cc12_recap_2_4m_512

A 4,000,000-sample, quality-filtered selection from the 512-base buckets of data-archetype/cc12_imagenet21k_recap_hq_bucketed, a bucketed export of CC12M and ImageNet-21K images with synthetic captions.

It complements data-archetype/cc12_imagenet21k_512_subset, an earlier subset of the same source:

	cc12_imagenet21k_512_subset	cc12_recap_2_4m_512 (this dataset)
Samples	1,649,276	4,000,000
Selection	every 512-capable sample with a Gemini caption	the best of the remaining 512-capable samples (see below)
Captions	Gemini	Ministral 14B or the upstream recaption; 3,202 defective captions repaired with Gemma 4 31B

The two datasets share no samples, and images that nearly duplicate one in cc12_imagenet21k_512_subset were removed here, so they can be combined without repeats. Both are drawn from the source's r512_* buckets: use them instead of those buckets, not in addition to them.

Contents

• Images: 4,000,000
• Shards: 3,932 uncompressed TAR files of up to 1,024 samples (repaired shards carry a .caption-<sha256> suffix: shard-000123.caption-<sha256>.tar)
• Size: 299.2 GB
• Base resolution: 512
• Buckets: 40 aspect-ratio buckets, dimensions divisible by 32
• Format: bucketed_shards_v1
• Images: source JPEG bytes, not resized or re-encoded
• Captions: UTF-8 text, one English caption per image, copied from the source except for 3,202 defective captions replaced by the caption repair below
• Metadata: one JSON sidecar per sample

Source

Candidates are the 9,363,532 samples in the source's r512_* buckets, whose images cover a 512-base bucket without upscaling. Removing the 1,649,276 samples of cc12_imagenet21k_512_subset leaves 7,714,256 candidates, of which this dataset keeps 4,000,000 (52%).

The source export re-buckets gmongaras/CC12M_and_Imagenet21K_Recap_Highqual (CC12M and ImageNet-21K, images with a side under 256px removed) and adds captions from newer models for part of it.

Selection

Every candidate was scored with:

• an image embedding from google/siglip2-base-patch16-224;
• a watermark probability from prithivMLmods/Watermark-Detection-SigLIP2;
• an aesthetic score: a small MLP on that embedding, trained to predict aesthetic-predictor-v2.5 (on google/siglip-so400m-patch14-384) from 234,496 samples where the exact score was computed (cross-validated Spearman 0.88);
• a non-photo probability: logistic regression on the embedding, trained on labels from caption keywords (merchandise mockups, posters and text graphics, documents, slides and screenshots, maps, charts and floor plans, logos and clipart); cross-validated AUC 0.90 against those labels.

Filters, applied in order to the 7,714,256 candidates:

filter	removed
Caption mentions a watermark	209,162
Watermark probability >= 0.8	771,441
Non-photo probability >= 0.35	591,337
In one of the 1,773 embedding clusters whose mean non-photo probability is >= 0.25	232,064
Aesthetic score below 3.34 (15th percentile of candidates)	572,678
Near duplicate (embedding cosine >= 0.95) of a higher-scoring candidate or of a cc12_imagenet21k_512_subset image	545,101
Not selected by the diversity quotas	792,473

Diversity quotas: the embeddings were grouped into 16,384 spherical k-means clusters. Each cluster gets a quota proportional to the square root of its remaining size, filled with its highest-scoring samples. This caps over-represented content such as stock interiors and product shots while keeping rare content; 14,414 clusters are represented.

Aesthetic score of the selected samples: median 4.40 (all candidates: 4.10), 5th-95th percentile 3.57-5.63.

All filters are automatic. Some watermarked, graphic or low-quality images remain.

Captions

Each sample carries the caption the source export chose for it, or its repair, recorded in caption_variant:

caption_variant	origin	samples
caption_ministral_14b_2512	long descriptive caption by mistralai/ministral-14b-2512	1,002,200
caption_original	upstream recaption (LLaVA-NeXT, shortened with Llama 3), usually one sentence	2,994,384
caption_gemma_4_31b_it	repair caption by google/gemma-4-31b-it (see below)	3,199
caption_gemini	Gemini caption	216
caption_original_short	upstream short recaption, used as one repair fallback	1

Caption repair

3,202 captions (0.08%) were defective and were replaced in place; images, sample keys, sample order and every other caption are unchanged:

• 1,822 degenerate captions, mostly generation loops: the same sentence, phrase, number sequence or character run repeated until the caption reached up to 51,000 characters (27,000 tokens), plus captions cut off after a few words, captions ending mid-sentence, captions with tag meta-language, and 12 hand-reviewed captions (templated enumerations, a binary-digit stream, a caption preamble);
• 1,380 upstream recaptions that leaked the captioning prompt or its inputs: 1,005 opened with a preamble such as "A detailed but succinct caption for the image:" or "Here is the distilled caption:", and 379 cited the source page's alt text ("as indicated by the alt-text attribute", "given the context provided"); 4 also had other defects.

They were found by an automatic detector census of all 4,000,000 captions. Replacement captions were written by google/gemma-4-31b-it through OpenRouter on fp8 or bf16 endpoints, with a descriptive prompt that opens with the image type and puts every transcribed piece of visible text in double quotes. Each was checked before use: the response stopped normally; the detectors found no loop, truncation, malformed ending, tag or prompt leakage, unquoted transcribed text, unbalanced quote or repetition; the caption has at least two sentences or 25 words; and it is at most 1,024 tokens of the T5Gemma 2 tokenizer. A caption that failed was retried once. Three captions flagged only for repeated words quoted from the image (the band name "The The", for example) were accepted on review. Three captions that failed twice fell back to the sample's own upstream recaption, which passed the same checks. The JSON sidecar of each repaired sample records the new caption_variant and caption_source_id, and manifest.json lists the caption sources.

Layout

manifest.json
buckets/<bucket_id>/shard-*.tar

Each TAR shard contains three files per sample key:

<key>.jpg
<key>.txt
<key>.json

<key> is the sample key of the source export. The JSON sidecar holds the bucket target (w, h, bucket_id, proto_w_1024, proto_h_1024), the caption provenance fields, and three fields added by this selection: source_bucket_id, aesthetic_v25_probe (predicted aesthetic score) and watermark_prob. The global manifest.json lists buckets, shard paths, sample counts, checksums and the selection settings and statistics.

Buckets

Bucket names use 1024-base dimensions: p1216x832, for example, targets 608x416. Images are stored as-is, so a JPEG is usually larger than its bucket target; resize to cover the target and crop at load time (manifest.json records the source export's crop convention). The largest buckets are:

bucket	target size	samples
p1216x832	608x416	1,140,329
p1152x832	576x416	598,957
p1024x1024	512x512	432,218
p832x1216	416x608	376,654
p1344x768	672x384	276,159
p832x1152	416x576	248,771
p1280x768	640x384	156,525
p1152x896	576x448	146,880
p896x1152	448x576	130,603
p1088x896	544x448	62,736

See manifest.json for the full bucket list.

Loading

Using webdataset:

import glob

import webdataset as wds

urls = sorted(glob.glob("buckets/*/shard-*.tar"))
dataset = (
    wds.WebDataset(urls, shardshuffle=100)
    .decode("pil")
    .to_tuple("jpg", "txt", "json")
)

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

Using the Python standard library:

import json
import tarfile
from pathlib import Path

tar_path = next(Path("buckets").rglob("shard-*.tar"))

with tarfile.open(tar_path, "r") as tf:
    for member in tf:
        if not member.isfile() or not member.name.endswith(".txt"):
            continue
        key = member.name[:-4]
        caption = tf.extractfile(member).read().decode("utf-8").strip()
        meta = json.loads(tf.extractfile(tf.getmember(key + ".json")).read())
        image_bytes = tf.extractfile(tf.getmember(key + ".jpg")).read()
        break

Verification

Checked after export:

• Manifest sample count: 4,000,000
• Shard sample count: 4,000,000
• Unique sample keys: 4,000,000
• Empty captions: 0
• Undecodable images: 0
• Caption repair: every repaired shard differs from its original only in the repaired captions and their sidecars; JPEG payloads are byte-identical

License

The images come from CC12M and ImageNet-21K and remain subject to the terms of their original sources. Captions were generated by the models named above. Check that your use complies with those terms.