PixelPerv/danbooru-tags-20260518-duckdb

Danbooru Tags Analysis (DuckDB Edition)

This dataset is an optimized, streamlined derivative of u-haru/danbooru-tags-20260518 (thanks so much for making it available, you rock!). It is specifically engineered to facilitate tag analysis for Illustrious and Anime-focused AI models, described in the Civitai article https://civitai.red/articles/30556.

Key Improvements

• Size Optimization: Retains only a limited subset of columns essential for analytical workloads, drastically reducing the file size and download times.
• Strict Data Typing: Columns are cast (e.g., BIGINT, TIMESTAMP WITH TIME ZONE, and structured arrays VARCHAR[]) to accelerate query processing times.
• Single-File DuckDB Format: All data is packaged into a single .duckdb database file. This enables local exploration, easy webUI downloading.

---

Database Schema & Architecture

Below is the structural visualization of the post table schema:

Column Name	Data Type	Description
id	BIGINT	Unique identifier for the Danbooru post (Primary Key).
created_at	TIMESTAMP WITH TIME ZONE	The exact date and time the post was submitted.
uploader_id	BIGINT	Identifier for the user who uploaded the post.
score	BIGINT	Net post score (upvotes minus downvotes).
upvotes	BIGINT	Total number of positive votes.
downvotes	BIGINT	Total number of negative votes.
fav_count	BIGINT	Total number of times users added the post to favorites.
rating	VARCHAR	Content rating code (g, s, q, or e).
general_tags	VARCHAR[]	Array of general tags detailing content (e.g., 1girl, solo).
character_tags	VARCHAR[]	Array of fictional character names depicted.
copyright_tags	VARCHAR[]	Array of series, franchises, or intellectual properties.
artist_tags	VARCHAR[]	Array of creator or artist names.
meta_tags	VARCHAR[]	Array of technical/meta tags (e.g., highres, absurdres).

The table can be recreated or referenced using the following SQL query:

create or replace table post as
select id, created_at::timestamptz as created_at, uploader_id, score, up_score as upvotes, down_score as downvotes, fav_count, rating,
       case when nullif(trim(tag_string_general), '') is null then array[]
            else split(trim(tag_string_general), ' ') end as general_tags,
       case when nullif(trim(tag_string_character), '') is null then array[]
            else split(trim(tag_string_character), ' ') end as character_tags,
       case when nullif(trim(tag_string_copyright), '') is null then array[]
            else split(trim(tag_string_copyright), ' ') end as copyright_tags,
       case when nullif(trim(tag_string_artist), '') is null then array[]
            else split(trim(tag_string_artist), ' ') end as artist_tags,
       case when nullif(trim(tag_string_meta), '') is null then array[]
            else split(trim(tag_string_meta), ' ') end as meta_tags,
from 'data\*.parquet'

Rating Reference Definition

The rating column uses single-character values to denote content safety:

• g = General (Safe for work, general audiences)
• s = Sensitive (Mildly suggestive/borderline content)
• q = Questionable (Explicit themes, but not fully explicit)
• e = Explicit (NSFW, fully explicit content)

---

Analytical Query Example: Extracting Most Popular Characters

To find the top 10 most frequent characters in highly-rated explicit posts along with their average scores:

SELECT 
    character,
    COUNT(*) as post_count,
    ROUND(AVG(score), 2) as avg_score
FROM (
    SELECT score, unnest(character_tags) as t(character)
    FROM post 
    WHERE rating = 'e'
)
GROUP BY character
ORDER BY post_count DESC
LIMIT 10;