bulk-cc12m-features — teacher-tower image features for CC12M

Precomputed image-tower features for 10,968,539 CC12M images (all 2,176 shards of pixparse/cc12m-wds), from two teachers: CLIP ViT-B/16 (LAION-2B) (laion/CLIP-ViT-B-16-laion2B-s34B-b88K, 512-d projection space, with captions) and SigLIP base-patch16-384 (google/siglip-base-patch16-384, 768-d, see the tower section for its distinct preprocessing). Built for CLIP distillation research — a student can train against these targets with zero teacher inference, cutting distillation compute by roughly a third. Captions are included per sample, so contrastive and text-side objectives work from this file set alone.

~30 GB total across both towers (vs 1.18 TB for the source images): fp16, one .pt file per source shard per tower.

Exact preprocessing — CLIP tower (the part banks usually leave undocumented)

Feature parity requires bit-level preprocessing agreement. Everything below was verified against live-tower recomputation at cosine 1.00000 before extraction started, and the extractor re-proved it at startup on reference images:

• Model: laion/CLIP-ViT-B-16-laion2B-s34B-b88K (open_clip weights, deterministically converted to transformers CLIPModel format)
• Activation: plain GELU — NOT QuickGELU. (The open_clip default for some checkpoints silently differs; the mismatch reads as cosine ~0.975 and is easy to miss.)
• Geometry: torchvision.transforms.Resize(224, InterpolationMode.BICUBIC) (shortest side) → CenterCrop(224) on the PIL image
• Normalization: CLIP mean (0.48145466, 0.4578275, 0.40821073), std (0.26862954, 0.26130258, 0.27577711)
• Compute: fp32, TF32 disabled; readout get_image_features (CLS → visual projection)
• Storage: fp16, UNNORMALIZED (L2-normalize at load if you need unit vectors; mean norm ≈ 12.76)

Layout

clip_b16_laion2b/features_0000.pt ... features_2175.pt   # one per wds shard
clip_b16_laion2b/train-00000-of-00064.parquet ...        # parquet (viewer/datasets)
siglip_b16_384/features_0000.pt ... features_2175.pt     # 2nd tower, 768-d
siglip_b16_384/train-00000-of-00064.parquet ...          # parquet (viewer/datasets)
siglip_b16_384/ledger.jsonl                              # its per-shard counts
ledger.jsonl                                             # per-shard counts
extract.log                                              # full run log

Each tower ships in TWO forms with identical content: compact fp16 .pt shards (2,176, matching the wds shard numbering) and consolidated parquet (64 files, columns key, caption (CLIP config only), features as fp32 lists carrying fp16 precision — the same convention as the companion COCO bank) for the dataset viewer and datasets loading:

from datasets import load_dataset
ds = load_dataset("AbstractPhil/bulk-cc12m-features",
                  "clip_b16_laion2b", split="train", streaming=True)
row = next(iter(ds))          # {'key', 'caption', 'features': [512 floats]}

Each .pt (loadable with weights_only=True) contains:

field	type	meaning
keys	list[str]	sample keys, exactly the pixparse/cc12m-wds keys
captions	list[str]	the paired CC12M captions
emb	float16 (N, 512)	unnormalized projection features, row-aligned
tower / precision / normalized	str/str/bool	provenance markers

Shards hold 5,040–5,041 samples; row order within a shard is decode-completion order (keys are the join handle, not position).

Second tower: siglip_b16_384/ (added 2026-07-29)

SigLIP base-patch16-384 (google/siglip-base-patch16-384) image features for the same 10,968,539 images: 768-d, fp16, unnormalized, same shard numbering and keys. Preprocessing differs from the CLIP tower and is the SigLIP AutoProcessor default: 384×384 warp resize (bicubic, no crop), mean/std 0.5. Extraction was verified before spend by two gates: the worker path vs the reference processor (cos 1.000000) and reproduction of this tower's stored COCO features from the companion bank (cos 0.999996+). Captions are not duplicated here — join them from clip_b16_laion2b/ by key. Fields per shard: keys, emb (N, 768), provenance markers (no captions).

Usage

import torch
import torch.nn.functional as F
from huggingface_hub import hf_hub_download

p = hf_hub_download("AbstractPhil/bulk-cc12m-features",
                    "clip_b16_laion2b/features_0000.pt", repo_type="dataset")
d = torch.load(p, map_location="cpu", weights_only=True)
z = F.normalize(d["emb"].float(), dim=-1)      # (5041, 512) unit vectors
print(d["keys"][0], d["captions"][0])

To pair features with pixels, stream the matching tar from pixparse/cc12m-wds and join on keys (shard numbering is identical).

Provenance and quality

• Extracted 2026-07-27 in a single 8.1-hour streaming pass (download → decode → embed → discard images), ~376 img/s sustained on one consumer GPU.
• Zero decode errors and zero download failures across all 2,176 shards (ledger.jsonl has per-shard counts; truncated-JPEG tolerance was enabled).
• Full-bank integrity sweep after extraction: every file loads, key/caption/ embedding row counts agree with the ledger, no non-finite values, feature norms stable (12.75–12.77) across the whole run.
• Startup parity gate: 8 held-out reference images through the extraction path vs independently stored features of the same tower — cosine 1.00000.

Licensing note

These are derived features and captions, not images. CC12M imagery remains the property of its owners; captions and the underlying URL list are provided by Google's Conceptual 12M under its stated terms, and the image snapshot mirrored by pixparse/cc12m-wds. Intended for research use.

References

• Changpinyo et al., Conceptual 12M: Pushing Web-Scale Image-Text Pre-Training To Recognize Long-Tail Visual Concepts — https://arxiv.org/abs/2102.08981
• Radford et al., Learning Transferable Visual Models From Natural Language Supervision (CLIP) — https://arxiv.org/abs/2103.00020
• Cherti et al., Reproducible scaling laws for contrastive language-image learning (OpenCLIP / LAION-2B) — https://arxiv.org/abs/2212.07143
• Companion COCO-2017 bank (34 towers): AbstractPhil/bulk-coco-features
• First consumer of this bank: AbstractPhil/clip-vitb-mini-distilled