AI inside Launches “Customized SLM” for “PolySphere-2,” the LLM Specialized for Japanese Document Processing, to Learn Company-Owned Data

Realizes Utilization of Business-Focused Advanced Generative AI

Tokyo, August 20, 2024 –  AI inside Inc., a provider of an AI platform, announces the launch of the installation and operation service of “Customized SLM (Small Language Model)” (hereinafter referred to as the “Service”) for the enterprise market.

The Service is based on AI inside’s original “PolySphere-2,” a 47.3 billion-parameter LLM specialized for Japanese document processing, with “Customized SLM” installed by training and fine-tuning of company-owned data. It provides an AI infrastructure that enables the use of generative AI specific to the organization and operations of individual companies.

It is also provided in an on-premise environment utilizing “AI inside Cube Pro,” an edge computer developed by AI inside for optimal AI operations. For companies and organizations that cannot take data outside the company for information security reasons, such as local governments, financial, medical, and manufacturing industries, or for operations that handle confidential or sensitive information, such as personal data, we can provide highly specialized AI applications respecting confidentiality and privacy.

[For inquiries]
This Service is quoted on an individual basis. Please contact us for more information.
Contact form: https://inside.ai/en/contact/contact-form/

Provided Values

1. High-performance output

“PolySphere-2,” an LLM originally developed by AI inside that specializes in Japanese document processing, performs training and fine-tuning on a closed data set that includes specific domain knowledge and terminology owned by the company, enabling SLM that demonstrates high performance for particular tasks to be built. This allows the construction of an SLM that performs well on specific tasks.

“PolySphere-2” evolved from AI inside’s original LLM “PolySphere-1” launched in June 2023 and further improved its performance by being trained with the data that we hold. The results of a performance test*1 of “PolySphere-2” against other LLMs showed that it outperformed the latest LLMs by up to 30% in terms of accuracy in structuring unstructured data. “PolySphere-2” also technically addresses the issue of hallucination, a challenge unique to generative AI, and achieves a hallucination reduction score of 0.25% in the hallucination occurrence rate, significantly better than other companies’ LLMs.

*1 Based on AI inside’s evaluation method comparing the OCR function of semi-structured forms in the AI-OCR service “DX Suite.” This score does not guarantee performance.

2. Secure and flexible AI infrastructure

The “AI inside Cube Pro” edge computer enables its training and inference within its own environment, allowing users to securely use generative AI and SLM in an on-premise environment. “AI inside Cube Pro” can be used immediately on a subscription model, with no complicated configuration required. The infrastructure is scalable, enabling operation in the optimal system configuration according to increased or decreased operational volume and usage.

Global trends and background of the launch

With the rise of generated AI services such as ChatGPT, anyone can now easily apply innovative technologies. As global Big Tech companies successively release high-performance LLMs, there emerges a view that SLM based on training on closed data will become a new trend in the AI industry, especially in business use, where customizability is required to respond to specific operations.

Meanwhile, as generative AI makes rapid progress and models grow enormous, demand is increasing in all industries, and the resulting increase in power consumption is also a global challenge. Technological innovations to improve power efficiency are required to build a sustainable AI society.

Given this global trend, AI inside has launched this Service to promote the use of generative AI in more specialized and mission-critical tasks. We leverage SLM, a lightweight model that requires less resources and power consumption than LLM, to train and fine-tune closed data containing knowledge from a specific domain.

Comment by Representative Director, President & CEO, Taku Toguchi

AI inside has developed the hardware “AI inside Cube” and software “PolySphere-2” in-house and integrated them seamlessly to achieve unprecedented performance and robust data security. A particularly symbolic achievement is that we have successfully solved the hallucination issue of generative AI, which is considered an unsolved issue worldwide, even though it is for specific applications. These synergies give “PolySphere-2” and “Customized SLM” an overwhelming power. The era of SLM has now arrived, and its impact will fundamentally transform the companies’ future. We are truly convinced of the possibilities this technology can unleash.

[Performance comparisons: Accuracy in structuring unstructured data]

[Performance comparisons: Hallucination occurrence rate]

About “PolySphere-2”

“PolySphere-2” is a 47.3 billion parameter LLM specialized for processing Japanese documents. It was configured by “X Research,” AI inside’s research and development team, which conducts R&D and social implementation of generative AI and LLM. “PolySphere-2” evolved from “PolySphere-1” through repeated training utilizing data held by AI inside.

About “AI inside Cube Pro”

“AI inside Cube Pro” is an edge computer that optimizes any AI operation, with well-designed hardware and AI inside’s proprietary AI management OS and software. Subscription allows for prompt implementation.
https://inside.ai/en/service/ai-inside-cube-series/

About AI inside Inc.

AI inside is a tech company that conducts R&D and social implementation of cutting-edge technologies, such as generative AI, LLM, and autonomous AI. We provide the AI agent “Heylix,” the AI-OCR service “DX Suite,” and the AI infrastructures “AnyData” and “AI inside Cube” to a wide range of government agencies, local governments, and private companies. Our AI services have been used more than 5.5 billion times by over 50,000 users. Through these initiatives, we will promote the collaboration between humans and AI, and realize the “VALUE SHIFT,” by shifting the time saved through improved productivity and operational efficiency to higher-value-added, advanced operations.
https://inside.ai/en/

*The service names appearing on this site are trademarks or registered trademarks of AI inside.