3 ms·Pretraining with hierarchical memories separating long-tail and common knowledge5 points by dataminer 1y ago