BootsofLagrangian/cc12m-jpeg-hybrid-resolution

CC12M JPEG hybrid-resolution image materialization

This gated repository contains 11,983,828 readable CC12M images in 1,467 WebDataset tar shards and one combined metadata Parquet. It was assembled for research reproducibility because the image-bearing materialization required by the research workflow was not otherwise available in one place.

The release combines the inherited 512px pixparse/cc12m-wds surface with medium/high-resolution images retrieved from the recorded CC12M URLs using img2dataset. Consequently, this is a hybrid-resolution JPEG materialization rather than a uniform resize. The materialization_tranche column records which path supplied each asset, while normalized URL and content-hash fields support reconciliation.

The source metadata retains the original CC12M text, URL, dimensions, download status, error field, EXIF field, digest, stable URL identity, byte size, and WebDataset binding. It adds only the explicit image_shard and image_member locators needed to open the paired image.

Frozen accounting

check	count
audited tar shards	1,467
publishable tar shards	1,467
image/JSON pairs	11,983,828
image members missing JSON	0
JSON members missing images	0
globally unique source keys	11,469,078
globally unique normalized URLs	8,837,750
rows missing normalized URLs	0
rows missing a source SHA-256	7,183,625
retained images below 256 px on the short side	9,036

Native download status

status	rows	share
success	11,983,828	100.00%

Materialization tranche

tranche	rows	share
hf_inherited_lower_resolution	11,172,325	93.23%
img2dataset_mid_high_resolution	811,503	6.77%

Only complete image/JSON pairs from full-stream-valid tar shards are published. The original acquisition had already selected successful materializations, so failed URL attempts that left no complete image/JSON pair cannot be reconstructed from these tars and are not presented as image rows.

Users who need the inherited source surface should obtain it directly from pixparse/cc12m-wds. Users who need a fresh URL materialization should use the retained source_url field with img2dataset; failed or stale URLs may remain unavailable.

Loading metadata or images

import os
from datasets import load_dataset

repo = "BootsofLagrangian/cc12m-jpeg-hybrid-resolution"
read_token = os.environ["HF_READ_TOKEN"]

metadata = load_dataset(repo, "metadata", split="train", streaming=True, token=read_token)
metadata_row = next(iter(metadata))
print(metadata_row["source_url"], metadata_row["image_shard"], metadata_row["image_member"])

images = load_dataset(
    "webdataset",
    data_files={"train": [f"hf://datasets/{repo}@main/images/*.tar"]},
    split="train",
    streaming=True,
    token=read_token,
    storage_options={"token": read_token},
)
sample = next(iter(images))
image_field = next(name for name in ("webp", "jpg", "jpeg", "png") if name in sample)
image = sample[image_field]
source_record = sample["json"]

Replace main with a commit hash for a frozen run. The metadata config does not download the full image payload; the WebDataset example streams gated TAR samples.

Access and use are limited to research reproducibility. Each image remains subject to the rights held by the owner of its source URL. By downloading or using an image, the user acknowledges those third-party rights and accepts responsibility for determining whether the intended research use is permitted. Gating grants no image rights, and no dataset-wide image-license claim is made. Takedown requests should be sent through the repository owner's Hugging Face contact channel.