Weblab-10B

The Matsuo Lab's first large open Japanese LLM and the lab's from-scratch legacy anchor: a 10B-parameter GPT-NeoX-architecture bilingual model pretrained on ~600B tokens of Japanese C4 plus The Pile, released with an instruction-SFT variant under CC-BY-NC-4.0. At release (August 2023) among the largest LLMs pretrained in Japan; direct ancestor of the GENIAC-program Tanuki models.