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-556k

本資料集收錄中華民國(臺灣)政府公開網站的中文(含中英對照)文本,涵蓋國史、機關沿革、政府公開資料等內容,總字數約 556,000 字(556k),以「text」單欄位的純文本格式提供,作為繁體中文持續預訓練(continued pretraining)語料的補充。

Dataset Details

Dataset Description

資料來源以中華民國政府公開頁面為主,內容偏向具有時間軸/傳記/沿革性質的中文敘事文本(部分樣本含中英對照)。文本經過簡單清理(移除 HTML 標籤、空白、合併段落等),但保留段落原貌,方便用於語言建模任務。

資料集名稱中的「556k」指原始彙整文本之總字數約 556,000 字,並非樣本數;實際樣本數落在 1K–10K 區間。

  • Curated by: Huang Liang Hsun
  • Language(s) (NLP): Traditional Chinese(部分樣本含英文對照)
  • License: cc-by-nc-sa-4.0

Dataset Sources

Uses

Direct Use

  • 補充繁體中文語言模型對「臺灣政府/中華民國史」相關語境的覆蓋。
  • 作為政府文體、公文體、傳記體的小型參考語料。

Out-of-Scope Use

  • 不適用作為政治宣傳、特定立場操作之素材;本資料的內容反映原始來源網站之敘事,而非平衡論述。
  • 不適用作為事實查證的單一來源;應與其他史料互相核對。

Dataset Structure

{
  "text": "西元 1866 年清同治 5 年十一月十二日 Nov. 12 誕生於廣東省香山縣(今中山縣)翠亨村。Sun was born in Cuiheng Village..."
}

僅有單一欄位 text

Dataset Creation

Curation Rationale

臺灣政府公開網站包含大量具歷史與制度價值的繁體中文文本,但格式分散、版面雜訊多。本資料集的目的在於將此類文本整理為可直接餵入語言模型的純文字段落,作為較通用的預訓練語料補充。

Source Data

Data Collection and Processing

  1. 蒐集目標政府機關公開頁面之 HTML。
  2. 移除版面雜訊(導覽列、頁尾、廣告區塊)。
  3. 合併段落,輸出為單欄 text 的 JSONL。

Who are the source data producers?

原始文本由中華民國政府機關撰寫並公開於官方網站。

Annotations

Annotation process

無人工標註。

Who are the annotators?

無標註者。

Personal and Sensitive Information

本資料集源自公開政府文本,不含個人隱私資料;如果有侵犯特定人士或機構之隱私情形,請聯繫排除。

Bias, Risks, and Limitations

  • 政府網站文本本身具特定立場與敘事方式,使用者應理解其視角侷限。
  • 規模偏小(556k 字),不適合作為主訓練語料,建議僅作為 sub-mix 使用。

Recommendations

建議與其他繁體中文預訓練語料(如 fineweb-edu-zhtw)一同混訓,並依需求調整其權重。

Citation

@misc{tw_gov_556k,
  title        = {tw-gov-556k: Traditional Chinese Text from Taiwan Government Sources},
  author       = {Huang, Liang Hsun},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-gov-556k}}
}

Dataset Card Authors

Huang Liang Hsun

Dataset Card Contact

Huang Liang Hsun

Downloads last month
16

Models trained or fine-tuned on lianghsun/tw-gov-556k