DoFIT: Domain-aware Federated Instruction Tuning with Alleviated Catastrophic Forgetting (NeurIPS 2024)

Requirements

To install requirements:

pip install -r requirements.txt

Training

To train the model(s) on F&G, M&G, run this command:

bash training_scripts/FG_run_sft.sh
bash training_scripts/MG_run_sft.sh

Evaluation

To evaluate my model on FPB, FiQA-SA, TFNS, and NWGI, run:

cd evaluation/financial/
sh scripts/fin_all.sh ../../output/fingpt-sentiment-train/full-50 <gpu>

To evaluate my model on MedQA and MedMCQA, run:

medical domain environment - LMFlow:
git clone -b v0.0.5 https://github.com/OptimalScale/LMFlow.git
cd LMFlow
conda create -n lmflow python=3.9 -y
conda activate lmflow
conda install mpi4py
bash install.sh
cd data && ./download.sh all && cd -

cd evaluation/medical/
bash scripts/medqa.sh $path_to_your_model $gpu_ids $deepspeed_port
bash scripts/medmcqa.sh $path_to_your_model $gpu_ids $deepspeed_port

Citation

@inproceedings{xudofit,
  title={DoFIT: Domain-aware Federated Instruction Tuning with Alleviated Catastrophic Forgetting},
  author={Xu, Binqian and Shu, Xiangbo and Mei, Haiyang and Bai, Zechen and Fernando, Basura and Shou, Mike Zheng and Tang, Jinhui},
  booktitle={The Thirty-eighth Annual Conference on Neural Information Processing Systems}
}

Acknowledgements

This repo is based on OpenFedLLM, thanks to the original authors for their works!