Implementation:Open compass VLMEvalKit VGRPBench Common Puzzle Factory

Field	Value
source	VLMEvalKit
domain	Vision, Evaluation, Puzzle_Solving

Overview

Provides the base PuzzleFactory class that all VGRPBench puzzle types inherit from.

Description

This module defines the abstract base PuzzleFactory class used by all puzzle-specific factories in VGRPBench. It provides shared infrastructure including a constraint list, hint sampling via random cell selection, puzzle saving with train/val/ablation dataset splitting based on unique solutions, and a unified check method that iterates through all registered constraints. The hint_type utility function parses command-line hint arguments accepting either "random" or integer values.

Usage

Called internally by VGRPBench evaluation pipeline for puzzle constraint verification.

Code Reference

• Source: vlmeval/dataset/utils/vgrpbench/puzzles/common_puzzle_factory.py, Lines: L1-136
• Import: from vlmeval.dataset.utils.vgrpbench.puzzles.common_puzzle_factory import PuzzleFactory

Key Classes:

class PuzzleFactory:
    def __init__(self) -> None: ...
    def sample_hints(self, board, num_sample_hints): ...
    def save_puzzles(self, puzzles, save_path, filename): ...
    def check(self, game_state): ...
    def get_possible_values(self, game_state, row, col): ...

Key Functions:

def hint_type(value): ...

I/O Contract

Direction	Description
Inputs	Puzzle grid state with board and constraint parameters; puzzle lists for saving
Outputs	Boolean validation result from constraint checking; JSON dataset files for train/val/ablation splits

Usage Examples

# Internal usage - subclassed by specific puzzle factories
from vlmeval.dataset.utils.vgrpbench.puzzles.common_puzzle_factory import PuzzleFactory
class MyPuzzleFactory(PuzzleFactory):
    def __init__(self, size):
        super().__init__()
        self.constraints = [...]

Related Pages

• Principle:Open_compass_VLMEvalKit_Benchmark_Dataset_Construction