Model Overview

ganchengguang/Yoko-7B-Japanese-v1 is a 7 billion parameter language model built upon the vanilla LLaMA2-7B architecture. Developed with contributions from Yokohama National University Mori Lab, this model has undergone fine-tuning using QLoRA.

Key Capabilities

• Enhanced Multilingual Performance: Demonstrates improved performance specifically in Chinese and Japanese language tasks.
• Extensive Training Data: Fine-tuned on a comprehensive dataset, including 49,000 chat samples and 280,000 non-chat samples from the Guanaco dataset.
• LLaMA2 Base: Benefits from the robust foundation of the LLaMA2-7B model.

Good For

• Applications requiring strong performance in Chinese and Japanese language generation and understanding.
• Chatbot implementations and general text generation tasks in these languages.

Recommended Generation Parameters

For optimal results, consider using the following generation parameters:

• Temperature: 0.5 - 0.7
• Top P: 0.65 - 1.0
• Top K: 30 - 50
• Repeat Penalty: 1.03 - 1.17