// the find
ymcui/Chinese-LLaMA-Alpaca
中文LLaMA&Alpaca大语言模型+本地CPU/GPU训练部署 (Chinese LLaMA & Alpaca LLMs)
Chinese-LLaMA-Alpaca extends Meta's original LLaMA-1 with an expanded Chinese vocabulary and Chinese pretraining data, then fine-tunes with Chinese instruction data to produce the Alpaca variants. It targets researchers and hobbyists who want a Chinese-capable LLM running locally on consumer hardware. This is generation 1 of a three-generation project — the repo itself opens by telling you to go use v3 instead.
The vocabulary expansion is the real technical contribution here: they merged ~20k Chinese tokens into the LLaMA tokenizer, which meaningfully reduces token count for Chinese text and speeds up inference. Training scripts (pretraining and SFT) are included, not just weights, so you can reproduce or extend the work. Deployment coverage is thorough — llama.cpp, transformers, text-generation-webui, LangChain, and Colab notebooks are all documented with separate wiki pages. C-Eval benchmark numbers are published with reproducible eval code, which is more honest than most Chinese LLM repos of this era.
This is LLaMA-1-based, which was already two generations behind when Chinese-LLaMA-Alpaca-3 (LLaMA-3) launched in April 2024. The C-Eval scores tell the story plainly: the best model here (Plus-33B) hits 46.5% zero-shot on a task with 25% random baseline, which is mediocre even by 2023 standards. Distribution is LoRA-only because of LLaMA's license terms, meaning you have to independently obtain the original LLaMA weights and run a merge script before you can do anything — a real barrier for casual users. The non-commercial restriction from LLaMA-1 flows through the whole project, so it's research-only by default.