patvessel/danbooru-rag-G-v3

About the Full-size data set containing the character category:

To maintain versatility and lightweight design, the "character" category has been removed by default in this dataset. If you require the full-size dataset including the "character" category, please visit the patvessel/danbooru-rag-v3 dataset. For more information about the dataset, please refer to the dataset card for that dataset.

---

This dataset is a deeply cleaned and structured version of the isek-ai/danbooru-wiki-2024 (Train Split) dataset.

The original Danbooru Wiki contains a massive amount of noise, including raw URLs, complex Textile markup, Wiki syntax, and redundant/malformed aliases. This linguistic noise significantly degrades the performance of embedding models and confuses Large Language Models (LLMs).

To solve this, I filtered out low-information tags and engineered two specialized text fields tailored for Retrieval-Augmented Generation (RAG) and anime-domain semantic search: embed_text and display_text.

Key Enhancements

Syntax Noise Removal: Used regular expressions (Regex) to precisely strip out Wiki strikethroughs [s], Textile links, raw URLs, HTML headings (h1.), and table syntaxes.

Embedding-Friendly Structure: Automatically converted bulleted lists (* list) into single-line, comma-separated formats to maintain semantic continuity during vectorization.

Data Filtering:

General Tags: Retained only entries where the description body is longer than 30 characters.

Alias Text Sanitization: Cleaned the other_names field by discarding aliases that are over 100 characters, pure numbers, or composed solely of punctuation and whitespace.

Dataset Structure

The final processed dataset contains the following core fields:

tag: The original Danbooru tag identifier (e.g., tasting_plate).

category: The category of the tag (primarily general).

embed_text: [For Embedding Models] Natural language format. Connects the tag name, aliases, and description body into a fluent paragraph, optimized for generating high-quality text embeddings.

display_text: [For LLM Prompt Context] A structured, clean Key-Value format. It clearly separates fields so that when retrieved via RAG, it can be seamlessly injected into the LLM's prompt context for accurate reading.

Data Example

JSON

{ "tag": "tasting plate", "category": "general", "embed_text": "Danbooru tag 'tasting plate' (also known as 味見皿). Ajimisara. A small plate like mamezara but for tasting soups.", "display_text": "tag: tasting plate\category: general\display: tasting plate\also known as: 味見皿\description: Ajimisara. A small plate like mamezara but for tasting soups." }

How to Use

You can easily load this dataset using the Hugging Face datasets library:

from datasets import load_dataset
dataset = load_dataset("patvessel/anbooru-rag-G-v3")
print(dataset["train"][0])

File Formats

This dataset is provided in multiple formats to suit different workflows:

Hugging Face Dataset (Default):

Optimized for memory-efficient loading using the datasets library.

JSONL Format:

A raw, line-delimited JSON file (anbooru-rag-G-v3.jsonl) is available in the repository files. This is recommended for users who prefer using pandas or custom scripts without relying on the datasets library. Usage with Pandas: python import pandas as pd df = pd.read_json("anbooru-rag-G-v3.jsonl", lines=True)

Markdown(.MD) Format:

Designed specifically for users using out-of-the-box, no-code RAG services like AnythingLLM, Open WebUI, or Dify.

Since these platforms usually rely on simple, length-based chunking strategies and do not allow custom splitting logic, a standard JSON or raw file often gets fragmented, causing LLMs to lose track of which tag belongs to which description.

To solve this, this Markdown file employs a "Chunk-Proof Semantic Anchoring" strategy:

• Every single line of description and alias is explicitly prepended with its corresponding tag identifier (e.g., Tag: breast_smother | Description: ...).
• This ensures that no matter how aggressively the platform's naive splitter slices the document, every vectorized chunk inherently carries its Tag identity.

Licensing & Acknowledgements

Source Dataset: isek-ai/danbooru-wiki-2024

The licensing and usage restrictions of this derivative dataset follow the guidelines and policies set by the original Danbooru dataset creators.

Note: This dataset is intended for research and creative AI applications.