That'd be sweet. If one who can count 19x19 perfectly at a glance, I expect he'd be mislead on 17x17 because he'll undervalue the territory near the edge. That said, maybe it's best to only include 19x19 in the training set.

Also, some games are much easier to count than others. Perhaps have increasing difficulty levels. Some endgame positions require much reading to discern which groups are dead. But having to read these large kills distracts from the goal of practicing counting (esp. counting at a glance). But this reading is not required in our games -- we are intimately familiar with the life/death status of every group already.

Also, in pro games, there are often many empty points (technically dame, but they don't look like textbook examples of dame), which don't show up in the endgame of 10k games.

So maybe it'd be best to explicitly mark dead groups. And maybe it'd be best to make three difficulty levels. At 20k, the games should include large uninvaded moyos. 10k boards should look kind of reasonable, with some dead groups in places where a 10k might lose a group due to carelessness or poor reading. The 1d boards should look more like professional games.

Anyway, that's my two cents.