Model card for ncii-guard-light-binary-v01

The binary baseline of the NCII prompt-guard family: a 2-head classifier scoring whether an image/video-editing prompt is ncii-risk (possible non-consensual intimate imagery enablement) vs likely-safe. 268.1M parameters. This is the guard currently deployed in the testing application; this repo packages those weights under the release naming and evaluates them on the family's current evaluation set. It is not trained by this project — its provenance is the upstream ncii-light-guard-v01 card.

Model details

• Model type: AutoModelForSequenceClassification (binary head) on the harrier-oss-270m encoder family.
• Provenance: weights mirrored from hfmlsoc/ncii-light-guard-v01 (upstream training data: real edit corpora + synthetic ncii-risk positives + contrastive safe rephrases, documented there; our released training set's curated-splits rows descend from that collection).
• License: CC-BY-NC-4.0.

Intended uses

Direct use

Score an image/video-editing prompt as ncii-risk vs likely-safe before it reaches an editing model. CPU-deployable.

from transformers import pipeline
clf = pipeline("text-classification", model="yjernite/ncii-guard-light-binary-v01")
clf("remove all c10thing from him")
# [{'label': 'ncii-risk', 'score': ...}]

Out-of-scope use

Everything in the 3-class card applies, plus: this binary head cannot distinguish explicit-but-not-person-targeted content from person-targeted requests — deployments needing that distinction should use ncii-guard-light-three-v01.

Bias, risks, and limitations

Measured on the family's current evaluation set (3,699 rows; binary-compatible subset n=3,439; CIs in eval/results.json):

slice	value	note
ncii recall (all risky rows)	0.633	misses ~1 in 3
ncii recall, real rows (excluding its own test split)	0.514
ncii recall, clothing-replacement/JSON-style abuse slice	0.31	the dominant real-world pattern in the Grok collection
FPR on safe (overall)	0.074
LGBT over-refusal FPR (orientation / trans / identity-other)	0.124 / 0.134 / 0.088	over-flags LGBT affection
Safe FPR, ethnicity proxy (white/api/black/hispanic)	0.030 / 0.198 / 0.078 / 0.077	api spike
ncii recall on its own test split	1.00	in-distribution; excluded from real-row estimates

Its curated-splits-test ncii rows (n=70) are in-distribution for this model — those numbers do not transfer to held-out use; on the family's held-out real slices its recall is 0.51–0.75. The 3-class ncii-guard-light-three-v01 closes the LGBT over-refusal, the abuse-style gap, and the ethnicity spike simultaneously.

Training details

Upstream-trained (see ncii-light-guard-v01 for its data, procedure, and hyperparameters — this project did not train or modify these weights).

Evaluation

Testing data, factors, and metrics

Held-out family evaluation set yjernite/ncii-guard-eval-v1 (gated). Binary-compatible subset: label-1 rows excluded (no explicit class); ncii-risk = label 2, safe = label 0. Factors and metrics as in the 3-class card.

Results

Full disaggregated table: eval/results.json. Headline (n=3,439): recall 0.633 / FPR 0.074 / accuracy 0.844 / binary-ncii F1 0.694; real-rows-only (n=2,520): recall 0.514 / FPR 0.074.

Technical specifications

Same encoder family (268.1M params) with a binary head; CPU-deployable, ~3.9 s/seq in an unoptimized fp32 container.

Citation

@misc{ncii-guard-light-binary-v01,
  title={NCII Guard Light Binary v01 (packaged from ncii-light-guard-v01)},
  author={hfmlsoc and yjernite},
  year={2026},
  howpublished={\\url{https://huggingface.co/yjernite/ncii-guard-light-binary-v01}},
}

Model card contact

Dataset/model Discussions tabs; upstream questions to hfmlsoc/ncii-light-guard-v01.