LoRA

Low-Rank Adaptation • Interactive Demo

LoRA freezes the original weights W and learns a low-rank update ΔW = B × A. This dramatically reduces the number of trainable parameters.
1.5
A (8×r)
Low-rank adapter A
B (r×8)
Low-rank adapter B
W
Original
+
αΔW = α(A × B)
Scaled update
↓
W′ = W + αΔW
W′
Adapted weights
Low
High
Edit values in A and B to instantly see how the low-rank update changes ΔW and the final model.