Do LLMs estimate uncertainty well in instruction-following?