FLoRA: Federated Fine-Tuning Large Language Models with Heterogeneous Low-Rank Adaptations

Authors: Ziyao Wang, Zheyu Shen, Yexiao He, Guoheng Sun, Hongyi Wang, Lingjuan Lyu, Ang Li

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate FLORA s superior performance in both homogeneous and heterogeneous settings, surpassing stateof-the-art methods.
Researcher Affiliation Collaboration University of Maryland, College Park 2. Rutgers University 3. Gen Bio.ai 4. Sony Reasearch
Pseudocode No The paper includes diagrams and mathematical equations to describe the method, but does not present a formal pseudocode or algorithm block.
Open Source Code Yes Our code is available at https://github.com/ATP-1010/Federated LLM.
Open Datasets Yes We use the Databricks-dolly-15k [28] instruction dataset, Alpaca dataset [19], and Wizard dataset [14] for the question-answering (QA) task, and Wizard and Share GPT for the chat assistant task.
Dataset Splits No The paper states the datasets used for training (Dolly, Alpaca, Wizard, Share GPT) and evaluation (MMLU, MT-bench), and mentions a non-IID client sampling setting. However, it does not provide explicit train/validation/test dataset splits (e.g., percentages or counts) for the training data or specify how a validation set was derived or used.
Hardware Specification Yes We use a 256GB AMD EPYC 7763 64-Core Processor on Linux v4.18.0 to run the experiments. For Lo RA fine-tuning on all the models, we use 4 NVIDIA RTX A6000 GPUs.
Software Dependencies No The paper does not explicitly list software dependencies with specific version numbers (e.g., Python 3.x, PyTorch 1.x).
Experiment Setup Yes In all our experiments, the learning rate of fine-tuning is set to 0.0003; the batch size is 128 and the micro batch size is 16. Table 2 shows the fine-tuning rounds and local epochs we selected.