The top submission was the same on for the public and the private scores and consisted of a the arithmetic average of other submissions. It was created using blender.py (see ensemble folder).
The submission consisted of 5 runs created using run.py (see code folder) with different seeds and slight variations in the model parameters: learning rate, feature_fraction and regularization (L1 and L2).