A light-weight codebase that enables memory-efficient and performant finetuning of Mistral's models. It is based on LoRA, a training paradigm where most weights are frozen and only 1-2% additional weights in the form of low-rank matrix perturbations are trained.
"The paper introduces a technique called LoReFT (Low-rank Linear Subspace ReFT). Similar to LoRA (Low Rank Adaptation), it uses low-rank approximations to intervene on hidden representations. It shows that linear subspaces contain rich semantics that can be manipulated to steer model behaviors."
