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-society-88M

本資料集收錄臺灣社會、生活、觀光、旅遊景點等繁體中文文本,總 token 數約 88M(88 百萬),可作為繁中模型在「日常社會生活」語境下的補充預訓練語料。

Dataset Details

Dataset Description

資料來自繁體中文旅遊/生活/社會議題類公開報導,文本特色:

  • 描述各縣市景點、街廓、店家
  • 介紹地方文化、節慶、特產
  • 含 URL 與 updated_at metadata,便於追溯

每筆樣本包含 text 主文以及 token_countword_counturlupdated_at 等 metadata,方便後續清理與追蹤。

  • Curated by: Huang Liang Hsun
  • Language(s) (NLP): Traditional Chinese, English(部分英文夾雜)
  • License: cc-by-nc-sa-4.0

Dataset Sources

Uses

Direct Use

  • 繁中模型對「臺灣社會生活」語境的補強(地名、店家、景點、民俗)。
  • 旅遊/在地推薦 chatbot 之 RAG 知識來源。

Out-of-Scope Use

  • 不適用作為即時資訊(營業時間、票價、活動)之來源;資料反映抓取當下狀態。
  • 不適用於不實宣傳、商業廣告。

Dataset Structure

{
  "text": "基隆市 中正區 基隆海洋廣場 ... 「海洋廣場」為基隆港內側海域的海上原木平台...",
  "token_count": 292,
  "word_count": 309,
  "url": "https://example.com/...",
  "updated_at": "2024-12-24"
}

Dataset Creation

Curation Rationale

臺灣社會生活相關語料品質參差,但對在地化模型來說又非常重要。本資料集集中蒐集多個社會生活來源,並保留 URL 與時間戳,方便後續更新與品質追蹤。

Source Data

Data Collection and Processing

  1. 蒐集繁體中文旅遊/生活類公開文章。
  2. 移除版面雜訊與廣告區塊。
  3. 計算字元數與 token 數,附上原 URL 及 updated_at 時間戳。

Who are the source data producers?

原始文章由各原作者撰寫並公開於網路。

Annotations

Annotation process

無人工標註。

Who are the annotators?

無標註者。

Personal and Sensitive Information

資料為公開網頁內容,所提及之店家名稱、地址等屬商業公開資訊。

Bias, Risks, and Limitations

  • 來源多為觀光、生活類報導,敘述常帶有正向行銷語氣,未必中立。
  • 部分內容會隨營業狀況更新而失效;建議搭配 updated_at 做新鮮度判斷。

Recommendations

建議在訓練前再做一次去重;下游若用於旅遊推薦,務必標註資訊時效。

Citation

@misc{tw_society_88m,
  title        = {tw-society-88M: Traditional Chinese Society and Tourism Corpus from Taiwan (88M tokens)},
  author       = {Huang, Liang Hsun},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-society-88M}}
}

Dataset Card Authors

Huang Liang Hsun

Dataset Card Contact

Huang Liang Hsun

Downloads last month
14

Models trained or fine-tuned on lianghsun/tw-society-88M