[Eval] Robust CPU-based Efficient MATH grader

[Nativeatom]

TL; DR

Extending the existing CPU-based MATH grader to more broader cases with larger testing samples (considering the string match nature of many libraries). Specifically

• 1000 new MATH samples introduced (10x more).
• White space insensitive to expressions. For example (1,2) = ( 1, 2 ), 1+2i = 1 + 2 i.
• Robust interpretation of numerical value separators. For example, $1,000,000$, $1,\\!000\\!000$ and $1000000$ are identical.
• Insensitive to leading zeros. For example, $.12$ = $0.12$.
• Fraction and decimal with identical values. For example, $\frac{1}{2}$ = $0.5$.
• Redundant symbol wrappers. For example, $$2$$ and $2$.
• Insensitive to redundant text wrappers. For example, \\text{Amy}, $Amy$ are equivalent to Amy.
• Latex equivalent variants. For example, \\dfrac and \\tfrac should be reduced and equivalent to \\frac.
• The latex bracket wrapper. For example, \\left[ X, Y \\right], \\left(X, Y\\right) should be reduced to [X, Y] and (X, Y) respectively. Besides, it also obey the previous white-space insensitivity: \\left( X, Y \\right) is equivalent to (X, Y).
• Extra multiple choice variants. For example, $E$, (E), $(E)$, \\text{E}, \\text{(E)} should be reduced and equivalent to E after extraction.
• Tolerate the missing base if the value matches the reference. For example, $40$ is correct if the reference is $40_{9}$.

Performance

grader	cold cache	warm	agreement
minerva	1438.06 ms	907.24 ms	999/1000
hendrycks	2002.91 ms	775.03 ms	999/1000

The missing agreement is due to evaluating answers with multiple values, see Notes

Notes

• The order agnostic comparison of product is still missing due to the string-match limitation. For example, (a+2)(b+1) and (b+1)(a+2).
• For answer with multiple values, due to the parse_latex design of sympy, the ANTLR parser parses the first valid sub-expression and silently ignores the trailing input. The grammar has no top-level comma production. And the remaining answer would be dropped.

Testing

New testing script in

python scripts/benchmarks/benchmark_math_graders_extend.py

Part of marin-community/marin#7463, extended from #58.

[penfever]

🤖 Requesting changes: the revised graders are not wired into shipped math benchmarks, which still import lm-eval graders. Integrate them into the intended task path, or narrow the PR to a non-production utility with accurate scope. Required code CI also has not run.