Chinese-Llama-2-7b

Why we included this project

Teams building Chinese-language chat applications on Llama 2 will find this a workable base rather than a research curiosity. It was the first Chinese LLaMA 2 model the open-source community could actually download and run, and the release still works as a practical starting point. The 7B chat weights come with the SFT training and inference scripts used to produce them, and the model was fine-tuned on a bilingual instruction set of roughly 10 million samples, so the training recipe is visible instead of hidden. Input formatting follows the original llama-2-chat spec, which means tooling and optimizations written for the English model tend to carry over with little adjustment. Docker images and a 4-bit quantized build get you from download to a running endpoint on your own hardware fairly quickly, and the Chinese and English fine-tuning data is released openly for anyone who wants to keep training. Everything is commercially usable, and a live demo on HuggingFace Spaces lets you check the output before you commit.