MSTS-Japanese

Japanese localization of the Multimodal Safety Test Suite (MSTS) for vision-language models — safety prompts where the harm emerges only from the image+text combination. Translates 330 of the 400 original cases and augments 82 with Japan-local images and reference answers. Gated for safety research (answer-carefully terms); ANLP 2026 paper. Part of llm-jp's safety-evaluation line.