ez-trainer

Train Llama Loras Easily

Extremely ugly interface for easy lora training for Llama models. Put your models in the models folder, put your datasets in the datasets folder. Wow!

Datasets are expected to be json or jsonl format right now. Models should be the unquantized huggingface format. More info soon, maybe.

Features:

Training:

Train Loras for Llama. See wiki.

Merging

Merge your Loras back into Llama if you want. Supports strength-based merging.

Check out the wiki for usage info

https://github.com/CoffeeVampir3/ez-trainer/wiki/Train-the-Stuff-and-Stuff#crash-course

Install:

pip install -r requirements.txt

(This does not install torch, you should install that yourself. https://pytorch.org/get-started/locally/)

(Step by step instructions for exactly how to install in the wiki https://github.com/CoffeeVampir3/ez-trainer/wiki/Train-the-Stuff-and-Stuff#crash-course)

Run:

python train_module.py