ganchengguang/Yoko_13B_Japanese_QLoRA

This model is traned with llm-japanese-dataset dataset. And this model used a few of dataset by 50000 chat samples and 280000 non chat samples. Improved performance in Chinese and Japanese. Use the QLoRA to fine-tune the vanilla Llama-2-13b-chat-hf. And you can use test.py to test the model.

Recommend Generation parameters:

• —temperature: 0.5~0.7
• —top p: 0.65~1.0
• —top k: 30~50
• —repeat penalty: 1.03~1.17

Contribute by Yokohama Nationaly University Mori Lab.