MixedPEFT: Combining Multiple PEFT Methods with Mixed Objectives for Unsupervised Domain Adaptation
This study introduces a novel parameter-efficient strategy for unsupervised domain adaptation by combining invertible adapters and LoRA into a custom union. The method simultaneously optimizes classification on labeled source data and masked language modeling (MLM) on unlabeled target data using a dynamically weighted mixed-objective loss function, achieving state-of-the-art results with only 7% of the model's parameters.