Crunch 3 eval random forest input: raw or log counts? normalized or unnormalized counts? what sequencing depth?

Competition Broad Institute Crunch #3
1

My understanding is that Crunch 3 will be scored primarily based on the performance of random forest classification trained on 50 genes per submission. Will the random forest be trained on raw counts, or on log-transformed counts? Will the counts be normalized per cell, and/or standardized per gene? At approximately what depth will the libraries be sequenced (approximately how many reads per cell)?

Thanks in advance!