README.md

---
license: other
license_name: cc12m-derived
license_link: https://github.com/google-research-datasets/conceptual-12m
pretty_name: bulk-cc12m-features (10 teacher towers + consensus target)
tags:
- clip
- siglip
- dinov3
- features
- embeddings
- cc12m
- knowledge-distillation
size_categories:
- 100M<n<1B
configs:
- config_name: clip_b16_laion2b
  data_files:
  - split: train
    path: clip_b16_laion2b/train-*.parquet
- config_name: clip_b16_openai
  data_files:
  - split: train
    path: clip_b16_openai/train-*.parquet
- config_name: clip_b32_openai
  data_files:
  - split: train
    path: clip_b32_openai/train-*.parquet
- config_name: clip_b32_laion2b
  data_files:
  - split: train
    path: clip_b32_laion2b/train-*.parquet
- config_name: clip_b32_datacomp
  data_files:
  - split: train
    path: clip_b32_datacomp/train-*.parquet
- config_name: clip_l14_openai
  data_files:
  - split: train
    path: clip_l14_openai/train-*.parquet
- config_name: clip_l14_laion2b
  data_files:
  - split: train
    path: clip_l14_laion2b/train-*.parquet
- config_name: clip_l14_datacomp
  data_files:
  - split: train
    path: clip_l14_datacomp/train-*.parquet
- config_name: dinov3_l16
  data_files:
  - split: train
    path: dinov3_l16/train-*.parquet
- config_name: siglip_b16_384
  data_files:
  - split: train
    path: siglip_b16_384/train-*.parquet
- config_name: consensus_gpa
  data_files:
  - split: train
    path: consensus_gpa/train-*.parquet
---

# bulk-cc12m-features — ten teacher towers over CC12M, plus their consensus

Precomputed image-tower features for **10,968,539 CC12M images** (all 2,176
shards of
[pixparse/cc12m-wds](https://huggingface.co/datasets/pixparse/cc12m-wds))
from **ten independent teacher extractions** — eight CLIP variants across
three pretraining corpora and two model scales, SigLIP, and DINOv3 — plus
one **derived consensus target**.
About **110 million feature vectors**, roughly 130 GPU-hours of extraction,
so that a student can be distilled against any of these teachers (or a
consensus of them) with **zero teacher inference at training time**.

Captions are included with the `clip_b16_laion2b` config and join to every
other tower by key.

## The towers

| config | teacher | dim | space | mean ‖v‖ |
|---|---|---|---|---|
| `clip_b16_laion2b` | [CLIP ViT-B/16 LAION-2B](https://huggingface.co/laion/CLIP-ViT-B-16-laion2B-s34B-b88K) | 512 | CLIP projection | 12.77 |
| `clip_b16_openai` | [CLIP ViT-B/16](https://huggingface.co/openai/clip-vit-base-patch16) | 512 | CLIP projection | 10.61 |
| `clip_b32_openai` | [CLIP ViT-B/32](https://huggingface.co/openai/clip-vit-base-patch32) | 512 | CLIP projection | 10.52 |
| `clip_b32_laion2b` | [CLIP ViT-B/32 LAION-2B](https://huggingface.co/laion/CLIP-ViT-B-32-laion2B-s34B-b79K) | 512 | CLIP projection | 10.60 |
| `clip_b32_datacomp` | [CLIP ViT-B/32 DataComp-XL](https://huggingface.co/laion/CLIP-ViT-B-32-DataComp.XL-s13B-b90K) | 512 | CLIP projection | 14.63 |
| `clip_l14_openai` | [CLIP ViT-L/14](https://huggingface.co/openai/clip-vit-large-patch14) | 768 | CLIP projection | 18.80 |
| `clip_l14_laion2b` | [CLIP ViT-L/14 LAION-2B](https://huggingface.co/laion/CLIP-ViT-L-14-laion2B-s32B-b82K) | 768 | CLIP projection | 19.45 |
| `clip_l14_datacomp` | [CLIP ViT-L/14 DataComp-XL](https://huggingface.co/laion/CLIP-ViT-L-14-DataComp.XL-s13B-b90K) | 768 | CLIP projection | 21.49 |
| `siglip_b16_384` | [SigLIP base-384](https://huggingface.co/google/siglip-base-patch16-384) | 768 | SigLIP projection | — |
| `dinov3_l16` | [DINOv3 ViT-L/16](https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m) | 1024 | **CLS hidden state** | 14.74 |
| `consensus_gpa` | *derived* — GPA mean of the five B-tier CLIP towers | 512 | consensus (own frame) | 1.00 (pre-normalized) |

**`consensus_gpa` is a derived target, not an extraction.** It is the
generalized-Procrustes mean shape of the five B-tier CLIP towers
(`clip_b16_laion2b`, `clip_b32_openai`, `clip_b32_laion2b`,
`clip_b32_datacomp`, `clip_b16_openai`): each tower L2-normalized, then
five iterations of *{orthogonally align every tower to the running mean on
a fixed 10,000-row subsample via fp64 SVD, rotate the full tower, re-mean,
renormalize}*, converging at mean alignment ≈ 0.91. Rows are already unit
norm. This is the exact tensor the published consensus-distilled students
were trained against — see
[clip-vitb-mini-distilled](https://huggingface.co/AbstractPhil/clip-vitb-mini-distilled).
Note that the mean has **no privileged frame**: a student trained on it
lands in the consensus frame, not any teacher's, which is why that model
ships a fitted rotation alongside its weights.

**`dinov3_l16` is different in kind.** DINOv3 has no projection head and no
text tower: its 1024-d vector is the CLS hidden state, not a
language-aligned projection. It is a full, valid extraction over the same
images, but it is not a drop-in member of a CLIP consensus and has no text
side to evaluate against. Fusing it with the CLIP towers needs an explicit
cross-space alignment (learned projectors, or a whitened-Procrustes fit),
not a plain average.

All extracted towers are stored **fp16 and UNNORMALIZED** — L2-normalize at
load if you want unit vectors. (`consensus_gpa` is the exception: it is
already unit norm by construction.)

## Join by key, never by position

Every tower covers the **identical key set**, shard for shard (verified on
first, middle and last shards). But **row order differs between towers**:
extraction ran decode-completion-ordered, so shard *n* of two towers holds
the same 5,041 images in different order.

```python
# correct
order = {k: i for i, k in enumerate(tower_a["keys"])}
b_aligned = tower_b["emb"][[order[k] for k in tower_b["keys"]]]

# WRONG — silently misaligns every row
pairs = zip(tower_a["emb"], tower_b["emb"])
```

## Forms

| form | path | use |
|---|---|---|
| parquet | `{config}/train-*.parquet` (64 files) | dataset viewer, `load_dataset`, streaming |
| shards | `{config}/features_0000.pt … features_2175.pt` | 1:1 with the source wds shards; fp16, compact |
| concat | `{config}/all_concat.pt` | whole tower in one tensor — `{keys, emb}` |

The ten extracted towers carry all three forms. `consensus_gpa` is derived
rather than extracted, so it has no per-wds-shard form — parquet and concat
only.

```python
from datasets import load_dataset
ds = load_dataset("AbstractPhil/bulk-cc12m-features",
                  "clip_l14_laion2b", split="train", streaming=True)
row = next(iter(ds))          # {'key', 'features': [768 floats]}
```

Parquet stores features as fp32 lists carrying fp16 precision (the
convention of the companion COCO bank). The `.pt` shards and concats load
with `weights_only=True`.

## Exact preprocessing (the part banks usually leave undocumented)

Feature parity requires bit-level preprocessing agreement, and it is **per
tower, not per family** — every extraction was gated before spending
compute: the fast decode path against the tower's reference processor, and
reproduction of that tower's stored COCO features from the companion bank,
both at cosine ≥ 0.9999.

| tower | geometry | normalization |
|---|---|---|
| CLIP B/16, B/32, L/14 (OpenAI + DataComp) | `Resize(224, bicubic)` shortest edge → `CenterCrop(224)` | CLIP mean/std `(0.48145466, 0.4578275, 0.40821073)` / `(0.26862954, 0.26130258, 0.27577711)` |
| `clip_l14_laion2b` | same as above | **mean/std 0.5** — its own `open_clip_config` declares it; using CLIP norm reads 0.935 instead of 0.999987 |
| `siglip_b16_384` | `Resize((384, 384), bicubic)` warp, no crop | mean/std 0.5 |
| `dinov3_l16` | `Resize((224, 224), bilinear)` warp | ImageNet mean/std `(0.485, 0.456, 0.406)` / `(0.229, 0.224, 0.225)` |

Other constants that matter: **plain GELU, not QuickGELU**, for the CLIP
towers (the open_clip default for some checkpoints differs, and the
mismatch reads as cosine ≈ 0.975 against stored features — close enough to
miss); fp32 compute; readout `get_image_features` (CLS → visual
projection) for CLIP/SigLIP and **CLS** for DINOv3. Two checkpoints ship
open_clip weights only (LAION B/16, DataComp B/32) and were converted
deterministically to `transformers` format; hub-hosted conversions of
open_clip checkpoints do **not** reproduce these banks — convert from the
open_clip original.

## Provenance and quality

- Extracted 2026-07-27..31. The primary tower ran as a single 8.1-hour
  streaming pass (download → decode → embed → discard; the source imagery
  was never resident) at ~376 img/s; the remaining towers followed on
  rented A40s.
- **Zero decode errors and zero download failures** across all 2,176 shards
  of the primary pass; per-shard counts in each tower's `ledger.jsonl`.
- Full-bank integrity sweep on the primary: every file loads, key/caption/
  embedding counts agree with the ledger, no non-finite values, feature
  norms stable across the run.
- Shards hold 5,040–5,041 samples each.

## Licensing note

These are *derived features and captions*, not images. CC12M imagery
remains the property of its owners; captions and the underlying URL list
come from Google's
[Conceptual 12M](https://github.com/google-research-datasets/conceptual-12m)
under its stated terms, and the image snapshot mirrored by
[pixparse/cc12m-wds](https://huggingface.co/datasets/pixparse/cc12m-wds).
Intended for research use.

## References

- Changpinyo et al., *Conceptual 12M* — https://arxiv.org/abs/2102.08981
- Radford et al., *CLIP* — https://arxiv.org/abs/2103.00020
- Cherti et al., *Reproducible scaling laws for contrastive language-image
  learning* (OpenCLIP) — https://arxiv.org/abs/2212.07143
- Gadre et al., *DataComp* — https://arxiv.org/abs/2304.14108
- Zhai et al., *SigLIP* — https://arxiv.org/abs/2303.15343
- Siméoni et al., *DINOv3* — https://arxiv.org/abs/2508.10104
- Companion COCO bank (34 towers):
  [AbstractPhil/bulk-coco-features](https://huggingface.co/datasets/AbstractPhil/bulk-coco-features)
- First consumer of this bank:
  [AbstractPhil/clip-vitb-mini-distilled](https://huggingface.co/AbstractPhil/clip-vitb-mini-distilled)