myanmar-llm-data

📌 Overview

myanmar-llm-data is a comprehensive Burmese (Myanmar) conversational dataset designed for training language models. This dataset contains multi-turn conversations in Burmese and English, covering various domains including:

• 💻 Coding: Programming conversations (Python, JavaScript, React, etc.)
• 🌐 Translation: English to Burmese translations
• 📚 General: General knowledge Q&A in Burmese
• 👋 Greetings: Burmese greeting phrases

📊 Dataset Statistics

Split	Samples	Description
Train	~20,327	Training data
Test	~17,155	Testing data
Validation	~17,071	Validation data
Total	~54,553	Combined dataset

Tags Distribution

Tag	Count	Percentage
coding	~49,000	~90%
translation	~600	~1%
general	~580	~1%
greeting	~500	~1%

📝 Data Format

Each row is a JSON object with:

• messages: List of message objects (role + content)
• tags: Category tag

{
    "messages": [
        {"role": "system", "content": "သင်သည် မြန်မာစာကျွမ်းကျင်သော AI အကူအညီပေးသူဖြစ်သည်။"},
        {"role": "user", "content": "မင်္ဂလာပါ။"},
        {"role": "assistant", "content": "မင်္ဂလာပါ၊ ဘာများကူညီပေးရမလဲ။"}
    ],
    "tags": "greeting"
}

Example Entries

Greeting

{
    "messages": [
        {"role": "system", "content": "သင်သည် မြန်မာစာကျွမ်းကျင်သော AI အကူအညီပေးသူဖြစ်သည်။"},
        {"role": "user", "content": "မင်္ဂလာပါ။"},
        {"role": "assistant", "content": "မင်္ဂလာပါ၊ ဘာများကူညီပေးရမလဲ။"}
    ],
    "tags": "greeting"
}

Translation

{
    "messages": [
        {"role": "system", "content": "သင်သည် ဘာသာပြန်ကျွမ်းကျင်သော AI အကူအညီပေးသူဖြစ်သည်။"},
        {"role": "user", "content": "'Thank you' ကို မြန်မာလို ဘယ်လိုပြန်မလဲ။"},
        {"role": "assistant", "content": "ကျေးဇူးတင်ပါတယ် လို့ ပြန်ဆိုနိုင်ပါတယ်။"}
    ],
    "tags": "translation"
}

Code Generation

{
    "messages": [
        {"role": "system", "content": "You are a coding assistant."},
        {"role": "user", "content": "Write a Python function to calculate Fibonacci numbers"},
        {"role": "assistant", "content": "def fibonacci(n):\n    if n <= 1:\n        return n\n    return fibonacci(n-1) + fibonacci(n-2)"}
    ],
    "tags": "coding"
}

🚀 Usage

Load Dataset

from datasets import load_dataset

# Load full dataset
dataset = load_dataset("amkyawdev/myanmar-llm-data")
print(dataset)

# Load specific split
train_data = load_dataset("amkyawdev/myanmar-llm-data", split="train")

Filter by Tag

from datasets import load_dataset

dataset = load_dataset("amkyawdev/myanmar-llm-data")

# Filter coding examples
coding_data = dataset["train"].filter(lambda x: x["tags"] == "coding")

# Filter translation examples
translation_data = dataset["train"].filter(lambda x: x["tags"] == "translation")

# Filter greeting examples
greeting_data = dataset["train"].filter(lambda x: x["tags"] == "greeting")

Use with Transformers

from datasets import load_dataset
from transformers import AutoTokenizer

# Load dataset
dataset = load_dataset("amkyawdev/myanmar-llm-data")

# Load tokenizer
tokenizer = AutoTokenizer.from_pretrained("microsoft/phi-2")

# Prepare for training
def preprocess(examples):
    texts = []
    for msg_list in examples["messages"]:
        text = "\n".join([f"{msg['role']}: {msg['content']}" for msg in msg_list])
        texts.append(text)
    return tokenizer(texts, truncation=True, padding=True, max_length=512)

tokenized_dataset = dataset.map(preprocess, batched=True)

🛠️ Use Cases

1. Fine-tune LLMs: Train models like phi-2, Llama, Mistral
2. RAG: Build retrieval systems for Myanmar content
3. Code Generation: Train coding assistants in Myanmar
4. Translation: English ↔ Myanmar translation
5. Chatbots: Build Myanmar conversational AI

📁 Files

File	Size	Rows
train.jsonl	~6.9 MB	~20,327
test.jsonl	~5.8 MB	~17,155
validation.jsonl	~5.8 MB	~17,071

🧪 Data Quality

• ✅ Consistent 3-message format (system, user, assistant)
• ✅ Tags manually verified
• ✅ Burmese text in Unicode
• ✅ No PII (Personally Identifiable Information)
• ✅ Clean and formatted

📜 License

MIT License - See LICENSE file for details.

---

🇲🇲 Built for Myanmar NLP Community