Datasets:
The dataset viewer is not available for this dataset.
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Dataset Card for c4-zhtw
c4-zhtw 是一個繁體中文(zh-tw)版本之 C4 子集之小型實驗用語料集,合計 2,783 筆,於原始文本之外額外附帶分類器預測(pred_label / pred_conf)與人工標註(label / reason)欄位,用於訓練或評估「繁中高品質文本分類器」(例如用於篩選 c4 / fineweb 繁中子集之 classifier)。
Dataset Details
Dataset Description
繁體中文之高品質預訓練語料長期不足,過濾現有大規模語料(如 C4、FineWeb)成為常見解法。本資料集為其中一個小型實驗子集,記錄了:
- 原始 C4 之 zh 子集中之文本、時間戳、URL;
- 分類器自動預測之標籤(
pred_label與信心pred_conf); - 人工覆核之最終標籤(
label)與標註理由(reason); - 文本長度之 word / token 計數。
主要用途為訓練或評估「繁中品質分類器」之 labeled corpus,亦可作為繁中預訓練語料篩選 pipeline 之 validation set。對應之大規模版本可參考 fineweb-edu-zhtw。
- Curated by: Liang Hsun Huang
- Language(s) (NLP): Traditional Chinese
- License: ODC-BY(繼承自 C4)
- Upstream Dataset: allenai/c4
Dataset Sources
- Repository: lianghsun/c4-zhtw
- Upstream Dataset: allenai/c4
- Related: lianghsun/fineweb-edu-zhtw
Uses
Direct Use
本資料集主要設計用於:
- 訓練或評估繁中文本之品質分類器(binary / multi-class quality classification);
- 作為大規模繁中語料篩選 pipeline 之 validation / test set;
- 研究分類器預測結果與人工標註之差異;
- 作為實驗用之標註樣本,提供 pred_label 與 label 之雙欄對照。
Out-of-Scope Use
本資料集不適用於下列用途:
- 作為大規模繁中預訓練之主力語料(僅 2,783 筆);
- 作為商業分類器之唯一訓練資料;
- 作為 simplified Chinese 之訓練資料;
- 作為純淨文本語料(因內容含 c4 之雜訊,pred_label=0 之樣本代表「不建議納入訓練」)。
Dataset Structure
{
"text": "知識閉合原則(Epistemic Closure Principle) - 紫煙亭 Home 知識論 ...",
"timestamp": "2017-12-17 21:32:10",
"url": "http://thiseven.blogspot.tw/2013/06/epistemic-closure-principle.html",
"zh_id": 1,
"word_count": 727,
"token_count": 727,
"pred_label": 1,
"pred_conf": 0.87,
"label": 1,
"reason": "具教育價值之哲學論文段落",
"source_file": "c4-zh-train.00000-of-01024.json.gz",
"row": 42
}
| 欄位 | 型別 | 說明 |
|---|---|---|
text |
string | 原始 C4 文本 |
timestamp |
timestamp[s] | 原始 crawl 時間 |
url |
string | 原始 URL |
zh_id |
int64 | 繁中樣本序號 |
word_count |
int64 | 文字計數 |
token_count |
int64 | token 計數 |
pred_label |
int64 | 分類器預測之標籤 |
pred_conf |
float64 | 分類器預測信心 |
label |
int64 | 人工覆核之最終標籤 |
reason |
string | 人工標註之理由描述 |
source_file |
string | 原始 C4 檔名 |
row |
int64 | 原始 C4 檔案內之行號 |
| 統計項目 | 數值 |
|---|---|
| 總筆數 | 2,783 |
| 切分 | 無(使用者自行切分) |
| 基礎語料 | C4 中之 zh 子集 |
Dataset Creation
Curation Rationale
繁中語料篩選 pipeline 需要一份「有人工標註」之 golden set,用來:
- 訓練 classifier(supervised);
- 評估 classifier 與人工標註之一致性;
- 做 error analysis,找出 classifier 錯判之模式。
本資料集即為此目的建立之小型 labeled set。
Source Data
Data Collection and Processing
- 原始來源:C4 之 zh 子集;
- 分類器預測:以 pipeline 之分類器對文本進行預測,記錄
pred_label與pred_conf; - 人工覆核:curator 對樣本進行覆核,填入最終
label與reason; - 輔助欄位:附上原始 C4 之
source_file、row,便於追溯原始來源。
Who are the source data producers?
原始文本來自 C4 涵蓋之各網站作者,本資料集之標註由 curator 進行。
Annotations
Annotation process
Curator 依其對「適合訓練語料」之判準進行人工覆核,標準包含是否具可讀性、是否涉及敏感或低品質內容、是否具知識密度等。reason 欄位記錄覆核之簡短理由。
Who are the annotators?
Personal and Sensitive Information
本資料集為 C4 原始文本之子集,未進行額外匿名化。使用者應注意 C4 原始內容即可能包含公開網頁之個人資訊。
Bias, Risks, and Limitations
- 規模小:僅 2,783 筆,作為 classifier 訓練集時容易 overfit;
- 單一標註者:僅 curator 一人標註,存在標註者偏差;
- 原始 C4 雜訊:C4 含大量廣告、模板、低品質網頁,本資料集部分樣本即為「不建議納入訓練」之負例;
- 繁中 / 簡中邊界:C4 之 zh 子集可能含少量簡體中文,使用時需留意。
Recommendations
建議使用者:
- 搭配更大規模之弱監督或半監督資料訓練 classifier;
- 使用時以
label為準,pred_label僅作為 classifier 診斷用; - 若需大規模繁中 pretrain 語料,參考 fineweb-edu-zhtw;
- 遵循 C4 之 ODC-BY 授權條款。
Citation
@misc{c4-zhtw,
title = {c4-zhtw: Labeled Traditional Chinese Subset of C4},
author = {Liang Hsun Huang},
year = {2024},
howpublished = {\url{https://huggingface.co/datasets/lianghsun/c4-zhtw}},
note = {2,783 labeled zh-tw examples from C4 with classifier predictions and manual labels.}
}
Dataset Card Authors
Dataset Card Contact
- Downloads last month
- 67