weight-decay-plasticity

Improve downstream task performance by increasing weight decay during pretraining (0.3-1.0 vs. default 0.1), enabling better model plasticity and representation structure.

Independent GitHub snapshot · Automated classification · Original documentation governs