Nemotron Content Safety V2: 33,416 Annotated Human-LLM Interactions

NVIDIA's Nemotron Content Safety Dataset V2 contains 33,416 annotated interactions between humans and LLMs, released in June 2025. It provides structured training, validation, and test splits curated from human preference data to support safety alignment and toxicity detection.

Use Cases

• Training text classification models to identify unsafe content using the harmlessness labels
• Benchmarking LLM safety performance using the 1,964 test samples
• Fine-tuning safety-aligned models using the 30,007 training interactions

Strengths

• 33,416 annotated interactions
• Pre-split into training, validation, and test sets
• NVIDIA-curated human preference data

Limitations

• Specific column names are not provided in the source metadata
• Dataset focus is limited to harmlessness, which may not cover all moderation categories like misinformation

Provenance

Source

NVIDIA

Collection Method

Curated from HuggingFace human preference data regarding harmlessness

Freshness

Updated June 2025