masterofaudio2077/manga-moderation-bench

Manga Moderation Benchmark

Dataset Summary

A multi-label content-moderation benchmark for manga panels. 8,092 black-and-white panels from 1,110 series, rated for 12 sensitive themes on a 0-3 scale, each rating with a short written reason. It is meant for evaluating and training content-safety classifiers and vision-language moderators on manga.

Panels rated as sensitive by the model are shown blurred in the dataset viewer and in the image column (3,809 of 8,092). The blur is irreversible; the full-size originals are kept in the originals/ folder and unblur.py restores them (see Restoring the originals).

The ratings are silver labels written by a vision-language model and have not been reviewed by people. Treat them as a starting point for a human audit.

Supported Tasks and Leaderboards

• Multi-label image classification: predict which themes a panel contains (<tag> >= 1), or their severity (0-3).
• Calibration of a moderator: compare a classifier's severity with the model's reasons.
• No leaderboard yet.

Languages

The panels contain Japanese ,Arabic ,and English text in speech bubbles; the card and the tags are in English.

Versions of the labels

The same panels carry two independent sets of silver labels. Both configs have the same columns (the sonnet55 config adds model), and id is the same in both, so they can be joined. The full-size originals in originals/ serve both.

Config	Files	Annotator	Prompt	Notes
sonnet55 (default)	sonnet55/w<worker>-NNNNN.parquet	Claude Sonnet 5.5, read through the claude.ai web interface	up to 4 panels per request (the originals resized to a long side of 1,280 px), a JSON answer per panel with a reason and severity per tag	the blur in image follows Sonnet's own ratings (same blur rule); a panel whose answer was not complete is valid=false and blurred. Panels Qwen rated as explicit were sent one per request.
qwen3vl	viewer/*.parquet	Qwen3-VL-8B-Instruct	one panel per request, JSON schema constrained	the version described in the rest of this card; the counts below (8,092 panels, 3,809 blurred, the per-tag table) are for this config.

Both are model labels without human review. Compare them panel by panel to find disagreements worth a human audit.

Dataset Structure

Data Instances

One row is one panel.

Data Fields

field	description
id	shard index and panel path in the source
blurred	whether the panel was blurred in image (see the blur rule below)
image	the black-and-white panel: blurred when blurred is true, otherwise a downscaled copy (long side 768 px)
(originals/ folder)	originals/train-*.parquet holds id and original, the full-size original of every panel; unblur.py writes them out
panel, source, series, chapter	origin of the panel in the source collection (chapter is empty when the source had no chapter folder)
valid	whether the model's answer was a complete valid rating (otherwise all of its tags are 0)
<tag>	severity of each tag: 0 absent, 1 mild or implied, 2 clear, 3 graphic or the main focus
reasons	JSON object {tag: reason} with the model's one-sentence justification for every tag rated 1 or higher

Tags (violence, gore, blood, nudity, sexual_content, suggestive, self_harm, death, weapons, substance_use, disturbing_horror, crude_humor), with the number of panels rated 1 or higher and 2 or higher:

tag	what counts	rated 1+	rated 2+
violence	physical fighting, attacks, beatings, threats with force, or the clear result of such acts on people or creatures	3,501	1,977
gore	injuries shown in graphic detail: open wounds, exposed organs, dismemberment, severed body parts	910	429
blood	visible blood, bleeding or blood stains, whatever the amount	524	219
nudity	bare breasts, buttocks or genitals, or characters with no clothing covering those areas	1,025	824
sexual_content	sexual acts, sexual contact, or poses and framing whose clear purpose is sexual	1,013	908
suggestive	revealing outfits, fan-service framing, or erotic tension without nudity or a sexual act	1,612	1,178
self_harm	a character hurting themselves, suicide or its attempt, or tools and scenes that depict it	103	62
death	a dead body, a killing, or a character dying in the panel	970	459
weapons	guns, blades, bombs or other weapons being held, aimed or used	1,113	617
substance_use	smoking, drinking alcohol, drugs, or intoxication	337	30
disturbing_horror	frightening or unsettling imagery: monsters, body horror, torture, distressed or terrorised characters	3,202	1,447
crude_humor	toilet humor, bodily-function jokes or vulgar gestures	280	109

Blur rule

A panel is blurred when, nudity, sexual_content, gore or self_harm is rated 1 or higher, or suggestive, violence, blood, death or disturbing_horror is rated 2 or higher, or the model's answer was invalid. weapons, substance_use and crude_humor never blur a panel.

Restoring the originals

pip install huggingface_hub pyarrow pillow
python unblur.py --repo masterofaudio2077/manga-moderation-bench --out originals

Data Splits

One train split, meant to be used as an evaluation set. Invalid answers: 0.

Dataset Creation

Curation Rationale

Manga moderation needs test material that is varied across series and shows the themes moderators actually meet. Panels are drawn at random, at most 100 per series, so that no series dominates.

Source Data

Panels come from a private collection of black-and-white manga panels. The panels belong to their respective rights holders.

Annotations

Annotator: Qwen/Qwen3-VL-8B-Instruct, run with vLLM (fp8 weights, fp16 compute), greedy decoding, answers constrained to a JSON schema with one integer per tag, one panel per request, longest image side 1024 px. For every tag it gives a short written reason before the rating, kept in reasons so a reviewer can see why a level was chosen. No human reviewed the labels.

Considerations for Using the Data

Social Impact of Dataset

The data is intended to make moderation tools more accurate and more consistent, including for younger audiences who may be protected by those tools.

Discussion of Biases

Ratings reflect one model's judgement and its cultural defaults. Stylised art and dark shading can move a rating up or down. Series with more panels in the source are capped, but the choice of series follows the source collection, not any balanced catalogue.

Other Known Limitations

Silver labels, no human audit, one annotator model, one prompt. Panels are cropped views, so a theme that depends on the surrounding page can be missed.

Citation Information

@misc{manga_moderation_bench,
  title = {Manga Moderation Benchmark},
  year = {2026},
  howpublished = {Hugging Face dataset, masterofaudio2077/manga-moderation-bench}
}