Discovering Preference Optimization Algorithms with and for Large Language Models

🤗 Model | 📚 Paper | 📝 Blog

This repository contains the code for our paper "Discovering Preference Optimization Algorithms with and for Large Language Models".

The code for training is largely taken and adapted from huggingface/alignment-handbook.

Setup and Evolution

To run the code in this project, first, create a Python virtual environment using e.g. Conda:

conda create -n handbook python=3.10 && conda activate handbook

Next, install PyTorch v2.1.2 - the precise version is important for reproducibility! Since this is hardware-dependent, we direct you to the PyTorch Installation Page.

You can then install the remaining package dependencies as follows:

python -m pip install .

You will also need Flash Attention 2 installed, which can be done by running:

python -m pip install flash-attn==2.5.7 --no-build-isolation

Note If your machine has less than 96GB of RAM and many CPU cores, reduce the MAX_JOBS arguments, e.g. MAX_JOBS=4 pip install flash-attn==2.5.7 --no-build-isolation

Next, log into your Hugging Face and Wandb accounts as follows:

huggingface-cli login
wandb login

Finally, install Git LFS so that you can push models to the Hugging Face Hub:

sudo apt-get install git-lfs

Then, install FastChat for MT-Bench as follows (in the same directory that you cloned this repo):

cd ../
git clone https://github.com/lm-sys/FastChat.git
cd FastChat
pip install -e ".[model_worker,llm_judge]"

Make sure that it is loading the correct chat template for Zephyr-Gemma.

See this issue for the template.

To launch the evolution script:

python3 scripts/launch_evo.py --wandb

Evaluations

Chat Evals

Finally, you need to install Alpaca Eval 2.0. Annoyingly, alpaca_eval uses openai>1.5.0 and mt-bench uses openai==0.28, which is not backward compatible. Therefore we need to create a second conda environment, that is a copy of the first.

conda create --name handbook_alpaca --clone handbook
conda activate handbook_alpaca

Subsequently we install alpaca_eval as follows:

pip install alpaca-eval

I have also created an extra folder in this repo named alpaca_eval, where we store all the model and api config files

Whenever you want to run an mt-bench model evaluation, you can do this with the following command:

conda activate handbook
python scripts/run_evaluations.py \
    --model-id <name_of_your_model> \
    --model-path <path_to_model_weights_or_HF> \
    --num-generations 1 \
    --mt-bench \

Whenever you want to run an alpaca_eval model evaluation, you can do this with the following command:

conda activate handbook_alpaca
python scripts/run_evaluations.py \
    --model-id <name_of_your_model> \
    --num-generations 1 \
    --alpaca-eval \
    --alpaca-model <path_to_your_model_config>/configs.yaml \
    --alpaca-reference-model path_to_ref_model_config>/configs.yaml \
    --alpaca-openai-configs <path_to_your_client_config>/openai_configs.yaml

TL;DR

If you want to run both together, We have prepared bash scripts:

source scripts/train_tldr.sh 

source scripts/eval_tldr.sh 

IMDb

source scripts/train_eval_imdb.sh 

Citation

@article{lu2024discopop,
  title={Discovering Preference Optimization Algorithms with and for Large Language Models},
  author={Lu, Chris and Holt, Samuel and Fanconi, Claudio and Chan, Alex J and Foerster, Jakob and van der Schaar, Mihaela and Lange, Robert Tjarko},
  journal={arXiv preprint arXiv:2406.08414},
  year={2024}
}