RT @Dahan_Noam: Look at the CRAZY domain gap we found in summarization datasets: while English resources are diverse, other languages are m…

Interestingly - all models had a distinct point where they no longer relied on information from previous tokens! Joint work with @royschwartzNLP 8/8

RT @DariaLioub: 📢Paper release📢 What computation is the Transformer performing in the layers after the top-1 becomes fixed (a so called "sa…