Nemotron Content Safety Audio: 1,928 Adversarial Prompts for Multimodal AI Safety

1,928 audio files generated from the test set of the Nemotron Content Safety Dataset V2. This multimodal extension provides spoken versions of adversarial and safety-critical prompts across 23 violation categories, enabling research in AI safety. The dataset was created by NVIDIA and was last updated on the platform in December 2025.

Use Cases

• Training audio-based content moderation models based on adversarial prompts.
• Benchmarking multimodal AI safety systems on spoken safety-critical inputs.
• Researching the transfer of text-based safety violations to the audio domain.

Strengths

• Contains 1,928 audio files for multimodal research.
• Covers prompts across 23 distinct violation categories.
• Extends a known safety dataset (Nemotron Content Safety Dataset V2/Aegis 2.0).

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

NVIDIA

Collection Method

Audio files generated from the test set prompts of the Nemotron Content Safety Dataset V2.

Freshness

Last updated 2025-12-01 19:29:23; freshness should be verified.