well to be honest what you're aiming for is the brightest part of the flat to be as bright as possible while still remaining in the linear range of the sensor.
i think people go for the median at 50%, or visually, the middle of the histogram hump at 50%.
people use 50% as a rule of thumb since probably every sensor out there is still linear at 1/2 well capacity, with enough headroom to allow the right half of the histogram to still fall into the linear range of the sensor. to really know for sure you need to use a light source for which you can control the brightness accurately, then increase the exposure times and plot the ADU vs. exposure time. the graph will be linear up to a point, and then diverge. where it 'breaks' is the ADU you need to stay below for proper flats.
again if the flat is from an OSC and one channel is not exposed well the statistics are going to look funny. that's why it probably makes sense to debayer to see what's going on with each channel individually.
rob
