KSAFE-MM · Datasets at Hugging Face

KSAFE-MM is a Korean-language benchmark for evaluating safety risks in multimodal large language models (MLLMs), which process both images and text. Most existing safety benchmarks are in English and reflect Western contexts, so they may miss risks specific to other languages and cultures. KSAFE-MM addresses this gap by testing how models respond to harmful or sensitive requests set in a Korean cultural context. It was released in June 2026 for academic research on AI safety, multimodal model evaluation and content moderation.

The benchmark contains 14,135 query-image pairs across 11 safety risk categories, grouped into three domains. Content safety risks cover hate and unfairness, violence, sexual content and self-harm. Socio-economic risks cover political and religious neutrality, anthropomorphism and sensitive uses. Legal and rights-related risks cover privacy, illegal or unethical activity, copyright and weaponisation.

The dataset has two subsets. KSAFE-MM-G (1,650 entries) is a general safety subset. It presents each harmful query as an image, as text rendered in an image (typography), or both. KSAFE-MM-C (12,485 entries) uses synthetic images to test Korean culture-specific risks. Its queries also apply ten jailbreak strategies, such as role-play, research framing and translation, to test whether models can be manipulated into unsafe responses. Researchers run models on the queries and assess whether responses are safe. Access requires agreeing to the dataset's terms. It is released under a CC BY-NC 4.0 licence.

Auto-discovered on 2026-07-01 by OECD Catalogue Automation