CLEVR-HOPE

The CLEVR Held-Out Pair Evaluation (CLEVR-HOPE) dataset is a diagnostic dataset for testing the systematicity of VQA models.

CLEVR-HOPE is a controlled setting to test whether VQA models generalize to pairs of attribute values that were not seen during either training or fine-tuning.

Within CLEVR-HOPE, we refer to an unseen pair of attribute values as a Held-Out Pair (HOP). The dataset is composed of 29 sub-datasets, each for a different HOP.

For each of the 29 HOPs, we have the train split, and 4 test splits: minimal-IID, minimal-OOD, complex-IID, and complex-OOD.

A list of the 29 HOPs can be found at Listing of HOPs; each HOP is represented as a 4-tuple of the form (Size, Colour, Shape, Material). If the HOP does not constrain an attribute, then None is used in the corresponding position in the 4-tuple. Note that each HOP is numbered according to their index on this list. For example HOP0 is rubber cylinder, a.k.a. (None, None, 'cylinder', 'rubber'), and HOP28 is small purple, a.k.a. ('small', 'purple', None, None).

For further design details, please see the paper: https://doi.org/10.48550/arXiv.2311.08695

The full data set is available for download at the Internet Archive, that variant contains additional metadata and is deduplicated for reduced storage space.

Listing of HOPs

• HOP00: (None, None, 'cylinder', 'rubber')
• HOP01: (None, 'cyan', None, 'rubber')
• HOP02: ('large', None, None, 'rubber')
• HOP03: (None, 'cyan', 'cylinder', None)
• HOP04: ('large', None, 'cylinder', None)
• HOP05: ('large', 'cyan', None, None)
• HOP06: (None, None, 'sphere', 'rubber')
• HOP07: (None, 'brown', None, 'rubber')
• HOP08: ('small', None, None, 'rubber')
• HOP09: (None, 'brown', 'sphere', None)
• HOP10: ('small', None, 'sphere', None)
• HOP11: ('small', 'brown', None, None)
• HOP12: (None, None, 'cylinder', 'metal')
• HOP13: (None, 'red', None, 'metal')
• HOP14: ('small', None, None, 'metal')
• HOP15: (None, 'red', 'cylinder', None)
• HOP16: ('small', None, 'cylinder', None)
• HOP17: ('small', 'red', None, None)
• HOP18: (None, None, 'cube', 'metal')
• HOP19: (None, 'gray', None, 'metal')
• HOP20: ('large', None, None, 'metal')
• HOP21: (None, 'gray', 'cube', None)
• HOP22: ('large', None, 'cube', None)
• HOP23: ('large', 'gray', None, None)
• HOP24: (None, None, 'cube', 'rubber')
• HOP25: (None, 'purple', None, 'rubber')
• HOP26: (None, 'purple', 'sphere', None)
• HOP27: ('small', None, 'cube', None)
• HOP28: ('small', 'purple', None, None)

If you found this data useful to your work, then please cite the original paper:

@inproceedings{berlot-attwell-etal-2024-attribute,
    title = "Attribute Diversity Determines the Systematicity Gap in {VQA}",
    author = "Berlot-Attwell, Ian  and
      Agrawal, Kumar Krishna  and
      Carrell, Annabelle Michael  and
      Sharma, Yash  and
      Saphra, Naomi",
    editor = "Al-Onaizan, Yaser  and
      Bansal, Mohit  and
      Chen, Yun-Nung",
    booktitle = "Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2024",
    address = "Miami, Florida, USA",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2024.emnlp-main.537/",
    doi = "10.18653/v1/2024.emnlp-main.537",
    pages = "9576--9611",
    abstract = "Although modern neural networks often generalize to new combinations of familiar concepts, the conditions that enable such compositionality have long been an open question. In this work, we study the systematicity gap in visual question answering: the performance difference between reasoning on previously seen and unseen combinations of object attributes. To test, we introduce a novel diagnostic dataset, CLEVR-HOPE. We find that the systematicity gap is not reduced by increasing the quantity of training data, but is reduced by increasing the diversity of training data. In particular, our experiments suggest that the more distinct attribute type combinations are seen during training, the more systematic we can expect the resulting model to be."
}