single mention · training objective
LK loss
The loss used for MTP boosting in place of a temperature-scaled forward-KL objective.
The sources on this page only mention it in passing. The line above is the extraction's one-line gloss. See multi-token prediction objective for the methods filed beside it.
Also called LK loss for MTP boosting, we now use the LK loss.
Filed under training objective › multi-token prediction objective