Atlas: Few-shot Learning with Retrieval Augmented Language Models

In this work we present Atlas, a carefully designed and pre-trained retrieval augmented language model able to learn knowledge intensive tasks with very few training examples.