yolo5s model size is not equal to sparseml's yolov5s baseline version

[kuonumber]

I follow codes below with custom dataset to train model.
%run ./train.py --img 640 --batch-size 12 --epochs 3 --data "D:\yolov5\data\custom.yaml" \ --cfg ../models/yolov5s.yaml --weights "" \ --workers 0 --name 20220321_640_20b_3e_digitals \ --multi-scale --device 0 --hyp ".\data\hyp.scratch_2.yaml"
But I can not reproduce the size of weights(about 24.8 MB), what I get is around size 14 MB.

Do you have any idea about this?