openai/CLIP
1
0
Fork 0
mirror of https://github.com/openai/CLIP.git synced 2026-11-05 01:37:26 +00:00
Code Issues 274 Projects Packages Wiki Activity

[GH-ISSUE #254] What exactly are the sources of WebImageText dataset? #220

New issue
Open
opened 2026-10-31 21:12:37 +00:00 by bsakaguchi · 5 comments
bsakaguchi commented 2026-10-31 21:12:37 +00:00
Owner
Copy link

Originally created by @andrey-savov on GitHub (Jun 19, 2022).
Original GitHub issue: https://github.com/openai/CLIP/issues/254

In this paper there is only a vague description about the WIT dataset:

... we constructed a new dataset of 400 million (image,
text) pairs collected form a variety of publicly available
sources on the Internet.

Can you enumerate specific sources and methods used to acquire the images?

Originally created by @andrey-savov on GitHub (Jun 19, 2022). Original GitHub issue: https://github.com/openai/CLIP/issues/254 In [this paper](https://arxiv.org/pdf/2103.00020.pdf) there is only a vague description about the WIT dataset: > ... we constructed a new dataset of 400 million (image, > text) pairs collected form a variety of publicly available > sources on the Internet. Can you enumerate _specific_ sources and _methods_ used to acquire the images?
bsakaguchi commented 2026-10-31 21:12:38 +00:00
Author
Owner
Copy link

@ShilpaGopal commented on GitHub (Aug 10, 2022):

How are texts engineered? Paper talks about prompt engineering for tasks datasets but not for the 400M training dataset!!

<!-- gh-comment-id:1210641546 --> @ShilpaGopal commented on GitHub (Aug 10, 2022): How are texts engineered? Paper talks about prompt engineering for tasks datasets but not for the 400M training dataset!!
bsakaguchi commented 2026-10-31 21:12:38 +00:00
Author
Owner
Copy link

@placeforyiming commented on GitHub (Apr 23, 2024):

Looking for more information here

<!-- gh-comment-id:2071336279 --> @placeforyiming commented on GitHub (Apr 23, 2024): Looking for more information here
bsakaguchi commented 2026-10-31 21:12:38 +00:00
Author
Owner
Copy link

@haardikdharma10 commented on GitHub (Feb 8, 2025):

The paper Demystifying CLIP data does that for you

<!-- gh-comment-id:2645906246 --> @haardikdharma10 commented on GitHub (Feb 8, 2025): The paper [Demystifying CLIP data](https://arxiv.org/pdf/2309.16671v4) does that for you
bsakaguchi commented 2026-10-31 21:12:38 +00:00
Author
Owner
Copy link

@ShilpaGopal commented on GitHub (Feb 10, 2025):

The paper Demystifying CLIP data does that for you

The authors did not publish the full 400M CLIP dataset due to the dataset’s proprietary nature, which is owned by OpenAI. However, they provided key insights and partial data reconstructions based on their research.

Data Published:

Subset or Proxy Dataset: They attempted to reproduce a dataset similar to the 400M CLIP dataset, referred to as a "pseudo-CLIP dataset," using web-scraped data and public sources to approximate the original dataset’s composition and characteristics.
Analysis Data: The paper includes detailed statistics, comparisons, and benchmarks of their reconstructed dataset against the original CLIP dataset to evaluate performance gaps and biases.

Did they successfully reproduce the 400M dataset?

Partially: While they didn’t fully recreate the 400M dataset, they built a large-scale proxy dataset that captures some of the original dataset's properties.
Performance Comparison: Their experiments revealed that models trained on this reconstructed dataset performed reasonably well, though with some limitations and noticeable performance differences compared to models trained on the original dataset.
<!-- gh-comment-id:2647576974 --> @ShilpaGopal commented on GitHub (Feb 10, 2025): > The paper [Demystifying CLIP data](https://arxiv.org/pdf/2309.16671v4) does that for you The authors did not publish the full 400M CLIP dataset due to the dataset’s proprietary nature, which is owned by OpenAI. However, they provided key insights and partial data reconstructions based on their research. **Data Published:** Subset or Proxy Dataset: They attempted to reproduce a dataset similar to the 400M CLIP dataset, referred to as a "pseudo-CLIP dataset," using web-scraped data and public sources to approximate the original dataset’s composition and characteristics. Analysis Data: The paper includes detailed statistics, comparisons, and benchmarks of their reconstructed dataset against the original CLIP dataset to evaluate performance gaps and biases. **Did they successfully reproduce the 400M dataset?** Partially: While they didn’t fully recreate the 400M dataset, they built a large-scale proxy dataset that captures some of the original dataset's properties. Performance Comparison: Their experiments revealed that models trained on this reconstructed dataset performed reasonably well, though with some limitations and noticeable performance differences compared to models trained on the original dataset.
bsakaguchi commented 2026-10-31 21:12:38 +00:00
Author
Owner
Copy link

@andrey-savov commented on GitHub (Jul 29, 2025):

The paper Demystifying CLIP data does that for you

No it does not answer the original question, can @haardikdharma10 how did you figure it does?

<!-- gh-comment-id:3130754878 --> @andrey-savov commented on GitHub (Jul 29, 2025): > The paper [Demystifying CLIP data](https://arxiv.org/pdf/2309.16671v4) does that for you No it does not answer the original question, can @haardikdharma10 how did you figure it does?
Sign in to join this conversation.
No Branch/Tag specified
Branches Tags
main
codex/pin-github-workflow-refs-20260325-132838
jongwook/issue-396
jongwook-patch-1
No results found.
Labels
Clear labels   
pull-request
Mirrored from GitHub Pull Request

No labels
pull-request
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
openai/CLIP#220
No description provided.