Imbalance dataset

[Huy-Ngo]

The dataset has more train example with false value than true. Modify the dataset to resolve this.

[trahoa]

One solution is to double true examples: the json2tsv script should make a copy of any true example and shuffle it into a random position. This is for the training set only!

[Huy-Ngo]

I've just done that. The training dataset now has 6350 trues and 6835 falses. That seems balanced enough, so I'm closing this issue

[McSinyx]

Since no benchmark done on this naïve solution, I'm reopening this just to discuss better alternatives if necessary.

[McSinyx]

Apparently it is better handled by cbc4a09.