AI Dynamics

Global AI News Aggregator

Cross-Document Attention Leakage in Sequence Packing Solutions

This paper tries to solve a non-issue actually . Their claim is that when you do packing (they call it concat and chunk lol) you get cross document attention leakage. The truth is that if your infra is decent you'll have segmentation masks that prevent this from happening in

→ View original post on X — @yitayml,

Commentaires

Leave a Reply

Your email address will not be published. Required fields are marked *