#Japan #Tokyo #Qlean Dataset #Visual Bank #msts-japanese

Introduction

In a significant milestone for AI safety evaluation, Visual Bank Inc. has announced that its subsidiary, amana images, has rolled out the Qlean Dataset, now featured in the "msts-japanese" dataset. This initiative is the result of a collaboration with Japan’s National Institute of Informatics (NII) and marks a substantial leap in multimodal safety evaluation adapted to Japanese language and culture.

Background on AI Development in Japan

With the rapid advancement of AI technologies, particularly since the introduction of models like ChatGPT, most leading AI systems have evolved into vision-language models (VLM). However, assessing safety in these complex multimodal frameworks has posed unique challenges, especially in Japan. Historically, there has been a lack of evaluation datasets that focus on the risks associated with combining images and text, particularly under the context of Japanese culture.

The msts-japanese Dataset

The newly released msts-japanese dataset is a transformative step. Launched on July 7, 2026, on Hugging Face, this dataset includes 330 translated items from the MSTS framework, renowned for its multilingual capabilities. Out of these, 82 items feature specially curated Japanese-localized images that correspond to the context, in partnership with amana images.

• - Content Overview:

The dataset leverages images that are rights-cleared, ensuring that they are authentic and non-AI-generated. This is crucial as safety evaluations require images that genuinely reflect real-life scenarios.

• - Utilization Terms:

Registration is necessary, and the dataset is intended solely for the purpose of enhancing safety in AI systems, while also permitting commercial use, yet prohibiting redistribution.

• - Access and More Information:

Users can access the dataset here.

Importance of Cultural Context in AI Safety

NII's Center for Large Language Model Research has highlighted a critical insight: the nuances of language and culture can significantly impact AI models' safety outcomes. Their findings indicate that switching prompts from English to Japanese can escalate harmful response rates by two- to threefold. This underscores the urgency for a comprehensive Japanese-language evaluation framework that reliably assesses AI safety within its cultural context.

Tackling the Challenges of Safety Evaluation in Japan

The issues extend beyond mere language translation. The challenge lies in the inherent risks associated with multimodal safety—where both images and text may appear innocuous in isolation but can pose hazards when combined. Prior international safety test suites, such as MSTS, do not incorporate Japanese prompts or relevant cultural imagery, which necessitated the development of a localized dataset to ensure safe AI deployment in Japan.

Collaborating for a Secure Future

The partnership between Visual Bank and NII is a vital step toward addressing these safety evaluation deficiencies. Through the Qlean Dataset, amana images has secured a wide array of rights-cleared photographs that meet the specific needs for safety evaluations. Their ability to provide clear provenance of data allows organizations to utilize these images confidently without worrying about legal complexities.

• - Leading the Charge in AI Safety Improvements:

Hisami Suzuki from NII emphasized the necessity of utilizing authentic images for safety evaluations, stating that rights-cleared visuals are critical for building reliable datasets. The msts-japanese initiative sets a benchmark for future endeavors in ensuring AI safety across various domains in Japan.

Future Directions

Moving forward, NII plans to expand the scope of the multimodal safety evaluation dataset while continuing collaborations focused on improving AI safety. As part of this initiative, they will also participate in the esteemed MIRU2026 symposium from August 3-6, 2026, in Nagasaki, showcasing these advancements.

Conclusion

In summation, the collaboration between Visual Bank and NII signifies a momentous achievement in establishing the first Japanese multimodal AI safety evaluation dataset. The msts-japanese dataset marks a pivotal development not only in enhancing AI safety but also in understanding cultural nuances vital in AI applications. As the field of AI rapidly evolves, initiatives like this become essential pathways for ensuring its secure and effective deployment in society.