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-math-ocr

tw-math-ocr 是一個台灣數學題目之圖片資料集,包含 894 張數學題目圖片,依章節結構組織為 4 章 10 節。圖片來源為台灣數學教材,適用於數學 OCR(光學字元辨識)模型之訓練與評估,特別是包含數學符號、公式與繁體中文之混合辨識場景。

Dataset Details

Dataset Description

本資料集收錄台灣數學教材中的題目圖片,以章節目錄結構組織。每張圖片包含一道或多道數學題目,內容涵蓋數學符號、公式、圖表與繁體中文說明文字。

章節分布如下:

章節 圖片數
ch1/1-1 63
ch1/1-2 72
ch2/2-1 88
ch2/2-2 69
ch3/3-1 137
ch3/3-2 63
ch3/3-3 118
ch4/4-1 96
ch4/4-2 94
ch4/4-3 94

圖片格式為 GIF,典型尺寸約 470×300 像素。

  • 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:

  • 訓練與評估數學 OCR 模型,特別是繁體中文與數學符號混合之辨識任務;
  • 建構數學題目之自動化數位化流程(圖片轉結構化文字/LaTeX);
  • 視覺語言模型(VLM)在數學圖片理解任務上之微調與評測。

Out-of-Scope Use

本資料集不適用於下列用途:

  • 作為通用 OCR 之訓練資料,因內容僅限於數學題目圖片。
  • 直接作為數學教學素材使用,因不包含答案或解題過程。
  • 用於非繁體中文之數學 OCR 任務。

Dataset Structure

欄位 說明
id 圖片唯一識別碼
filename 原始檔名(如 10043.jpg
relpath 章節相對路徑(如 ch1/1-1/10043.jpg
image 圖片內容(HuggingFace Image 格式)
統計項目 數值
圖片總數 894
章數 4
節數 10
圖片格式 GIF
典型尺寸 ~470×300 px

Dataset Creation

Curation Rationale

現有的數學 OCR 資料集多以英文與 LaTeX 為主,缺乏繁體中文數學教材之圖片資料。本資料集填補此空缺,提供台灣數學教材中的題目圖片,使 OCR 與 VLM 模型能學習辨識繁體中文數學內容。

Source Data

Data Collection and Processing

圖片擷取自台灣數學教材,依原始章節結構保留目錄組織,轉換為 Parquet 格式儲存於 HuggingFace Hub。

Who are the source data producers?

原始圖片來自台灣數學教材出版物。

Annotations

Annotation process

本資料集目前不包含文字標註(ground truth transcription)。圖片僅以章節路徑組織,尚無對應之 OCR 文字或 LaTeX 輸出。

Who are the annotators?

不適用。

Personal and Sensitive Information

本資料集僅包含數學題目圖片,不涉及個人資訊或敏感資料。

Bias, Risks, and Limitations

  • 目前不包含 ground truth 文字標註,無法直接用於監督式 OCR 訓練,需搭配自行標註或半自動標註流程。
  • 圖片來源為特定教材,章節覆蓋與難度分佈可能不具全面代表性。
  • GIF 格式與調色盤模式(mode=P)可能需轉換為 RGB 後才能用於部分模型。
  • 894 張圖片規模較小,建議搭配其他數學 OCR 資料集使用。

Recommendations

建議使用者先對圖片進行文字標註(如 LaTeX 或結構化文字),再用於監督式訓練。可搭配 VLM 進行半自動標註以加速流程。

Citation

@misc{tw-math-ocr,
  title        = {tw-math-ocr: Traditional Chinese Math OCR Image Dataset},
  author       = {Liang Hsun Huang},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-math-ocr}},
  note         = {894 math problem images from Taiwan textbooks, organized by chapter and section.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang

Downloads last month
19