So was Caml (and its precursor Le-ML), originally :)

That it grew so far beyond that I think is a consequence of it not having a formal definition like SML. This allows the developers to extend the language willy-nilly without going through the grueling task of extending a formal semantics. On top of that, as another commentor pointed out, is the fact that there's a single "reference implementation", which means that users don't need to worry about getting locked-in when opting to use new functionality.

That said, SML is still a fantastic base for new research projects in the PL space, mostly for the very same reasons it hasn't seen nearly as much adoption in the "practical" space as Caml: it's a small, simple, well-understood, and formally-specified language that grants researchers an excellent starting point for their inquiries and experiments.