Grio43/Tag_cleaning

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

Column correction (2026-09-09): renamed image_id to post_id. All 1,739,622 rows and their values are unchanged. Update code that references image_id to use post_id.

Column	Type	Description
post_id	string	Danbooru post ID, or a generated-sample identifier
add	list[string]	Tags to add
remove	list[string]	Tags to remove
updated_at	string	Original correction timestamp in ISO-8601 format

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

• 1,739,622 correction rows
• 1,915,583 add actions
• 474,953 remove actions
• 9,363 unique source tags examined
• 301 dead or unmapped source tags removed
• 39,371 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. Additional data cleaning was applied to the correction fields, and rows with no remaining actions were omitted. 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.