Spectral Adapter: Fine-Tuning in Spectral Space

Authors: Fangzhao Zhang, Mert Pilanci

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

Reproducibility Variable Result LLM Response
Research Type Experimental We show through extensive experiments that the proposed fine-tuning model enables better parameter efficiency and tuning performance as well as benefits multi-adapter fusion.
Researcher Affiliation Academia Fangzhao Zhang Electrical Engineering Stanford University zfzhao@stanford.edu Mert Pilanci Electrical Engineering Stanford University pilanci@stanford.edu
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Code is released at https://github.com/pilancilab/spectral_adapter.
Open Datasets Yes Llama3 8B model with Orca Math dataset [38] and evaluation score on GSM8K benchmark [8].
Dataset Splits No The paper mentions training data (Orca Math) and test data (GSM8K) but does not provide explicit training/validation/test splits with percentages or sample counts for all experiments, nor does it specify the method for creating such splits beyond "10K Orca Math data (shuffled)".
Hardware Specification Yes All experiments are done with NVIDIA RTX A6000 GPU.
Software Dependencies No The paper mentions software like Pytorch [44] and Hugging Face PEFT library but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes Table 4 shows the hyperparameter setting for our Spectral Adapter A used for fine-tuning De BERTa V3base model in Section 4.1. We set number of diagonal blocks to be 4 and enable block sharing for OFT to maintain similar amount of trainable parameters.