Benchmark locally returns worse values than published

I have noticed that my own benchmarks gave slightly worse numbers than published in SOTA, to check why, and how - as i need to publish some numbers - in installed torchbench in my local docker/conda
Ubuntu 18.04, TitanRTX, pytorch 1.5.1
I have local ImageNetILSVRC2012 with val (also train and test) subfolders and ILSVRC2012_devkit_t12.tar.gz

To run a simple validation test. I ran pretrained ResNet-18, both from torchvision and from pytorch-image-models. only modifying the model source and data_root in script from https://github.com/paperswithcode/torchbench
ie, same transforms

Why am I getting *** Acc@1 0.695 Acc@5 0.891** but SOTAbench reports. 73.3% 91.8%

I do have the good data. val folder has 1000 subolder classes, each with 50 JPEGs and 50 xmls with names in this format ILSVRC2012_val_00000293.JPEG

Any ideas?
thank you

Please Withdraw topic_ I noticed my error.
ResNet-18 benchmark that I was citing uses, SSL weights. pytorch-image-models ssl_resnet18

thanks