OpenFedLLM is an open-source research-use codebase for training Large Language Models (LLM) via federated learning. Please check our paper for details and the corresponding empirical study.

OpenFedLLM includes the following key features:

• 7 federated learning algorithms (e.g., FedAvg, FedProx, SCAFFOLD, FedAvgM, etc.).
• 2 LLM training algorithms, including instruction tuning (i.e. SFT) and value alignment (i.e., DPO).
• 30+ evaluation metrics covering general capabilities, medical QA, financial QA, code generation, math solving, and more.

Clone the repo and install the required packages.

git clone https://github.com/rui-ye/OpenFedLLM.git
cd OpenFedLLM
conda create -n fedllm python=3.10
conda activate fedllm
pip install -r requirements.txt

We provide training scripts under training_scripts/. Try them out from the top-level directory of this repository.

The training script is in training_scripts/run_sft.sh.

CUDA_VISIBLE_DEVICES=1 python main_sft.py \
 --model_name_or_path "meta-llama/Llama-2-7b-hf" \
 --dataset_name "vicgalle/alpaca-gpt4" \
 --dataset_sample 20000 \
 --fed_alg "fedavg" \
 --num_clients 20 \
 --sample_clients 2 \
 --max_steps 10 \
 --num_rounds 200 \
 --batch_size 16 \
 --gradient_accumulation_steps 1 \
 --seq_length 512 \
 --peft_lora_r 32 \
 --peft_lora_alpha 64 \
 --use_peft \
 --load_in_8bit \
 --output_dir "./output" \
 --template "alpaca" \

Key arguments:

• model_name_or_path: the name or local location of your base model
• template: template for chatting. Define your own template in utils/template.py.
• dataset_name: the name of dataset. You may modify utils/process_dataset.py if your interested dataset has not been supported.
• dataset_sample: needed if you want to sample a specific number of samples from the original dataset.
• fed_alg: the name of federated learning algorithm
• num_clients/sample_clients: num_clients clients in total, sample_clients clients for each round
• max_steps: the number of model update steps for one client at each round.

The training script is in training_scripts/run_dpo.sh.

python main_dpo.py --template "vicuna_v1.1"

Note that the main difference between the usage of main_sft.py and main_dpo.py lies in the template argument. We plan to make them consistent in the future.

• For SFT, templates are defined in utils/template.py
• For DPO, templates are defined in utils/conversation.py

Evaluation codes are put in evaluation/ directory. Most of our evaluations follow existing high-incluence open-source repos. Please refer to each sub-directory for the corresponding detailed README and running script.

For example, evaluation/open_ended/ include open-ended evaluations on three benchmarks, covering MT-Bench, Vicuna Bench, and AdvBench; see README.md.

Please cite our paper if you find the repository helpful.

@article{ye2024openfedllm,
  title={OpenFedLLM: Training Large Language Models on Decentralized Private Data via Federated Learning},
  author={Ye, Rui and Wang, Wenhao and Chai, Jingyi and Li, Dihan and Li, Zexi and Xu, Yinda and Du, Yaxin and Wang, Yanfeng and Chen, Siheng},
  journal={arXiv preprint arXiv:2402.06954},
  year={2024}
}