cc12_imagenet21k_recap_hq_bucketed

• Title: cc12_imagenet21k_recap_hq_bucketed
• Description: This ~18M rows dataset is a re upload of https://huggingface.co/datasets/gmongaras/CC12M_and_Imagenet21K_Recap_Highqual where the images have been pre bucketed into SDXL style aspect ratio buckets for target training at ~512^2 and ~256^2 pixels, and where about 7M rows were recaptioned with either Gemini or Ministral. To avoid re encoding the images they have been left untouched so cropping and resizing must be done at loading time

---

Technical details

This repository contains a bucketed-shards export (uncompressed TAR shards).

Format

• Format: bucketed_shards_v2
• Created: 2026-01-10T15:53:34.486914+00:00
• Export ID: export-2026-01-10T15:53:34.486914+00:00
• Manifest: manifest.json
• Image mode: passthrough_jpeg

Directory layout:

• manifest.json (global metadata + per-bucket shard listing)
• buckets/<bucket_id>/shard-*.tar

Each TAR shard contains 3 files per sample:

• <key>.jpg (JPEG bytes; either re-encoded RGB JPEG or source JPEG passthrough depending on image_mode)
• <key>.txt (caption text, UTF-8, newline-terminated)
• <key>.json (per-sample metadata: w, h, jpeg, image_mode, caption_variant, caption_selector_index, caption_source_id)

Image preprocessing

Unlike other datasets available in this repo, the images have been left unprocesed and are only pre bucketed into target aspect ratio buckets.

Resize/crop intended usage details at load time:

• Cover scale is scale = max(target_w / src_w, target_h / src_h); if scale > 1, the sample is skipped.
• After resize, a crop box is chosen deterministically from the sample key (sha256 of image_id).
  • Corner strategy chooses a corner from allowed_corners where 0=TL, 1=TR, 2=BL, 3=BR (optional small jitter for corner_jitter).

Buckets / resolutions

• Buckets follow SDXL-style proto buckets defined at a 1024×1024 base.
• Base resolution(s): [512, 256]
• In single-res exports, bucket_id is the proto (1024-base) bucket, e.g. p1024x1024.
• In multi-res exports, buckets are namespaced by base resolution: r<base>_<proto>, e.g. r512_p1024x1024.
• The actual target resolution for each bucket (scaled by the per-bucket base resolution and divisible=32) is stored in:
  • manifest.json → buckets[<bucket_id>].scaled.w/h (and base_resolution)
  • each sample’s <key>.json → w/h

Bucket IDs (preview): r256_p1024x1024, r256_p1088x896, r256_p1152x896, r256_p1216x832, r256_p1344x704, r256_p1344x768, r256_p1472x704, r256_p1600x640, r256_p1728x576, r256_p1856x512, r256_p1984x512, r256_p2048x512, r256_p512x1920, r256_p512x2048, r256_p576x1664, r256_p576x1792, r256_p640x1536, r256_p704x1408, r256_p768x1280, r256_p832x1152, … (+42 more)

Bucket distribution:

bucket_id	target_w×h	aspect	count
r256_p1152x896	288×224	1.286	2,917,754
r256_p1216x832	288×192	1.500	2,194,582
r512_p1216x832	608×416	1.462	2,103,325
r512_p1152x832	576×416	1.385	1,670,669
r512_p1024x1024	512×512	1.000	1,371,409
r256_p1024x1024	256×256	1.000	1,076,075
r256_p896x1152	224×288	0.778	973,834
r256_p832x1152	192×288	0.667	852,564
r512_p832x1216	416×608	0.684	776,575
r512_p832x1152	416×576	0.722	600,705
r512_p1344x768	672×384	1.750	598,503
r512_p1152x896	576×448	1.286	347,970
r256_p1088x896	256×224	1.143	333,188
r512_p1280x768	640×384	1.667	327,133
r512_p896x1152	448×576	0.778	310,259
r256_p960x1024	224×256	0.875	237,259
r256_p1344x768	320×192	1.667	210,921
r512_p1088x896	544×448	1.214	158,051
r512_p896x1088	448×544	0.824	151,487
r512_p960x1024	480×512	0.938	151,345
r512_p768x1280	384×640	0.600	110,424
r512_p1344x704	672×352	1.909	107,761
r512_p1088x960	544×480	1.133	103,963
r512_p1024x960	512×480	1.067	101,368
r512_p960x1088	480×544	0.882	93,788
r256_p768x1280	192×320	0.600	88,633
r256_p1344x704	320×160	2.000	84,077
r512_p768x1344	384×672	0.571	71,153
r512_p1408x704	704×352	2.000	67,854
r512_p1472x704	736×352	2.091	41,942
r512_p1536x640	768×320	2.400	32,080
r256_p704x1408	160×352	0.455	29,786
r256_p1600x640	384×160	2.400	28,130
r512_p704x1408	352×704	0.500	26,929
r256_p1472x704	352×160	2.200	24,696
r256_p1728x576	416×128	3.250	14,366
r512_p704x1472	352×736	0.478	13,622
r256_p640x1536	160×384	0.417	11,144
r512_p640x1536	320×768	0.417	8,459
r512_p1600x640	800×320	2.500	7,885
r256_p576x1664	128×416	0.308	4,034
r256_p2048x512	512×128	4.000	3,650
r512_p640x1600	320×800	0.400	3,129
r256_p1856x512	448×128	3.500	2,886
r256_p1984x512	480×128	3.750	2,369
r512_p1664x576	832×288	2.889	1,293
r512_p576x1664	288×832	0.346	913
r512_p1792x576	896×288	3.111	732
r256_p512x2048	128×512	0.250	730
r256_p576x1792	128×448	0.286	729
r256_p512x1920	128×480	0.267	565
r512_p1856x512	928×256	3.625	522
r512_p576x1792	288×896	0.321	518
r512_p512x1856	256×928	0.276	476
r512_p1728x576	864×288	3.000	450
r512_p576x1728	288×864	0.333	313
r512_p1920x512	960×256	3.750	136
r512_p512x1920	256×960	0.267	114
r512_p1984x512	992×256	3.875	88
r512_p512x1984	256×992	0.258	74
r512_p512x2048	256×1024	0.250	59
r512_p2048x512	1024×256	4.000	56

Caption selection

Available caption variants

selected	variant	images_with_ok_caption
✓	caption_original	18,655,051
✓	caption_ministral_14b_2512	4,000,718
✓	caption_gemini	3,299,328

Missing caption policy: drop

Export summary

• images_seen: 18,655,051
• images_exported: 18,455,504
• skipped_no_caption: 0
• skipped_too_small: 199,547
• decode_errors: 0
• encode_errors: 0

Efficient loading

Recommended

Treat this as a webdataset-style collection of tar shards:

• Prefer sequential reads of tar files for throughput.
• Shuffle at the shard level (and optionally within-shard) for good randomness without expensive random I/O.
• Use manifest.json to list buckets and shards.

Python (webdataset)

import webdataset as wds

urls = "buckets/*/shard-*.tar"  # glob; adjust if you want a single bucket only
ds = (
    wds.WebDataset(urls)
    .decode("pil")            # decodes .jpg to PIL.Image
    .to_tuple("jpg", "txt", "json")
)
for jpg, caption, meta in ds:
    ...

Python (tarfile, no extra deps)

import io, json, tarfile
from pathlib import Path

tar_path = next(Path("buckets").rglob("shard-*.tar"))
with tarfile.open(tar_path, "r") as tf:
    members = tf.getmembers()
    for m in members:
        if not m.name.endswith(".txt"):
            continue
        key = m.name[:-4]
        caption = tf.extractfile(m).read().decode("utf-8").strip()
        meta = json.loads(tf.extractfile(tf.getmember(key + ".json")).read().decode("utf-8"))
        jpg_bytes = tf.extractfile(tf.getmember(key + ".jpg")).read()
        ...