Most downstream taggers and training pipelines suffer from long-tail label noise that hides in rare tags; this dataset surfaces targeted, image-level corrections so you can patch labels without rehosting images. The release is a manifest of add/remove actions normalized against a Danbooru snapshot, not a relabeled image dump.
What Sets It Apart
- Scope and granularity: 1,739,622 correction rows with explicit per-post
addandremovelists, enabling surgical fixes rather than blanket relabels. This makes it practical to apply changes to existing metadata pipelines. - Normalization and provenance: corrections were reconciled to a Danbooru metadata snapshot taken on 2026-08-30, with alias resolution and documented splits expanded where available. Rows with no remaining actions were omitted, and dead/unmappable source tags were removed.
- Lightweight manifest format: schema is simple (post_id, add[], remove[], updated_at) and distributed in JSON/Parquet, compatible with pandas/polars and Hugging Face Datasets for easy joins against image sidecars.
Who It's For and Tradeoffs
Great fit if you maintain or train multi-label anime/illustration taggers and need a vetted corrections layer to improve label quality without reannotating images. Look elsewhere if you require fully relabeled image files or ground-truth annotations independent of Danbooru metadata—this dataset is a correction manifest and must be applied to the original image metadata to take effect. Also note some structural label noise in the underlying source corpus remains; the corrections are a targeted, AI-assisted plus human-reviewed pass, not a complete relabel.
