Understanding set_transform

Great, thanks for confirming that it doesn’t cache to disk. That’s exactly what I was hoping.

I guess now I’ll have to update to the latest masters and start testing how much on-the-fly tokenization and other data transforms slow my training down.