Verified project record

ymcui/Chinese-LLaMA-Alpaca-2

chinese-llama-alpaca-2 provides open-source Chinese LLaMA-2 and Alpaca-2 large language models that address insufficient Chinese understanding in the original models. It features an optimized Chinese vocabulary, FlashAttention-2 training, and long-context extensions up to 64K.

Project overview

By redesigning the Chinese vocabulary to 55,296 tokens for better encoding efficiency and applying FlashAttention-2 training, this project delivers an improved and hardware-efficient Chinese language modeling experience built on the Llama-2 architecture.

Project type: Model Runtime
Use cases: Chat Assistants
Deployment: Refer to project documentation
License: Apache-2.0

Best for

AI engineers and developers needing open-source large language models with optimized Chinese language understanding.
Users looking to run models locally with hardware flexibility using quantization and speculative sampling.

Key capabilities

Redesigned Chinese vocabulary (size 55,296) to improve Chinese text encoding/decoding efficiency and coverage.
All models were trained using FlashAttention-2 for faster speed and better memory usage.
Uses PI and YaRN techniques to offer 16K and 64K context length models.
Alpaca-2-RLHF series models tuned with RLHF for improved value alignment.
Provides GGUF (imatrix) and AWQ quantized models compatible with frameworks like llama.cpp and transformers.
Uses 1.3B models as draft models to accelerate inference of 7B/13B target models.
Provides pre-training and instruction fine-tuning scripts for further model training.

Limitations and risks

Standard models support only 4K context length, while long-context versions support 16K or 64K.
LoRA models cannot be used directly and must be merged with the original Llama-2 model.
It is not recommended to use the 1.3B model alone; it should be paired with larger models via speculative sampling.
The models may generate unpredictable harmful content or content misaligned with human preferences.
Training may be insufficient due to compute and data limits, meaning Chinese understanding needs further improvement.

Getting started

To begin, download a complete model or LoRA weights; if using LoRA, merge them with the original Llama-2 model. Then select a deployment framework like llama.cpp, quantize the model if necessary (such as to 4-bit), and run it locally on your chosen hardware.
Ensure sufficient storage for the target model size: the 13B model requires 24.7 GB.

Evidence and sources

README: 开源了**中文LLaMA-2基座模型和Alpaca-2指令精调大模型**。这些模型**在原版Llama-2的基础上扩充并优化了中文词表**，使用了大规模中文数据进行增量预训练，进一步提升了中文基础语义和指令理解能力
README: 支持[🤗transformers](https://github.com/huggingface/transformers), [llama.cpp](https://github.com/ggerganov/llama.cpp), [text-generation-webui](https://github.com/oobabooga/text-gene…
README: Chinese-LLaMA-Alpaca-3 已正式发布，开源基于Llama-3的Llama-3-Chinese-8B和Llama-3-Chinese-8B-Instruct，推荐所有一期、二期项目用户升级至三代模型
Release: 中文羊驼大模型二期 v4.1
README: 基座模型（4K上下文）：Chinese-LLaMA-2 (1.3B, 7B, 13B) - 聊天模型（4K上下文）：Chinese-Alpaca-2 (1.3B, 7B, 13B) - 长上下文模型（16K/64K）：

Reviewed at: 2026-07-16T07:35:54Z