---
dataset_info:
- config_name: tqa
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: oracle_answer
    dtype: string
  - name: oracle_option
    dtype: string
  - name: oracle_full_answer
    dtype: string
  splits:
  - name: test
    num_bytes: 4723238
    num_examples: 4635
  download_size: 804261
  dataset_size: 4723238
- config_name: vqa
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: image
    dtype: image
  - name: oracle_answer
    dtype: string
  - name: oracle_option
    dtype: string
  - name: oracle_full_answer
    dtype: string
  splits:
  - name: test
    num_bytes: 733091578.0
    num_examples: 4635
  download_size: 712137895
  dataset_size: 733091578.0
- config_name: vtqa
  features:
  - name: id
    dtype: string
  - name: text
    dtype: string
  - name: image
    dtype: image
  - name: oracle_answer
    dtype: string
  - name: oracle_option
    dtype: string
  - name: oracle_full_answer
    dtype: string
  splits:
  - name: test
    num_bytes: 736109315.0
    num_examples: 4635
  download_size: 712879771
  dataset_size: 736109315.0
configs:
- config_name: tqa
  data_files:
  - split: test
    path: tqa/test-*
- config_name: vqa
  data_files:
  - split: test
    path: vqa/test-*
- config_name: vtqa
  data_files:
  - split: test
    path: vtqa/test-*
---

## 🤔 About SpatialEval
SpatialEval is a comprehensive benchmark for evaluating spatial intelligence in LLMs and VLMs across four key dimensions:
- Spatial relationships
- Positional understanding
- Object counting
- Navigation

### Benchmark Tasks
1. **Spatial-Map**: Understanding spatial relationships between objects in map-based scenarios
2. **Maze-Nav**: Testing navigation through complex environments
3. **Spatial-Grid**: Evaluating spatial reasoning within structured environments
4. **Spatial-Real**: Assessing real-world spatial understanding

Each task supports three input modalities:
- Text-only (TQA)
- Vision-only (VQA)
- Vision-Text (VTQA)

![spatialeval_task.png](https://cdn-uploads.huggingface.co/production/uploads/651651f5d93a51ceda3021c3/kpjld6-HCg5LXhO9Ju6-Q.png)

## 📌 Quick Links
Project Page: https://spatialeval.github.io/

Paper: https://arxiv.org/pdf/2406.14852

Code: https://github.com/jiayuww/SpatialEval

Talk: https://neurips.cc/virtual/2024/poster/94371

## 🚀 Quick Start

### 📍 Load Dataset

SpatialEval provides three input modalities—TQA (Text-only), VQA (Vision-only), and VTQA (Vision-text)—across four tasks: Spatial-Map, Maze-Nav, Spatial-Grid, and Spatial-Real. Each modality and task is easily accessible via Hugging Face. Ensure you have installed the [packages](https://huggingface.co/docs/datasets/en/quickstart):

```python
from datasets import load_dataset

tqa = load_dataset("MilaWang/SpatialEval", "tqa", split="test")
vqa = load_dataset("MilaWang/SpatialEval", "vqa", split="test")
vtqa = load_dataset("MilaWang/SpatialEval", "vtqa", split="test")
```

## ⭐ Citation

If you find our work helpful, please consider citing our paper 😊

```
@inproceedings{wang2024spatial,
title={Is A Picture Worth A Thousand Words? Delving Into Spatial Reasoning for Vision Language Models},
author={Wang, Jiayu and Ming, Yifei and Shi, Zhenmei and Vineet, Vibhav and Wang, Xin and Li, Yixuan and Joshi, Neel},
booktitle={The Thirty-Eighth Annual Conference on Neural Information Processing Systems},
year={2024}
}
```

## 💬 Questions
Have questions? We're here to help!
- Open an issue in the github repository
- Contact us through the channels listed on our project page