How random is the division of images?

[chrisgoringe]

Making use of your excellent collection of scored images, I noticed that almost any model I trained on the train_set images performed better on the test_set than it performed during training.

Using excel on the csv files, I found this:

train_set.csv - stdev of scores = 1.646 
eval_set.csv - stdev of scores = 1.095 
test_set.csv - stdev of scores = 0.6074 
combined - stdev of all scores = 1.548

It appears that some hidden factor has biased the division of the images into train, eval, and test sets?