---
license: cc-by-nc-4.0
task_categories:
- visual-question-answering
- multiple-choice
language:
- en
pretty_name: SAT (Spatial Aptitude Test)
size_categories:
- n<1K
configs:
- config_name: default
  data_files:
  - split: test
    path: data/test-*.parquet
---

# SAT (Spatial Aptitude Test) — re-hosted for lmms-eval

This is a re-hosted copy of [`array/SAT`](https://huggingface.co/datasets/array/SAT)
prepared for upstream-friendly use with
[`EvolvingLMMs-Lab/lmms-eval`](https://github.com/EvolvingLMMs-Lab/lmms-eval).

## What changed vs. `array/SAT`

The original parquet uses a nested `list<binary>` schema for image bytes that
fails pyarrow chunked-array conversion when `load_dataset` is called without
`streaming=True`. lmms-eval's `api/task.py` calls `load_dataset` with
`num_proc=1`, which is incompatible with streaming, so neither path works
out-of-the-box.

The re-host fixes this by:

1. **Storing images as `Sequence(Image())`** instead of raw nested binary —
   `datasets` handles the parquet encoding correctly and non-streaming loads
   succeed.
2. **Pre-shuffling the answer order** with a fixed seed (`random.Random(42)`)
   and storing `correct_answer_idx` directly. The upstream fork's
   `api/task.py` originally shuffled answer order at load time with a
   non-deterministic seed; baking this in makes the eval reproducible and
   removes the need for any framework patch.

All other fields (`question`, `question_type`, `correct_answer`) are passed
through unchanged.

## Schema

| Field                 | Type                       | Description                              |
|-----------------------|----------------------------|------------------------------------------|
| `image_bytes`         | `Sequence(Image())`        | 1 or 2 RGB JPEGs per item                |
| `question`            | `Value("string")`          | Question text                            |
| `answers`             | `Sequence(Value("string"))`| Two answer choices, **shuffled** seed=42 |
| `correct_answer_idx`  | `Value("int32")`           | 0 or 1, index into `answers`             |
| `correct_answer`      | `Value("string")`          | Full text of the correct answer          |
| `question_type`       | `Value("string")`          | One of `obj_movement`, `ego_movement`, `action_conseq`, `perspective`, `goal_aim` |

## Stats

- 150 test items, single `test` split
- Image count per item: 1 (104 items) or 2 (46 items)
- Answer count per item: 2 (binary MCQ)
- Question types: `action_conseq` 37, `goal_aim` 34, `perspective` 33,
  `obj_movement` 23, `ego_movement` 23

## License

Inherits from the original `array/SAT` release. Use under the original
licensing terms.

## Citation

```bibtex
@article{ray2025sat,
  title={SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models},
  author={Ray, Arijit and Duan, Jiafei and Tan, Reuben and Bashkirova, Dina and Hendrix, Rose and Ehsani, Kiana and Kembhavi, Aniruddha and Plummer, Bryan A and Krishna, Ranjay and Zeng, Kuo-Hao and others},
  journal={arXiv preprint arXiv:2412.07755},
  year={2025}
}
```