This project fine-tunes and distributes Chinese versions of Meta’s Llama-2 foundation models, including base models and instruction-tuned variants. It extends the original Llama-2 vocabulary with additional Chinese tokens and uses large-scale Chinese data for continued pre-training. The repository offers both standard 4K context models and extended context variants supporting 16K and 64K sequences, along with RLHF-aligned models for improved human preference alignment.