Chinese-LLaMA-Alpaca-2

Why we included this project

Anyone building or fine-tuning models that need to handle Chinese well gets a real starting point here, not just an API to call. The project takes Meta's Llama-2, expands the vocabulary so it covers Chinese properly, and runs additional pretraining over large Chinese corpora, which shows up as noticeably better Chinese comprehension and instruction following than the original weights. The 16K and 64K long-context checkpoints are worth a look if your documents or retrieval tasks routinely blow past the standard 4K window. You also get the pretraining and instruction fine-tuning scripts, so this works as a base for continued training on your own data. Because the models stay inside the LLaMA ecosystem, they run in transformers, llama.cpp, vLLM, and most of the usual tooling with minimal setup, including local quantization.