deepghs/site_tags

Site Tags Dataset Collection

Summary

This comprehensive dataset collection provides structured tag metadata from 18 popular image hosting and booru websites, offering a unified repository for multilingual tag analysis and content classification. The dataset encompasses over 2.5 million unique tags across various platforms including Danbooru, Gelbooru, Pixiv, Sankaku Complex, and Wallhaven, making it one of the most extensive collections of image annotation metadata available. Each platform's tags are provided in multiple formats including JSON, CSV, Parquet, and SQLite databases, ensuring compatibility with diverse data processing workflows.

The dataset captures rich semantic relationships between tags through various metadata fields such as tag categories, post counts, aliases, and hierarchical relationships. For platforms like Danbooru and Gelbooru, the collection includes comprehensive tag alias mappings that facilitate cross-platform tag normalization and synonym resolution. The multilingual nature of the dataset is particularly valuable, with tags available in English, Japanese, and Russian across different platforms, enabling research in cross-lingual information retrieval and multicultural content analysis.

Key technical features include structured categorization systems where tags are classified into types such as character, copyright, artist, and general tags, providing granular control over content classification. The inclusion of usage statistics like post counts and view metrics allows for popularity-based analysis and trend identification. This dataset serves as a foundational resource for machine learning applications in content recommendation, automated tagging systems, and semantic search engines, while also supporting academic research in digital humanities and web content analysis.

Dataset Structure

The repository is organized by source website, with each directory containing tag data in multiple formats:

• JSON files: Complete tag metadata with full structural information
• CSV files: Tabular format for easy data analysis
• Parquet files: Optimized columnar storage for large-scale processing
• SQLite databases: Relational database format for complex queries

Supported Platforms

• Anime-focused: anime-pictures.net, danbooru.donmai.us, safebooru.donmai.us
• Booru networks: gelbooru.com, konachan.com, konachan.net, rule34.xxx, e621.net
• Art communities: pixiv.net, en.pixiv.net, wallhaven.cc
• Specialized: sankakucomplex.com, hypnohub.net, lolibooru.moe, xbooru.com, zerochan.net

Usage

The dataset can be accessed through multiple interfaces depending on your preferred data format:

Using JSON Format

import json
import pandas as pd

# Load tags from Danbooru
with open('danbooru.donmai.us/tags.json', 'r', encoding='utf-8') as f:
    danbooru_tags = json.load(f)

# Convert to DataFrame for analysis
df = pd.DataFrame(danbooru_tags)
print(f"Danbooru contains {len(df)} tags")
print(f"Most popular tags: {df.nlargest(5, 'post_count')['name'].tolist()}")

Using Parquet Format

import pandas as pd

# Load efficient parquet format
df = pd.read_parquet('gelbooru.com/tags.parquet')
print(f"Gelbooru tag categories: {df['type'].value_counts()}")

Using SQLite Database

import sqlite3

# Query the SQLite database
conn = sqlite3.connect('pixiv.net/tags.sqlite')
cursor = conn.cursor()

# Get tag statistics
cursor.execute("SELECT COUNT(*), AVG(posts) FROM tags")
count, avg_posts = cursor.fetchone()
print(f"Pixiv has {count} tags with average {avg_posts:.1f} posts")

Accessing Tag Aliases

import json

# Load tag aliases for synonym resolution
with open('danbooru.donmai.us/tag_aliases.json', 'r') as f:
    aliases = json.load(f)

# Create alias mapping
alias_map = {item['antecedent_name']: item['consequent_name'] for item in aliases}
print(f"Found {len(alias_map)} tag aliases")

Data Schema

Common Tag Fields

• id: Unique identifier for the tag
• name: Primary tag name (often in English)
• post_count: Number of posts using this tag
• category/type: Classification (0=general, 1=artist, 3=copyright, 4=character)
• created_at/updated_at: Timestamps for tag lifecycle

Platform-Specific Extensions

• Pixiv: wiki_url, views, checklists, content type flags
• Sankaku: trans_en, trans_ja, trans_ru multilingual translations
• Wallhaven: category_name, subscriptions, view statistics
• e621: related_tags with co-occurrence information

Original Content

Directory Structure

datasets/deepghs/site_tags@main/
├── anime-pictures.net
│   ├── tags.csv
│   ├── tags.json
│   ├── tags.parquet
│   └── tags.sqlite
├── booru.allthefallen.moe
│   ├── tag_aliases.csv
│   ├── tag_aliases.json
│   ├── tags.csv
│   ├── tags.json
│   └── tags.sqlite
├── chan.sankakucomplex.com
│   ├── tags.csv
│   ├── tags.json
│   └── tags.sqlite
├── danbooru.donmai.us
│   ├── tag_aliases.csv
│   ├── tag_aliases.json
│   ├── tag_aliases.parquet
│   ├── tags.csv
│   ├── tags.json
│   ├── tags.parquet
│   └── tags.sqlite
# ... and 14 more platforms

Sample Data Examples

The dataset includes comprehensive tag information from each platform. For example, Danbooru tags include metadata such as:

• Tag categories (general, artist, copyright, character)
• Post counts indicating popularity
• Timestamp information
• Deprecation status
• Word segmentation for compound tags

Pixiv data includes additional content classification flags:

• Media type indicators (anime, manga, novel, game, etc.)
• View and engagement statistics
• Wiki integration URLs

Applications

This dataset enables various applications including:

• Content recommendation systems based on tag similarity
• Multilingual tag normalization across platforms
• Trend analysis through post count temporal data
• Semantic search enhancement using tag relationships
• Machine learning training for automated tagging
• Cultural analysis of content preferences across regions

Citation

@misc{site_tags_dataset,
  title        = {Site Tags Dataset Collection: Comprehensive Tag Metadata from Image Hosting Platforms},
  author       = {deepghs},
  howpublished = {\url{https://huggingface.co/datasets/deepghs/site_tags}},
  year         = {2024},
  note         = {A unified collection of structured tag metadata from 18 popular image hosting and booru websites, enabling multilingual tag analysis and content classification research},
  abstract     = {This comprehensive dataset collection provides structured tag metadata from 18 popular image hosting and booru websites, offering a unified repository for multilingual tag analysis and content classification. The dataset encompasses over 2.5 million unique tags across various platforms including Danbooru, Gelbooru, Pixiv, Sankaku Complex, and Wallhaven, making it one of the most extensive collections of image annotation metadata available. Each platform's tags are provided in multiple formats including JSON, CSV, Parquet, and SQLite databases, ensuring compatibility with diverse data processing workflows. The dataset captures rich semantic relationships between tags through various metadata fields such as tag categories, post counts, aliases, and hierarchical relationships.},
  keywords     = {tags, metadata, image-tagging, booru, dataset, multilingual}
}