Chinese-LLaMA-Alpaca-3 provides open-source Chinese Llama-3 base and instruct models, with multiple releases (v1.0, v1.1, v2.0, v3.0) and ecosystem integrations for deployment. It supports multiple inference tools and detailed model comparisons.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
Chinese-LLaMA-Alpaca-3 is an open-source project that provides Chinese Llama-3 base and Llama-3-Instruct models, including v1, v2, and v3 variants. It includes pretraining and instruction-fine-tuning scripts, instruction data, and deployment guidance. It supports transformers, llama.cpp, text-generation-webui, vLLM, Ollama, ModelScope, and Hugging Face ecosystems.
How it works
The project combines Chinese incremental pretraining on Llama-3 with instruction fine-tuning data (approximately 500万条) and provides model download guidance for base and instruct variants. It offers multiple model flavors (8B base and 8B instruct) across v1.0, v1.1, v2.0, and v3.0, with instructions on merging LoRA adapters where applicable and quantized GGUF formats for deployment via llama.cpp and other tools. It maintains a comparison table of models (Llama-3-Chinese-8B-Instruct-v3, -v2, -v1, and base) across training type, context length, and input templates.
Getting started
INSTALL/USAGE commands are included in the README, such as the model download sections:
- https://huggingface.co/hfl/llama-3-chinese-8b-instruct-v3
- https://huggingface.co/hfl/llama-3-chinese-8b-instruct-v2
- https://huggingface.co/hfl/llama-3-chinese-8b-instruct
- https://huggingface.co/hfl/llama-3-chinese-8b and references to LoRA variants and GGUF download pages. It also notes how to merge LoRA with base models and provides a wiki page for model conversion steps.
Recent releases
Latest four releases include:
- v3.0 中文羊驼大模型三期 v3.0 (2024-05-30) introducing Llama-3-Chinese-8B-Instruct-v3 and noting improved downstream task performance.
- v2.0 中文羊驼大模型三期 v2.0 (2024-05-08) introducing Llama-3-Chinese-8B-Instruct-v2, trained on 500万指令数据 on Meta-Llama-3-8B-Instruct.
- v1.1 中文羊驼大模型三期 v1.1 (2024-05-07) adding pretrained/fine-tuning scripts and output examples.
- v1.0 中文羊驼大模型三期 v1.0 (2024-04-30) released base Llama-3-Chinese-8B and Llama-3-Chinese-8B-Instruct.
Traction
Stars: 1974
Behind the repo
No startup or company link provided in FACTS.
Caveats
License: Apache-2.0. Created 2023-11-09. Last push 2026-04-19. Issues: 2 open. Language: Python. Topics include alpaca, large-language-models, llama, llama-2, llama-3, llm, nlp.