Add comprehensive README with paper info, loading code, and circular eval instructions
Browse files
README.md
CHANGED
|
@@ -38,3 +38,63 @@ configs:
|
|
| 38 |
- split: test
|
| 39 |
path: data/test-*
|
| 40 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
- split: test
|
| 39 |
path: data/test-*
|
| 40 |
---
|
| 41 |
+
|
| 42 |
+
# SAT-v2 Dataset
|
| 43 |
+
|
| 44 |
+
## Paper
|
| 45 |
+
|
| 46 |
+
**SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models**
|
| 47 |
+
|
| 48 |
+
This dataset is part of the SAT (Spatial Aptitude Training) project, which introduces a dynamic benchmark for evaluating and improving spatial reasoning capabilities in multimodal language models.
|
| 49 |
+
|
| 50 |
+
- **Project Page**: [https://arijitray.com/SAT/](https://arijitray.com/SAT/)
|
| 51 |
+
- **Paper**: [arXiv:2412.07755](https://arxiv.org/abs/2412.07755)
|
| 52 |
+
|
| 53 |
+
## Dataset Description
|
| 54 |
+
|
| 55 |
+
SAT-v2 is a comprehensive spatial reasoning benchmark containing over 300,000 questions across multiple splits. The dataset tests various aspects of spatial understanding including perspective-taking, object relationships, and dynamic scene understanding.
|
| 56 |
+
|
| 57 |
+
## Loading the Dataset
|
| 58 |
+
|
| 59 |
+
```python
|
| 60 |
+
from datasets import load_dataset
|
| 61 |
+
|
| 62 |
+
# Load the training split
|
| 63 |
+
dataset = load_dataset("array/SAT-v2", split="train")
|
| 64 |
+
|
| 65 |
+
# Or load a specific split
|
| 66 |
+
val_dataset = load_dataset("array/SAT-v2", split="val")
|
| 67 |
+
static_dataset = load_dataset("array/SAT-v2", split="static")
|
| 68 |
+
test_dataset = load_dataset("array/SAT-v2", split="test")
|
| 69 |
+
|
| 70 |
+
# Access a sample
|
| 71 |
+
sample = dataset[0]
|
| 72 |
+
print(sample["question"])
|
| 73 |
+
print(sample["answers"])
|
| 74 |
+
print(sample["correct_answer"])
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
## Dataset Splits
|
| 78 |
+
|
| 79 |
+
- **train**: 172,384 examples - Dynamic training questions
|
| 80 |
+
- **static**: 127,405 examples - Static spatial reasoning questions
|
| 81 |
+
- **val**: 4,001 examples - Validation set
|
| 82 |
+
- **test**: 150 examples - Test set
|
| 83 |
+
|
| 84 |
+
**Important Note on Test Set Evaluation:** When evaluating on the test set, please use circular evaluation by switching the position of the correct answer to avoid position bias. If you're using lmms-eval, refer to the implementation here: [https://github.com/arijitray1993/lmms-eval/tree/main/lmms_eval/tasks/sat_real](https://github.com/arijitray1993/lmms-eval/tree/main/lmms_eval/tasks/sat_real)
|
| 85 |
+
|
| 86 |
+
## Citation
|
| 87 |
+
|
| 88 |
+
If you use this dataset, please cite:
|
| 89 |
+
|
| 90 |
+
```bibtex
|
| 91 |
+
@misc{ray2025satdynamicspatialaptitude,
|
| 92 |
+
title={SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models},
|
| 93 |
+
author={Arijit Ray and Jiafei Duan and Ellis Brown and Reuben Tan and Dina Bashkirova and Rose Hendrix and Kiana Ehsani and Aniruddha Kembhavi and Bryan A. Plummer and Ranjay Krishna and Kuo-Hao Zeng and Kate Saenko},
|
| 94 |
+
year={2025},
|
| 95 |
+
eprint={2412.07755},
|
| 96 |
+
archivePrefix={arXiv},
|
| 97 |
+
primaryClass={cs.CV},
|
| 98 |
+
url={https://arxiv.org/abs/2412.07755},
|
| 99 |
+
}
|
| 100 |
+
```
|