LoRA

Formula

$$h=(W+BA)x,B\in {\mathbb{R}}^{d\times r},A\in {\mathbb{R}}^{r\times k},r\ll \min (d,k)$$

$W$ is frozen; only $A$ and $B$ are trained: $r(d+k)$ instead of $dk$ parameters (100× to 1000× fewer).

Adapters can be swapped, merged or ablated at inference. For safety this means changing an open model is cheap: LoRA removes Llama 2 70B’s safety for <$200, and a single rank-1 LoRA on the MLP down-projections induces emergent misalignment.

Appears in

• Lecture 4, LoRA
• Lecture 4, EM with rank-1 LoRA

Related concepts

• Emergent Misalignment
• Supervised Fine-Tuning