MUM, or Multitask Unified Model, is trained on large volumes of text from the Internet but not on click-and-query data; it focuses on raw text to enhance understanding of language and world knowledge [1]. 

Additionally, MUM is also trained on a large corpus of task-specific data, which may incorporate user feedback and interactions to improve its effectiveness in search capabilities [9]. The potential of new language models like MUM lies in their ability to transform the understanding of language and information, suggesting a complementary role in Google's search architecture rather than rendering existing systems obsolete [9].

What's MUM trained on?

Related Content You May Like