04 · CAPACITY

Verified

Rank, Alpha, and Target Modules

Separate capacity, scale, and intervention location; change one variable at a time.

QUICK LOOK
  • r = capacity
  • alpha = scale
  • α/r standard, α/√r rsLoRA
  • Change one variable at a time

KNOW THESE FIRST

B (d_out × r)A (r × d_in)r = rank | α = scale (α/r LoRA, α/√r rsLoRA)CONCEPT DIAGRAM

The budget, window, and …Rank r determines adapter capacity and the r × (d_in + d_out) parameter budget. Alpha controls alpha/r in standard LoRA and alpha/√r in rsLoRA.

Target modules decide which real attention and MLP matrices receive adapters. Names are architecture-specific and must be verified on the model.

01

First thought

The same alpha/r means the same adapter capacity.

02

Correction

Effective scale may match while a higher rank still carries more parameters and capacity.

03

Decision rule

Start at r=8/alpha=8; if capacity is insufficient, hold everything else fixed and test r=16/alpha=16.

CONCEPT DEPTH

Read the same concept at three depths. Pick a level, switch instantly. Recommended starting point for university students: UNDERGRAD.

UNDERGRAD

Rank r is the inner dimension of A ∈ R^(r×d_in) and B ∈ R^(d_out×r). Parameter count scales as r × (d_in + d_out). Alpha is the scale factor: α/r in standard LoRA, α/√r in rsLoRA. Even if effective scale is held constant, higher rank carries more capacity and increases overfitting risk.