Sn0w123/booru-characters

Booru Characters

Overview

A line-oriented JSON dataset of character tag metadata extracted from Danbooru using the Danbooru API. Each record contains tag-level metadata and simple relationships between tags.

Contents

• —hf_dataset/characters.jsonl: one JSON object per line. Each object contains the fields described below.
• —hf_dataset/dataset_info.json: minimal metadata describing the exported features.

Fields (per record)

• —id (int): internal database id
• —name (string): tag name
• —post_count (int): number of posts associated with the tag at extraction time
• —gender (string): inferred majority gender tag (e.g., 1girl / 1boy)
• —clothing (list[string]): tags classified as clothing
• —characteristics (list[string]): tags classified as character attributes
• —copyright (list[string]): most common copyright tags from posts
• —relationships (dict): lists of related tag names under parents, children, and siblings

Methods

• —Tag selection: character-category tags are queried from the Danbooru API. Tags are filtered by a minimum post count of 100 and ordered by post count.
• —Relationship extraction: parent/child relations are derived from tag implication fields returned by the API (antecedent_implications / consequent_implications). Siblings are inferred by grouping tags that share the same parent.
• —Post sampling: for each character tag the script fetches posts matching [<character> solo] (up to 5 pages), requesting only score, rating and tag fields. A minimum post score threshold is applied for high-volume tags to reduce noise.
• —Tag classification: general tags from posts are split into characteristics and clothing using pattern lists manually created.
• —Frequency thresholds: characteristics are included when they appear in >=35% of sampled posts; clothing tags use a 15% threshold; copyright tags use a 50% threshold.