Dataset Card for CoDa

Dataset Summary

The Color Dataset (CoDa) is a probing dataset to evaluate the representation of visual properties in language models. CoDa consists of color distributions for 521 common objects, which are split into 3 groups. We denote these groups as Single, Multi, and Any, which represents the typical object of each group.

The default configuration of CoDa uses 10 CLIP-style templates (e.g. "A photo of a [object]"), and 10 cloze-style templates (e.g. "Everyone knows most [object] are [color]." )

Supported Tasks and Leaderboards

This version of the dataset consists of the filtered and templated examples as cloze style questions. See the GitHub repo for the raw data (e.g. unfiltered annotations) as well as example usage with GPT-2, RoBERTa, ALBERT, and CLIP.

Languages

The text in the dataset is in English. The associated BCP-47 code is en-US.

Dataset Structure

Data Instances

An example looks like this:

{
  "text": "All rulers are [MASK].",
  "label": [
    0.0181818176, 0.0363636352, 0.3077272773, 0.0181818176, 0.0363636352,
    0.086363636, 0.0363636352, 0.0363636352, 0.0363636352, 0.086363636,
    0.301363647
  ],
  "template_group": 1,
  "template_idx": 0,
  "class_id": "/m/0hdln",
  "display_name": "Ruler",
  "object_group": 2,
  "ngram": "ruler"
}

Data Fields

• text: The templated example. What this is depends on the value of template_group.
  • template_group=0: A CLIP style example. There are no [MASK] tokens in these examples.
  • template_group=1: A cloze style example. Note that all templates have [MASK] as the last word, but in most cases, the period should be included.
• label: A list of probability values for the 11 colors. Note that these are sorted by the alphabetic order of the 11 colors (black, blue, brown, gray, green, orange, pink, purple, red, white, yellow).
• template_group: Type of template, 0 corresponds to A CLIP style template (clip-imagenet), and 1 corresponds to A cloze style templates (text-masked).
• template_idx: The index of the template out of all templates
• class_id: The Corresponding OpenImages v6 ClassID.
• display_name: The Corresponding OpenImages v6 DisplayName.
• object_group: Object Group, values correspond to Single, Multi, and Any.
• ngram: Corresponding n-gram used for lookups.

Data Splits

Object Splits:

Group	All	Train	Valid	Test
Single	198	118	39	41
Multi	208	124	41	43
Any	115	69	23	23
Total	521	311	103	107

Example Splits:

Group	All	Train	Valid	Test
Single	3946	2346	780	820
Multi	4146	2466	820	860
Any	2265	1352	460	453
Total	10357	6164	2060	2133

Dataset Creation

Curation Rationale

[More Information Needed]

Source Data

Initial Data Collection and Normalization

[More Information Needed]

Who are the source language producers?

[More Information Needed]

Annotations

Annotation process

[More Information Needed]

Who are the annotators?

[More Information Needed]

Personal and Sensitive Information

[More Information Needed]

Considerations for Using the Data

Social Impact of Dataset

[More Information Needed]

Discussion of Biases

[More Information Needed]

Other Known Limitations

[More Information Needed]

Additional Information

Dataset Curators

[More Information Needed]

Licensing Information

CoDa is licensed under the Apache 2.0 license.

Citation Information

@misc{paik2021world,
      title={The World of an Octopus: How Reporting Bias Influences a Language Model's Perception of Color},
      author={Cory Paik and Stéphane Aroca-Ouellette and Alessandro Roncone and Katharina Kann},
      year={2021},
      eprint={2110.08182},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contributions

Thanks to @github-username for adding this dataset.