Honestly, I was expecting this to turn out badly, the dataset was horrible.

Works best at 0.75 weight