You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset contains anime illustrations collected from Danbooru and Pixiv (third-party artworks, redistributed here at reduced resolution for research on aesthetic-quality assessment) together with AI-generated negatives and human quality votes. By requesting access you agree to use it for research / model evaluation only, to respect the rights of the original artists, and not to redistribute the images.
Log in or Sign Up to review the conditions and access this dataset content.
noobreward-acgq-data
Training and validation data of the acgq ACG aesthetic-quality reward model (model: Yidhar/noobreward-acgq; code:
GitHub Yidhar/noobreward). 120,087 images with human good/bad votes, per-dimension human scores, a Gemini teacher, WD-tagger
priors, metadata labels (artist / rating / year / score / characters) and the derived validation set.
Images (images/)
aligned-000.tar … aligned-011.tar (about 1 GB each, plain tar, 120,491 JPEG files) hold every image referenced by the
manifest, already aligned: long side resized to 768 px (the P set is 768 wide), RGB, alpha composited over white. Paths
inside the tars are the manifest uri values (aligned/<source>/…); file_list.txt lists them. Originals are not included.
| source | images | role |
|---|---|---|
| danbooru | 38,542 | U / X pool (Danbooru posts, ≥ 2015, sampled by score and tags) |
| pixiv_top50 | 38,493 | curated Pixiv set (mostly good, a minority bad) |
| p20k | 19,298 | P set: curated high-quality positives |
| danbooru_style | 14,451 | artist style set (≥ 20 images per artist, 444 artists) |
| nai3 / gen_niji5_* / gen_anime_synthetics / gen_nai3_artists / gen_animagine30_artists | 8,600 | AI-generated negatives |
| user_bad_ai / user_good_gen / user_hack / user_eval | 703 | user-supplied badcases, verified-good sets, hack set |
Roles (pu_role in the manifest): P curated positives (19,298), U unlabelled pool (32,520), X extra / style /
generated images (67,247), H hard set (1,000), N (22).
Labels
| file | rows | content |
|---|---|---|
labels/h_labels.parquet |
8,855 | human good/bad votes aggregated per image: soft_label (0–1, 0.5 = disagreement), n_votes, weight, round_id, guideline_version |
analysis/h_labels_debiased.parquet |
8,858 | the same labels after annotator-leniency × era/rating debiasing (used for training) |
labels/human_dims.parquet |
7,221 | per-annotator dimension scores: composition, color, anatomy, detail (1–4; −1 = not applicable) |
labels/human_dims_images.parquet |
693 | images of the dimension-scoring rounds |
labels/joint.parquet |
29,571 | joined view of human votes, human dims and effective dims per image |
labels/val_set.parquet |
2,164 | the fixed validation set (artist-disjoint), with reason |
labels/extra/*_labels.parquet / *_dims.parquet |
9,299 / 550 | labels of the extra sets (generated negatives = bad, NAI3 = bad, user-verified good sets, hack set); never used for validation |
labels/meta/meta.parquet |
120,087 | metadata labels: artist_id (444 classes, style_split train/val), rating (g/s/q/e → 0–3, −1 unknown), year, era (5 buckets), score, score_z (year-adjusted z-score), character_ids (762 classes), typed tag strings |
labels/meta/rating_soft.parquet |
67,094 | PixAI tagger v1.0 soft ratings + top style predictions for images without a site rating |
labels/meta/artist_index.json, character_index.json |
class index → name | |
teacher/teacher_labels_merged_v2.parquet |
29,707 | Gemini teacher labels: dims, dims_fine, defects, style, evidence, refused (JSON strings) |
tagger/tagger.parquet |
120,785 | WD-tagger outputs: tag_prior (256 selected tags, see vocab.json), defect tags, style, top tags |
completeness/completeness.parquet |
120,087 | deepghs anime_completeness teacher: polished / rough / monochrome |
signals/signals.parquet (+ normalizer.json) |
120,785 | low-level image signals (sharpness, noise, contrast, …) and a 12-bin hue histogram |
manifests/coldstart.parquet |
120,087 | the manifest: image_id, uri, source, split, pu_role, sizes, sha256, phash, artist(s), tags, rating, year, score, danbooru_id |
manifests/danbooru.parquet, danbooru_plan.parquet |
184,456 | the full Danbooru sampling pool the U/X images were drawn from (metadata only) |
image_id is the join key everywhere. SHA256SUMS covers every file.
Validation protocol
Validation images are the manifest val split ∪ the H set ∪ a 15% artist hash bucket; artists of validation images are
excluded from training (leak guard), and the extra sets are never used for validation. The deployed model reaches 0.729
AUROC on the 2,154 real-image validation labels (0.847 on ≥2-vote unanimous labels).
Licensing and intended use
The illustrations are third-party artworks from Danbooru and Pixiv, redistributed at reduced resolution for research on
aesthetic-quality assessment under gated access. Rights remain with the original artists; do not redistribute the images or
use them to train generative models without the artists' permission. Human votes were collected with the project's own
annotation tool by a small group of annotators (pseudonymous annotator_id). Teacher labels were produced with Gemini and
the WD / PixAI / deepghs taggers.
- Downloads last month
- 9