MODEL_CARD.md

mm-llm-coder-lite-v1 Model Card

📌 Overview

mm-llm-coder-lite-v1 is a Lite version of the Myanmar Large Language Model, specifically optimized for efficiency in Myanmar (Burmese) programming tasks. This model is designed for developers in Myanmar who need a lightweight, fast model for code generation and conversational AI.

Key Design Goals

• 🚀 Efficient: Optimized for low-resource environments
• 💻 Code-focused: Specialized in programming tasks
• 🌍 Myanmar-first: Built for Myanmar developers

📊 Model Specifications

Specification	Value
Parameters	~2.7B (base), ~2.6M (trainable with LoRA)
Base Model	microsoft/phi-2
Fine-tuning Method	LoRA (Low-Rank Adaptation)
Training Data Type	Myanmar code + conversation dataset
LoRA Rank (r)	16
LoRA Alpha	32
Max Length	512 tokens
Training Epochs	3
Learning Rate	2e-4

🚀 Quick Start

Installation

pip install torch transformers peft accelerate

Basic Usage (Python)

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load the model
model_name = "amkyawdev/mm-llm-coder-lite-v1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16,
    device_map="auto"
)

# Set pad token
tokenizer.pad_token = tokenizer.eos_token

Generate Response

# Create prompt in Myanmar format
prompt = """System: သင်သည် မြန်မာစာကျွမ်းကျင်သော AI အကူအညီပေးသူဖြစ်သည်။

User: Python နဲ့ Fibonacci function ရေးပေးပါ။

Assistant:"""

# Generate
inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=512)
inputs = {k: v.to(model.device) for k, v in inputs.items()}

outputs = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.95,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Using Gradio Space

# Visit: https://huggingface.co/spaces/amkyawdev/mm-llm-coder-lite-v1
# Or use via API
from gradio_client import Client

client = Client("amkyawdev/mm-llm-coder-lite-v1")
result = client.predict(
    "Python နဲ့ list sort လုပ်နည်း",  # user message
    fn_index=0
)
print(result)

📝 Sample Prompts (Myanmar)

Example 1: Code Generation

User: Python နဲ့ Fibonacci function ရေးပေးပါ။
Assistant: def fibonacci(n):
    if n <= 1:
        return n
    else:
        return fibonacci(n-1) + fibonacci(n-2)

Example 2: Translation

User: Hello ပါတ်မှားပါ။
Assistant: မင်္ဂလာပါ။ သင့်အား ကူညီပါသည်။

Example 3: Data Cleaning

User: မြန်မာစာသားအမှားမှားပြင်ပါ။
Assistant: import re

def clean_myanmar_text(text):
    # Remove extra spaces
    text = re.sub(r'\s+', ' ', text)
    # ... (more cleaning logic)
    return text

⚠️ Limitations (Lite Version)

This is a Lite version with intentional trade-offs:

Performance Limitations

Limitation	Description
Smaller Context	Max 512 tokens (vs 2048+ in full version)
Limited Knowledge	Trained on ~20K samples
Code Complexity	Best for simple to intermediate tasks
Language Coverage	Primarily Myanmar, limited English

Expected Behavior

1. Fast Inference: optimized for speed over quality
2. Simple Tasks: Good for basic code generation
3. Complex Tasks: May struggle with advanced algorithms
4. Long Conversations: Context may degrade after ~3-4 turns

Recommendations for Developers

• Use for: Simple scripts, code translation, learning
• Avoid: Production-grade complex systems, long context tasks
• Fine-tune: For your specific use case if needed

📁 Training Data

• Dataset: amkyawdev/myanmar-llm-data
• Training Samples: ~20,327
• Test Samples: ~17,155
• Categories: Code (90%), Translation, General, Greetings

🏷️ Tags

myanmar burmese llm code-generation fine-tuned lora phi-2 transformers

📜 License

MIT License - See LICENSE file for details.

🙏 Acknowledgments

• Microsoft for phi-2 base model
• Hugging Face community
• Myanmar developers

---

🇲🇲 Made for Myanmar Developers