🤖 amk-coder-v2 — Myanmar Coding Agent

Myanmar Coding Assistant — Fine-tuned from Qwen2.5-Coder-1.5B using LoRA (PEFT)

---

📋 Table of Contents

• Model Overview
• Training Details
• Quick Start
• Usage Examples
• API Deployment
• Limitations
• License

---

Model Overview

amk-coder-v2 is a Myanmar-localized coding assistant fine-tuned from Qwen2.5-Coder-1.5B using LoRA (PEFT) technique.

Attribute	Value
Base Model	Qwen2.5-Coder-1.5B
Parameters	2B (2,000M)
Architecture	Qwen2ForCausalLM
Training Method	LoRA (PEFT) fine-tuning
Dataset	amkyawdev/mm-llm-coder-agent-dataset (4M rows)
Context Length	32,768 tokens
Format	Safetensors (BF16)
License	Apache-2.0
Languages	Burmese + English

Features

Feature	Description
🇲🇲 Myanmar Support	Full support for Myanmar Unicode text
💻 Code Generation	Python, JavaScript, C++, Java, and more
🐛 Debugging	Bug detection and fixes
📖 Code Explanation	Line-by-line explanations

---

Training Details

Parameter	Value
Framework	Transformers + PEFT
Training Method	LoRA fine-tuning
Target Modules	q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Optimizer	paged_adamw_8bit
Learning Rate	3e-5
Epochs	3
Batch Size	8
Max Length	2048
Precision	FP16 mixed
Hardware	Kaggle Dual T4 GPU
Training Time	~3-5 hrs

Chat Template (ChatML)

<|im_start|>system
You are an expert Myanmar AI coding agent with tool access.<|im_end|>
<|im_start|>user
{Instruction}
Tools available: {Tools}<|im_end|>
<|im_start|>assistant
Thought & Code:

---

Quick Start

Using Transformers (Python)

# Method 1: Pipeline (Recommended for beginners)
from transformers import pipeline

pipe = pipeline("text-generation", model="amkyawdev/amk-coder-v2")
messages = [
    {"role": "user", "content": "Python function တစ်ခုရေးပါ။ list comprehension နဲ့ sorting လုပ်ပေးပါ။"}
]
result = pipe(messages, max_new_tokens=512, temperature=0.2)
print(result[0]['generated_text'])

# Method 2: Direct Model Loading
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

tokenizer = AutoTokenizer.from_pretrained("amkyawdev/amk-coder-v2")
model = AutoModelForCausalLM.from_pretrained(
    "amkyawdev/amk-coder-v2",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)

messages = [
    {"role": "system", "content": "You are a helpful coding assistant."},
    {"role": "user", "content": "Write a Python function to reverse a string"}
]

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=512, temperature=0.2)
response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
print(response)

Using vLLM (Production)

# Install vLLM
pip install vllm

# Start server
vllm serve "amkyawdev/amk-coder-v2" --tensor-parallel-size 1

# API call
curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amkyawdev/amk-coder-v2",
    "messages": [
      {"role": "user", "content": "Hello, write Python code"}
    ],
    "max_tokens": 512,
    "temperature": 0.2
  }'

Using SGLang

# Install SGLang
pip install sglang

# Start server
python -m sglang.launch_server --model-path "amkyawdev/amk-coder-v2" --port 30000

# API call
curl -X POST "http://localhost:30000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "amkyawdev/amk-coder-v2",
    "messages": [{"role": "user", "content": "Write a hello world in Python"}]
  }'

---

Usage Examples

🇲🇲 Myanmar Prompts

messages = [
    {"role": "user", "content": "Python function တစ်ခုရေးပါ။ ဂဏန်းတွေကို sorting လုပ်ပေးပါ။"}
]
# Output: def sort_numbers(numbers): return sorted(numbers)

🇬🇧 English Prompts

messages = [
    {"role": "user", "content": "Explain this code:\nfor i in range(10):\n    print(i)"}
]
# Output: This is a for loop that prints numbers 0 to 9

🐛 Debugging

messages = [
    {"role": "user", "content": "Fix this Python code:\nprint('Hello' + 5)"}
]
# Output: TypeError fix suggestion with corrected code

---

API Deployment

Backend Server

cd backend
pip install -r requirements.txt
export HF_TOKEN=hf_your_token
uvicorn app.main:app --host 0.0.0.0 --port 8000

Endpoints

Method	Endpoint	Description
GET	/	Health check
GET	/health	Service health status
POST	/chat	Streaming chat (SSE)
GET	/demo	Demo HTML interface
GET	/models	Model information

Request Format

# Streaming chat
curl -X POST "http://localhost:8000/chat" \
  -H "Content-Type: application/json" \
  -d '{
    "messages": [
      {"role": "user", "content": "Write a Fibonacci function in Python"}
    ],
    "stream": true
  }'

Docker Deployment

# Using Docker Model Runner
docker model run hf.co/amkyawdev/amk-coder-v2

# Using vLLM Docker
docker run --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  --rm \
  vllm/vllm-openai:latest \
  --model amkyawdev/amk-coder-v2

---

⚠️ Limitations

1. Context Length - Maximum 32,768 tokens
2. Code Quality - May generate incorrect code; verify outputs
3. Myanmar Unicode - Best results with proper Zawgyi-to-Unicode conversion
4. Domain Knowledge - Limited to common programming languages
5. Safety - May produce harmful content; use responsible AI practices

---

📖 Resources

• Qwen2.5-Coder Documentation
• Transformers Library
• HuggingFace Hub

---

🙏 Acknowledgments

• Alibaba Cloud Qwen Team - Base model Qwen2.5-Coder
• HuggingFace - Model hosting and infrastructure
• Myanmar Developer Community - Testing and feedback

---

📝 License

Apache License 2.0 - See LICENSE file for details.

---

📧 Contact

• Author: amkyawdev
• HuggingFace: amkyawdev/amk-coder-v2
• GitHub: github.com/amkyawdev

---

Made with ❤️ for Myanmar Developers