You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Dataset Card for tw-gov-news-90M

本資料集收錄中華民國(臺灣)政府機關公開網站/新聞稿之繁體中文文本,總 token 數約 90M(90 百萬),可作為繁中模型在政府公文體與時事公告語境上的補充預訓練語料。

Dataset Details

Dataset Description

資料來自中華民國政府公開網頁與新聞稿,內容涵蓋:

  • 政府網站開放資料宣告與授權條款
  • 政策說明與新聞稿
  • 公告事項、施政成果報告

文本依政府公開資訊公開原則整理,可作為政府公文體、政策論述語境的訓練語料。

  • Curated by: Huang Liang Hsun
  • Language(s) (NLP): Traditional Chinese
  • License: cc-by-nc-sa-4.0

Dataset Sources

Uses

Direct Use

  • 補強模型對臺灣政府公文體、政策語言的覆蓋。
  • 衍生政府公告 chatbot、政策摘要、檢索系統的 RAG 知識來源。

Out-of-Scope Use

  • 不適用作為現行政策/法規的事實依據;資料反映抓取當下狀態。
  • 不應用於政治宣傳;本資料集為政府公開資訊整合,並非任一政黨之文宣。

Dataset Structure

{
  "text": "一、政府網站資料開放宣告 1、授權方式及範圍 ..."
}

僅有單一 text 欄位(部分樣本另含 token_count、word_count 等 metadata)。

Dataset Creation

Curation Rationale

通用語料中,政府公文體比例偏低,但對部署在臺灣公部門場景的模型而言這類文體是高頻情境。本資料集集中蒐集政府新聞與公告,補強對應的覆蓋。

Source Data

Data Collection and Processing

  1. 蒐集中華民國政府機關之公開網頁與新聞稿。
  2. 移除版面雜訊、合併段落為純文字。
  3. 控制累計 token 數至約 90M。

Who are the source data producers?

原始公告/新聞稿由各政府機關發布。多數內容依「政府資料開放授權條款」開放。

Annotations

Annotation process

無人工標註。

Who are the annotators?

無標註者。

Personal and Sensitive Information

資料屬政府公開資訊;如有遺漏 PII,請於 discussion 區提出排除。

Bias, Risks, and Limitations

  • 政府公告反映其發布時點之政策,可能與最新狀況不一致。
  • 不同機關的公文風格不一,模型可能學到風格混雜。
  • 本資料集 license 為 cc-by-nc-sa-4.0,但部分原文依政府資料開放授權條款另有規範,下游使用前請以原始來源為準。

Recommendations

建議與通用繁中網頁語料一同混訓並控制比重;下游若用於政府服務 chatbot,請結合 RAG 取最新公告再回答。

Citation

@misc{tw_gov_news_90m,
  title        = {tw-gov-news-90M: Taiwan Government News and Public Announcement Corpus (90M tokens)},
  author       = {Huang, Liang Hsun},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-gov-news-90M}}
}

Dataset Card Authors

Huang Liang Hsun

Dataset Card Contact

Huang Liang Hsun

Downloads last month
13

Models trained or fine-tuned on lianghsun/tw-gov-news-90M