Feature extraction is expensive, and sometimes we want to re-run training and data collection with only slight variations. In these cases, it's convenient that any existing w_cache file be supplied to extraction, and we can reuse most of the data.
I'm imagining something like,
aawiki.labeled.w_cache: aawiki.labeled revscoring extract --cache=$@ --input=$^ > .tmp.$@ && mv .tmp.$@ $@