imhmdf/taiwan-culture-llama-70b-training-v3

🌟 Dataset Description

This is a comprehensive Taiwanese cultural and governmental knowledge dataset specifically designed for fine-tuning models. The dataset combines multiple high-quality sources to create a rich training corpus focused on Taiwan's culture, government services, and conversational patterns.

🔥 Key Features:

• ✅ 12,982 high-quality samples across 4 diverse data sources
• ✅ Traditional Chinese focus with Taiwan-specific terminology
• ✅ Unsloth compatible for efficient fine-tuning
• ✅ Quality filtered with robust data processing
• ✅ Multi-domain coverage from culture to government services

📊 Dataset Composition

Data Source	Samples	Domain	Format
Taiwan Culture MCQ	3,828	Educational/Cultural Knowledge	Multiple Choice → Answer
Government QA	4,844	Government Services	Question → Response
Taiwan Chat	2,648	Conversational AI	Chat/Dialogue
Local Culture	1,662	Cultural Heritage	Instruction → Description
Total	12,982	Mixed Domains	Instruction-Response

🗂️ Data Sources

1. 📚 Taiwan Culture MCQ

• Source: aqweteddy/Taiwan-Curlture-MCQ
• Content: Multiple-choice questions about Taiwanese culture, history, and knowledge
• Format: Questions with options → Correct answer text (not letter)
• Example:

  Q: 為防止胚胎發育，種雞蛋不得高於攝氏幾度保存？
  A: 23.9
  

2. 🏛️ Government QA

• Source: renhehuang/TaiwanGovQA-CN-Responses
• Content: Government service questions and official responses
• Note: Uses only 'answer' field, excludes 'answer_zh_cn' for consistency
• Example:

  Q: 國人於海外出生之子女回臺設籍，要如何申請？
  A: [Detailed government procedure explanation]
  

3. 💬 Taiwan Chat

• Source: jasonhuang3/sft_taitung_geminiv1v3_taiwan_chat_merged_243
• Content: Conversational exchanges about Taiwan locations and culture
• Format: Natural dialogue with Taiwan-specific context
• Example:

  User: 加路蘭遊憩區以前是做什麼用的啊？
  Assistant: 跟你說喔，加路蘭以前其實是個不起眼的地方啦！...
  

4. 🎭 Local Culture Data

• Source: Local processed cultural texts
• Content: Traditional Taiwanese cultural artifacts and heritage information
• Format: Cultural instruction → Detailed description

🎯 Format Structure

Each sample follows the Unsloth-compatible Llama-2 chat format:

{
    "text": "<s>[INST] {instruction} [/INST] {response}</s>",
    "source": "dataset_source_identifier"
}

Example formatted text:

<s>[INST] 請介紹關於台灣傳統藝術的相關資訊。 [/INST] 台灣傳統藝術包含豐富的文化內涵...</s>

🚀 Usage

Basic Loading

from datasets import load_dataset

# Load the complete dataset
dataset = load_dataset("imhmdf/taiwan-culture-llama-70b-training-v3")

# Access train/validation splits
train_data = dataset["train"]
validation_data = dataset["validation"]

print(f"Training samples: {len(train_data)}")
print(f"Validation samples: {len(validation_data)}")

Unsloth Integration

from unsloth import FastLanguageModel
from datasets import load_dataset

# Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="unsloth/llama-3-70b-bnb-4bit",
    max_seq_length=2048,
    dtype=None,
    load_in_4bit=True,
)

# Load dataset
dataset = load_dataset("imhmdf/taiwan-culture-llama-70b-training-v3", split="train")

# Apply LoRA adapters
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=3407,
)

📈 Dataset Statistics

• Total Samples: 12,982
• Train Split: 11,684 samples (90%)
• Validation Split: 1,298 samples (10%)
• Average Instruction Length: ~127 characters
• Average Response Length: ~312 characters
• Language: Traditional Chinese (Taiwan variant)
• Encoding: UTF-8

🔧 Data Processing

Quality Assurance

• ✅ Length filtering: Removed samples with instructions <5 chars or responses <5 chars
• ✅ Deduplication: Processed to avoid duplicate content
• ✅ Encoding consistency: All text properly UTF-8 encoded
• ✅ Field standardization: Consistent instruction-response format across sources

Version History

• v1: Initial dataset compilation
• v2: Fixed MCQ answer format (letters → actual answers)
• v3: Excluded 'answer_zh_cn' field, improved data consistency

🏷️ Use Cases

This dataset is ideal for:

• 🎯 Fine-tuning models for Taiwan-specific knowledge
• 💬 Building Taiwan cultural chatbots
• 🏛️ Government service AI assistants
• 📚 Educational AI for Taiwanese culture
• 🌐 Traditional Chinese language models

⚖️ Ethical Considerations

• Cultural Sensitivity: Content focuses on accurate representation of Taiwan culture
• Government Information: Official government responses included for factual accuracy
• Language Preservation: Supports Traditional Chinese and Taiwan-specific terminology
• Educational Purpose: Designed for legitimate educational and research applications

📝 Citation

If you use this dataset, please cite:

@misc{taiwan-culture-llama-dataset-v3,
    title={Taiwan Culture Training Dataset v3},
    author={imhmdf},
    year={2026},
    publisher={Hugging Face},
    url={https://huggingface.co/datasets/imhmdf/taiwan-culture-llama-70b-training-v3}
}

🔗 Related Resources

• Original MCQ Dataset: Taiwan-Curlture-MCQ
• Government QA Dataset: TaiwanGovQA-CN-Responses
• Taiwan Chat Dataset: sft_taitung_geminiv1v3_taiwan_chat_merged_243
• Unsloth: Efficient Fine-tuning Library

📞 Contact

For questions about this dataset, please open an issue in the repository or contact the maintainer.

---

Dataset created with ❤️ for advancing Taiwan cultural AI research