ELYZA Japanese Llama 3

About

The ELYZA Japanese Llama 3 family consists of large language models specifically fine-tuned for Japanese language processing. Derived from Meta's Llama 3 architecture, these models have been further enhanced by ELYZA, Inc. through additional pre-training and instruction tuning using high-quality Japanese corpora and proprietary datasets 15. This has led to marked improvements in Japanese language generation, with the 70B parameter model surpassing the performance of models such as GPT-4, Claude 3 Sonnet, and Gemini 1.5 Flash in specific benchmarks 5. The 8B parameter model, accessible on Hugging Face, offers performance on par with GPT-3.5 Turbo and Claude 3 Haiku 5. Licensed under the Llama 3 Community License, ELYZA's models accommodate both research and commercial applications 15. They are optimized for efficiency, utilizing techniques like Speculative Decoding to address inference speed challenges common with larger model sizes 5.