duongve/dataset_danbooru_new_fork

Danbooru Krea2 2M Metadata

This gated dataset is the image-free, portable definition of a frozen Danbooru training selection. It contains no images, latent tensors, text-encoder tensors, or visual embeddings.

Contents

• records: 2,133,904 unique selected posts in deterministic manifest_order, with original categorized tags, supplemental general tags, final training captions, alternate captions, repaired VLM caption fields, OCR, quality/duplicate signals, exact source archive locators, hashes, and sample_count.
• tag_edges: 80,437,362 normalized anonymous hard-label edges used to audit concept exposure.
• data/auxiliary: entity coverage, rating/capacity, repeat schedule, caption locks, exposure audits, and compact wiki context cards.
• audits: frozen selection/caption/exposure summaries and policies.

One training cycle contains 2,782,172 samples after applying each row's sample_count (1-3). Do not duplicate rows before latent caching: cache each unique record once, then apply repeats in the training sampler.

Reconstructing images

For exact pixels, use source_hub_endpoint, source_repo, source_archive_file, and source_member_name, then verify source_member_sha256. Access to the source image repository is separate from access to this metadata repository.

danbooru_post_url and original_file_url are provided as fallbacks. Images in WebP source repositories were re-encoded and capped at 4 megapixels, so fetching the Danbooru original reproduces the post but not necessarily the exact training bytes. The danbooru_newest_original source points to hub.deepghs.org; the other source repositories are hosted on Hugging Face.

Loading

from datasets import load_dataset

records = load_dataset(
    "SumomoLee/danbooru-krea2-2m-metadata",
    "records",
    split="train",
    token=True,
)

The main records deliberately omit local filesystem paths and tar byte offsets. Those values were machine-specific; archive filenames/member names and content hashes are the portable identifiers.

Integrity

• Source training-manifest SHA-256: 30a3038a7f027e16153082d3c52cdb8209f3bd2807b65c01d2526fe3132e68b1
• Release audit: release_audit.json
• Per-file hashes: file_manifest.json
• Schema version: danbooru-krea2-image-free-metadata-v1

Access is gated because captions, OCR, source metadata, and mature-content ratings may be sensitive. Users must also comply with Danbooru's terms and the terms/licenses of the original image creators and source repositories.