ce-models | ce-models plus questions and answers to support Hudson | Machine Learning library

z3's model printing around arrays have been in flux recently; and different versions do print the models in different ways. (For instance, see https://github.com/Z3Prover/z3/issues/5604)

If I try your program using the latest z3 from GitHub master, then it prints:

I have to admit that I'm not completely sure that I completely understand what you want to achieve here, but here are the answers to your questions:

It seems each proof obligation is in a different format; Lisp and OCaml? What exactly is the format?

These files represent the formulas that are given to the provers you ask Frama-C to launch. The format depends on the prover. If I recall correctly, for many provers, this will be smtlib, or tptp, but some provers such as Alt-Ergo can also enjoy a custom output. The generation of the file is described by Why3's driver files, as mentioned (quite briefly) in section 12.4 of the Why3 manual.

Is it correct that these are proof ?

No, these are formulas to be validated by the prover they have been generated for.

and are sufficient to show the contract/proof is correct without proving that Z3, alt-ergo, etc are correct?

No. If the provers you use have a bug, they might mistakenly tell you that a given proof obligation is valid. Some provers are able to provide a proof trace (e.g. if you tweak a driver to use the get-proof command of smtlib), but as far as I know, the format of such a trace is prover-specific, so that it would probably be difficult to have it checked by an external tool.

SMTLib language does not have a mechanism to retrieve "all-models." So, if you're bound to be using SMTLib only, you cannot do this; at least not easily.

However, most solvers (definitely including cvc4 and z3) can be scripted from higher-level languages. The idea is to make a check-sat call, and if you get a solution, you add an additional assertion that disallows that model, and query for a new one. See this answer for how to do this in z3, as scripted from Python: Trying to find all solutions to a boolean formula using Z3 in python. You can do the same from C/Java etc.; or use a higher-level binding that provides such a command out-of-the box.

You didn't tell us what exact error you're getting, so I'm assuming it comes from z3 and not from some other part of your toolchain. Let's run the SMTLib program you posted through z3. When I do that, z3 tells me that there's a "sort mismatch:"

Your repository does not contain the required files to create a tokenizer. It seems like you have only uploaded the files for your model. Create an object of your tokenizer that you have used for training the model and save the required files with save_pretrained():

You cannot rely on the model if the solver returned unknown. That is, the model you get is not "partial" by any means: It may or may not satisfy some of the constraints, but otherwise, there's not much you can do with it. It's at best a representation of the internal state of the solver. But in general, it is not guaranteed to be a representation of anything.

Also, when I run your SMTLib script with z3, I get:

Got an answer from Andrew Reynolds on the CVC4 mailing list:

First, notice that we have updated to a new API (https://github.com/CVC4/CVC4/blob/master/src/api/cvc4cpp.h).
Coincidentally, I just submitted a PR that adds the interface for getting the required constructor you are looking for, in the new API: https://github.com/CVC4/CVC4/pull/4817

If you are interested in the old API, your code is almost correct, however, there is a subtle difference in what we expect to be cast.

In particular, in SMT-LIB casts are applied to constructor operators, not terms (you may be interested in this discussion https://github.com/Z3Prover/z3/issues/2135). This means that the AST of a casted nil term in CVC4 is: (APPLY_CONSTRUCTOR (APPLY_TYPE_ASCRIPTION None T)) not (APPLY_TYPE_ASCRIPTION (APPLY_CONSTRUCTOR None) option[Int]) Unfortunately, there is a complication in the old API about what T is. It is not "option[Int]", instead it is "the type of constructors of type option[Int]", or what we call a "constructor type".

Here is the corrected code for creating the expression:

This seems to be a bug in SBV. Reporting it at the github repo is the right thing to do.

Note: Should be fixed as of this commit. Please give it a try!