Found 1 document(s)
Page 1 of 1
-
In this work, we propose a unified framework for training algorithms across different stages of modern large language model development, based on the perspective of distribution matching. We show that a broad class of la...