README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: choices
      dtype: string
    - name: category
      dtype: string
    - name: source
      dtype: string
    - name: type
      dtype: string
  configs:
    - config_name: default
      data_files:
        - split: test
          path: data-*.parquet
license: mit
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - misleading
  - vqa
  - hallucination
  - benchmark
size_categories:
  - 1K<n<10K

MVI-Bench

Misleading Visual Input Benchmark with 1,248 expert-annotated VQA instances across 6 visual illusion categories (resemblance, representation, material, mirror, occlusion, illusion).

Fields

Field	Description
image	Input image (potentially misleading)
question	Question about the image
answer	Ground truth answer
choices	JSON-encoded list of options
category	Visual illusion category
source	Image source type (edited/synthetic)
type	Question type (normal/adversarial)

Source: kittyhy/MVI-Bench