Teaching LLaMA Chinese by shipping LoRA patches, not weights

What it does

The project releases LoRA adapter weights and an expanded Chinese vocabulary that retrofit Meta’s LLaMA into two distinct Chinese-speaking models: a base model for text continuation and an Alpaca-style instruction model for chat and Q&A. The adapters are trained on large Chinese corpora and instruction data, and the repo includes scripts for pre-training and fine-tuning so you can keep tuning on your own data.

The interesting bit

Because Meta’s LLaMA license prohibits redistribution and commercial use, the authors do not publish full model weights. Instead they ship LoRA “patches” and a merge script; you bring your own LLaMA base, bolt on the adapters, and obtain a complete Chinese-capable model without the authors ever distributing Meta’s bytes. It is a pragmatic workaround that also happens to make the download much smaller.

Key highlights

• Expanded Chinese vocabulary to ~49,954 tokens, improving encoding efficiency for Chinese text compared to the original LLaMA tokenizer
• Multiple model tiers (Base, Plus, Pro) at 7B, 13B, and 33B; Pro versions specifically address the “too-short reply” problem in earlier Alpaca releases
• Designed for local, quantized inference on consumer CPUs and GPUs via llama.cpp, transformers, text-generation-webui, LangChain, and privateGPT
• Open-sourced training and fine-tuning scripts for continued pre-training or domain adaptation
• Backed by a technical report on efficient Chinese text encoding for LLaMA

Caveats

• The released LoRA files are useless without the original Meta LLaMA weights, which are not freely redistributable and carry a non-commercial restriction
• The authors explicitly recommend new users migrate to their successor Chinese-LLaMA-Alpaca-3 project based on Llama-3, making this repo largely a legacy snapshot
• Alpaca and LLaMA variants use slightly different vocabularies (an extra pad token), so mixing them up produces gibberish

Verdict

Grab it if you want a proven recipe for Chinese-tuning LLaMA on modest hardware and can tolerate the non-commercial license. If you need a modern, ready-to-run Chinese model or plan to ship a product, look at the authors’ Llama-3-based successor instead.