Datasets:
Dataset Card for fineweb-edu-zhtw-magistral-annotations
本資料集是訓練 fineweb-edu-zhtw-classifier 所使用的「教育性評分」標註資料:以 Magistral-Small-2506 對 FineWeb-zhtw 樣本文本給予 0–5 分的教育性評分,用於訓練分類器、進而過濾出 fineweb-edu-zhtw。
Dataset Details
Dataset Description
資料集規模龐大(train ≈ 4.7M、validation ≈ 521k),每筆樣本包含:
text:原始繁中網頁段落。score:Magistral-Small-2506 對該段落給的 0–5 分教育性評分(int64)。annotator:標註模型名稱(用於可重現性與後續再標註)。
這份資料的角色是 fineweb-edu-zhtw pipeline 的標註層:基於這些評分訓練出的分類器,再對全量 FineWeb-zhtw 進行推論,最終輸出 fineweb-edu-zhtw。
- Curated by: Huang Liang Hsun
- Language(s) (NLP): Traditional Chinese
- License: MIT
Dataset Sources
- Repository: lianghsun/fineweb-edu-zhtw-magistral-annotations
- Downstream classifier: lianghsun/fineweb-edu-zhtw-classifier
- Final filtered corpus: lianghsun/fineweb-edu-zhtw
- Annotator model: mistralai/Magistral-Small-2506
- Paper: TBA
Uses
Direct Use
- 訓練教育性分類器(label 為 0–5;下游 fineweb-edu-zhtw classifier 再整併為 c0/c1/c2 三類)。
- 研究 LLM-as-Annotator 在繁中內容上的標註行為與一致性。
- 作為大規模 weakly-supervised 文本分類任務的素材。
Out-of-Scope Use
- 不適合直接用於語言模型預訓練;本集是「網頁段落 + 分數」結構,並非乾淨的純文本。
- 不適合作為人工教育性標準;分數來自 LLM 判斷,並非教育專家評分。
Dataset Structure
{
"text": "...",
"score": 4,
"annotator": "Magistral-Small-2506"
}
split:
train:~ 4.7M 筆validation:~ 521k 筆
Dataset Creation
Curation Rationale
要過濾出教育導向的繁中網頁文本,最具規模化的方法是訓練分類器。本資料集即為訓練該分類器所需的標註層:藉由 LLM 對大量段落打分,再以這些 LLM 標註訓練 encoder-based 分類器,達成「LLM 知識 + 小模型推論成本」的折衷。
Source Data
Data Collection and Processing
- 從 FineWeb-zhtw 抽樣段落作為標註對象。
- 以 Magistral-Small-2506 為每段文本給 0–5 分教育性評分,並紀錄標註模型名稱於
annotator欄位。 - 切分為 train / validation 以利分類器訓練。
Who are the source data producers?
- 原始文本來自 Common Crawl → FineWeb(zh-tw 子集)。
- 評分由 Magistral-Small-2506 自動產生。
Annotations
Annotation process
標註以 prompt-based 方式由 Magistral 自動完成。每筆樣本只有一個標註者,未做多人複核。
Who are the annotators?
mistralai/Magistral-Small-2506。
Personal and Sensitive Information
資料源自公開網頁,與 fineweb-edu-zhtw 母集處理一致。
Bias, Risks, and Limitations
- 標註偏誤:Magistral 對「教育性」的定義不一定與人類專家一致,分數分布可能偏高或偏低。
- 規模龐大但分布不一定平衡:高分(4–5)樣本可能比低分樣本少。
- 資料量大(>15GB),下游處理需注意記憶體與儲存空間。
Recommendations
- 訓練分類器時建議做 class balancing 或 class-weighted loss。
- 若要重新標註以對齊不同 LLM 風格,可參考本集流程,將
annotator換為其他模型。
Citation
@misc{fineweb_edu_zhtw_magistral_annotations,
title = {fineweb-edu-zhtw-magistral-annotations: LLM-annotated Educational Score Dataset for FineWeb-zhtw},
author = {Huang, Liang Hsun},
year = {2025},
howpublished = {\url{https://huggingface.co/datasets/lianghsun/fineweb-edu-zhtw-magistral-annotations}}
}
Dataset Card Authors
Dataset Card Contact
- Downloads last month
- 10
Models trained or fine-tuned on lianghsun/fineweb-edu-zhtw-magistral-annotations
Text Classification • 0.3B • Updated • 1