Calculate the number of trainable parameters in a model with frozen layers.
When fine-tuning, you often freeze most layers and only train a small subset. LoRA typically trains <1% of total parameters. This calculator shows how many parameters are actually updated during training based on the frozen percentage.
Trainable Params
trainable = total_params × (1 - frozen_pct/100)
Freezing pretrained layers reduces compute, memory, and risk of catastrophic forgetting. Only training top layers or adapter modules (LoRA) achieves good results with far less resources.
LoRA adds small low-rank matrices alongside frozen weight matrices. With rank=8 on a 7B model, you might only train 0.1% of total parameters while achieving near-full fine-tuning quality.