opentau.datasets.vqa.vsr

VSR (Visual Spatial Reasoning) dataset for true/false statement vqa.

This module provides the VSR dataset implementation for training vision-language models on visual spatial reasoning tasks. The dataset contains images with statements about spatial relationships, and models must determine whether each statement is true or false based on the image content.

Classes

VSRDataset(cfg[, consecutive_bad_tolerance])

Visual Spatial Reasoning (VSR) dataset for true/false statement vqa.

class opentau.datasets.vqa.vsr.VSRDataset(cfg: TrainPipelineConfig, consecutive_bad_tolerance=100)

Bases: VQADataset

Visual Spatial Reasoning (VSR) dataset for true/false statement vqa.

Loads the cambridgeltl/vsr_random dataset from HuggingFace and formats it for visual reasoning tasks where models must determine if statements about images are true or false.

__init__(cfg: TrainPipelineConfig, consecutive_bad_tolerance=100)