Inconsistency in libsvm formats

[mhaseebtariq]

I found out that when I trained a model using the distributed version of the JVM package - the libsvm format for the indexes has to be one-based ascending order. Whereas, when I used the saved model in the Python package it expected the libsvm format to be zero-based - which resulted in different predictions in the Python package. To work around this you have to modify your libsvm format accordingly.

[tqchen]

Mainly due to the fact that spark does the internal conversion, and do not accept format that are 0-based. So the labeledVector transformed from spark's input was already transformed. While xgboost itself do not run any transformation

[mhaseebtariq]

@tqchen Thank you for the clarification. So this should be mentioned in the README, right? I spent some time before realizing this - as I was getting very different results while using the saved model (distributed training) in the Python package. I think mentioning this somewhere might help new users. I have made another change in README with your clarification.