Open Models Almanac
Methodsoptimizationtraining stability

single mention · optimization

Loss masking for excessively stale tokens

Removes tokens with excessive staleness from the loss contribution to mitigate stale-sample effects on gradient updates.

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

Filed under optimization › training stability