Neura Market
Facico/Chinese-Vicuna logo

Facico/Chinese-Vicuna

Free

Chinese-Vicuna: A Chinese Instruction-following LLaMA-based Model —— 一个中文低资源的llama+lora方案，结构参考alpaca

FreeFree tier
Inputs: textOutputs: text
Type
Open Source
Visit Website GitHub (38 stars)

About Facico/Chinese-Vicuna

Chinese-Vicuna is an open-source project for instruction-tuning LLaMA models on Chinese language tasks using parameter-efficient LoRA (Low-Rank Adaptation). It is designed to run on consumer GPUs such as Nvidia RTX-2080TI (11GB) for 7B models and RTX-3090 (24GB) for 13B models, including multi-turn chatbot training with context length 2048. The repository provides code for fine-tuning, generation, CPU inference (fp16/int4), and tools to download/convert/quantize original LLaMA weights. It supports 4-bit QLoRA for reduced memory usage. The project includes pre-trained checkpoints and example fine-tuning on medical and legal domains. Training data includes 30k Chinese ShareGPT conversations and 20k Alpaca instruction Chinese dataset.

Key Features

Fine-tune LLaMA-7B on a single Nvidia RTX-2080TI (11GB VRAM)
Fine-tune LLaMA-13B on a single Nvidia RTX-3090 (24GB VRAM)
Multi-turn chatbot training with context length 2048 on 3090
Parameter-efficient LoRA (Low-Rank Adaptation) tuning
Support for 4-bit QLoRA training and inference
CPU inference with fp16 or int4 support
Multi-GPU inference to reduce VRAM usage
Pre-trained checkpoints for instruction-following and domain-specific (medical, legal) tasks
Training data: 30k Chinese ShareGPT + 20k Alpaca instruction Chinese dataset
Tools to download, convert, and quantize original LLaMA weights

Pros & Cons

Pros
  • Can be trained on commonly available consumer GPUs (2080Ti, 3090)
  • Parameter-efficient LoRA reduces memory and compute requirements
  • Supports 4-bit quantization for even lower VRAM usage
  • Open-source with complete code, checkpoints, and example datasets
  • Easy deployment with CPU and multi-GPU inference options
Cons
  • Requires original LLaMA weights separately (not included in repo)
  • Primarily focused on Chinese language, limiting use for other languages
  • Documentation is mostly in the README with limited structured guides

Best For

Chinese instruction-following tasksMulti-turn Chinese chatbotFine-tuning LLaMA on domain-specific Chinese corpora (e.g., medical, legal)Low-resource fine-tuning of large language models on consumer GPUs