HYDRA

This repository contains the code used in the paper HYDRA: Model Factorization Framework for Black-Box LLM Personalization, which will appear at NeurIPS 2024. It also provides a framework for factorizing the model to personalize the black-box LLMs.

Framework

Experiments

You need to first download the LaMP benchmark data into the directory data/.

Data Generation

For both reranker and adapter training, you need to generate the data and annotation first:

python main-reranker.py --seed 40 --debug generation --config configs/bbox-train/LaMP_2.yaml

or:

python main-adapter.py --seed 40 --debug generation --config configs/bbox-train/LaMP_2.yaml

We will update a detailed guidance of data generation in the following versions.

Model Training

Use reranker-gen option to train the reranker and also generate training data for the adapter.

CUDA_VISIBLE_DEVICES=2 python main-reranker.py --config configs/reranker/train-topk-final/LaMP_5.yaml --debug reranker-gen

For Adapter:

CUDA_VISIBLE_DEVICES=1 python main-adapter.py --config configs/bbox-p/LaMP_5.yaml --debug trainer

Release Updates

• Release core training code;
• Release data generation code and guidance;
• Update camera-ready version of code;
• Polish and prettier the current code base.

Citation

If you find this repository valuable for your research, we kindly request that you acknowledge our paper by citing the following paper. We appreciate your consideration.

@misc{zhuang2024hydramodelfactorizationframework,
      title={HYDRA: Model Factorization Framework for Black-Box LLM Personalization}, 
      author={Yuchen Zhuang and Haotian Sun and Yue Yu and Rushi Qiang and Qifan Wang and Chao Zhang and Bo Dai},
      year={2024},
      eprint={2406.02888},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2406.02888}, 
}