Skip to content
/ koza Public
forked from tloen/alpaca-lora

Instruct-tune LLaMA on consumer hardware

License

Notifications You must be signed in to change notification settings

bqpro1/koza

 
 

Repository files navigation

🐐🌲KOZA: Low-Rank LLaMA Instruct-Tuning

KOZA 🐐 is an instruct model for Polish language with similar capabilities for LLMs like ChatGPT, but it can be run on a single machine 💻 with a strong GPU (RTX 4080/4090). This repo was originally forked from alpaca-lora.

Using this repo, you can reproduce the Stanford Alpaca for the Polish language. Original paper from Stanford: low-rank adaptation (LoRA).

1. Quick start

In your terminal clone repo:

git clone git@github.com:bqpro1/alpaca-koza.git

From repo folder install requirements.txt:

pip isntall -r requirements.txt

And run:

python generate.py

Go to http://127.0.0.1:7860 in your browser. Enjoy KOZA!

2. Translation

I used translation_instructions.ipynb for auto-translation instructions from data/alpaca_data.json to data/alpaca_data_pl_verified.json. Translation is far from perfect, might be improved. Translation took 12h. You can also find translated data on huggingface🤗

3. Training

Just run

python finetune.py --base_model='decapoda-research/llama-7b-hf'

Weights for Polsih language are also on huggingface🤗.

About

Instruct-tune LLaMA on consumer hardware

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 71.6%
  • Jupyter Notebook 28.4%