/burmese-human-eval},
  dataset_mbpp = {https://huggingface.co/datasets/WYNN747/burmese-mbpp}
}
```

Find the full research project here: [Burmese-Coder-4B]
- https://www.waiyannyeinnaing.com/projects/burmese-coder-4b
- https://www.waiyannyeinnaing.com/whitepaper/burmese-coder-4b-paper.pdf


## 📜 License
This model is released under the **Apache 2.0** license.