PragmaSafe-Beavertails: Vision-Language Safety Evaluation Data in 19 Harm Categories

An augmented version of the PKU-Alignment/BeaverTails-V dataset, organized into 19 categories of harmful content. The dataset was created by SII-fleeeecer and last updated on March 31, 2026. It includes train and validation splits for categories such as animal_abuse, dangerous_behavior, and discriminatory_depictions.

Use Cases

• Training safety classifiers for multimodal models based on the 19 defined harm categories.
• Evaluating the robustness of vision-language models against deceptive or dangerous content.
• Benchmarking model performance on specific safety risks like environmental damage or discriminatory depictions.
• Fine-tuning models to refuse generating harmful image-text pairs based on the provided categories.

Strengths

• Organized into 19 distinct categories of harmful content, providing structured evaluation targets.
• Includes both train and validation splits, suggesting a design for model development and testing.
• Based on an established safety dataset (PKU-Alignment/BeaverTails-V), indicating a foundation in prior research.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the source platform and underlying dataset.

Provenance

Source

SII-fleeeecer on Hugging Face, based on PKU-Alignment/BeaverTails-V.

Collection Method

Augmented version of an existing dataset; specific augmentation methods are not detailed in the provided input.

Freshness

Last updated 2026-03-31 09:31:43; freshness should be verified.