VisualPun-Bench

A bilingual visual-pun benchmark: 376 pun images (276 English, 100 Mandarin), each with one question, one reference punchline, and the annotated chain of reasoning hops (1,589 hops) that leads from the image to the pun. A model sees the image and the question and must produce the pun phrase the image encodes; the hop questions probe each component skill separately.

column	content
id	item id
language	EN or CN
image	the pun image (PIL image)
question	the question asked about the image
answer	the reference punchline
hops	list of annotated hops, in chain order: hop (index), type (object recognition, OCR, spatial, action recognition, context, punch linguistic), question, answer, annotation (the annotator's original note)

Usage

from datasets import load_dataset
ds = load_dataset("Icey444/VisualPun-Bench", split="train")
row = ds[0]
row["image"].size            # PIL.Image
row["question"], row["answer"]
for h in row["hops"]: print(h["hop"], h["type"], h["question"], "->", h["answer"])

Hop questions are meant to be asked independently of one another; each supplies in its wording the context it needs from earlier hops, so errors do not propagate. Open-ended answers are graded against the reference as correct / partial / incorrect by the LLM judge described in the paper; partial answers do not count as correct.

License

Data and annotations are released under CC BY-NC 4.0. Images are included for research use only.