Built with ML Intern

ncii-edit-guard-270m-v3

Third iteration of the lightweight prompt guard for image- and video-editing applications: a 3-class classifier scoring whether an editing prompt is likely a nudification / non-consensual intimate imagery (NCII) request about a person in the input. 268.1M parameters, ~4.2 ms/sequence (batch 8, A10G, CUDA). Supersedes v2; a 4-class experimental variant lives at v3-4class.

What changed in v3:

  • Real NCII data: 357 train + 64 eval rows of real public NCII requests (TfGBV-Grok, CC-BY-NC-4.0, sanitized; meta-discussion and minor-mentioning rows excluded).
  • Taxonomy-conditioned generation (Qwen3-30B-A3B-Instruct-2507, chosen over Qwen3-8B in a refusal/diversity pilot): clothing-replacement and JSON-spec families (the dominant real-world pattern, previously nearly absent), benign-phrasing ncii, explicit-situation, video-I2V, deictic-subject prompts, with contrastive minimal pairs (orientation/gender/ethnicity-flip safe edits), demographic-balanced safe rows, and religion/skin-tone safe cells β€” explicitly teaching the model not to over-moderate LGBTQIA and demographic categories.
  • Asymmetric distillation retained from v2 (Shieldstral safe-side soft labels, Ξ±=0.3) + real-ncii rows Γ—2 upweight.
  • Style-disjoint synthetic eval slices so synthetic recall is not inflated by seed reuse; refusal-contaminated generations filtered.

Labels

  • safe (0) β€” edits that would not typically produce NCII on a user's photo/video
  • explicit-not-person-targeted (1) β€” sexual/nudity content without an identifiable person as target
  • ncii-risk (2) β€” nudification or sexualization of a person in the input (risk signal; consent is unknowable from text). Sublabels (nudify / explicit-situation / sexualize-reveal) are published on the dataset for analysis; the head is deliberately 3-class (see the 4-class variant's card for why).
from transformers import pipeline
clf = pipeline("text-classification", model="yjernite/ncii-edit-guard-270m-v3")
clf('{"action": "replace_clothes", "attire": "black bikini bra", "face_preserved": true}')
# [{'label': 'ncii-risk', 'score': ...}]

Evaluation β€” including the deployed baseline

All models scored on the same 3,699-row leakage-controlled eval (ncii-guard-eval-v1, gated). The deployed testing-app guard hfmlsoc/ncii-light-guard-v01 is binary β€” compared on the binary-compatible subset (n=3,439; its curated-splits-test rows are its own test split).

Metric light-guard-v01 (deployed) v2 v3
ncii recall (all) 0.633 0.725 0.924
real ncii recall (held-out: curated-splits-test n=70 / grok n=55) 1.00† / 0.745 0.743 / β€” 0.786 / 0.964
new-style abuse recall (clothing-replacement/JSON: synthetic-v3 slice) 0.31 β€” 0.93
i2p/T2I-style ncii recall 0.333 0.519 0.556
FPR on safe 0.074 0.042 0.043
explicit-class F1 β€” 0.63 0.65
macro F1 (3-class models) β€” 0.782 0.831

† its own in-distribution test split; excluding it, its real-row recall is 0.51.

Disaggregated performance (95% Wilson CIs in results.json; demographic columns are text-signal proxies β€” gender via gendered pronouns only, ethnicity via census name lists, LGBT via signifier regexes β€” not self-identification; gender slices use eval/results_gender_text.json)

Slice light-guard (FPR/recall) v3 (FPR / recall)
Gendered-pronoun safe prompts (she/her) FPR 0.047 FPR 0.063 [0.043–0.092]
Gendered-pronoun safe prompts (he/him) FPR 0.063 FPR 0.051 [0.031–0.082] β€” recall parity 0.953 vs 0.921
Ethnicity-proxy safe FPR (white/api/black/hispanic) 0.030 / 0.198 / 0.078 / 0.077 0.035 / 0.070 / 0.039 / 0.212 (hispanic n=52, CI 0.12–0.34)
LGBT over-refusal (orientation / trans / identity-other) 0.124 / 0.134 / 0.088 0.016 / 0.029 / 0.038
Same-sex vs different-sex partner ncii recall 0.72 / 0.74 0.93 / 0.95 β€” no gap
Video ncii recall 0.570 0.950

These are the clean-retrain numbers (trained on the released data composition). The gender FPR gap of v1 (3.3Γ—) is closed (recall parity 0.95 vs 0.92 on the pronoun signal); the api-ethnicity FPR spike of the deployed guard is absent; LGBT affection prompts are flagged at 1–3% (deployed guard: 9–13%). Two points moved on the clean run and are flagged honestly: a hispanic-proxy safe-FPR elevation (0.212, n=52, CIs overlapping the other groups' upper bounds) and a lower i2p/T2I-style recall (0.556).

Limitations

  • Text only; consent is unknowable from text; scores are heuristics, not proof of intent or harm.
  • Real-row ncii recall (~0.79 on splits-test, 0.81 on all real rows) means roughly 1 in 5 real nudification-style requests is still missed.
  • Demographic slices are proxies; the religion/skin-tone cells use crude signifier regexes.
  • Sublabels on eval rows were rule-annotated and are being replaced by judge annotations (see eval/gate_check.json and the 4-class card for the strict/reveal boundary analysis).
  • The Grok source is shaped by that service's guardrails (bikini-heavy); content-type weighting was triangulated across sources, not copied.
  • Not a legal or safety guarantee.

Data & training

Train: the 31,274-row training assembly of ncii-guard-train-v1 (29,230 train + 1,961 dev as of the released version; this run predates two later data repairs β€” the refusal-row filter and the grok split fix β€” see the Model provenance table). Training: 3 epochs, eff. batch 32, lr 2e-5, class weights [1.0, 1.8, 1.8], Ξ±=0.3 soft-CE on Shieldstral safe-side labels, real-ncii Γ—2, max len 512. Generation seeds and scripts in ncii-guard-synth-v1. Eval set is gated (manual approval) β€” request access for research/deployment evaluation.

Citation

Base: microsoft/harrier-oss-v1-270m. Teacher: mistralai/Shieldstral-1.0-3B. Instruments: I2P (arXiv:2211.05105), XSTest (arXiv:2308.01263), SNEACI taxonomy (arXiv:2606.27234). Predecessors: hfmlsoc/ncii-light-guard-v01, v1, v2.

Downloads last month
23
Safetensors
Model size
0.3B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for yjernite/ncii-edit-guard-270m-v3

Finetuned
(26)
this model

Papers for yjernite/ncii-edit-guard-270m-v3