Nemotron 3.5: Hybrid Content Safety Classification Dataset

Nemotron 3.5 Content Safety Dataset is a hybrid real and synthetic supervised instruction dataset for content-safety classification. It contains text-only and image-grounded single-turn conversations, with examples asking a classifier to determine user safety, response safety, and harmful categories. The dataset was authored by NVIDIA and was last updated on the Hugging Face platform in June 2026.

Use Cases

• Training classifiers to assess user safety based on conversational text inputs.
• Training classifiers to assess response safety based on assistant text outputs.
• Identifying harmful content categories within human and AI interactions.
• Evaluating topic-following classification performance in a subset of the data.

Strengths

• Dataset is designed for supervised instruction tuning, indicating a structured training objective.
• Includes both text-only and image-grounded conversations, providing multimodal scope.
• Covers multiple classification tasks: user safety, response safety, and harmful categories.

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

Hybrid of real and synthetically generated supervised instruction data.

Freshness

Last updated 2026-06-04 01:41:37; freshness should be verified.