README.md

---
pretty_name: MOAT
configs:
- config_name: default
  data_files:
  - split: test
    path: MOAT.parquet
task_categories:
- image-text-to-text
---

<h1>MOAT: Evaluating LMMs for Capability Integration and Instruction Grounding</h1>

<div align="center">
    Zhoutong Ye, Mingze Sun, Huan-ang Gao, Chun Yu, Yuanchun Shi
</div>

<div align="center">
<a href="https://arxiv.org/abs/2503.09348" target="_blank">
    <img alt="arXiv" src="https://img.shields.io/badge/arXiv-MOAT-red?logo=arxiv" height="20" />
</a>
<a href="https://cambrian-yzt.github.io/MOAT/" target="_blank">
    <img alt="Website" src="https://img.shields.io/badge/🌎_Website-MOAT-blue.svg" height="20" />
</a>
<a href="https://huggingface.co/datasets/waltsun/MOAT" target="_blank">
    <img alt="HF Dataset: MOAT" src="https://img.shields.io/badge/%F0%9F%A4%97%20_HuggingFace-MOAT-yellow" height="20" />
</a>
<a href="https://github.com/Cambrian-yzt/MOAT" target="_blank">
    <img alt="GitHub: MOAT" src="https://img.shields.io/badge/GitHub-MOAT-yellow?logo=github" height="20" />
</a>
</div>

## Overview

**MOAT** (**M**ultimodal model **O**f **A**ll **T**rades) is a challenging benchmark for large multimodal models (LMMs). It consists of vision language (VL) tasks that require the LMM to integrate several VL capabilities and engage in human-like generalist visual problem solving. Moreover, many tasks in **MOAT** focus on LMMs' capability to ground complex text and visual instructions, which is crucial for the application of LMMs in-the-wild. Developing on the VL capability taxonomies proposed in previous benchmark papers, we define 10 fundamental VL capabilities in **MOAT**. 

Please check out our [GitHub repo](https://github.com/Cambrian-yzt/MOAT) for further information.

## Usage

Please check out our [GitHub repo](https://github.com/Cambrian-yzt/MOAT) for detail usage.

**Run Your Own Evaluation**

You can access our dataset with the following code:

```python
from datasets import load_dataset
dataset = load_dataset("waltsun/MOAT", split='test')
```

As some questions are formatted as interleaved text and image(s), we recommend referring to the `./inference/eval_API.py` file in our [GitHub repo](https://github.com/Cambrian-yzt/MOAT) for the correct way to query the LMM.

**Column Description**

- `index`: The index of the question in the dataset.
- `question`: The question text.
- `choices`: A list of the answer choices. Can be empty.
- `images`: The list of PIL images.
- `outside_knowledge_text`: The essential information for answering the question. Optional.
- `outside_knowledge_images`: The list of PIL images that are essential for answering the question. Can be empty.
- `answer`: The correct answer.
- `capability`: The VL capabilities required to answer the question. A list of strings.
- `human_cot`: The human annotation for the CoT reasoning process.

## Future Work

Going forward, we intend to further increase the diversity of the tasks in **MOAT**, involving more capability combinations and encompassing more domains and scenarios. Stay tuned!