Ensembling models beats training longer when data runs out

Original: A Scaling Theory of Multi-Epoch Pretraining: Late-Time Overfitting and Ensembling

Source: arxiv.org ↗