Unsupervised Identification and Removal of Spurious Correlations During Fine-Tuning
The paper addresses how fine-tuning a language model can introduce bias through spurious correlations entangled with curation data. It proposes GRASP, an unsupervised method that identifies these hidden factors by analyzing weight changes and then projects future gradients to remove the correlation. This allows bias reduction without losing the model's actual task performance on unrelated topics.