You should learn a little bit more about machine learning training. The “training” you describe is not particularly useful for machine learning. Data scientists want to see whether the actual device at the outlet or socket is on or off (ground truth) at every sample point (with Sense, every microsecond). Better yet, add in a second piece of data, the rough power that the device is consuming. The on/off data can be offset in time by a second or so, since the machine learning network will have some “short-term memory” built into it.
The good news is that there are lots of readily network accessible sources of power usage ground truth data to feed machine learning. Hue is one good example. Other sources include smart thermostats (Ecobee, Nest), smart plugs (provide power information as well), smart chargers for/in EVs, smart appliances, etc. And, not every device needs to report in for Sense to improve with this additional supervision / feedback. Other devices are likely to be better defined if Sense more accurately pegs devices with ground truth input. So ground truth feedback should be good for everybody. The bad news it that you need a machine to provide training input, not yourself.