🦙🧋🇹🇼 Finetune LLaMA-7B with Chinese instruction datasets

This repository is a tutorial for finetuning LLaMA-7B with Chinese datasets! I survey and combine the dataset & method for finetuning my own LLM for complex NLP tasks such as summarization, question answering, text generation, custom data augmentation, etc.

Since the original Stanford Alpaca-7B finetune need lots of GPU resources, I focus on surveying the method with low GPU consumption.

So here's how to reproduce:

Installation

1. Install requirements

$ pip install -r requirements.txt

2. Install PyTorch at compatible version with CUDA

$ pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1+cu116 --extra-index-url https://download.pytorch.org/whl/cu116

Datasets

This repository combined all datasets using English-instruction, Chinese-output construction:

1. alpaca_data.json: Original dataset from Stanford Alpaca
2. alpaca_data_cleansed.json: Cleansing by gururise/AlpacaDataCleaned
3. alpaca-zhCN.json: Translate by carbonz0/alpaca-chinese-dataset
4. alpaca-zhTW.json: Translate to Traditional Chinese using OpenCC
5. alpaca-en-zh.json: Combine the English instruction/input and Chinese output by ntunlplab/traditional-chinese-alpaca: (Traditional Chinese dataset translate by ChatGPT API (gpt-3.5-turbo) by ntunlplab/traditional-chinese-alpaca (Update at 2023.03.29))

Finetune

Reference finetune method provide by tloen/alpaca-lora

1. Run on 1 GPU with Colab: https://colab.research.google.com/drive/1QvtrJpikkkNKSbwwG766SIGbBw2TQRd5?usp=sharing

2. Use torchrun for distributed training on Multi-GPUs

$ cd finetune/
$ torchrun --standalone --nnodes=1 --nproc_per_node=4 finetune.py

Finetune Domain Tasks

(In progress, welcome to discuss together: jiunyi.yang.abao@gmail.com. I'd like to try tasks from different domains such as investment, fraud, e-commerce, law, healthcare, ...)

Model Serving

To serve your own model service through API & simple website UI!

1. Model API

   

   $ cd serve/
   $ python api.py

2. demo UI

   

   $ cd serve/
   $ python ui.py

Learn More

I curated lots of method that try to run large language models with fewer GPU resources:

• PEFT
• LoRA
• FlexGen ...

See full list: chatgpt-alternatives

@misc{alpaca-7b-chinese,
  author = {JiunYi Yang},
  title = {Alpaca-7B Chinese: Finetune LLaMA-7B with Chinese instruction datasets},
  year = {2023},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/A-baoYang/alpaca-7b-chinese}},
}