Dataset Viewer
The dataset viewer is not available for this subset.
Cannot get the split names for the config 'default' of the dataset.
Exception:    SplitsNotFoundError
Message:      The split names could not be parsed from the dataset config.
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 286, in get_dataset_config_info
                  for split_generator in builder._split_generators(
                                         ~~~~~~~~~~~~~~~~~~~~~~~~~^
                      StreamingDownloadManager(base_path=builder.base_path, download_config=download_config)
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 81, in _split_generators
                  first_examples = list(islice(pipeline, self.NUM_EXAMPLES_FOR_FEATURES_INFERENCE))
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/webdataset/webdataset.py", line 32, in _get_pipeline_from_tar
                  fs: fsspec.AbstractFileSystem = fsspec.filesystem("memory")
                                                  ~~~~~~~~~~~~~~~~~^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 302, in filesystem
                  cls = get_filesystem_class(protocol)
                File "/usr/local/lib/python3.14/site-packages/fsspec/registry.py", line 239, in get_filesystem_class
                  raise ValueError(f"Protocol not known: {protocol}")
              ValueError: Protocol not known: memory
              
              The above exception was the direct cause of the following exception:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/split_names.py", line 71, in compute_split_names_from_streaming_response
                  for split in get_dataset_split_names(
                               ~~~~~~~~~~~~~~~~~~~~~~~^
                      path=dataset,
                      ^^^^^^^^^^^^^
                      config_name=config,
                      ^^^^^^^^^^^^^^^^^^^
                      token=hf_token,
                      ^^^^^^^^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 340, in get_dataset_split_names
                  info = get_dataset_config_info(
                      path,
                  ...<6 lines>...
                      **config_kwargs,
                  )
                File "/usr/local/lib/python3.14/site-packages/datasets/inspect.py", line 291, in get_dataset_config_info
                  raise SplitsNotFoundError("The split names could not be parsed from the dataset config.") from err
              datasets.inspect.SplitsNotFoundError: The split names could not be parsed from the dataset config.

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

CC12M 1MP+ Realistic Bucketed 1024

This dataset is a self-contained, recaptioned export of the train split from opendiffusionai/cc12m-1mp_plus-realistic. The upstream dataset provides metadata and image URLs; this release contains the downloaded image bytes, so training does not require fetching images from the original URLs.

It contains 573,052 image-caption pairs packaged as aspect-ratio-bucketed TAR shards for text-to-image training. Images are assigned to buckets targeting a 1024px base resolution, with 768px and 512px fallback buckets for smaller source images. Source JPEGs are stored without resizing, cropping, or re-encoding. Non-JPEG inputs are converted to JPEG using the dataset's JPEG export settings. Each sample contains an image, one recaptioned English caption, and a small JSON metadata file.

Dataset Updates

2026-05-06

The dataset was cleaned. A total of 9,720 samples were removed from the original 582,772-sample export, leaving 573,052 image-caption pairs. The removed samples include missing-image placeholders, blank/uniform images, exact duplicate pHash groups, and visually reviewed low-quality images such as heavily compressed, pixelated, upsampled, or otherwise degraded sources.

Contents

  • Images: 573,052
  • Shards: 1,839 uncompressed TAR files
  • Buckets: 75 aspect-ratio buckets
  • Format: bucketed_shards_v2
  • Image encoding: JPEG
  • Image storage: source JPEG passthrough; non-JPEG converted to JPEG
  • Captions: one recaptioned UTF-8 English caption per image
  • Manifest: manifest.json

Directory layout:

manifest.json
buckets/
  r1024_p1024x1024/
    shard-000000.tar
    shard-000001.tar
  r1024_p1216x832/
    shard-000000.tar
    ...

Each TAR shard contains grouped sample triplets:

<sample_key>.jpg
<sample_key>.txt
<sample_key>.json

The .txt file is the caption. The .json file contains per-sample metadata such as bucket target width and height, image encoding information, and caption provenance fields. Bucket dimensions describe the intended training bucket; for source-JPEG passthrough samples, the stored JPEG may retain its original pixel dimensions.

Bucket Summary

The full bucket and shard listing is in manifest.json. The largest buckets are:

Bucket Resolution Images
r1024_p1216x832 1216x832 128,497
r1024_p832x1216 832x1216 77,532
r1024_p1024x1024 1024x1024 77,354
r1024_p1152x832 1152x832 66,226
r1024_p832x1152 832x1152 51,108
r1024_p1344x768 1344x768 41,767
r1024_p896x1088 896x1088 25,455
r1024_p896x1152 896x1152 19,333
r1024_p1280x768 1280x768 17,245
r1024_p1152x896 1152x896 14,826
r1024_p768x1344 768x1344 7,834
r1024_p1088x896 1088x896 7,687

The remaining 63 buckets contain 38,188 images.

Captions

The original captions have been replaced with a generated descriptive caption set intended for image-model training. Captions are generally visual and factual, but may contain errors, omissions, uncertain identity guesses, or subjective wording. They should not be treated as human-verified annotations.

Loading

WebDataset

from pathlib import Path

import webdataset as wds

root = Path("/path/to/dataset")
urls = [str(p) for p in sorted((root / "buckets").glob("*/*.tar"))]

dataset = (
    wds.WebDataset(urls, shardshuffle=True)
    .decode("pil")
    .to_tuple("jpg", "txt", "json")
)

for image, caption, metadata in dataset:
    ...

Plain Python

import json
import tarfile
from pathlib import Path

tar_path = next(Path("/path/to/dataset/buckets").glob("*/*.tar"))

with tarfile.open(tar_path, "r") as tar:
    members = {member.name: member for member in tar if member.isfile()}
    for name in sorted(members):
        if not name.endswith(".txt"):
            continue
        key = name[:-4]
        caption = tar.extractfile(members[f"{key}.txt"]).read().decode("utf-8").strip()
        metadata = json.loads(
            tar.extractfile(members[f"{key}.json"]).read().decode("utf-8")
        )
        image_bytes = tar.extractfile(members[f"{key}.jpg"]).read()
        ...

Manifest

manifest.json is the source of truth for:

  • bucket IDs and target dimensions,
  • shard filenames and sample counts,
  • per-shard byte sizes and SHA-256 hashes,
  • export settings,
  • original URL/source metadata,
  • caption provenance summary.

Consumers should use the manifest rather than assuming a fixed bucket set.

Intended Use

This dataset is intended for research and training workflows involving image-text models, especially systems that benefit from aspect-ratio buckets and WebDataset-style sequential reads.

Limitations

The dataset is derived from web-sourced imagery referenced by the upstream metadata. Some upstream rows are not present because the image URL could not be fetched or decoded, or because the image did not meet export requirements. The remaining data may contain duplicates, watermarks, compression artifacts, OCR errors, incorrectly matched captions, unsafe content, or other dataset noise. Captions are automatically generated and may be inaccurate.

Users are responsible for ensuring that their use of the dataset complies with applicable laws, licenses, platform policies, and model training requirements.

Downloads last month
395