# punctuation_restoration

    ## Dataset Summary

    This dataset is designed for **instruction fine-tuning** of large language models (LLMs), especially for the **Qwen3** family, to perform **punctuation restoration** on Mandarin Chinese text.

    It is derived from the [`AWeirdDev/zh-tw-articles-6k`](https://huggingface.co/datasets/AWeirdDev/zh-tw-articles-6k) dataset. The `context` field is processed to create input-output pairs in the Qwen3-style message format.

    - 🔧 **User message**: A cleaned version of the article with all punctuation removed, spaces added where punctuation was, and an instruction prefix.
    - ✍️ **Assistant message**: The original article with punctuation.
    - 🧠 **Use Case**: Fine-tuning Qwen3 or similar chat-style LLMs to restore punctuation from flat text.

    ## Format

    Each sample in the dataset follows this structure:

    ```json
    {
      "messages": [
        {
          "role": "user",
          "content": "Your task is to restore the punctuations of this article:\nZhe li shi yi ge mei you biao dian de wen zhang ..."
        },
        {
          "role": "assistant",
          "content": "這裡是一個沒有標點或者標點不全的文章，請你恢復正確的標點符號。"
        }
      ]
    }