Multi-TW: Traditional Chinese Language Learning Dataset

Dataset Description

Multi-TW is a Traditional Chinese language learning and assessment dataset containing 900 multiple-choice questions with multimedia content. This dataset is designed for evaluating multi-modal language models on Traditional Chinese comprehension tasks.

Dataset Structure

The dataset contains 900 samples in the validation split, suitable for benchmarking purposes.

Data Fields

• id: Unique identifier for each question
• instruction: Task instructions in Chinese
• question: The question text in Chinese
• option1: Multiple choice option A
• option2: Multiple choice option B
• option3: Multiple choice option C
• option4: Multiple choice option D (may be empty)
• answer: Correct answer (A, B, C, or D)
• image: PIL Image object (for visual questions)
• audio: Audio data with sampling rate (for audio questions)

Data Composition

• Total samples: 900
• Samples with images: 450
• Samples with audio: 450
• Answer distribution: A: 249, B: 261, C: 263, D: 127
• Question types: L (Listening): 660, R (Reading): 240

Usage

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("ntuai/multi-tw")
validation_data = dataset["validation"]

# Access a sample
sample = validation_data[0]
print(f"Question: {sample['question']}")
print(f"Options: {sample['option1']}, {sample['option2']}, {sample['option3']}")
print(f"Answer: {sample['answer']}")

# Check if sample has image or audio
if sample['image'] is not None:
    # Process image
    image = sample['image']
    
if sample['audio'] is not None:
    # Process audio
    audio_array = sample['audio']['array']
    sampling_rate = sample['audio']['sampling_rate']

Dataset Statistics

The dataset covers various aspects of Chinese language learning:

• Visual comprehension: Questions requiring image understanding
• Audio comprehension: Questions requiring audio understanding
• Multiple choice format: 3-4 options per question
• Balanced distribution: Relatively even distribution across answer choices

License

MIT License

Citation

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

@dataset{multi_tw_2025,
  title={Multi-TW: Benchmarking Multimodal Models on Traditional Chinese Question Answering in Taiwan},
  author={Jui-Ming Yao, Bing-Cheng Xie, Sheng-Wei Peng, Hao-Yuan Chen, He-Rong Zheng, Bing-Jia Tan, Peter Shaojui Wang, and Shun-Feng Su},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/ntuai/multi-tw}
}