I would build a more balanced training set and you have to realize that your resultant probabilities for churn will be artifically higher.

To validate, use a testing set with the original distribution of data to get an accurate portrayal of the performance of the model.

You could also try both and compare the results.
[quoted text, click to view]