SafeVid-350K: 350,000 Video Safety Preference Pairs

SafeVid-350K is a large-scale dataset containing 350,000 preference pairs designed to instill Helpful, Honest, Harmless principles in Video Large Multimodal Models. It covers 30 scene categories and 29 fine-grained safety sub-dimensions. The dataset was created by yxwang and was last updated on Hugging Face in November 2025.

Use Cases

• Training safety classifiers based on the 29 fine-grained safety sub-dimensions.
• Fine-tuning VLMMs for harmless video responses using the 350,000 chosen vs. rejected preference pairs.
• Evaluating model performance across the 30 distinct scene categories mentioned in the description.
• Aligning multimodal AI systems with 3H (Helpful, Honest, Harmless) principles using video-centric data.

Strengths

• Contains 350,000 preference pairs, providing a substantial volume of training data.
• Covers 30 scene categories and 29 fine-grained safety sub-dimensions, suggesting detailed annotation.

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

yxwang on Hugging Face.

Collection Method

Likely curated from YouTube video segments via YoutubeID and timestamps.

Freshness

Last updated 2025-11-27 11:30:31; freshness should be verified.