🇲🇲 Myanmar LLM Coder Dataset (mm-llm-coder-dataset)

မြန်မာဘာသာ Coding LLM များ training အတွက် ရည်ရွယ်ထားသော dataset

A bilingual (Myanmar + English) coding instruction dataset designed primarily for training Myanmar language Coder LLMs.

---

🎯 ရည်ရွယ်ချက် / Purpose

ဤ dataset သည် မြန်မာဘာသာ programming/coding LLM များ training လုပ်ရန်အတွက် အဓိက ရည်ရွယ်ထားပါသည်။ မြန်မာ developer များ၏ မိခင်ဘာသာစကားဖြင့် coding အကူအညီပေးနိုင်သော AI assistant များကို ဖန်တီးနိုင်စေရန် Myanmar (my) နှင့် English (en) ဘာသာစကား နှစ်မျိုးဖြင့် pair training data ထည့်သွင်းထားပါသည်။

This dataset is primarily intended for training Myanmar (Burmese) language Coder LLMs — enabling AI coding assistants that natively understand and respond in မြန်မာဘာသာ. Both English and Myanmar examples share the same schema for parallel/cross-lingual training.

📊 Dataset Statistics

Metric	Value
Total Samples	4,000,000
Myanmar (my)	2,000,000
English (en)	2,000,000
Format	Parquet (Snappy compressed)
Files	data/train-00000-of-00004.parquet … data/train-00003-of-00004.parquet
Splits	train (single split)

🗂️ Schema

{
    "instruction": str,          # Topic/title of the Q&A
    "messages": [                # Multi-turn conversation
        {"role": "user", "content": str},
        {"role": "assistant", "content": str},
        ...                      # 2 / 4 / 6 messages per sample
    ],
    "category": str,             # See categories below
    "language": str,             # "my" or "en"
    "difficulty": str,           # "beginner" | "intermediate" | "advanced"
    "response": str,             # JSON-serialized messages (string form)
    "task_type": str,            # "qa_conversation"
}

Categories (၉ မျိုး)

code_debugging, mongodb_project, telegram_bot, ui_ux_design, cloudflare_worker, deployment_help, api_integration, firebase_project, website_builder

Difficulty Levels

beginner · intermediate · advanced

📝 Sample Examples

Myanmar (my)

{
  "instruction": "React useEffect ကို StrictMode မှာ နှစ်ကြိမ် run တာ ဘယ်လိုဖြေရှင်းမလဲ?",
  "messages": [
    {"role": "user", "content": "ကျွန်တော့် useEffect က development မှာ နှစ်ကြိမ် run နေပါတယ်။ ဘာကြောင့်လဲ?"},
    {"role": "assistant", "content": "React 18 StrictMode က side effects တွေကို detect လုပ်ဖို့အတွက် effects တွေကို နှစ်ကြိမ်ခေါ်ပါတယ်။\n\n**ဖြေရှင်းနည်း**: cleanup function ထည့်ပါ..."}
  ],
  "category": "code_debugging",
  "language": "my",
  "difficulty": "beginner",
  "task_type": "qa_conversation"
}

English (en)

{
  "instruction": "React useEffect runs twice in StrictMode - how to fix?",
  "messages": [
    {"role": "user", "content": "My useEffect is running twice in development. Why?"},
    {"role": "assistant", "content": "React 18 StrictMode intentionally double-invokes effects to detect side effects..."}
  ],
  "category": "code_debugging",
  "language": "en",
  "difficulty": "beginner",
  "task_type": "qa_conversation"
}

🚀 Usage

Load full dataset

from datasets import load_dataset

ds = load_dataset("amkyawdev/mm-llm-coder-dataset")
print(ds)
# DatasetDict({ train: Dataset(num_rows=4000000, ...) })

Filter by language

# Myanmar only — for Myanmar-focused fine-tuning
my_data = ds["train"].filter(lambda x: x["language"] == "my")

# English only — for cross-lingual / parallel training
en_data = ds["train"].filter(lambda x: x["language"] == "en")

Filter by category & difficulty

debugging_advanced = ds["train"].filter(
    lambda x: x["category"] == "code_debugging" and x["difficulty"] == "advanced"
)

Streaming (recommended for large-scale training)

ds = load_dataset("amkyawdev/mm-llm-coder-dataset", streaming=True)
for sample in ds["train"]:
    print(sample["language"], sample["instruction"])
    break

🎓 Use Cases

1. 🇲🇲 Myanmar Coder LLM training — fine-tune base models (Llama, Qwen, Mistral, etc.) into Myanmar-language coding assistants
2. Cross-lingual code Q&A — train models that handle both Myanmar and English coding queries
3. Instruction tuning — multi-turn conversation format suitable for chat models
4. Code debugging assistants — error fixing patterns across React, Node.js, MongoDB, WebSocket, etc.
5. Topic-specific fine-tuning — filter by category (e.g., MongoDB-only, Firebase-only)

🔗 Related Datasets

This dataset is part of the combined Myanmar LLM dataset collection by @amkyawdev:

• chat-skill → amkyawdev/myanmar-llm-data — conversational data, translations, general Q&A
• agent-skill → amkyawdev/mm-llm-coder-agent-dataset — agentic coding tasks
• code-skill → this dataset — code generation, debugging, and Q&A

⚠️ Notes / Caveats

• The dataset is template-based: the 4M samples are produced by combining a curated set of coding instructions with category × difficulty × conversation-length variations. This makes the dataset large and structurally consistent, but with limited semantic diversity per topic.
• For higher-quality, more diverse Myanmar samples, you may consider augmenting with LLM-generated translations of curated English programming Q&A.
• Both messages (list) and response (JSON string) fields contain the same conversation — use whichever your training pipeline prefers.

📄 License

Apache 2.0

🙏 Citation

If you use this dataset in your work, please cite:

@dataset{amkyawdev_mm_llm_coder_2025,
  author    = {amkyawdev},
  title     = {Myanmar LLM Coder Dataset (mm-llm-coder-dataset)},
  year      = {2025},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/amkyawdev/mm-llm-coder-dataset}
}