"So, in contexts where the top token is 6%, a Min P of 0.1 will only consider tokens that are at least 0.6% probable. But if the top token is 95%, it will only consider tokens at least 9.5% probable."
Isn't there a big post on automatic temperature scaling?