Cluster-Former: Clustering-based Sparse Transformer for Long-Range Dependency Encoding

Benchmark Model Rank Results
language-modelling-on-enwiki8Cluster-Former (#C=512)Bit per Character (BPC): 1.22
open-domain-question-answering-on-searchqaCluster-Former (#C=512)EM: 68.0
question-answering-on-natural-questions-longCluster-Former (#C=512)F1: 76.5
question-answering-on-quasart-tCluster-Former (#C=512)EM: 54