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-news-551M

tw-news-551M 是一個台灣繁體中文之長文本預訓練語料集,包含 648,576 篇文章,總計約 5.63 億 tokens。涵蓋政治、社會、文化、環境、科技等多元主題,適用於語言模型在繁體中文語境之持續預訓練。

Dataset Details

Dataset Description

本資料集彙整台灣繁體中文之公開文章,涵蓋政治、社會、文化、環境、科技等多元主題。每筆資料包含文章全文與 token/字數等統計元資料。

  • Curated by: Liang Hsun Huang
  • Language(s) (NLP): Traditional Chinese
  • License: CC BY-NC-SA 4.0

Dataset Sources

Uses

Direct Use

本資料集主要設計用於:

  • 語言模型在繁體中文之持續預訓練(continued pretraining),提升對台灣時事議題與中文長文本之理解;
  • 建構檢索增強生成(RAG)系統之語料庫;
  • 訓練繁體中文之文本分類、摘要或主題模型。

Out-of-Scope Use

本資料集不適用於下列用途:

  • 直接作為指令微調資料,因不包含問答或對話結構。
  • 作為事實查核之權威來源,因文章內容可能帶有個人立場或觀點。
  • 用於非繁體中文語境之語言模型訓練。

Dataset Structure

{
  "text": "...",
  "token_count": 2610,
  "word_count": 2972,
  "url": "https://example.com/article/12345",
  "updated_at": "2024-12-23"
}
欄位 說明
text 文章全文
token_count 文章 token 數
word_count 文章字數
url 原始來源識別字串
updated_at 資料更新日期
統計項目 數值
文章數 648,576
總 tokens ~562.9M
平均 tokens/篇 867
平均字數/篇 974

Dataset Creation

Curation Rationale

繁體中文語言模型在台灣語境之表現受限於高品質預訓練資料之不足。本資料集彙整台灣繁體中文之公開長文本,補充現有預訓練資料在繁體中文領域之缺口。

Source Data

Data Collection and Processing

文章為台灣繁體中文之公開內容彙整,並計算每篇文章之 token 數與字數等統計值。

Who are the source data producers?

原始文章由不同作者獨立撰寫,不屬於單一組織或機構。

Annotations

Annotation process

本資料集不包含額外標註。

Who are the annotators?

不適用。

Personal and Sensitive Information

本資料集內容為公開發表之長文本。文章中可能提及公眾人物姓名與社會議題觀點,均為原始文章內容之一部分。

Bias, Risks, and Limitations

  • 文章可能帶有特定立場或個人觀點,模型訓練時應注意偏差問題。
  • 資料內容之主題分佈可能不均衡,無法完整代表台灣繁體中文之全貌。
  • 文章內容具時效性,部分內容可能已過時。
  • 部分內容可能涉及爭議性議題。

Recommendations

建議使用者在預訓練時搭配多元來源之語料,以降低單一語料偏差之影響。涉及事實性內容時,應搭配其他權威資料來源進行交叉驗證。

Citation

@misc{tw-news-551M,
  title        = {tw-news-551M: Taiwan Traditional Chinese Long-form Pre-training Corpus},
  author       = {Liang Hsun Huang},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-news-551M}},
  note         = {648,576 long-form articles (~562.9M tokens) in Traditional Chinese for pre-training.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang

Downloads last month
8

Models trained or fine-tuned on lianghsun/tw-news-551M