Large Pretraining Datasets May Reduce Robustness After Fine‑Tuning

Fine‑tuning vision models, even those pretrained on LAION‑2B, can sharply cut out‑of‑distribution robustness; the ImageNet‑RIB benchmark quantifies this drop. getnews.me/large-pretraining-datase... #visionmodels #robustness #laion2b