Meerkat-Safe: Implicit Cross-Modal Risk Dataset for MLLMs

The first training dataset for implicit cross-modal risks in Multimodal LLMs, introduced in the ICML 2026 paper 'Meerkat-VL: Implicit Risk Safety Alignment in MLLMs via Perceptual Reasoning and Self-Verification'. It targets implicit risks by pairing benign images with potentially harmful text, contrasting with existing datasets focused on explicit risks. The dataset was uploaded by Tunanzzz on June 16, 2026.

Use Cases

• Training MLLMs for safety alignment based on the concept of implicit cross-modal risks.
• Evaluating model robustness against perceptual reasoning challenges described in the paper.
• Benchmarking self-verification capabilities of MLLMs using the paired benign image and text samples.

Strengths

• First dataset specifically for implicit cross-modal risks in MLLMs.
• Introduced in a peer-reviewed ICML 2026 conference paper.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Sample data and file formats are unavailable for preview.

Provenance

Source

huggingface

Freshness

Last updated 2026-06-16 04:58:20; freshness should be verified.