Open Models Almanac
Methodstraining objectivedistillation objective

single mention · training objective

Forward-KL distillation with a sink token

Forward KL is used with an extra sink token to represent probability mass outside the retained top-128 tokens.

The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See distillation objective for the methods filed beside it.

Filed under training objective › distillation objective