MELoRA: Mini-Ensemble Low-Rank Adapters for Parameter-Efficient Fine-Tuning

MELoRA, a mini-ensemble low-rank adapters that uses fewer trainable parameters while maintaining a higher rank, thereby offering improved performance potential.The core idea is to freeze original pretrained weights and train a group of mini LoRAs with only a small number of parameters. This can capture a significant degree of diversity among mini LoRAs, thus promoting better generalization ability.

Method Overview

Quickstart

1. Install dependencies

   conda create -n MELoRA python=3.10
   conda activate MELoRA
   pip install torch==2.0.1
   pip install -r requirements.txt
   

   cd peft-0.5.0
   pip install -e .
   

2. Run experiments

fill in the --model_name_or_path --wandb_project and --output_dir in llama_finetune.sh and glue_finetune.sh with the path to the model and the output directory.

Instruction Tuning

bash llama_finetune.sh

NLU

bash glue_finetune.sh

Thanks

Code is largely based on AGI-Edgerunners/LLM-Adapters, huggingface/peft, huggingface/transformers

Cite

If you find this method or code useful, please cite

@article{melora,
  title={Mini-Ensemble Low-Rank Adapters for Parameter-Efficient Fine-Tuning},
  author={Ren, Pengjie and Shi, Chengshun and Wu, Shiguang and Zhang, Mengqi and Ren, Zhaochun and de Rijke, Maarten and Chen, Zhumin and Pei, Jiahuan},
  journal={arXiv preprint arXiv:2402.17263},
  year={2024}
}