Falcon: A Large-Scale Multimodal Safety Benchmark for AI Models

Falcon is a large-scale dataset introduced to address the scarcity of multimodal safety evaluation benchmarks. It contains 71,000 multimodal samples across 13 harmful categories, including illegal activities and adversarial jailbreaking prompts. The dataset was created by author zhangrjjj and was last updated on Hugging Face in May 2026.

Use Cases

• Safety fine-tuning of multimodal models based on the 69.3k-sample training set.
• Quantifying risks in MLLMs based on the dataset's 13 defined harmful categories.
• Evaluating model robustness against adversarial jailbreaking prompts.
• Studying cross-modal disinformation generation and detection.

Strengths

• Contains 71,000 multimodal samples, providing a substantial scale for evaluation.
• Covers 13 distinct harmful categories, including illegal activities and adversarial prompts.
• Includes a dedicated training set of 69.3k samples specifically curated for safety fine-tuning.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset's specific composition, file formats, and license are unknown from the provided metadata.
• Freshness should be verified as the last update timestamp is in the future (2026-05-03).

Provenance

Source

huggingface

Collection Method

Specifically curated for safety fine-tuning and risk quantification, as described.

Freshness

Last updated 2026-05-03 17:51:17