Cohere Labs - Konstantin Dobler, Ph.D. Student, Hasso Plattner Inst.

In this talk, I’ll highlight shortcomings of existing methods for embedding initialization, present our recent paper Token Distillation on how to tackle some of these, and give my thoughts on the future of embedding initialization research.

I‘m a PhD student at the ELLIS Unit Potsdam and Hasso Plattner Institute advised by Gerard de Melo and Desmond Elliott (University of Copenhagen). My research focuses on multilingual LLMs and embeddings, particularly embedding initialization for tokenizer transfer.

Add event to calendar

Apple Google Office 365 Outlook Outlook.com Yahoo