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-public-figures-vision

tw-public-figures-vision 是一個以台灣公眾人物與地景為主題之繁體中文視覺語言資料集,合計約 19,000 筆,分為 383 個 parquet shard。每筆包含一至多張圖片(以二進位 bytes 儲存)、對應之中文說明文字、以及(可選之)對話格式 messages,可用於訓練繁中 vision-language model(VLM)對台灣人物、地點與社群貼文風格之理解。

Dataset Details

Dataset Description

繁體中文 VLM 之訓練資料多以英文圖文對為主,缺乏對台灣本土人物、地景、活動之涵蓋。本資料集由 curator 整理來自公開社群媒體與公開來源之圖文內容,內容包含:

  • 台灣地景照片(如花東縱谷、夜市、活動現場);
  • 社群貼文之說明文字(含 emoji、hashtag);
  • 投稿者之識別碼(contributor)與每筆圖文之獨立 uuid
  • 圖片之二進位內容與原始 path

資料以 parquet 分 shard 儲存,每 shard 約 50 筆。

  • Curated by: Liang Hsun Huang
  • Language(s) (NLP): Traditional Chinese
  • License: Other — 圖片與文字皆來自公開社群媒體,授權依原始投稿者之社群使用條款為準

Dataset Sources

Uses

Direct Use

本資料集主要設計用於:

  • 繁中 VLM 之訓練,加強對台灣人物、地景與社群貼文之理解;
  • 訓練模型生成社群貼文風格之中文說明;
  • 作為繁中圖文對齊(image-text alignment)之補強資料;
  • 研究台灣本土視覺元素之語言描述。

Out-of-Scope Use

本資料集不適用於下列用途:

  • 作為商業產品之訓練資料,商業使用前應確認每筆圖文之原始社群授權;
  • 作為個人識別、監控或人臉辨識之訓練資料;
  • 作為生成深偽(deepfake)內容之素材;
  • 未經原始投稿者同意之再散布。

Dataset Structure

{
  "images": [{"bytes": b"<png bytes>", "path": "..."}],
  "text": "慶祝我們的ig上線\n那就來一張富里農會花海!...",
  "messages": [],
  "contributor": "<contributor id>",
  "img_license": "social_media",
  "text_license": "social_media",
  "uuid": "1f7526d7-7417-4caa-8d62-de1243e44ed0",
  "model": "",
  "source": "<original social media URL>"
}
欄位 型別 說明
images list[{bytes, path}] 一至多張圖片(bytes 二進位 + 原始 path)
text string 對應之說明文字(含 emoji / hashtag)
messages list[{role, content}] 可選之對話格式(多數為空)
contributor string 投稿者識別碼
img_license string 圖片授權(通常為 social_media
text_license string 文字授權(通常為 social_media
uuid string 每筆之唯一 ID
model string 若由模型生成,記錄模型名稱
source string 原始社群來源 URL
統計項目 數值
總筆數 ~19,000
Parquet shards 383
每 shard 筆數 ~50

Dataset Creation

Curation Rationale

台灣本土 VLM 在處理在地內容(人物、地景、活動、社群貼文)時能力有限。本資料集以社群貼文式之圖文對提供補強,讓 VLM 能更貼近台灣使用者實際接觸之視覺內容。

Source Data

Data Collection and Processing

  • 來源:公開社群媒體之投稿內容;
  • 處理:以 uuid 建立每筆之唯一鍵,將圖片 bytes 與說明文字配對,並分 shard 為 parquet;
  • 格式:每 shard 約 50 筆。

Who are the source data producers?

原始圖文由各投稿者於公開社群發布,以 contributor 欄位記錄。

Annotations

Annotation process

本資料集不包含額外人工標註;messages 欄位多為預留空白。

Who are the annotators?

不適用。

Personal and Sensitive Information

本資料集之圖文來自公開社群內容,可能包含人臉、地點、活動等可識別資訊。使用者不得將資料集用於個人識別、監控或未經同意之任何用途。

Bias, Risks, and Limitations

  • 授權不明確:社群貼文授權以原始平台規則為主,再散布風險需自行評估;
  • 投稿者偏態:內容以少數投稿者為主,多樣性有限;
  • 可識別人臉:部分圖片含可識別人物;
  • 風格偏態:內容以花東縱谷、農會花海等特定主題為主,非全台灣均勻分布。

Recommendations

建議使用者:

  • 商業使用前逐一確認每筆圖文之原始社群授權;
  • 訓練時避免強化人臉辨識能力,並於輸出階段過濾可識別資訊;
  • 與其他繁中圖文資料集(如 FineWeb-edu-zhtw 之圖文子集)搭配使用,降低風格偏態。

Citation

@misc{tw-public-figures-vision,
  title        = {tw-public-figures-vision: Taiwan Public Figures Vision Dataset},
  author       = {Liang Hsun Huang},
  year         = {2024},
  howpublished = {\url{https://huggingface.co/datasets/lianghsun/tw-public-figures-vision}},
  note         = {~19,000 Taiwanese image-text pairs from public social media posts for VLM training.}
}

Dataset Card Authors

Liang Hsun Huang

Dataset Card Contact

Liang Hsun Huang

Downloads last month
10