Finetune an LLM, within a few clicks!
Model \ Method | SFT | DPO | ORPO | KTO | PRO |
---|---|---|---|---|---|
llama 2 | ✅ | ❌ | ❌ | ❌ | ❌ |
llama 3 | ✅ | ❌ | ✅ | ❌ | ❌ |
gguf | ✅ | ❌ | ❌ | ❌ | ❌ |
phi-3 | ✅ | ❌ | ❌ | ❌ | ❌ |
Mistral | ✅ | ✅ | ❌ | ❌ | ❌ |
... | ? | ? | ? | ? | ? |
This repo provides 3 modules, frontend
(react), server
(nodejs), and trainer
(python django)
You need CUDA for now, but once llama.cpp is integrated, this will no longer be required.
-
For Linux
Install CUDA from Nvidia installation guide
-
For Windows (with Nvidia GPU)
Enable WSL2 on your machine.
Install CUDA from Nvidia installation guide
Setup frontend
// copy .env.example to .env
cd frontend
yarn dev
Setup server
// copy .env.example to .env
cd server
npx prisma migrate dev --name add_datasets_model
yarn dev
Setup trainer
cd trainer
pip install -r requirements.txt
python manage runserver
- Support training at GGUF level
- Support
llama.cpp
- Expose a finetune script as API and GUI for CUDA (WIP)
- Add more finetune scripts for different OS, e.g. Apple Silicon with mlx
- Make finetune script configurable
- Provide a playground to instantly test the trained model
- Containerize backend and frontend