Parameter Efficient Quasi-Orthogonal Fine-Tuning via Givens Rotation

Authors: Xinyu Ma, Xu Chu, Zhibang Yang, Yang Lin, Xin Gao, Junfeng Zhao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various tasks and pretrained models validate the effectiveness of our methods. In this section, we conduct extensive experiments to evaluate the effectiveness of our methods.
Researcher Affiliation Academia 1School of Computer Science, Peking University, Beijing, China 2Key Laboratory of High Confidence Software Technologies, Ministry of Education, Beijing, China 3Center on Frontiers of Computing Studies, Peking University, Beijing, China.
Pseudocode Yes Algorithm 1 The fine-tuning and testing procedure of a pre-trained model with (q)GOFT.
Open Source Code Yes We implement GOFT and q GOFT for fine-tuning De BERTa V3-base (He et al., 2021) and LLa MA2-7B (Touvron et al., 2023), we also integrate our methods into the PEFT library (Mangrulkar et al., 2022) 1. 1https://github.com/ArthurLeoM/peft-givens
Open Datasets Yes Various downstream NLP tasks are applied to fine-tune the PLMs for conducting comparisons between baselines, including natural language understanding (Wang et al., 2018a, GLUE), instruction following (Hendrycks et al., 2021, MMLU) (Chiang et al., 2023, Vicuna-Eval), and question answering (Rajpurkar et al., 2016, SQu AD). We also validate the effectiveness of our method on visual tasks (Zhai et al., 2019, VTAB-1K) by fine-tuning VFMs like Vi T-B/16 (Dosovitskiy et al., 2021).
Dataset Splits Yes We present the detailed dataset statistics of GLUE benchmark (Wang et al., 2018a) in Table 6. (Table 6 shows #Dev column). SQu ADv1.1 consists of 87,599 training samples and 10,570 validation samples.
Hardware Specification Yes The experiments are conducted on a single NVIDIA-A100-80GB GPU or distributedly on a maximum of 4 NVIDIA-RTX3090-24GB GPUs.
Software Dependencies No The paper mentions software like Py Torch, Hugging Face transformers, PEFT library, and LLa MA-Factory, but does not specify their version numbers.
Experiment Setup Yes The specific tuned hyperparameters used in our experiments are presented in Table 5.