PEFT / LoRA Adapter Simulator

Parameter-efficient finetuning (PEFT) methods like LoRA let large language models be adapted to new tasks without touching most of their weights. This simulator visualizes the core LoRA mechanism directly: a frozen weight matrix W sits untouched while a trainable low-rank pair A (r×d) and B (d×r) is multiplied together every frame to form the correction ΔW = (α/r)·B·A, which is what actually gets added to W during a forward pass. Adjust the rank r, the base matrix dimension d and the scale α to see how the trainable-parameter count, its share of the full matrix, and the compression ratio change — the same trade-off that lets LoRA finetune a multi-billion-parameter model by training a few million numbers.