lightblue - Alibaba Cloud

Technology

Japan

SME/Startup

"Regarding language capabilities, we found that Alibaba Cloud’s Tongyi Qianwen (Qwen) not only performed well in English, but also proved to be the best publicly available option for supporting Japanese. We chose Qwen because our LLM’s accuracy significantly improved when fine-tuned with a base model capable of understanding Japanese."

Shunichi Taniguchi

Director | Senior Researcher, Lightblue Co., Ltd.

About
Lightblue

Lightblue Co., Ltd. is a startup dedicated to democratizing AI. It has launched LLab, a specialized team focused on the research and development of generative AI and large language models (LLMs), prioritizing safety and transparency. Lightblue's mission is to broaden the applications of AI technology and drive transformative, positive change in society.

Challenges

Following the release of ChatGPT, the popularity of generative AI surged. Despite the growth, no domestically produced LLMs existed in Japan, leading Lightblue to initiate its own LLM development.

Lightblue introduced its LLMs, Karasu and Qarasu, which stood out due to the rigorous efforts of engineers who refine and optimize data sets daily, ensuring exceptional performance in Japanese. However, the major challenge was finding the most suitable base LLM specifically for training in the Japanese language. Lightblue wanted a technology partner capable of addressing this challenge and providing the necessary expertise and support.

Why Alibaba Cloud

The support of Alibaba Cloud’s Tongyi Qianwen (Qwen) to Lightblue was crucial for the release of its Karasu and Qarasu models around December 2023. Lightblue explored prominent models like LLama2 and Mistral Large to assess the best foundation for development. While both showed strong performance in English, Qwen emerged as the best choice for handling Japanese. Its advanced architecture and extensive training in East Asian languages provided outstanding accuracy when fine-tuned for Japanese, ensuring clear and relevant interactions. Alibaba Cloud’s Qwen was crucial in giving Lightblue the capabilities needed to succeed in Japanese language processing.

Architecture

Lightblue primarily utilized the open-source edition of Tongyi Qianwen (Qwen), Alibaba Cloud's advanced foundation LLM, known for its proficiency in complex natural language processing tasks and strong multilingual capabilities.

In addition, Lightblue employed several other Alibaba Cloud solutions. The Elastic Compute Service (ECS) was used to offer scalable, on-demand computing resources for rapid deployment of virtual servers, featuring various instance types and storage options for optimal performance. Alibaba Cloud's Server Load Balancer (SLB) was also used to distribute incoming traffic across multiple servers, ensuring the high availability and reliability of the application.

Lightblue also utilized Alibaba Cloud Object Storage Service (OSS), which provides scalable and secure storage for large amounts of unstructured data, ensuring reliable data management.

Key Results

Qwen offers a range of model sizes, from lightweight to large, which proved highly convenient during development. While a 72b model is more accurate for achieving high scores, a 7b model is easier to manage and comes with various parameter sizes, including output speed, making it user-friendly from a development standpoint.

As Qwen was released on Hugging Face, it was in a state that allowed for seamless learning, enabling Lightblue to develop its LLM without any obstacles.

Looking Forward

Lightblue and Alibaba Cloud anticipate collaborating on more projects in the near future. Lightblue aims to enhance the accuracy of its SaaS service, Lightblue Assistant, a RAG-based solution, to eventually replace the parts that currently rely on APIs provided by other vendors. Additionally, Lightblue is considering developing a user-friendly LLM model tailored for customers who need to run it in a local environment.

A Free Trial That Lets You Build Big!

Sales Support

Technical Support

Connect & Report Abuse