Danbooru 2026 Tag Cleaning Corrections

This dataset contains image-level tag corrections for an anime-image tagging corpus. It contains correction instructions only; it does not contain images, captions, or the original sidecar metadata.

Schema

Tag values use the prefixes from the reference sidecars:

• gen: — general
• char: — character
• copyright: — copyright/series
• artist: — artist
• meta: — metadata

For example, gen:holding_toothbrush is a general tag.

Dataset statistics

• 1,740,318 correction rows
• 1,916,768 add actions
• 474,953 remove actions
• 9,364 unique source tags examined
• 301 dead or unmapped source tags removed
• 38,675 rows with no remaining actions removed

Normalization

The corrections were normalized against a Danbooru metadata snapshot taken on 2026-08-30. Active tag aliases were resolved to canonical names, explicit wiki-documented splits were expanded when available, and tags with no current record or no active replacement were removed. No fuzzy tag-name guesses were used.

The normalization used 1 serialized API request at a minimum interval of one second, reusing cached responses for the remaining tag records. The original merged.json and the original image sidecars were not modified.

Usage

from datasets import load_dataset

dataset = load_dataset("Grio43/Tag_cleaning", split="train")
print(dataset[0])

Limitations

This is a correction manifest, not a fully relabeled image dataset. Applying the corrections requires access to the corresponding image metadata. The corrections reflect the source cleaning process and should be reviewed before being used as ground truth.

Source and audit

The source manifest was merged.json. The local normalized release was validated for JSON/Parquet row parity before publication. The Danbooru site and its tag documentation are available at https://danbooru.donmai.us.