Implementation:Open compass VLMEvalKit MEGABench LaTeX Expr Equality

Field	Value
source	VLMEvalKit
domain	Vision, Evaluation, Mathematics, LaTeX Parsing

Overview

Determines equality between LaTeX mathematical expressions using symbolic parsing for the MEGA-Bench evaluation framework.

Description

The `LatexExprEquality` class provides a `match` method that normalizes and parses two LaTeX expression strings using sympy's `parse_latex`, then checks symbolic equality. It includes a configurable timeout (default 15 seconds) using signal-based alarm to prevent hanging on complex expressions. On parsing failure (LaTeXParsingError, SympifyError, TypeError), it falls back to `SimpleStrMatch` for direct string comparison. The `separate_text_and_latex` helper splits mixed text-and-LaTeX content for individual comparison.

Usage

Called internally by the corresponding dataset class during evaluation.

Code Reference

• Source: vlmeval/dataset/utils/megabench/scoring/latex_expr_equality.py, Lines: L1-96
• Import: from vlmeval.dataset.utils.megabench.scoring.latex_expr_equality import LatexExprEquality

Key Functions:

class LatexExprEquality:
    @classmethod
    def match(cls, responses, targets, timeout_duration=15) -> int: ...

def separate_text_and_latex(text): ...

I/O Contract

Direction	Description
Inputs	Two LaTeX expression strings for comparison
Outputs	Integer 1 (equal) or 0 (not equal)

Usage Examples

from vlmeval.dataset.utils.megabench.scoring.latex_expr_equality import LatexExprEquality

score = LatexExprEquality.match("\\frac{1}{2}", "0.5")

Related Pages

• Principle:Open_compass_VLMEvalKit_Benchmark_Dataset_Construction