I recommend the 6502[1] if you're just starting. What you learn applies almost everywhere, and it's very much a what-you-write-is-what-happens sort of processor. No branch prediction, no caches, no fancy features. Its op-set is very small and highly expressive. They still make them, in fact.
That said, whether you choose to start with it or RISC-V, from personal experience SiFive's implementations are phenomenal. I used one of their free-implementations, and I was blown away by how easy it was to use their automated tools to build a core exactly the way I wanted. Not as easy as generating a NIOS core, but much more capable. With an Artix-7, I had a very capable low-range processor in about an hour.