Personal choice, is what I'd say. You can get a lot of mileage out of implementing a dynamic language with NaN boxing[1].
It really depends on the kind of language you're trying to build an interpreter for, and what purpose it could serve. For dynamic languages, I'd say looking at the core types of Erlang is a great place to start (Integers, Symbols, Functions, etc.). For a statically typed language things get more complex, but even with just numeric types, characters, and some kind of aggregating type like a Struct you can build up more complex data structures in your language itself.
It was my first time making a language, so I built into the language whatever was needed to build something cool with Crumb.
Wanted first class functions to simplify the parse step (they can be treated like any other value), but I needed a different mechanism to invoke “native code” vs user-defined methods, so there’s two different types for that.
Needed some kind of compound data type, but I didn’t want to deal with side effects from pass by reference, so Crumb implements lists, but they are always pass by value :)
It really depends on the kind of language you're trying to build an interpreter for, and what purpose it could serve. For dynamic languages, I'd say looking at the core types of Erlang is a great place to start (Integers, Symbols, Functions, etc.). For a statically typed language things get more complex, but even with just numeric types, characters, and some kind of aggregating type like a Struct you can build up more complex data structures in your language itself.
[1]: https://leonardschuetz.ch/blog/nan-boxing/