Longformer: The Long-Document Transformer

Benchmark Model Rank Results
question-answering-on-wikihopLongformer-large#2Test: 81.9
language-modelling-on-hutter-prizeLongformer Large#5Bit per Character (BPC): 0.99Number of params: 102M
language-modelling-on-hutter-prizeLongformer Small#6Bit per Character (BPC): 1.00Number of params: 41M
language-modelling-on-text8Longformer Small#7Bit per Character (BPC): 1.10
text-summarization-on-arxivLED Large#9ROUGE-1: 46.63ROUGE-2: 19.62ROUGE-L: 41.83
language-modelling-on-enwiki8Longformer (30 layers, h=512)#14Bit per Character (BPC): 0.99Number of params: 102M
language-modelling-on-enwiki8Longformer (12 layers, h=512)#17Bit per Character (BPC): 1.00Number of params: 41M