Datasets:

aaronyee
/

sat-train-dynamic

license: mit
pretty_name: SAT train split, dynamic spatial questions
language:
  - en
task_categories:
  - visual-question-answering
tags:
  - spatial-reasoning
  - multi-image
  - synthetic
size_categories:
  - 10K<n<100K

SAT train split: dynamic spatial questions

A re-packaged subset of the train split of SAT (Spatial Aptitude Training, Ray et al.). It holds the 41,619 dynamic questions of SAT_train.parquet with their 58,919 images, extracted from the parquet file into image files. Questions, answers and images are unchanged; nothing is added.

Source

Dataset: array/SAT, file SAT_train.parquet, revision bda5dde942d7f7b41bff7935f086ed9a9e348ae3
Project page: https://arijitray1993.github.io/SAT/
Paper: https://arxiv.org/abs/2412.07755
The images are synthetic renders of procedurally generated indoor scenes (ProcTHOR / AI2-THOR).

Path	Content
`questions.jsonl`	41,619 rows, one per question
`images-000.tar`, `images-001.tar`	58,919 images (56,465 JPEG, 2,454 PNG), member paths `images/<row>_<k>.<jpg or png>` (uncompressed tar)
`files.tsv`	every image: path, bytes, sha256, tar file
`LICENSE`	the MIT license text

Fields of questions.jsonl:

Field	Meaning
`id`	`sat-train-<row>`
`row`	0-based row index in `SAT_train.parquet` at the revision above
`question_type`	`action_consequence` (15,100), `action_sequence` (10,370), `obj_movement` (6,930), `goal_aim` (6,765), `perspective` (2,454)
`question`	upstream `question`, unchanged
`answers`	upstream `answers` (the answer choices), unchanged and in upstream order
`correct_answer`	upstream `correct_answer`, unchanged
`images`	paths of the row's images (1 or 2), in upstream order

The upstream card asks users to randomize the order of the answer choices when presenting them to a model; this copy keeps the upstream order.

Processing

Downloaded SAT_train.parquet at the pinned revision.
Kept the rows whose question_type is one of the five dynamic types above: 41,619 of the 172,384 rows. The 130,765 rows of type other (static questions) are left out.
Wrote each element of image_bytes to images/<row>_<k>.<ext> (<row> zero-padded to 6 digits, <k> the position in the list, <ext> from the image's format). The bytes are the upstream bytes: no re-encoding, resizing or filtering. Identical images that upstream stores under different rows stay separate files.
Copied question, answers, question_type and correct_answer unchanged, and added id, row and images.
Packed the images, sorted by path, into two uncompressed tar files of at most 1 GB each.

No question or answer was edited, generated or filtered by a model.

Split

Training data only: every row comes from the SAT train split (SAT_train.parquet). Nothing from SAT_static.parquet, SAT_val.parquet or the real-image test split (SAT_test.parquet) is included.

License

MIT, the license declared by the upstream dataset card (array/SAT). The LICENSE file carries the MIT text with the copyright notice of the SAT project (Copyright (c) 2024 Arijit Ray, from github.com/arijitray1993/SAT). This copy is distributed under the same license.

Attribution and citation

The questions, answers and images are the work of the SAT authors. If you use this data, please cite their paper:

@misc{ray2025satdynamicspatialaptitude,
      title={SAT: Dynamic Spatial Aptitude Training for Multimodal Language Models},
      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},
      year={2025},
      eprint={2412.07755},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2412.07755},
}