maujim/vsr-sample-500

VSR Sample 500

This repository is a derivative sample of the Visual Spatial Reasoning (VSR) dataset. It contains 500 records and 483 unique COCO images in one train split. It is not the complete VSR corpus and is not a replacement for the upstream dataset.

The records were sampled without replacement from the upstream random-train split with deterministic seed 20260905. The sample preserves the source fields and values; the image field points to the bundled local file at images/<basename>.jpg. image_link retains the corresponding COCO URL.

Offline viewer

Download the entire repository and keep its folders together, including the images/ directory. Then double-click index.html to open the viewer locally—no installation, local server, or internet connection is needed after download. Browse records with Previous/Next or the keyboard arrow keys, filter and search by label or relation, jump to a random record, and inspect captions, truth labels, image filenames, and metadata. You can also copy a record's JSON for reuse.

Schema

The data/train.jsonl file has exactly 500 records with these features:

Feature	Type	Description
image	image	Bundled COCO image, loaded from images/*.jpg.
image_link	string	Original COCO image URL.
caption	string	Statement describing a spatial relation between two objects.
label	int64	1 means the caption is true of the image; 0 means it is false.
relation	string	Spatial relation asserted by the caption.
annotator_id	int64	Identifier of the original annotator.
vote_true_validator_id	string	Original serialized validator IDs voting true.
vote_false_validator_id	string	Original serialized validator IDs voting false.
reference_frame	float64	Original reference-frame value; it may be null.

Loading

From a local checkout of this repository:

from datasets import load_dataset

dataset = load_dataset("./vsr-sample-500", data_files={"train": "data/train.jsonl"})
print(dataset["train"][0])

The JSONL paths are relative to this repository, so the bundled images are available when the dataset is loaded locally. For a published Hub copy, use its dataset identifier instead of the local path; the train configuration maps to data/train.jsonl.

Provenance

VSR is a benchmark of caption-image pairs for judging whether a caption correctly describes a spatial relation in an image. This sample was prepared from the upstream random train split on 2026-09-05 using the deterministic seed stated above. See the upstream VSR repository for the original data, split definitions, implementation, and provenance.

Citation

If you use the original VSR work, please cite:

@article{Liu2022VisualSR,
  title={Visual Spatial Reasoning},
  author={Fangyu Liu and Guy Edward Toh Emerson and Nigel Collier},
  journal={Transactions of the Association for Computational Linguistics},
  year={2023},
}

Limitations

This is a small, intentionally sampled subset and may not represent the full distribution of VSR relations, captions, labels, or annotators. It contains only the random-train source split and should not be used to infer performance on VSR's other splits or on broader visual reasoning tasks. The labels represent the VSR caption-verification task and do not establish general image or object-recognition capability.

Licensing and image attribution

The VSR repository's code and dataset materials are released under the Apache-2.0 license, which is the license declared for this derivative repository's dataset-card materials. The bundled images originate from the COCO dataset; image rights, terms, and attribution obligations are governed by COCO and the applicable image creators' rights, not by this sample's Apache-2.0 declaration. Review the upstream VSR and COCO terms before redistributing or using the images.