wikipedia_tw / README.md
jslin09's picture
Update README.md
1156671 verified
metadata
license: cc-by-sa-4.0
language:
  - zh
multilinguality:
  - monolingual
pretty_name: wiki_tw
size_categories:
  - 100M<n<1B
source_datasets:
  - wikipedia
dataset_info:
  features:
    - name: title
      dtype: string
    - name: article
      dtype: string
extra_gated_prompt: >-
  You agree to not attempt to determine the identity of individuals in this
  dataset
extra_gated_fields:
  Company: text
  Country: text
  I agree to use this dataset for non-commercial use ONLY: checkbox

要搞自己的大型語言模型,最基本的基本,就是需要一大堆文字資料,從 Common Crawl 上頭抓回來慢慢清洗是一條路,清洗維基百科網站的週期性下載檔也是一個方法。本資料集是解析自維基百科於 20240420 發布的繁體中文版打包檔 bz2 檔案的內容,在解析出所需內容後,利用 wikitextparser 移除 Wiki 標記。解析後保留的欄位有兩個:條目名稱(title),條目內容(page article)。

原始的打包檔條目內容簡繁混雜,所以有利用 OpenCC 進行簡轉繁處理。

  • 原始總條目數: 4,451,426 條目。
  • 全部 4,451,426 個條目標題。
  • 無法自動去標記的條目數: 3,035,750
  • 有內容的條目數: 1,415,676

因為本資料集內容龐大,要塞進一般的個人電腦中進行計算,恐怕會有資源不足的情形。建議使用parquet格式下載使用。

資料集當中有不少內容為 #REDIRECT 的條目已經嘗試移除,如果移除的不乾淨,就等以後有空推出修正版再來清洗了。