ComLoRA: A Competitive Learning Approach for Enhancing LoRA

Authors: Qiushi Huang, Tom Ko, Lilian Tang, Yu Zhang

ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the superiority of Com Lo RA over Lo RA, without increasing inference overhead.
Researcher Affiliation Collaboration 1Southern University of Science and Technology, 2University of Surrey, 3Byte Dance
Pseudocode No The paper describes the methodology using natural language and mathematical equations in Section 3, but does not include any structured pseudocode or algorithm blocks.
Open Source Code Yes The Com Lo RA s code is available at https://github.com/hqsiswiliam/comlora.
Open Datasets Yes We evaluate our models using a comprehensive commonsense reasoning dataset, which includes eight sub-tasks: Bool Q (Clark et al., 2019), PIQA (Bisk et al., 2020), SIQA (Sap et al., 2019), Hella Swag (Zellers et al., 2019), Wino Grande (Sakaguchi et al., 2021), ARC-c (Clark et al., 2018), ARC-e (Clark et al., 2018), and OBQA (Mihaylov et al., 2018). MMLU. The MMLU (Massive Multitask Language Understanding) benchmark (Hendrycks et al., 2021b;a)... Personalized Conversation Task. We assess model performance on personalized conversational understanding using the CONVAI2 dataset (Dinan et al., 2019; Zhang et al., 2018).
Dataset Splits Yes We aggregate the training sets from all sub-tasks into a single corpus of 170,420 entries, from which we randomly select 120 entries for validation to identify the optimal model.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments. It only mentions the LLa MA-3-8B model as the base model.
Software Dependencies No The paper mentions using the Adam W optimizer and the LLa MA-3-8B model, but does not provide specific version numbers for any key software components or libraries.
Experiment Setup Yes The Adam W optimizer (Loshchilov & Hutter, 2019) is used to train Lo RA and Com Lo RA. The learning rates for both Lo RA and Com Lo RA methods are selected from [1e-3, 1e-4], and finetuning is conducted for 3 epochs for all the tasks. Each experiment runs with 5 different seeds, and the average results are reported.