セキュアなプライベート環境で高度なカスタマイズを実現し生成AI利活用を促進

当社は、Cohere Inc.（注1）（以下 Cohere）と共同開発したプライベート環境で使用可能な企業向け大規模言語モデル（以下 LLM）「Takane」（高嶺：タカネ）をAIサービス「Fujitsu Kozuchi（Generative AI）」に組み込み、クロスインダストリーで社会課題を解決する事業モデル「Fujitsu Uvance」のオファリングであるオールインワンオペレーションプラットフォーム「Fujitsu Data Intelligence PaaS（以下 DI PaaS）」を通じて、2024年9月30日よりグローバルに提供します。

「Takane」は、日本語言語理解ベンチマークJGLUEにおいて世界最高記録（注2）を達成しており、セキュアなプライベート環境で利用できる点が特徴です。「Takane」を、大規模な文章の参照を可能にする「ナレッジグラフ拡張RAG」や、法規制や企業ルールに準拠した出力が可能な「生成AI監査技術」などから成る「エンタープライズ生成AIフレームワーク」とあわせて提供することで、業務に特化したLLMを実現し、お客様の業務変革を支援します。

当社は、「Takane」を「Fujitsu Kozuchi」の「Generative AI」のLLMのラインナップに追加することで、お客様に最適な生成AIを全方位で提供していきます。さらに、お客様が意識せずに生成AIによるビジネス変革の恩恵を受けられるというBeyond Chatという概念に基づき、コンサルティングサービス「Uvance Wayfinders」や「Fujitsu Uvance」のオファリングとともにトータルに提供することで、お客様ビジネスの生産性と創造性の拡張、および社会課題の解決に貢献していきます。

Background

世界ではマーケティングや顧客対応など様々な分野で生成AIの業務活用が進んでいます。しかし、クラウドサービスで提供される大規模な汎用LLMは、金融や安全保障分野など、外部にデータを出せない業務領域や、法規制や業界ルールへの準拠が求められる業務においては、活用が難しい状況です。さらに、日本語は文字種の混在や主語の省略、敬語表現など、特有の難しさがあり、汎用LLMの出力精度には課題があります。そのため、言葉の間違いが大きな問題や損害を及ぼす可能性のある行政や金融、医療、法曹などの分野では、日本語能力に優れた業務特化型のLLMが求められています。

Takaneの特長

1. 日本語能力

"Takane" is a model developed by our company based on Cohere's LLM "Command R+", combining our rich knowledge gained from developing Japanese-specialized LLM and Cohere's expertise in developing business-specific LLM, exclusively available globally by our company. We have conducted additional learning and fine-tuning for Japanese strengthening, achieving performance that surpasses other companies in semantic understanding and syntactic analysis in the Nejumi LLM Leaderboard 3, a benchmark for Japanese language understanding, JGLUE, JNLI for natural language inference, and JSQuAD for machine reading comprehension tasks (note 3). In addition, it inherits the multilingual support of "Command R+" (supporting 10 languages) and the functionality to automate business processes.

2. A secure private environment

"Takane" is a LLM available in a secure private environment. This allows customers in the financial industry handling personal information, manufacturing dealing with sensitive data like design and development, and security sectors handling highly confidential information to confidently utilize LLM even in operations where concerns about data leakage make LLM implementation challenging.

3. Specialized LLM through fine-tuning

By using a company's own data for fine-tuning and customization, it is possible to advance LLM customized for customer operations. Furthermore, by leveraging Cohere's RAG technology alongside our "Knowledge Graph Extended RAG Technology" and "AI Audit Generation Technology," compliance with regulations and industry/company rules becomes easier. This ensures that even in industries like financial services with many specialized terms and frequent regulatory amendments, "Takane" can be utilized with peace of mind.

Furthermore, by offering 'Takane' through 'DI PaaS' of 'Fujitsu Uvance', we will create business applications integrating data and AI. 'DI PaaS' integrates vast amounts of data scattered inside and outside organizations into a form that can be understood, collaborates and analyzes fragmented data across departments and industries, deriving unprecedented insights and solutions, and activates data utilization between organizations and companies through a cloud-based all-in-one operation platform. We will support your business transformation with overwhelming productivity and creativity through the fusion of data and AI.

Comments from our Vice President and CTO, CPO Viviek Mahajan

As part of our AI strategy announced last year, we developed the enterprise-oriented LLM 'Takane'. We hope that this model suitable for private environments will contribute to AI utilization for customers in industries where secure environments are essential. We will continue to focus on AI technology development for enterprises, supporting the three growth drivers of modernization, 'Fujitsu Uvance', and consulting with technology.

Comments from our Vice President and COO, Minami Takahashi

The value proposition of 'Fujitsu Uvance' is to support the growth of customers' businesses and the resolution of social issues through advanced decision-making brought by Data×AI. With the addition of the new Japanese enhanced LLM 'Takane' to 'DI PaaS', we can now offer the most suitable LLM to a wider range of customers. We will continue to introduce the world's most advanced AI to the market by cooperating not only in proprietary development but also with various partners around the world, supporting customers' business transformation.

Comment from Cohere Inc. co-founder and CEO Aidan Gomez

We are looking forward to offering the Japanese-enhanced LLM 'Takane' to global companies. Through our partnership with Fujitsu, we will accelerate AI adoption in the critically important Japanese market by providing AI designed to include Japanese, among other languages, tailored for business use, safely and efficiently utilized with high performance.

Figure: Scheme for providing "Takane"

Comment from Kenbun Yamamoto, Group Deputy CIO, Executive Officer, Mizuho Financial Group

is conducting a joint demonstration experiment with Fujitsu in FY 2023 to utilize generated AI for system development and maintenance, and applying "Kozuchi" to a part of the system development design process from FY 2024 to work on improving quality.
The enterprise-focused LLM "Takane" announced this time is expected to make Mizuho's internal knowledge and know-how more effectively usable by using technologies such as the Knowledge Graph Expansion and RAG. We have great expectations. In the future, we will strive for further improvement in the quality and resilience of system development and maintenance processes.

Comment from Shoji Tanaka, Ph.D. (Engineering), Project Manager of AI Strategy Project at Mitsubishi Electric Corporation and Deputy Director of DX Innovation Center

"Takane", developed through the strategic partnership between Fujitsu and Cohere, is a Japanese-language compatible generated AI solution that combines the world's top-level technologies of both companies. We believe that the addition of such world-class solutions as new options will bring new possibilities to the Japanese companies' efforts in the generated AI market. We look forward to the future progress of both companies.

Comment from Yuki Shuto, Executive Officer, Chief Growth Officer (Head of Strategy, Innovation, and Alliance), Technology, Media, and Communications Asia Pacific Leader at Deloitte Tohmatsu Consulting LLC

Generated AI has completed the PoC phase in many companies and government agencies, and is now entering a transformation phase where it will be fully applied to products, services, and business processes. An important aspect at that stage is a secure environment where one can securely cooperate and utilize their own data, and the LLM, especially the Japanese-capable "Takane", holds the potential to accelerate the transformation of Japanese companies and government agencies. Deloitte Tohmatsu, together with Fujitsu, will unleash the potential of generated AI and contribute to the development of Japan's economy and society.

Comment from Takafumi Yano, President and CEO of RUTILEA, Inc.

Our company is committed to providing high-precision AI services with the mission of 'Make AI simple.' aiming to realize a society where AI is implemented in all business processes. Recently, in addition to vertical AI business expansion specialized in specific industries such as central ministries, local governments, the automobile industry, the pharmaceutical industry, and electrical utilities procurement optimization, we will start the AI development platform business in 2024 by establishing a GPU datacenter in Okuma-cho, Futaba-gun, Fukushima Prefecture. 'Takane' holds the potential to significantly promote the use of AI in various vertical businesses in Japan. We look forward to contributing to its development and utilization.

Trademarks

Proper nouns such as product names mentioned are trademarks or registered trademarks of their respective companies.

Notes

• Note 1
  Cohere Inc.:
  Headquarters: Toronto, Canada and San Francisco, California, USA; Co-founder and CEO: Aidan Gomez
• Note 2
  World record in the Japanese Language Understanding Benchmark, JGLUE:
  Surpassing the results of other companies' LLMs, including the large-scale general-purpose LLM provided in cloud services in JGLUE. Please note that for JNLI and JCoLA, there was uncertainty in the ground truth, so the correct data was revised and measured by multiple annotators as a reference value (measurement by Fujitsu and Cohere in September 2024).
• which utilizes generated AI, members who newly develop a manufacturing process for laundry detergent for clothing can easily search and utilize skilled workers' technology, knowledge, and know-how and, thus, support the efficient execution of the business task.
  Superior performance surpassing other companies:
  Achieving the highest performance with a score of 0.862 in the semantic understanding category and 0.773 in the syntactic analysis category in the Nejumi LLM Leaderboard 3 evaluating the Japanese language abilities of LLM models (measurement by Fujitsu and Cohere in September 2024).
• Note 4
  JGLUE:
  The Japanese version of GLUE (General Language Understanding Evaluation) by Kentaro Kurihara, Daisuke Kawahara, Tomohide Shibata, "JGLUE: Japanese Language Understanding Benchmark", 28th Annual Meeting of the Association for Natural Language Processing, 2022.
• Note 5
  JSTS:
  Task to estimate the semantic similarity of pairs of sentences
• Note 6
  JCoLA:
  Task to determine the acceptability of Japanese sentences
• Note 7
  JNLI:
  Recognizing the inferential relationship that the premise text holds towards the hypothesis text in a natural language inference task.
• Note 8
  JCommonsenseQA:
  Evaluation task of common sense reasoning ability.
• Note 9
  JSQuAD:
  Machine reading comprehension task to read a document and answer questions about it.

Related links

• Fujitsu and Cohere have established a strategic partnership to provide corporate AI solutions, initiating joint development (Press Release July 18, 2024).
• Automatically generate specialized AI tailored to corporate needs using world-first technology! Providing an Enterprise AI framework (Press Release - June 4, 2024)
• Fujitsu Uvance
• Fujitsu Data Intelligence PaaS.
• Fujitsu Kozuchi

Our contribution to SDGs

Sustainable Development Goals (SDGs), adopted by the United Nations in 2015, are common goals that should be achieved by the whole world by 2030. Our purpose, which is to bring trust to society through innovation and make the world more sustainable, promises to contribute to SDGs.

The main SDGs that this matter aims to contribute to

Inquiries regarding this matter: Fujitsu Contact Line (Comprehensive Window)

The price, specifications, service contents, etc. of the product described in the press release are as of the announcement date, and may be changed without notice. Thank you for your understanding.