NVIDIA AI Safety Report
Nemotron Content Safety moderation evidence.
Safety grade
?—
Not graded
Guardrail score
88.6%
1 configuration tested
Guardrail Moderation
1 configuration tested on 250 harmful and 313 benign generated texts · Antisemitism. The score is balanced accuracy: harmful texts caught and benign texts allowed, weighted equally. Not part of the safety grade.
Nemotron Content Safety (self-hosted)
BYOP policy · Self-hosted
88.6%
- Harmful missed
- 5/250 (2.00%)
- Benign flagged
- 65/313 (20.77%)
Configuration details
- Recall / precision
- 98.00% / 79.03%
- Model / version
- nvidia/Nemotron-Content-Safety-Reasoning-4B · runtime-00039
- Policy version
- baseline · recorded threshold, not tuned
- Weights
- Public; model-specific license
- Parameters (B)
- total not disclosed · active not disclosed
- Evidence
- frozen Sep 9, 2026 · released Sep 16, 2026
- Data
- Configuration · Anonymized decisions
Cohort sensitivity
Verified-termination representatives are reselected, so that sensitivity cohort is not necessarily a subset of the primary cohort. Scores describe generated English texts, judged unanimously by three AI models.