imhmdf/taiwan-culture-llama-70b-training-v2

Taiwan Culture Llama 70B Training Dataset

Dataset Description

This dataset is specifically prepared for fine-tuning Llama 70B models on Taiwan culture, combining multiple high-quality sources to create a comprehensive training resource.

Dataset Summary

• Total Samples: 8352
• Language: Traditional Chinese (Taiwan)
• Domains: Taiwan Culture, Government QA, Conversational AI, Cultural Knowledge
• Format: Instruction-following format optimized for chat models

Data Sources

1. Taiwan Culture MCQ - Multiple choice questions about Taiwan culture
2. Taiwan Government QA - Government-related question-answer pairs
3. Taiwan Chat - Conversational data with Taiwan context
4. Local Taiwan Culture Data - Rich cultural knowledge from processed datasets

Dataset Structure

{
    "text": "<s>[INST] instruction [/INST] response</s>",  // Llama-2 chat format
    "prompt": "### 指令:\ninstruction\n\n### 回應:",      // Alpaca format
    "completion": "response text",
    "instruction": "instruction text",
    "input": "input context (if any)",
    "output": "expected output",
    "source": "data_source_name",
    "category": "content_category"
}

Training Recommendations

• Model: Llama-2-70b-hf or similar 70B parameter models
• Method: LoRA fine-tuning recommended for efficiency
• Batch Size: 1 per device with gradient accumulation
• Learning Rate: 5e-5
• Epochs: 3-5 maximum to prevent overfitting

Hardware Requirements

• Minimum: 4 x A100 40GB GPUs
• Recommended: 8 x A100 80GB GPUs
• Memory: 320GB - 640GB total GPU memory

Usage

from datasets import load_dataset

dataset = load_dataset("imhmdf/taiwan-culture-llama-70b-training-v2")
train_data = dataset["train"]
val_data = dataset["validation"] 
test_data = dataset["test"]

Citation

If you use this dataset, please cite the original data sources and this compilation.

License

MIT License - Please respect individual dataset licenses from original sources.