This model was trained on the following commercially-licensed sources.
Only model weights are distributed; no source text is redistributed.

- nvidia/nemotron-pii  [CC-BY-4.0, yes-with-attribution]  55+ PII types, character-offset spans; synthetic text
- gretelai/synthetic_pii_finance_multilingual  [Apache-2.0, yes]  finance PII, char offsets, multilingual; synthetic text
- Faker  [MIT, yes]  synthetic PII generator (software MIT; output unrestricted)
- microsoft/presidio-research  [MIT, yes]  synthetic generator code; drive from Faker, NOT the bundled CC-BY-SA identity CSV
- google/civil_comments  [CC0-1.0, yes]  toxicity + sexual_explicit + identity_attack
- google/jigsaw_toxicity_pred  [CC0-1.0 (annotations), yes]  Jigsaw 6 labels; underlying text CC-BY-SA-3.0 (we ship weights only)
- ucberkeley-dlab/measuring-hate-speech  [CC-BY-4.0, yes-with-attribution]  violence / dehumanize / identity dimensions
- allenai/real-toxicity-prompts  [Apache-2.0, yes-with-attribution]  coarse sexual_explicit / threat signal
- redasers/difraud  [MIT, yes]  spam/scam/phishing (7 domains)
- ucirvine/sms_spam  [CC-BY-4.0, yes-with-attribution]  SMS spam (cite Almeida DOCENG'11)
- cyberec/llm-prompt-injection-attacks  [Apache-2.0, yes]  jailbreak multi-label (5 classes)
- deepset/prompt-injections  [Apache-2.0, yes]  prompt-injection binary
- jackhhao/jailbreak-classification  [Apache-2.0, yes]  jailbreak/benign
- Lakera/gandalf_ignore_instructions  [MIT, yes]  real injection attempts (PII-filtered)
- google/civil_comments  [CC0-1.0, yes]  nsfw sexual_explicit signal (shared with toxicity)
- mmathys/openai-moderation-api-evaluation  [MIT, yes]  nsfw eval only; minors class (S3) never read

Attribution required for: nvidia/nemotron-pii, ucberkeley-dlab/measuring-hate-speech, allenai/real-toxicity-prompts, ucirvine/sms_spam