Test giving better results than training set

I'm consistently getting better results on my test set than I am on my training set.

Split is about 80%/20%, and I'm running a classifier on about 600 classes.

The difference isn't huge. From memory, this means something, but I cannot for the life of me work out what it is.