linear ramp
If you just make a current, Vbg/R, and sum together 14 bits of switched binary currents into another R of same type, it will work.
However, I thought that super-duper layout technique could only get you 12 bits free. For 14 bits, they usually have to trim or use barrel-shifted LSB currents to spread the mismatch over the whole range. But that doesn't guarantee each bit is always accurate, it just spreads the mismatch.
What's the use of the ramp? Maybe a little varying mismatch won't bother it.