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.

台灣視覺知識庫 v3 (Formosa VLM Knowledge Base)

概述

台灣視覺知識庫 (Visual RAG Knowledge Base) 是 Formosa-VLM 的核心模組, 透過 CLIP 視覺檢索 + 知識注入,讓 VLM 能正確辨識台灣在地事物(建築、美食、寺廟、自然景觀等)。

此資料集為 **備援用 (backup)**,包含完整的實體 metadata、參考圖片、CLIP embeddings。

資料集統計

指標 數值
實體數 253
分類數 12
參考圖片 1648
CLIP Embeddings 1648 (ViT-B/16, 512-dim)
Schema 版本 v3

分類分布

分類 實體數
aboriginal 19
architecture 16
daily_life 18
festival_event 21
food 43
industry 12
landmark 51
nature 21
tea_beverage 9
temple_religion 21
traditional_craft 10
transport 12

Tier 分布

Tier 實體數
Tier 1 116
Tier 2 126
Tier 3 11

檔案結構

├── entities/                    # HF Dataset (253 rows)
│   └── train/
│       └── data-00000-of-00001.parquet
├── reference_images/            # 參考圖片 (~1,600+ 張, 按分類子資料夾)
│   ├── landmark/
│   ├── food/
│   ├── temple_religion/
│   └── ...
├── reference_embeddings.npy     # CLIP ViT-B/16 embedding matrix
├── entity_seed_list.yaml        # 實體種子清單 (schema v3)
├── reference_index.json         # 完整實體索引
├── entity_descriptions.json     # 實體描述
└── knowledge_base_report.json   # 建置報告

使用方式

from datasets import load_dataset
import numpy as np

# 載入實體 metadata
ds = load_dataset("renhehuang/formosa-vlm-knowledge-v3", name="entities")

# 載入 embeddings
from huggingface_hub import hf_hub_download
emb_path = hf_hub_download(
    "renhehuang/formosa-vlm-knowledge-v3",
    "reference_embeddings.npy",
    repo_type="dataset",
)
embeddings = np.load(emb_path)

Schema v3 欄位

欄位 說明
entity_id 實體唯一識別
category 主分類 (12 類)
secondary_categories 次標籤列表
description 繁體中文描述
aliases 別名列表
embedding_indices 對應 embeddings.npy 的 row indices
image_count 參考圖片數
source 資料來源 (wikipedia / manual)
granularity place_instance / entity_type / event_or_practice / brand_or_system
tier 1=高辨識度核心 / 2=重要但需上下文 / 3=文化價值高但資料稀疏
locale 地區脈絡 (taipei, tainan, nationwide...)
visual_anchor 主要視覺焦點

授權

此資料集為 Formosa-VLM 專案之備援,僅供學術研究與開源社群使用。 圖片來源包含 Wikimedia Commons (CC BY-SA) 與手動收集。

Downloads last month
5