CommonPool-128-DINOv2-small

This is a derived WebDataset version of quinnlue/commonpool-256-ssl with 128x128 stored images and precomputed facebook/dinov2-small image embeddings.

Provenance

• Source dataset: quinnlue/commonpool-256-ssl at revision 78cc60b3ba29bd3fe3bd212eb22faefc4906471d
• Source license/use terms: non-commercial research, inherited from the source dataset
• Model: facebook/dinov2-small at revision ed25f3a31f01632728cabb09d1542f84ab7b0056
• Embedding feature: last_hidden_state[:, 0]
• Embedding dtype/shape: float32, (384,)

Structure

Total rows: 5,684,754

Split	Rows	Shards
train	5,664,754	71
val	10,000	1
test	10,000	1

Each sample contains:

• <key>.jpg: 128x128 RGB JPEG, quality 95
• <key>.npy: DINOv2-small CLS embedding, float32 shape (384,)
• <key>.json: compact per-sample provenance and target hash metadata

manifest.parquet preserves the source manifest metadata and adds target image fields such as target_shard_path, target_index_in_shard, target_sha256, target_jpeg_sha256, target_width, and target_height. Embeddings are stored only in the tar shards, not duplicated in the manifest.

Image and Embedding Processing

Source 256x256 JPEGs are decoded as RGB, resized to 128x128 with Lanczos, and re-encoded as JPEG quality 95. DINOv2 embeddings are computed from the final stored 128x128 JPEG bytes after reopening them with PIL and applying the standard Hugging Face image processor for facebook/dinov2-small.

Usage

import io
import numpy as np
import webdataset as wds
from PIL import Image
from huggingface_hub import hf_hub_url

url = hf_hub_url(
    repo_id="quinnlue/commonpool-128-dinov2-small",
    repo_type="dataset",
    filename="data/train/train-000000.tar",
)

dataset = wds.WebDataset(url).decode()
sample = next(iter(dataset))
image = Image.open(io.BytesIO(sample["jpg"])).convert("RGB")
embedding = np.load(io.BytesIO(sample["npy"]), allow_pickle=False)