Atlas: Few-shot Learning with Retrieval Augmented Language Models

Benchmark Model Rank Results
multi-task-language-understanding-on-mmluAtlas (5-shot)#19Average (%): 47.9
question-answering-on-natural-questionsAtlas (full, Wiki-dec-2018 index)#1EM: 64.0
question-answering-on-natural-questionsAtlas (full, Wiki-dec-2021+CC index)#2EM: 60.4
question-answering-on-natural-questionsAtlas (few-shot, k=64, Wiki-Dec-2018 index)#11EM: 45.1
question-answering-on-natural-questionsAtlas (few-shot, k=64, Wiki-dec-2021+CC index)#15EM: 42.4