ELYZA Co., Ltd.
ELYZA develops and releases Japanese LLM with performance superior to “GPT-4” Developed based on open model, 70B model exceeds “GPT-4”, lightweight 8B model will be released to the public in commercially usable format ……
ELYZA Co., Ltd. (Representative Director: Yuya Soneoka, hereinafter referred to as ELYZA), which promotes social implementation of large-scale language models, has introduced Meta’s “Llama” as the latest model in the “ELYZA LLM for JP” series of large-scale language models provided by ELYZA. We developed “Llama-3-ELYZA-JP-70B” with 70 billion parameters and “Llama-3-ELYZA-JP-8B” with 8 billion parameters based on “3” and released their performance.
[Image 2: https://prtimes.jp/i/47565/46/resize/d47565-46-8c78335a1966ddb1f966-0.png&s3=47565-46-4bbaf98264986cedfb6003d86629e454-1200×630.png ]
“Llama-3-ELYZA-JP-70B” has a chat-style demo site.
ELYZA LLM for JP | Demo version (https://elyza.ai/lp/elyza-llm-for-jp) *Due to excessive access, there may be a waiting time until your request is processed.
The model of “Llama-3-ELYZA-JP-8B” will be released to the public. It is compliant with the LLAMA 3 COMMUNITY LICENSE and may be used for research and commercial purposes as long as it follows the Acceptable Use Policy.
Llama-3-ELYZA-JP
(https://huggingface.co/collections/elyza/llama-3-elyza-jp-667a311d51c8952e07778ecc) The more detailed status of the development process and evaluation is explained in the note article. Also check this.
https://note.com/elyza/n/n360b6084fdbd
By publishing and providing these models, ELYZA will support the promotion of social implementation of LLM in Japan and the development of LLM research and development.
■News summary
The latest model of the “ELYZA LLM for JP” series “Llama-3-ELYZA-JP” has been released.
The 70B model achieves performance that exceeds commercial global models such as “GPT-4” and “Claude 3 Sonnet.” A chat-style demo is also available. It is planned to be offered to companies as a high-performance domestically manufactured model.
The 8B model achieves performance that exceeds that of the “GPT-3.5 turbo” and achieves the highest level of performance among open models. Available to the public in a format available for research and commercial use.
ELYZA will promote the social implementation of LLM in Japan and support the development of LLM research and development by publishing and providing models.
■Overview of “Llama-3-ELYZA-JP-70B”
[Image 2: https://prtimes.jp/i/47565/46/resize/d47565-46-8c78335a1966ddb1f966-0.png&s3=47565-46-4bbaf98264986cedfb6003d86629e454-1200×630.png ]
“Llama-3-ELYZA-JP-70B” is a model developed based on Meta’s
“Llama-3-70B” by conducting additional learning (additional Japanese pre-learning and instructional learning).
This model has significantly improved Japanese language performance compared to the original “Llama-3-70B”, and in an automatic evaluation using two benchmarks (*1) to measure Japanese language performance, it was found to be superior to the domestic model. Among them, it has achieved the highest level of performance, achieving performance equivalent to or exceeding that of “GPT-4”, “Claude 3 Sonnet”, and “Gemini 1.5 Flash” (*2).
Figure 1: Performance comparison of “Llama-3-ELYZA-JP-70B” (vs major global model)
[Image 3: https://prtimes.jp/i/47565/46/resize/d47565-46-956f394e637397e5ec67-0.jpg&s3=47565-46-d3fe8ffaa2b2154694d94a09dff06347-1200×744.jpg] “Llama-3-ELYZA-JP-70B” will be gradually provided to companies in various forms through secure API services and joint development projects. We also have a chat-style demo site.
ELYZA LLM for JP | Demo version (https://elyza.ai/lp/elyza-llm-for-jp) *Due to excessive access, there may be a waiting time until your request is processed.
■Overview of “Llama-3-ELYZA-JP-8B”
[Image 4: https://prtimes.jp/i/47565/46/resize/d47565-46-b682368231582d9ca0ff-0.png&s3=47565-46-4123f2535b9924e8a7e088443309e7d1-1200×630.png ]
“Llama-3-ELYZA-JP-8B” is a model developed based on Meta’s
“Llama-3-8B” by conducting post-learning (additional Japanese pre-learning and instructional learning).
Similar to “Llama-3-ELYZA-JP-70B”, this model has greatly improved Japanese language performance from the original “Llama-3-8B”, and is a lightweight model with 8 billion parameters. In an automatic evaluation using two benchmarks (*1) to measure the performance of I am.
Figure 2: Performance comparison of “Llama-3-ELYZA-JP-8B” (vs. major lightweight global models)
[Image 5: https://prtimes.jp/i/47565/46/resize/d47565-46-e980fbcac7f363d228c1-0.jpg&s3=47565-46-d2fa09c9d27cb8c72f0e4156c7ebdd4a-1200×744.jpg] ELYZA will release the model “Llama-3-ELYZA-JP-8B” to the public. This model is compliant with the LLAMA 3 COMMUNITY LICENSE and may be used for research and commercial purposes as long as it follows the Acceptable Use Policy.
Llama-3-ELYZA-JP
(https://huggingface.co/collections/elyza/llama-3-elyza-jp-667a311d51c8952e07778ecc) The more detailed status of the development process and evaluation is explained in the note article. Also check this.
https://note.com/elyza/n/n360b6084fdbd
■About the “ELYZA LLM for JP” series
[Image 6: https://prtimes.jp/i/47565/46/resize/d47565-46-9ad5464dc72109e503c6-4.png&s3=47565-46-eabab8eb3bf3553ebc9fb57c4775ecc3-1274×636.png ]
“ELYZA LLM for JP” is a general term for the large-scale language model group provided by ELYZA. As a new option other than the global model, we are gradually providing various forms such as secure API services and joint development projects, mainly to companies that emphasize security and customizability, and companies that want to incorporate LLM into their services and businesses. Let’s start. The “Llama-3-ELYZA-JP-8B” and “Llama-3-ELYZA-JP-70B” released this time are the latest models of the “ELYZA LLM for JP” series. ELYZA LLM for JP | Demo version (https://elyza.ai/lp/elyza-llm-for-jp) *Due to excessive access, there may be a waiting time until your request is processed.
■Additional notes and future prospects
ELYZA has continued to work on Japaneseizing the “Llama” series since “ELYZA-japanese-Llama-2-7b”, and this time “Llama-3-ELYZA-JP” includes “GPT-4” and “Claude 3 Sonnet”. We were able to achieve performance that exceeds that of
Furthermore, we will continue to invest not only in the “Llama” series, but also in the Japanese translation of overseas open models and in the development of our own LLM. By publishing and providing these models, we will promote the social implementation of LLM in Japan and support the development of LLM research and development. *1 Two benchmarks:
1.ELYZA Tasks 100: A Japanese benchmark provided by ELYZA that aims to measure the ability to follow LLM instructions and return useful answers to users.
2. Japanese MT-Bench: A Japanese benchmark provided by Stability AI, which is a Japanese translation of the benchmark MT-Bench for evaluating the dialogue performance of LLM.
*2 “GPT-4” is “gpt-4 (0613)”, “Claude 3 Sonnet” is
“anthropic.claude-3-sonnet-20240229-v1:0”, “Gemini 1.5 Flash” is as of the end of June 2024 Refers to the latest stable version.
*3 “GPT-3.5 Turbo” is “GPT-3.5 Turbo (0125)”, “Claude 3 Haiku” is “anthropic.claude-3-haiku-20240307-v1:0”, “Gemini 1.0 Pro” is June 2024 Refers to the latest stable version as of the end of the month. ■ELYZA company profile
Based on the philosophy of “creating the norm in unexplored
territory,” ELYZA Co., Ltd. focuses on large-scale language models for the Japanese language, and conducts joint research with companies and develops cloud services. Through cutting-edge technology research and development and consulting, we promote the introduction and
implementation of large-scale language models in a way that
contributes to corporate growth.
-Company Profile-
[Image 7: https://prtimes.jp/i/47565/46/resize/d47565-46-8afba97ea0869f49a3d4-5.png&s3=47565-46-f9cf2b052b014134dcffe7d728ea1638-691×277.png ]
Company name: ELYZA Co., Ltd.
Address: SWT Building 6F, 3-15-9 Hongo, Bunkyo-ku, Tokyo 113-0033 Representative: Representative Director Yuya Soneoka
Established: September 2018
URL: https://elyza.ai/
More details about this release:
https://prtimes.jp/main/html/rd/p/000000046.000047565.html