metadata

pretty_name: Winoground
task_categories:
  - image-to-text
  - text-to-image
  - image-classification
extra_gated_prompt: >-
  By clicking on “Access repository” below, you also agree that you are using it
  solely for research purposes. The full license agreement is available in the
  dataset files.
language:
  - en
dataset_info:
  features:
    - name: id
      dtype: int32
    - name: image_0
      dtype: image
    - name: image_1
      dtype: image
    - name: caption_0
      dtype: string
    - name: caption_1
      dtype: string
    - name: tag
      dtype: string
    - name: secondary_tag
      dtype: string
    - name: num_main_preds
      dtype: int32
    - name: collapsed_tag
      dtype: string
  splits:
    - name: test
      num_bytes: 369475277
      num_examples: 400
  download_size: 367365762
  dataset_size: 369475277
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test-*

Dataset Card for Winoground

Dataset Description

Winoground is a novel task and dataset for evaluating the ability of vision and language models to conduct visio-linguistic compositional reasoning. Given two images and two captions, the goal is to match them correctly—but crucially, both captions contain a completely identical set of words/morphemes, only in a different order. The dataset was carefully hand-curated by expert annotators and is labeled with a rich set of fine-grained tags to assist in analyzing model performance. In our accompanying paper, we probe a diverse range of state-of-the-art vision and language models and find that, surprisingly, none of them do much better than chance. Evidently, these models are not as skilled at visio-linguistic compositional reasoning as we might have hoped. In the paper, we perform an extensive analysis to obtain insights into how future work might try to mitigate these models’ shortcomings. We aim for Winoground to serve as a useful evaluation set for advancing the state of the art and driving further progress in the field.

We are thankful to Getty Images for providing the image data.

Data

The captions and tags are located in data/examples.jsonl and the images are located in data/images.zip. You can load the data as follows:

from datasets import load_dataset
examples = load_dataset('facebook/winoground', use_auth_token=<YOUR USER ACCESS TOKEN>)

You can get <YOUR USER ACCESS TOKEN> by following these steps:

log into your Hugging Face account
click on your profile picture
click "Settings"
click "Access Tokens"
generate an access token

Model Predictions and Statistics

The image-caption model scores from our paper are saved in statistics/model_scores. To compute many of the tables and graphs from our paper, run the following commands:

git clone https://huggingface.co/datasets/facebook/winoground
cd winoground
pip install -r statistics/requirements.txt
python statistics/compute_statistics.py

FLAVA Colab notebook code for Winoground evaluation

https://colab.research.google.com/drive/1c3l4r4cEA5oXfq9uXhrJibddwRkcBxzP?usp=sharing

CLIP Colab notebook code for Winoground evaluation

https://colab.research.google.com/drive/15wwOSte2CjTazdnCWYUm2VPlFbk2NGc0?usp=sharing

Paper FAQ

Why is the group score for a random model equal to 16.67%?

Click for a proof!

Intuitively, we might think that we can multiply the probabilities from the image and text score to get 1/16 = 6.25%. But, these scores are not conditionally independent. We can find the correct probability with combinatorics:

For ease of notation, let:

a = s(c_0, i_0)
b = s(c_1, i_0)
c = s(c_1, i_1)
d = s(c_0, i_1)

The group score is defined as 1 if a > b, a > d, c > b, c > d and 0 otherwise.

As one would say to GPT-3, let's think step by step:

There are 4! = 24 different orderings of a, c, b, d.
There are only 4 orderings for which a > b, a > d, c > b, c > d:

a, c, b, d
a, c, d, b
c, a, b, d
c, a, d, b

No ordering is any more likely than another because a, b, c, d are sampled from the same random distribution.
We can conclude that the probability of a group score of 1 is 4/24 = 0.166...

Citation Information

https://arxiv.org/abs/2204.03162

Tristan Thrush and Candace Ross contributed equally.

@inproceedings{thrush_and_ross2022winoground,
  author = {Tristan Thrush and Ryan Jiang and Max Bartolo and Amanpreet Singh and Adina Williams and Douwe Kiela and Candace Ross},
  title = {Winoground: Probing vision and language models for visio-linguistic compositionality},
  booktitle = {CVPR},
  year = 2022,
}