text-ada-002

Findings

• IC-1017Vision and language models pre-trained on noisy data exhibit degraded OOD transfer that is partially recoverable via SVD-based feature-space regularization