DARE Dataset Overview

Dataset Information

Features

• id: String type
• instance_id: 64‑bit integer type
• question: String type
• answer: List of strings type
• A: String type
• B: String type
• C: String type
• D: String type
• category: String type
• img: Image type

Configurations

• 1_correct:
  • validation: 1_correct/validation/0000.parquet
  • test: 1_correct/test/0000.parquet
• 1_correct_var:
  • validation: 1_correct_var/validation/0000.parquet
  • test: 1_correct_var/test/0000.parquet
• n_correct:
  • validation: n_correct/validation/0000.parquet
  • test: n_correct/test/0000.parquet

Dataset Description

DARE (Diverse Visual Question Answering with Robustness Evaluation) is a meticulously created and curated multiple‑choice VQA benchmark. DARE evaluates visual‑language models on five distinct categories and incorporates four robustness assessments based on:

• Prompt
• Subset of answer options
• Output format
• Number of correct answers

The validation set includes images, questions, answer options, and correct answers. The test set’s correct answers are withheld to prevent contamination.