metadata

license: apache-2.0

📚 Traditional Chinese Examination Dataset 🤗

This project is an collection of exam-related files (PDFs and MP3s) that can be used to train document/audio understanding models, evaluate datasets for indexing question-answer pairs, or OCR preprocessing.

🗂️ Project Structure

project/
├── data/                     # Original exam files (PDFs, MP3s)
├── convert.py                # Script to generate metadata
├── main.py                   # Script to load dataset with datasets
├── metadata.jsonl            # Output metadata file
└── README.md                 # Project documentation

🛠️ Requirements

Python 3.7+
Install required packages:

pip install datasets

🔁 Generate Metadata

Run convert.py to extract metadata from filenames and generate a metadata.jsonl file.

python convert.py

This will read all files in the data/ directory and output a line-delimited JSON file (metadata.jsonl) describing each file with fields like:

{
  "id": "01-1131-2-一公民-題目",
  "serial": "01",
  "grade": "一",
  "subject": "公民",
  "variant": null,
  "type": "題目",
  "path": "data/01-1131-2-一公民-題目.pdf",
  "format": "pdf"
}

📥 Metadata

Run main.py to load the dataset using the Hugging Face datasets library.

python main.py

This will load and print a few samples from your dataset.

🔍 Dataset Fields

id: Unique identifier from the filename
serial: Serial code from the filename
grade: 一 (1st year), 二 (2nd year), 三 (3rd year)
subject: e.g., 公民, 英文, 數學
variant: Optional (e.g., 體, 音) from parentheses
type: 題目 (questions), 答案 (answers), 手寫卷 (handwritten)
path: File path
format: pdf or mp3

🧠 Tips

You can use this dataset for training document/audio understanding models, indexing question-answer pairs, or OCR preprocessing.
Extend convert.py to extract text from PDFs or audio features if needed.
Easily upload the dataset to the Hugging Face Hub via datasets.Dataset.push_to_hub.

📬 Future Work

PDF text extraction
Audio transcription
Tagging with more metadata (exam year, term, difficulty, etc.)
Train-test splitting

📝 License

Apache 2.0 License

🤝 Contributions Welcome!

Feel free to fork, improve, and PR!

🙏 Special Thanks

This project was made possible with the help of ChatGPT, an AI assistant by OpenAI, which provided:

Ez filename parsing and metadata extraction strategy
Dataset integration guidance
Project structuring and documentation
🧠 Motivation to keep going when the files got messy

Thanks, AI buddy! 🤖💡