Note 500K, too much, random sampling to 10K for training set.

Note Use subset="th", Using valid dataset as the training set (100 rows).