Datasets:

MM-Hallu
/

FIHA

dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_name
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: long_answer
      dtype: string
    - name: type
      dtype: string
    - name: evaluation
      dtype: string
    - name: root
      dtype: string
    - name: source
      dtype: string
  configs:
    - config_name: default
      data_files:
        - split: train
          path: fiha-*.parquet
license: mit
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - hallucination
  - fine-grained
  - foggy
  - coco
size_categories:
  - 10K<n<100K

FIHA

Fine-grained Hallucination Evaluation benchmark with 24,799 QA pairs across MSCOCO (500 images) and Foggy Cityscapes (150 images) datasets.

Fields

Field	Description
image	Input image
image_name	Original image filename
question	Question about the image
answer	Short answer (yes/no)
long_answer	Detailed answer
type	Question type (yes/no)
evaluation	Evaluation dimension (object_existence, attribute_existence, etc.)
root	Root object being queried
source	Image source (mscoco/foggy)

Source: confidentzzzs/FIHA