---
license: mit
---

<div align="center">
<h1>COLUMBUS: Evaluating <ins>CO</ins>gnitive <ins>L</ins>ateral <ins>U</ins>nderstanding through <ins>M</ins>ultiple-choice re<ins>BUS</ins>es</h1>

Inspired by the demonstration of creative problem-solving in the story of *Columbus' Egg*, this repository presents a benchmark of multiple-choice rebus puzzles. These are designed to challenge the lateral thinking skills of visual question answering (VQA) models. 

[**📄 Paper**](https://arxiv.org/pdf/2409.04053) | [**📚 arXiv**](https://arxiv.org/abs/2409.04053) | [**💻 Code**](https://github.com/koen-47/COLUMBUS) | [**📓 Notebooks**](https://github.com/koen-47/COLUMBUS/tree/main/notebooks) | [**🌐 Website**](https://columbus-vqa.github.io/)

<!-- <picture>
    <source media="(prefers-color-scheme: dark)" srcset="./visualizations/columbus_puzzle_examples_dark.png">
</picture> -->

![image/png](https://cdn-uploads.huggingface.co/production/uploads/644bd1e6f30d4ab8ea0d54a8/FVEut31UCBWzt28q5Gwcr.png)

</div>


## Uses
This dataset is used to evaluate the lateral thinking skills of vision-language models.


## Fields
- `options`: the four multiple-choice options for a puzzle.
- `correct`: the correct answer to the puzzle (part of `options`).
- `is_icon`: flag to denote if the puzzle belongs to COLUMBUS-<span style="font-variant:small-caps;">icon</span> (if true) or COLUMBUS-<span style="font-variant:small-caps;">text</span> (if false).
- `is_overlap`: flag to denote if the puzzle belongs to the subset of overlapping puzzles that have both a textual and icon variant.