Pacific-i64/complexity-atlas-images

Complexity Atlas Images

Complexity Atlas Images is a provenance-preserving CC0 image-text bank for image captioning and multimodal research. Version 1.0.0 contains 336,245 normalized museum images and English metadata-derived captions.

Splits

Split	Samples
Train	329,510
Validation	3,428
Test	3,307
Total	336,245

The dataset occupies approximately 6.67 GB across 68 WebDataset TAR shards.

Source

This release is derived from Mitsua/art-museums-pd-440k at the pinned revision fba945da78b36262eb9272067197cc28d06cffbf.

The retained records originate from the Metropolitan Museum of Art, the Smithsonian Open Access collection, the Cleveland Museum of Art and the Art Institute of Chicago. Every published row records CC0 as its source license.

Processing

• EXIF orientation is applied before processing;
• images are converted to RGB and center-cropped to 256 x 256;
• output images use deterministic WebP encoding at quality 90;
• exact duplicates are removed by SHA-256;
• near duplicates are removed with an 88-bit structure-plus-color perceptual hash;
• source images smaller than 128 pixels on either axis are rejected;
• no normalized caption is retained more than five times;
• identical normalized captions are assigned to one split, preventing target leakage;
• selection within repeated-caption groups is deterministic by image-ID SHA-256.

The source build examined 408,112 records. Before caption-frequency refinement, 398,532 images passed the image, license and duplicate filters. The final caption cap retained 336,245 samples.

Sample structure

Each WebDataset sample contains:

<image_id>.webp   normalized image
<image_id>.txt    caption target
<image_id>.json   provenance and integrity metadata

The JSON record includes the source URLs, museum, title, author, license, upstream revision, exact hash, perceptual hash, dimensions and split.

Loading

Load the complete training corpus directly from its WebDataset shards:

from datasets import load_dataset

dataset = load_dataset(
    "webdataset",
    data_files={
        "train": "hf://datasets/Pacific-i64/complexity-atlas-images/train/*.tar",
        "validation": "hf://datasets/Pacific-i64/complexity-atlas-images/validation/*.tar",
        "test": "hf://datasets/Pacific-i64/complexity-atlas-images/test/*.tar",
    },
    split="train",
    streaming=True,
)

sample = next(iter(dataset))
print(sample.keys())

The default preview configuration contains 100 representative samples per split in Parquet. It exists only to provide a responsive Dataset Viewer while the complete 336,245-sample corpus remains stored once in the TAR shards. The TAR paths are deliberately excluded from the Viewer configuration because of an upstream Dataset Viewer failure in its WebDataset/fsspec worker.

Audit artifacts

• build_manifest.json records filtering and split counts;
• audit_report.json records caption, source and license statistics;
• files_manifest.json contains the byte size, sample count and SHA-256 of every shard;
• review_sample.jsonl provides a deterministic 100-row human-review sample.

All 68 TAR archives were parsed during the final audit and their sample counts were checked against the build manifest.

Limitations

This is a museum-domain dataset rather than a balanced representation of the visual world. Captions are derived from museum titles and catalog metadata; they can describe materials, periods and collection information rather than only visible content. They were not generated by a captioning model. Some distinct images legitimately share one caption, although each normalized caption is capped at five occurrences.

Use this release as a clean first-stage image bank. Broader everyday-image coverage requires additional independently licensed sources.

Licensing

The published image and metadata records retain their upstream CC0 status. Source URLs and the pinned upstream revision are preserved for auditability. The builder code is maintained separately under Apache-2.0.