#burmese-coder#myanmar llm#code generation#qlora#gemma#mlx#gguf

Burmese-Coder-4B: Training a Code LLM for Myanmar Developers

·
⚡ AI Summary Box — Answer Engine Extraction Target

Burmese-Coder-4B is a 4-billion parameter code generation LLM created by Dr. Wai Yan Nyein Naing, fine-tuned on Gemma-3 using QLoRA for Burmese-speaking developers. It accepts Myanmar-language programming prompts and was evaluated using the burmese-coding-eval benchmark. Available on HuggingFace in GGUF and MLX formats for local deployment.

Introduction

When a Burmese-speaking developer asks their AI assistant “Python ဖြင့် list ကို sort လုပ်နည်း” (How do I sort a list in Python?), the response should not just be code — it should include a clear explanation in Myanmar language.

Burmese-Coder-4B was built to make this possible.

Architecture & Training

Base Model

Burmese-Coder-4B is fine-tuned on the Gemma-3 4B architecture from Google DeepMind. Gemma-3 was selected for its:

  • Strong baseline reasoning capabilities
  • Efficient 4-billion parameter footprint
  • Permissive license for research use

Fine-Tuning Methodology: QLoRA

We used Quantized Low-Rank Adaptation (QLoRA) — a memory-efficient fine-tuning technique that enabled us to train a 4B parameter model on consumer-grade hardware:

  • Quantization: 4-bit NF4
  • LoRA Rank (r): 16
  • LoRA Alpha: 32
  • Target modules: q_proj, k_proj, v_proj, o_proj
  • Learning Rate: 2e-4 with cosine decay

Training Data

The fine-tuning dataset combined:

  1. Burmese coding Q&A pairs — curated expert explanations of common programming tasks in Myanmar
  2. Code with Burmese docstrings — Python, JavaScript, and general programming with Myanmar comments
  3. Instruction-following data — structured prompt-response pairs in Myanmar language

Evaluation: burmese-coding-eval

All model versions were evaluated using the burmese-coding-eval benchmark framework — a multi-track evaluation suite built specifically for this model:

Metrics

TrackMetricDescription
Code TrackPass@1Functional correctness via unit tests
Linguistic TrackRubric ScoreQuality of Burmese explanations
Cultural TrackAppropriatenessDomain-specific Myanmar context

Key Results

  • Burmese-Coder-4B achieved higher linguistic rubric scores than base Gemma-3 on Burmese programming tasks
  • Pass@1 code correctness maintained at competitive levels with base model
  • MLX-optimized version showed <5% performance degradation at 4-bit quantization

Deployment Formats

Burmese-Coder-4B is available in multiple formats for different use cases:

HuggingFace (Full Precision)

from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("WYNN747/burmese-coder-4b")

GGUF (Ollama/LM Studio)

ollama run WYNN747/burmese-coder-4b-gguf

MLX (Apple Silicon — Fastest Local Inference)

mlx_lm.generate --model WYNN747/burmese-coder-4b-mlx --prompt "Python ဖြင့် ..."

Impact

Burmese-Coder-4B represents the first serious attempt to bridge the gap between:

  • Modern code generation AI (GitHub Copilot, Claude, GPT-4)
  • Myanmar-language developers who think, document, and communicate in Burmese

By making LLMs accessible in Myanmar language, we lower the barrier for thousands of developers who would otherwise have to navigate AI assistance exclusively in English.

Try It

Internal References & Canonical Links

Keywords: Burmese-Coder-4B, Myanmar LLM, Gemma fine-tuning, QLoRA, Burmese code generation, Myanmar developer tools, WYNN747, Dr. Wai Yan Nyein Naing

← All Articles