README.md

metadata

dataset_info:
  features:
    - name: images
      sequence: image
    - name: question
      dtype: string
    - name: label
      dtype: string
    - name: task
      dtype: string
    - name: num_images
      dtype: int64
    - name: image_names
      sequence: string
    - name: injected
      dtype: bool
    - name: object_counts
      dtype: string
  configs:
    - config_name: count
      data_files:
        - split: train
          path: data/count.parquet
    - config_name: existence_adversarial
      data_files:
        - split: train
          path: data/existence_adversarial.parquet
    - config_name: existence_popular
      data_files:
        - split: train
          path: data/existence_popular.parquet
    - config_name: existence_random
      data_files:
        - split: train
          path: data/existence_random.parquet
license: cc-by-4.0
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - multi-image
  - hallucination
  - benchmark
  - vision-language-model
  - multimodal
size_categories:
  - 1K<n<10K

MIHBench

Multi-Image Hallucination Benchmark for evaluating multi-image understanding in MLLMs. 3,200 samples across 4 tasks (800 each), with each sample containing 2-4 images from COCO.

Fields

Field	Description
images	2-4 images (list)
question	Natural language question about the images
label	Ground truth: "yes" or "no"
task	Task identifier
num_images	Number of images in the sample
image_names	Source image filenames

Additional fields for count task: injected (bool), object_counts (JSON string).

Tasks

Task	# Images	Description
count	2	Same number of target object in both images?
existence_adversarial	3	Target object exists in all images? (rare/confusing objects)
existence_popular	3	Target object exists in all images? (common objects)
existence_random	3	Target object exists in all images? (random objects)

Evaluation

metrics: Accuracy, Precision, Recall, F1
parser: yes/no binary

Source

Original data from MIHBench (ACM Multimedia 2025).