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 finepdf-filtered-zhtw

finepdf-filtered-zhtw 是以 Hugging Face FinePDF 為來源,過濾出繁體中文(zh-tw)為主的子集,保留 finepdf 的完整 metadata(包含語言判別分數、文件 dump、URL 等),可作為繁中持續預訓練語料的補充。

Dataset Details

Dataset Description

FinePDF 是 HuggingFace 從 Common Crawl 抽取出的 PDF 文本資料集,原本以英文為主、其他語言為輔。本資料集的處理方式:

  1. 以原始 metadata 中的 language 欄位先行篩選為 zh*
  2. 對每筆 PDF 文本以 fastText/lid 等語言判別工具重新計算 page_average_lidfull_doc_lid 等欄位,並依分數判定是否為繁體中文(zh-tw)為主。
  3. 對偵測為簡體為主的樣本進行排除。

最終資料切分為 290 個 parquet 分片,方便平行處理。

  • Curated by: Huang Liang Hsun
  • Language(s) (NLP): Traditional Chinese
  • License: ODC-By(沿用 finepdf 母資料集授權)

Dataset Sources

Uses

Direct Use

  • 繁體中文語言模型的持續預訓練語料補充(特別是來源為 PDF 的長段落、書面語)。
  • fineweb-edu-zhtw 等網頁類語料互補:finepdf 偏向書面、政府公文、研究報告風格。
  • 作為 PDF→純文字流程的下游品質評測來源。

Out-of-Scope Use

  • 不適用於需嚴格事實正確之領域(醫療、法律):PDF 內容繁雜、版面噪音多。
  • 使用時請遵守 ODC-By 授權,並標註原始 finepdf 來源。

Dataset Structure

{
  "text": "...",
  "id": "...",
  "dump": "CC-MAIN-...",
  "url": "https://example.com/foo.pdf",
  "date": "2024-..",
  "file_path": "...",
  "offset": 0,
  "token_count": 1234,
  "language": "zh",
  "page_average_lid": "zh",
  "page_average_lid_score": 0.95,
  "full_doc_lid": "zh",
  "full_doc_lid_score": 0.96,
  ...
}

完整欄位定義以原始 finepdf 為準,本資料集保留所有 metadata 欄位。

Dataset Creation

Curation Rationale

finepdf 母資料集是少見公開、可商用、且具規模的 PDF 文本資料集,但繁體中文比例偏低、且部分樣本實為簡體。本子集集中過濾為「以繁體中文為主」之內容,方便繁中模型訓練流程直接使用。

Source Data

Data Collection and Processing

  1. 自 HuggingFaceFW/finepdf 取得原始資料。
  2. 過濾 language 欄位非中文之樣本。
  3. 對中文樣本以 fastText/lid 等工具重新計算語言分數,並排除簡體為主者。
  4. 將結果切分為 290 個 parquet shard 輸出。

Who are the source data producers?

原始 PDF 文本由 Common Crawl 收錄之網際網路 PDF 文件作者所撰寫;finepdf 由 HuggingFace 整理。

Annotations

Annotation process

無額外人工標註;保留母資料集所有 metadata,並新增重算後的語言判別分數。

Who are the annotators?

無人工標註者。

Personal and Sensitive Information

資料來自公開 PDF;如包含特定人士隱私資料,請於 discussion 區提出排除。

Bias, Risks, and Limitations

  • PDF 解析品質參差:版面雜訊、表格破損、頁眉頁腳混入正文等情形不可避免。
  • 語言判別工具仍會誤判,尤其在簡繁夾雜或註腳區段。
  • 來源以可被 Common Crawl 抓到之公開 PDF 為主,學術/政府文件可能被 paywall 擋住未被收錄。

Recommendations

建議在訓練前再做:

  1. 簡體字偵測與第二輪過濾。
  2. 段落長度與重複片段去重。
  3. 對特定領域(教育/法規)做加權混訓。

Citation

@misc{finepdf_filtered_zhtw,
  title        = {finepdf-filtered-zhtw: Traditional Chinese Subset of FinePDF},
  author       = {Huang, Liang Hsun},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/finepdf-filtered-zhtw}},
  note         = {Filtered from HuggingFaceFW/finepdf under ODC-By.}
}

Dataset Card Authors

Huang Liang Hsun

Dataset Card Contact

Huang Liang Hsun

Downloads last month
34