Yes it's close the to mean with a little noise since it doesn't use the whole data but a subsample, after running the base notebook, thats what I get:

[ 70.79933333, 234.18273333, 160.7572 , 6.79858 ]

Simply either divide the labels by that during training or divide your submission preds, both work :)

I just want to point out that this is incorrect, the starter submission file uses:

70.3026558891455,227.9885103926097,159.28123556581986,6.782719399538106