---
license: apache-2.0
dataset_info:
  features:
  - name: id
    dtype: int32
  - name: x
    sequence: image
  - name: color
    dtype:
      class_label:
        names:
          '0': DarkSlateGray
          '1': OrangeRed
          '2': Yellow
          '3': SpringGreen
          '4': Cyan
          '5': Purple
          '6': DodgerBlue
          '7': DeepPink
          '8': NavajoWhite
  - name: shape
    dtype:
      class_label:
        names:
          '0': square
          '1': ellipse
          '2': heart
  - name: scale
    dtype:
      class_label:
        names:
          '0': '0.5'
          '1': '0.6'
          '2': '0.7'
          '3': '0.8'
          '4': '0.9'
          '5': '1.0'
  - name: orientation
    dtype:
      class_label:
        names:
          '0': '0.0'
          '1': '0.32221'
          '2': '0.64443'
          '3': '0.96664'
          '4': '1.28886'
  - name: posX_dynamic
    dtype:
      class_label:
        names:
          '0': right_1x
          '1': left_1x
          '2': right_2x
          '3': left_2x
          '4': mid_right
          '5': mid_left
  - name: posY_dynamic
    dtype:
      class_label:
        names:
          '0': down_1x
          '1': up_1x
          '2': down_2x
          '3': up_2x
          '4': mid_down
          '5': mid_up
  splits:
  - name: train
    num_bytes: 29330065.672
    num_examples: 20412
  - name: val
    num_bytes: 9822019.138
    num_examples: 4374
  - name: test
    num_bytes: 9787552.138
    num_examples: 4374
  download_size: 49141904
  dataset_size: 48939636.948
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: val
    path: data/val-*
  - split: test
    path: data/test-*
---

## MSD dSprites-Dynamic Dataset Attribution

The Multi-factor Sequential Disentanglement benchmark includes a **modified variant of the dSprites dataset**, adapted to support sequential multi-factor disentanglement.

In this modified sequential version, the object’s **color**, **shape**, **scale**, and **orientation** are fixed, while its **spatial position** may change over time.

- Original repository:
  https://github.com/deepmind/dsprites-dataset
```
@misc{dsprites17,
author = {Loic Matthey and Irina Higgins and Demis Hassabis and Alexander Lerchner},
title = {dSprites: Disentanglement testing Sprites dataset},
howpublished= {https://github.com/deepmind/dsprites-dataset/},
year = "2017",
}
```

⚠ **Note:** The dSprites dataset is licensed under the Apache License 2.0. We redistribute it here solely for non-commercial research purposes, following the original license terms. Please cite the above when using this dataset in your work.