Generalized Logit Adjustment: Improving Fine-tuning by Mitigating Label Bias in Foundation Models

Mitigating label bias in foundation models improves fine-tuning performance.