Open Models Almanac
Methodsoptimizationtraining stability

single mention · optimization

Per-token gradient-norm bounding

Bounds the gradient norm separately for individual tokens to prevent unusually large token-level gradients.

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.

Also called Per-token gradient-norm bound, gradient-norm bound.

Filed under optimization › training stability