Data is data. As long as it's appropriately labelled, it's probably better to keep it. I believe self-play games are the most expensive part of this project. Would be a shame to throw it away.
Is it kept out of the 250k training window if the network is too old though? (I agree no reason to through out, but we don't want it to dilute our window either)
This is also bad if network connectivity is out for the weekend (it keeps on generating self play from the last available network), and the data all gets uploaded 3 days later when it may be many networks behind the latest.
Most helpful comment
Data is data. As long as it's appropriately labelled, it's probably better to keep it. I believe self-play games are the most expensive part of this project. Would be a shame to throw it away.