Implementation:Open compass VLMEvalKit Physics Utils

Field	Value
source	VLMEvalKit
domain	Vision, Evaluation, Physics, LaTeX Answer Extraction

Overview

Provides physics question evaluation utilities with LaTeX boxed-answer extraction and symbolic equivalence checking.

Description

This module implements `build_physic_prompt` which constructs detailed physics problem prompts requiring answers in `\boxed{}` format. The `PHYSIC_auxeval` function evaluates physics answers by extracting boxed content from model responses using `extract_final_answer_allform`, then checking equivalence against ground truth via direct matching, symbolic comparison using sympy's `simplify`/`expand`/`trigsimp`, and LaTeX parsing. It supports both text and mathematical answers.

Usage

Called internally by the corresponding dataset class during evaluation.

Code Reference

• Source: vlmeval/dataset/utils/physic.py, Lines: L1-103
• Import: from vlmeval.dataset.utils.physic import build_physic_prompt, PHYSIC_auxeval

Key Functions:

def build_physic_prompt(line): ...
def PHYSIC_auxeval(model, line): ...

I/O Contract

Direction	Description
Inputs	A data line dict with 'question', 'prediction', and 'answer' fields
Outputs	Dict with 'log' (evaluation trace) and 'res' (boolean equivalence result)

Usage Examples

from vlmeval.dataset.utils.physic import PHYSIC_auxeval

result = PHYSIC_auxeval(model, line)

Related Pages

• Principle:Open_compass_VLMEvalKit_Benchmark_Dataset_Construction