llm-jp-13B v2.0

llm-jp-13b-v2.0 is a 13-billion-parameter Japanese language model from LLM-jp, a cross-organizational research project launched in Japan for fully open Japanese LLMs. It was trained from scratch on 256 billion tokens blending Japanese Common Crawl, Japanese and English Wikipedia, the Pile and the Stack, using a 100k-vocabulary Japanese-optimized tokenizer.

The model uses the Llama architecture with a 4096-token context window and is released under the Apache 2.0 license, with instruction-tuned variants available. It is developed openly with its training data and methods documented.