Dataset Card for COYO-Labeled-300M

Dataset Summary

COYO-Labeled-300M is a dataset of machine-labeled 300M images-multi-label pairs. We labeled subset of COYO-700M with a large model (efficientnetv2-xl) trained on imagenet-21k. We followed the same evaluation pipeline as in efficientnet-v2. The labels are top 50 most likely labels out of 21,841 classes from imagenet-21k. The label probabilies are provided rather than label so that the user can select threshold of their choice for multi-label classification use or can take top-1 class for single class classification use.

In other words, COYO-Labeled-300M is a ImageNet-like dataset. Instead of human labeled 1.25 million samples, it's machine-labeled 300 million samples. This dataset is similar to JFT-300M which is not released to the public.

Supported Tasks and Leaderboards

We empirically validated the quality of COYO-Labeled-300M dataset by re-implementing popular model, ViT. We found that our ViT implementation trained on COYO-Labeled-300M performs similar to the performance numbers in the ViT paper trained on JFT-300M. We also provide weights for the pretrained ViT model on COYO-Labeled-300M as well as its training & fine-tuning code.

Languages

The labels in the COYO-Labeled-300M dataset consist of English.

Dataset Structure

Data Instances

Each instance in COYO-Labeled-300M represents multi-labels and image pair information with meta-attributes.
And we also provide label information, imagenet21k_tree.pickle.

{
  'id': 315,
  'url': 'https://a.1stdibscdn.com/pair-of-blue-and-white-table-lamps-for-sale/1121189/f_121556431538206028457/12155643_master.jpg?width=240',
  'imagehash': 'daf5a50aae4aa54a',
  'labels': [8087, 11054, 8086, 6614, 6966, 8193, 10576, 9710, 4334, 9909, 8090, 10104, 10105, 9602, 5278, 9547, 6978, 12011, 7272, 5273, 6279, 4279, 10903, 8656, 9601, 8795, 9326, 4606, 9907, 9106, 7574, 10006, 7257, 6959, 9758, 9039, 10682, 7164, 5888, 11654, 8201, 4546, 9238, 8197, 10882, 17380, 4470, 5275, 10537, 11548],
  'label_probs': [0.4453125, 0.30419921875, 0.09417724609375, 0.033905029296875, 0.03240966796875, 0.0157928466796875, 0.01406097412109375, 0.01129150390625, 0.00978851318359375, 0.00841522216796875, 0.007720947265625, 0.00634002685546875, 0.0041656494140625, 0.004070281982421875, 0.002910614013671875, 0.0028018951416015625, 0.002262115478515625, 0.0020503997802734375, 0.0017080307006835938, 0.0016880035400390625, 0.0016679763793945312, 0.0016613006591796875, 0.0014324188232421875, 0.0012445449829101562, 0.0011739730834960938, 0.0010318756103515625, 0.0008969306945800781, 0.0008792877197265625, 0.0008726119995117188, 0.0008263587951660156, 0.0007123947143554688, 0.0006799697875976562, 0.0006561279296875, 0.0006542205810546875, 0.0006093978881835938, 0.0006046295166015625, 0.0005769729614257812, 0.00057220458984375, 0.0005636215209960938, 0.00055694580078125, 0.0005092620849609375, 0.000507354736328125, 0.000507354736328125, 0.000499725341796875, 0.000484466552734375, 0.0004456043243408203, 0.0004439353942871094, 0.0004355907440185547, 0.00043392181396484375, 0.00041866302490234375],
  'width': 240,
  'height': 240
}

Data Fields

name	type	description
id	long	Unique 64-bit integer ID generated by monotonically_increasing_id() which is the same value that is mapped with the existing COYO-700M.
url	string	The image URL extracted from the src attribute of the <img>
imagehash	string	The perceptual hash(pHash) of the image
labels	sequence[integer]	Inference results of EfficientNetV2-XL model trained on ImageNet-21K dataset (Top 50 indices among 21,841 classes)
label_probs	sequence[float]	Inference results of EfficientNetV2-XL model trained on ImageNet-21K dataset (Top 50 indices among 21,841 probabilites)
width	integer	The width of the image
height	integer	The height of the image

Data Splits

Data was not split, since the evaluation was expected to be performed on more widely used downstream task(s).

Dataset Creation

Curation Rationale

We labeled subset of COYO-700M with a large model (efficientnetv2-xl) trained on imagenet-21k. Data sampling was done with a size similar to jft-300m, filtered by a specific threshold for probabilities for the top-1 label.

Source Data

COYO-700M

Who are the source language producers?

Common Crawl is the data source for COYO-700M.

Annotations

Annotation process

The dataset was built in a fully automated process that did not require human annotation.

Who are the annotators?

No human annotation

Personal and Sensitive Information

The basic instruction, licenses and contributors are the same as for the coyo-700m.