EvalPort: portable interchange for the MATH dataset schema #34

Hi MATH dataset team — I maintain EvalPort (https://github.com/adhabnr-ux/evalport), an open JSON-Schema spec for portable LLM eval test cases, graders, and result sets.

The per-problem JSON schema your modeling/eval_math_gpt.py::get_level_type() reads (problem, level, type, solution) maps cleanly onto EvalPort's TestCase (input=problem, expected_output=boxed answer extracted from solution, metadata=level/type), and math_equivalence.is_equiv() is a ready-made Grader function (numeric/symbolic equivalence rather than exact string match — exactly the kind of grader EvalPort's GraderResult is meant to carry). A thin converter would let MATH-derived evals travel between tools without each one re-implementing is_equiv. Flagging as an option given how often MATH is used as a baseline.

Schemas: https://github.com/adhabnr-ux/evalport/tree/main/spec/schemas

— Sahi, independent contributor (not affiliated with the MATH dataset)