Open-source project providing Chinese LLaMA and Alpaca models with LoRA-based weights, plus guidance for local CPU/GPU deployment and multiple inference options.
Collecting history — the radar snapshots this repo daily. The trend line appears after 3 days of data (1 so far).
What it is
The repository hosts Chinese-LLaMA-Alpaca-3, focusing on Chinese LLaMA models and instruction-tuned Alpaca models, including LoRA-weighted variants for Chinese LLaMA and Alpaca. It documents model families (7B, 13B, 33B), their LoRA-based reconstructions, and deployment options for local inference.
How it works
The project offers model variants that are first built on original LLaMA weights with expanded Chinese vocabularies and subsequent instruction-tuning for Alpaca models. It provides guidance for merging LoRA weights with base models to form complete inference candidates and supports multiple inference backends, such as llama.cpp, transformers, and web demos. It also lists integration points with LangChain, privateGPT, and other ecosystems.
Getting started
- Uses LoRA-weighted models that must be merged with the original LLaMA weights before inference.
- It provides online and manual model merging/ conversion workflows. See merge instructions and wiki for details.
- The README references model download links and LoRA payloads, and mentions that some models require combining with base LLaMA models for full functionality.
Commands and prompts are included in the README and wiki for performing merges and deploying locally. For example, the repository notes that Online conversion is available via notebooks and that manual conversion is supported, with links to the relevant tutorials.
Recent releases
Latest releases include v5.0 (民间版中文羊驼模型 v5.0) and notes about new Chinese LLaMA-2/Alpaca-2 projects, Plus/Pro variants, 8K context support, and updated technical reports. The releases emphasize Chinese LLaMA-33B, Alpaca-33B, and Plus/Pro variants, along with 8K+ context capabilities without modifying model weights.
Traction
Stars: 18942 Forks: 1843 Open issues: 6 Created: 2023-03-15 Last push: 2026-04-19
安全性 and caveats
License: Apache-2.0 The project distributes LoRA payloads meant to be merged with original LLaMA weights, due to licensing restrictions on the base models. It highlights the requirement to verify SHA256 of downloaded archives and notes that some models are not recommended for certain use cases, with guidance to prefer specific Pro/Plus variants for longer replies.