#Japan #Tokyo #Weights & Biases #NTT DOCOMO #chakoshi

Partnership for Safer AI Experiences

Weights & Biases Japan (W&B Japan) has initiated a noteworthy collaboration with NTT Docomo Business, integrating their LLM guardrail technology, known as chakoshi, into the W&B Weave platform. This innovative partnership aims at enhancing the safety and reliability of AI applications tailored specifically for the Japanese language.

Overview of the Integration

The integration will allow users of W&B Weave to seamlessly incorporate moderation functionalities that are specially designed for Japanese. With chakoshi, developers will have access to high-speed and reliable guardrail logic, enabling them to effectively monitor and manage input and output processes within AI applications. This addresses a critical need in the generation AI landscape: ensuring safety alongside performance improvements.

By utilizing the advanced observability capabilities provided by W&B Weave, developers can now add robust safeguards that will significantly mitigate risks associated with inappropriate content during the development and deployment phases of AI applications. This move is crucial as it highlights the increasing importance of safety measures within the rapidly evolving AI sector.

Importance of Guardrails in Generative AI

As generative AI technology becomes increasingly prevalent, ensuring safe deployment has emerged as a fundamental challenge. The nature of generative AI applications necessitates the need for intervention in the input and output processes. Effective moderation can automatically detect and address harmful content, fostering a safer environment for users and developers alike.

Previously, Japanese developers faced challenges implementing tailored detection logic for their applications, often having to rely on their initiatives. However, the integration of chakoshi into W&B Weave simplifies this process, allowing a broader audience to benefit from these advanced safety features without needing extensive technical expertise.

Understanding chakoshi

The chakoshi guardrails are designed to protect businesses from threats posed by both input and output data. The risks include input attempts aimed at exploiting vulnerabilities in the LLM services and the potential dissemination of inappropriate or sensitive output information. By linking with customers' LLM applications via APIs, chakoshi effectively detects and blocks incidents, significantly reducing the overall risk.

Featured Quotes from Stakeholders

NTT Docomo Business

Akiko Fukuda, Executive Officer at NTT Docomo Business, emphasized the significance of safety in the emerging era of AI agents. She stated, "The need for AI to function autonomously and safely is becoming a primary focus. Through our partnership with W&B Japan, we aim to enhance safety in Japanese input/output processes, facilitating the wider adoption of generative AI in the Japanese market."

Weights & Biases Japan

Akira Shibata, Country Manager of W&B Japan, highlighted the recognition of guardrails in promoting generative AI applications. He remarked, "The realization of a partnership that allows for the seamless integration of high-level guardrails tailored for Japanese is set to accelerate AI adoption in Japan and bring substantial impacts."

Benchmarking Results

As part of this integration, we conducted a comparison of generative AI model performances, utilizing the popular Nejumi leaderboard for Japanese evaluations. We assessed both GPT-5 and OpenAI's OSS-120B models. Notably, significant disparities in safety-related evaluations became evident.

By implementing guardrails through W&B Weave and chakoshi, we observed remarkable improvements in both veracity and toxicity metrics for OSS-120B. This finding underscores the effectiveness of combining guardrails with larger models, enhancing reliability and addressing concerns prevalent during real-world deployments.

Upcoming Events

NTT Docomo Business will be a sponsor at the Fully Connected Tokyo 2025, a prominent user summit held on October 30-31. Attendees will have the opportunity to experience chakoshi firsthand at our booth in Toda Hall & Conference Tokyo. This summit represents one of Japan’s largest AI engineer gatherings, and we encourage everyone to participate in this free event.

Register Here

About Weights & Biases Japan

Weights & Biases Japan is committed to providing an enterprise-grade ML experiment management and end-to-end MLOps platform for developers. Our solutions, trusted by over 500,000 machine learning developers globally, embrace a range of deep learning use cases including LLM development and drug discovery.

For more information, visit our Japanese website.