LoRA: Low-Rank Decomposition, Rank r, and Alpha

~15 min read

LoRA decomposes a weight update into two small matrices A and B, governed by two key hyperparameters — rank r (how much capacity the adaptation has) and alpha (how strongly it's applied).

LoRA: Low-Rank Decomposition, Rank r, and Alpha is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • LoRA decomposes a weight update into two small matrices A (d×r) and B (r×k), training only these while the base weight W stays frozen
  • Rank r is the capacity dial — smaller r means far fewer trainable parameters than full fine-tuning
  • A is Gaussian-initialized, B is zero-initialized — guaranteeing A×B=0 at the start, so the model's original behavior is exactly preserved before any training
  • Alpha is a separate scaling factor controlling how strongly the adaptation affects the output, independent of r's capacity
  • r controls how expressive the adaptation CAN be; alpha controls how strongly that expressiveness is actually applied