BootsofLagrangian/pd12m-recap-qwen3p5-35b-a3b

PD12M recaptions with Qwen3.5-35B-A3B

Dataset pd12m: 12.316 Million primary train rows plus 8.141 Million bundled extension rows.

The default train split contains exactly the 12,316,334 successfully generated recaptions from the pinned full HF image surface. A user who loads only train receives one deterministic primary recap row per retained HF-full asset and does not need any local materialization. The bundle split contains 8,140,780 additional captions from the UUID-verified current local materialization, the historical local caption surface, the PD12M tranche of PD-Extended, and any additional distinct HF-full captions beyond the selected primary (0 in this freeze). It is an optional long-table extension for extra caption variants and audit use. Across both splits, only exact identity+caption duplicates are removed.

Both splits use the same columns. caption_text is always our generated recap. train is the simple primary consumption surface; bundle is not required to use it. When combining both splits, group by asset_instance_id and retain multiple caption_text values as caption variants.

Together the splits contain 20,457,114 caption rows over 12,379,180 asset identities. Of these, 1,495 are release-scoped local identities that occur only in bundle because their historical URL/UUID bridge could not be recovered. They are preserved with a null source_url and must not be joined to another materialization by numeric key. Other verified UUID identities may also occur only in bundle when no successful HF-full recap exists.

Layout note: the 2026-08-21 revision replaces the earlier all-variants-in-train layout. Existing consumers that want the former union must explicitly concatenate train and bundle; new consumers can load only train for the primary HF-full view.

The public audit/ files account for invalid/unresolved records, exact duplicates, and local captions whose URL bridge was unavailable. Numeric image_id values in those audit files are source-local bookkeeping keys, not stable cross-surface identities; do not use them to join datasets.

The historical surface contains 245 caption rows over 243 local assets whose original WDS JSON sidecars were unavailable. The current resume surface likewise contains 1,252 rows over 1,252 assets without a resume-range bridge. Those captions are preserved rather than discarded, with release-scoped pd12m_legacy_local: or pd12m_current_local: identities and a null source_url. They must not be joined to another image materialization by numeric key. All other retained rows use a verified PD12M UUID/URL identity.

PD12M's original caption field belongs to the upstream `Spawning/PD12M` metadata release and is mirrored in the pinned original_parquet/pd12m_full.parquet inside `Spawning/pd12m-full`. It is source text, not a caption produced by us. The caption_text field in this repository is our generated recap. We used the original metadata only as source context and identity/provenance; we do not relabel it as our output.

The stable identity is the UUID embedded in each PD12M S3 URL. The full-HF numeric key is used only as an asset locator: for key 001550016, shard 00155 and local index 0016 map to source metadata row 155 × 5000 + 16 = 775016. This rule was checked against the tar JSON sidecar and the pinned original metadata Parquet. Numeric keys from independent local materializations are never compared across surfaces.

image_shard and image_member bind 20,393,851 caption rows (99.691%) to the public image-bearing `Spawning/pd12m-full` release at revision 7bda6bfe13c6a39fd0a286adced5df1228041720. Other rows retain their original PD12M URL and UUID for direct retrieval or reconciliation with `Spawning/pd12m-full`. No image payload is duplicated here.

Of 20,457,117 scanned records, 2 invalid rows and 0 rows missing pinned HF source metadata were excluded; 1 exact identity-caption duplicates were removed. No non-empty valid local caption was dropped solely because its URL binding was unavailable.

Captions were produced with the same caption-generation method used in A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions: `Qwen/Qwen3.5-35B-A3B-FP8`, one long-caption stage, a 512-token output limit, Qwen3.5 release sampling defaults, free-form output, and vLLM 0.17.1 serving. Thinking was disabled. Output received whitespace stripping only; no semantic post-processing was applied. See `generation_config.yaml` for the system prompt, user prompt, image preprocessing, and reproduction hyperparameters.

This release is for research and audit reproducibility. It is not training-ready: users must perform semantic post-processing, safety and policy filtering, URL/content identity verification, and image-rights review before training use.

CC BY 4.0 applies only to the generated caption text in this repository. It does not alter the source image or third-party metadata terms.

Loading captions and images separately

from datasets import load_dataset

# Primary HF-full recaptions. This is the default, self-contained caption view.
captions = load_dataset(
    "BootsofLagrangian/pd12m-recap-qwen3p5-35b-a3b",
    split="train",
)
print(captions.column_names)
print(captions[0]["asset_instance_id"], captions[0]["caption_text"])

# Optional additional caption variants from the managed extension surfaces.
bundle = load_dataset(
    "BootsofLagrangian/pd12m-recap-qwen3p5-35b-a3b",
    split="bundle",
)
print(bundle[0]["asset_instance_id"], bundle[0]["caption_text"])

# The independently published upstream image-bearing dataset.
images = load_dataset(
    "Spawning/pd12m-full",
    split="train",
    revision="7bda6bfe13c6a39fd0a286adced5df1228041720",
    streaming=True,
)
image_example = next(iter(images))
print(image_example.keys())

The two datasets are not row-aligned. Join only through the published image_shard and image_member locators, or independently reconcile the PD12M UUID from asset_instance_id/source_url. Do not join by a local numeric download key from another materialization.

@misc{oh2026matchedbudget,
  title={A Matched-Budget Audit Framework for Recaptioned Image-Text Supervision Distributions},
  author={Giyeong Oh and Junghun Park and Yuhan Bae and Youngjae Yu},
  year={2026}
}