You can deploy this model and start earning money today!

haiefff/
nsfw-classifier-v2

Name: nsfw-classifier-v2
Availability: InStock

not-for-all-audiences

Views

No views yet

NSFW Classifier v2 — Production SigLIP2 Ensemble

A multi-label image moderation classifier developed for Bluesky content. The promoted production system is a validation-selected probability ensemble of three SigLIP2-base checkpoints.

Classes

real_sexual — photographic sexual/nude content; includes the former real_porn category
abdl_real — photographic ABDL/visible adult diaper content
abdl_illustrated — illustrated ABDL content
sexual_anime — sexual anime/manga/general illustrated content
sexual_furry — sexual furry/anthropomorphic content

The checkpoints retain a sixth tensor slot named real_porn for compatibility, but it has zero ensemble weight and is not a consumer-facing category.

Files

avg_pool.pt — legacy SigLIP2 global-average-pool component
map_pool.pt — SigLIP2 native learned-attention-pool component
abdl_clean.pt — native-pool component trained with focused Gemma-cleaned ABDL labels
ensemble.json — per-class component weights, preprocessing, and metadata
thresholds.json — promoted validation-frozen thresholds
SHA256SUMS — artifact integrity hashes

All three checkpoints are required to reproduce production probabilities. The abdl_clean.pt checkpoint is the strongest standalone compatibility model, but its probabilities are not equivalent to the ensemble.

Evaluation and production operating point

Evaluation used a 4,175-image held-out split grouped by artist/account. Historical F1-maximizing thresholds measured four-class macro F1 0.8345, but produced an unacceptable false-positive volume for automatic labeling and are not deployed.

The promoted thresholds target approximately 1% empirical FPR per class on validation negatives. Held-out test behavior at this conservative production operating point is:

Class	Threshold	Precision	Recall	F1	FPR
real_sexual	0.755	0.942	0.717	0.814	1.00%
abdl_real	0.7135	0.808	0.818	0.813	0.37%*
abdl_illustrated	0.785	0.857	0.579	0.691	0.92%
sexual_anime	0.930	0.955	0.378	0.542	0.38%
sexual_furry	0.895	0.921	0.340	0.497	1.13%

* ABDL-real uses the frozen clean-label protocol: validation labels were verified before selection and test labels only after the threshold was frozen. Other classes use weak labels and should be interpreted accordingly.

json

{
  "real_sexual": 0.755,
  "abdl_real": 0.7134765982627869,
  "abdl_illustrated": 0.785,
  "sexual_anime": 0.93,
  "sexual_furry": 0.895
}

The production policy prioritizes precision over recall. Do not substitute a universal 0.5 threshold or the historical F1-maximizing thresholds.

Usage

Use the implementation in haileyok/nsfw-classifier:

bash

hf download haiefff/nsfw-classifier-v2 \
  avg_pool.pt map_pool.pt abdl_clean.pt ensemble.json thresholds.json SHA256SUMS \
  --local-dir models/production
streamlit run src/nsfw_classifier/serve/bluesky_tester.py --server.port 8502

Input preprocessing is 256×256 RGB with mean/std [0.5, 0.5, 0.5]. Production probabilities are the per-class weighted average specified in ensemble.json. Use public_class_names for returned categories and checkpoint_class_names only to map those categories to the six checkpoint logits.

Limitations and intended use

Intended as a moderation-assistance signal, not the sole basis for punitive action.
The model may make errors on novel art styles, crops, text-heavy images, and domains underrepresented in training.
Bluesky anime/furry labels were obtained through weak supervision. Visual error audits found both missing positives and incorrect labels, so their measured F1 is conservative agreement with imperfect labels.
Pornographic and other photographic sexual/nude content share the real_sexual public category; the legacy checkpoint slot must not be returned.
Production thresholds target a conservative validation false-positive budget, but deployments should still validate them on their own traffic and policy.
The training/evaluation data contains adult content and should be handled accordingly.