RebusBench for Evaluating Cognitive Visual Reasoning

RebusBench contains 1,164 English rebus puzzles, each paired with a list of reference answers. Rebus puzzles encode words or phrases through the visual arrangement, repetition, color, position, or size of text and symbols. The dataset supports evaluating visual-language models on visual wordplay and reasoning.

Load with Hugging Face Datasets

pip install -U datasets Pillow

from datasets import load_dataset

# Load the evaluation split directly.
dataset = load_dataset("amirkasaei/rebusbench", split="test")

print(len(dataset))  # 1164
sample = dataset[0]
print(sample["puzzle_id"])
print(sample["answers"])  # Always a list, including for a single answer
image = sample["image"]  # PIL image
image.save("example_puzzle.png")

Without split="test", the loader returns a dataset dictionary:

dataset = load_dataset("amirkasaei/rebusbench")
puzzles = dataset["test"]

These Hub examples use the files published in the Hugging Face repository. Local edits must be uploaded before they are available through the Hub loader.

Dataset structure

Field	Type	Description
image	Hugging Face Image	Puzzle image, decoded as a PIL image when accessed
puzzle_id	int64	Unique puzzle identifier from 1 through 1164
answers	list[string]	One or more normalized English reference answers

puzzle_id identifies a puzzle independently of dataset row order. The dataset does not include a per-row question field; a suitable shared question is: "What English word or phrase is represented by this rebus puzzle?"

Multiple reference answers

Every answers value is a nonempty list. There are 1,131 puzzles with one reference and 33 with multiple references. Alternatives originally separated by / are stored individually and normalized as described below. Order follows the original annotation and does not indicate a preferred answer.

For example, puzzle 50 has answers = ["farm animal", "animal farm"], while puzzle 1 has answers = ["for once in my life"].

For exact-match evaluation, a prediction is correct if it matches any reference after applying the same text normalization to the prediction and each reference. For token F1, compute the score against each reference and take the maximum for that puzzle before averaging across puzzles. Do not join the references into one string or score only the first reference.

Normalized answers

answers contains normalized references. Alternatives are never merged or removed, even if normalization makes two entries identical.

Normalization applies Unicode NFKC normalization, Unicode case folding, and whitespace collapsing (including trimming). For example, ["Farm animal", "Animal farm"] becomes ["farm animal", "animal farm"]. Punctuation, accents, spelling, and word order are not otherwise changed. This conservative policy avoids silently treating different phrases as equivalent.

Use this same function to normalize model predictions:

import unicodedata

def normalize(text):
    return " ".join(unicodedata.normalize("NFKC", text).casefold().split())

def exact_match(prediction, answers):
    return normalize(prediction) in answers

assert exact_match(" ANIMAL   FARM ", ["farm animal", "animal farm"])

Use a consistent normalization policy and report it with evaluation results. The reference alternatives are preserved from the original annotations; they have not been independently revalidated.

Splits

Split	Examples	Purpose
test	1,164	Evaluation

The complete collection is packaged as one evaluation split. No training or validation split is provided. The test name does not establish that the puzzles were excluded from any model's training data.

Files

README.md                   Dataset card and loading configuration
data/
  metadata.jsonl            Image filenames, puzzle IDs, and reference answers
  puzzle_1.png
  ...
  puzzle_1164.png
docs/
  examples.png              Dataset illustration
examples/
  example_evaluation.ipynb   Original sample evaluation code, unchanged
  requirements.txt          Suggested dependencies for the sample code

The dataset uses the ImageFolder layout. data/metadata.jsonl contains one JSON object per line with file_name, puzzle_id, and answers. Each file_name is relative to data/. The loader combines each image with its metadata and exposes the image as the image field. The loading configuration restricts dataset files to data/, excluding the README illustration and supporting files.

To load a local checkout using the same dataset card configuration, run from the repository root:

from pathlib import Path
from datasets import load_dataset

dataset = load_dataset(str(Path.cwd()), split="test")

Intended use and limitations

The intended use is research and evaluation of models on rebus puzzle solving. Answers are English words or phrases, and interpretation can depend on familiarity with English idioms and visual conventions. The supplied references may not cover all valid phrasings. Exact-match evaluation can penalize unlisted alternatives.

Data preparation and provenance

The repository contains 1,164 PNG images and 1,164 corresponding answer records. data/metadata.jsonl was generated from the original answer workbook. Answer strings were split on / and each reference was normalized according to the policy above. Reference order and puzzle images were preserved. All 1,164 records and 1,198 references were checked against the workbook before removing the redundant spreadsheet from this distribution. The JSONL metadata is the distributed source of reference answers. Detailed original image sources, collection procedures, and annotation procedures have not yet been documented in this card.

Sample evaluation code

examples/example_evaluation.ipynb is an optional example, separate from dataset loading. Its original contents are retained, including its original paths and model settings; adapt those before use. It is not required to install or run the notebook to access this dataset with Hugging Face Datasets. examples/requirements.txt lists suggested notebook dependencies and is not a verified environment lockfile.

License

A dataset license has not yet been specified in this repository.

Citation

@inproceedings{kasaei2026rebusbench,
    title={Hidden Meanings in Plain Sight: RebusBench for Evaluating Cognitive Visual Reasoning},
    author={Seyed Amir Kasaei and Arash Marioriyad and Mahbod Khaleti and MohammadAmin Fazli and Mahdieh Soleymani Baghshah and Mohammad Hossein Rohban},
    booktitle={ICLR 2026 Workshop - From Human Cognition to AI Reasoning: Models, Methods, and Applications},
    year={2026},
    url={https://openreview.net/forum?id=LCc2CP4aS4}
}