[Preface: While the basic question may seem to be off-topic, as it's about today's use of COBOL, I like to see it as asking in help to educate about COBOL. The noble effort to keep the language alive and spoken might be worth an exception]
What is a good COBOL specification to target
Quick answer: The one your customer already uses.
In software engineering it's almost never about the tools or great ideas, but to fit in. This is especially true in long running settled environments. Here the use of COBOL may be a safe indicator that it's not only a long running one, but also a huge code base one got to fit in. In fact, it's usually way less effort to learn any new language for a project than it needs to lean the environment.
Having said that, I understand the question may be about what to suggest to students, or what to use as a base for teaching.
If that's the case, then it's all about sticking to the basics. Basic language structure, concepts and data definition.
Especially as the later is quite often done in custom, project specific tools, making it even more important to understand what the language (the compiler) sees (*1). The Mainframe/COBOL world is not ruled by hello-world type command line tools, but highly integrated packages sitting on top of what might be 'the language'. Macro like copy chains and preprocessors - of such I may have seen way more for COBOL than languages are on Rosetta Code.
So if it's strictly about teaching the basics to enable people to grow into understanding and maintaining COBOL code COBOL85 (including the 1993 revision) might be the best bet. It is a modest but useful extension of COBOL74 and made to cover everything learned since then without sacrificing compatibility.
Using COBOL85 will give the best over all base to understand the language and how features originated. Also, pre COBOL85 sources are rare, even in real old applications (read financial/accounting).
Beside that, a good peek should be taken at COBOL2002, as it does standardize many things essential today, even for COBOL, like wide character sets (Unicode) or new data types, which are quite important for interfaces to non COBOL applications.
A guideline about what newer feature is worth to be included in a course might be to separate between function and paradigm:
New Functionality (like charsets oder datatypes) is quite welcome by maintainer of existing applications, as it simplifies their life without making them throw away 50 years of codebase. Thus it helps if a student has basic knowledge.
New Paradigms may not be worth the time - or worse add confusion. True, Object COBOL new ways of functions and so on have added much, but are only worth it when doing rather large restructuring work or new projects - both not as common in COBOL. Students will learn the basics about OOP anyway, so no need to add the COBOL way. That's something they can do on their own when needed later on.
Long story short: Use a solid foundation of COBOL85 (1993 revision) and useful functional parts of later standard to give your students a fine base to ask the right questions and deliver the right answers when applying for a COBOL based job.
*1 - The relation of structures in a data division (and the way PERFORM works - the stackless structure in general) may be the most important part to teach. Spend more than just an hour on that, as it's absolute alien to today's students, grown up in an ALGOL mindset.