Datasets:
pretty_name: CaptionExam
license: other
license_name: research-only
license_link: LICENSE
language:
- en
task_categories:
- image-to-text
- visual-question-answering
tags:
- image-captioning
- dense-captioning
- benchmark
size_categories:
- n<1K
configs:
- config_name: default
data_files:
- split: test
path: images/**
CaptionExam
109 images and 3,099 fine-grained multiple-choice questions about them, for grading image captions: a text-only judge answers each question from a caption alone. Every question includes the option "not mentioned / indeterminable", which is never the answer. Evaluation code: Ki-Seki/CaptionExam.
The images are photos, illustrations, product shots, posters, book covers and text-heavy graphics. Each has 15–42 questions (median 27) about objects and counts, colors and materials, pose and orientation, layout, rendered text and its exact wording, typography, style and background.
Format
questions.jsonl has one question per line, and images/metadata.jsonl repeats each image's questions alongside the image.
{"id": "img_0001_q02", "image_id": "img_0001",
"question": "What color is the body of the pickup truck parked under the carport?",
"options": ["white or light silver", "dark blue or black", "red or maroon", "dark green",
"not mentioned / indeterminable"],
"answer": ["white or light silver"],
"bbox": [453, 552, 622, 804]}
bbox is the region the question is about, as [y_min, x_min, y_max, x_max] normalized to 0–1000. 120 questions cover the whole image.
from datasets import load_dataset
ds = load_dataset("Ki-Seki/CaptionExam", split="test") # one row per image, with its questions
License
The questions and annotations are for research use. The images were collected from the public web and their copyright stays with their owners; they are provided only for non-commercial research and evaluation. To have an image removed, open an issue at Ki-Seki/CaptionExam.