BLoB: Bayesian Low-Rank Adaptation by Backpropagation for Large Language Models

Authors: Yibin Wang, Haizhou Shi, Ligong Han, Dimitris Metaxas, Hao Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our empirical results verify the effectiveness of BLo B in terms of generalization and uncertainty estimation, when evaluated on both in-distribution and out-of-distribution data.
Researcher Affiliation Collaboration 1Rutgers University. 2MIT-IBM Watson AI Lab.
Pseudocode Yes Algorithm 1 Bayesian Low-Rank Adaptation by Backpropagation (BLo B)
Open Source Code Yes Code is available at https://github.com/Wang-ML-Lab/bayesian-peft.
Open Datasets Yes We fine-tune Llama2-7B on six common-sense reasoning tasks: Winogrande-small (WG-S), Winogrande-medium (WG-M) [82], ARC-Challenge (ARC-C) [18], ARC-Easy (ARC-E) [18], Open Book QA (OBQA) [65], and Bool Q [17].
Dataset Splits Yes For all baseline methods, using the same pre-trained LLM backbone, we maintain consistent hyperparameters across all datasets and do not use additional validation sets to achieve higher performance (See Appendix B.3 for detailed settings).
Hardware Specification Yes Our experiments on Llama2-7B were conducted using 2 NVIDIA RTX A5000 GPUs for parallel training, while experiments on Ro BERTa-base were conducted using 4 NVIDIA RTX A5000 GPUs for parallel training.
Software Dependencies No The paper mentions the use of specific software libraries like "PEFT library" and "Bayesian-Torch library" but does not provide specific version numbers for these dependencies within the main text or appendices to ensure reproducibility.
Experiment Setup Yes For hyperparameters, we strictly adhere to the default settings in the PEFT library and the original Lo RA paper [63, 41] to ensure maximal reproducibility. This includes the number of training steps, learning rate, and Lo RA rank r (see Appendix B.1 for details). ... Detailed hyperparameter settings are provided in the Table 4. Table 3 provides the hyperparameters for fine-tuning with Lo RA shared with other baselines.