Note Article published to situate the NCII Prompt Guard model and dataset release within the broader context of Non-Consensual Iintimate Imagery (NCII) harms, risk governance, and mitigations; and help outline the scope and limitations of text-based guardrails.

hfmlsoc/ncii-prompt-guard-3c

Text Classification • 0.3B • Updated 2 days ago • 60

Note 3-class classifier: 0 for low risk, 1 for less clothing without nudity, 2 for block

hfmlsoc/ncii-prompt-guard-5c

Text Classification • 0.3B • Updated 2 days ago • 41

Note 5-class classifier: 0 for safe, 1 for low risk ambiguous, 2 for less clothing without nudity, 3 for nudification, 4 for explicit sexualization

Note Testing space to evaluate the behavior of the different models at different filter thresholds

Note Training, validation, and test data for the NCII prompt guard family. Access gated by individual requests

Note Embedding-based dataset visualization for a subset of the training data and the full dev and test dataset

hfmlsoc/ncii-light-guard-v01

Text Classification • 0.3B • Updated Aug 6 • 8.23k • 2

Note Previous binary model, training data and disparate performance not available, provided as baseline only