Skip to content

Comment on Vera: Vector-Based Random Matrix Adaptation

Comments

VeRA makes LoRA ~10x more parameter efficient while retaining the same performance.

It's somewhat like a recursive LoRA scheme, where the LoRA A and B matrices are also decomposed using two small trainable vector parameters.

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.