README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: image_id
      dtype: string
    - name: question
      dtype: string
    - name: answer
      dtype: string
    - name: category
      dtype: string
  configs:
    - config_name: default
      data_files:
        - split: train
          path: repope-*.parquet
license: mit
task_categories:
  - visual-question-answering
language:
  - en
tags:
  - hallucination-detection
  - object-hallucination
  - pope
  - coco
size_categories:
  - 1K<n<10K

RePOPE

Re-annotated POPE benchmark with corrected ground-truth labels for object hallucination evaluation. 8,185 yes/no questions about object existence in MSCOCO images.

Fields

Field	Description
image	MSCOCO image
image_id	COCO image identifier
question	Yes/no question about object presence
answer	Ground truth (yes/no)
category	Sampling strategy (random/popular/adversarial)

Source: YanNeu/RePOPE