StanislavLev/tiny-clip-image-encoders-adapter

TinyCLIP Adapter Feature Repository

This repository stores precomputed feature tables for experiments that align visual encoder embeddings to a TinyCLIP text embedding space using contrastive losses and lightweight adapters.

Raw source images are not included. The files here are Parquet metadata, caption/text embeddings, image embeddings, and manifests that make the feature sets easy to join and reuse.

Layout

.
├── flickr30k/
│   ├── metadata/
│   ├── text_embeddings/
│   └── image_embeddings/
├── flickr30k_all/
│   ├── metadata/
│   ├── text_embeddings/
│   └── image_embeddings/
├── imagenet1k/
│   ├── metadata/
│   ├── text_embeddings/
│   └── image_embeddings/
├── laion1m/
│   ├── metadata/
│   ├── text_embeddings/
│   └── image_embeddings/
└── manifests/
    ├── datasets.yaml
    └── encoders.yaml

Use manifests/datasets.yaml and manifests/encoders.yaml as the top-level index. Each embedding folder also contains a manifest.json with the encoder, normalization, pooling, row count, image count, and dimensionality.

Dataset Folders

Folder	Split layout	Rows	Images	Notes
flickr30k	train, validation, test	155,070 captions	31,014	Original split-preserving Flickr30K metadata and embeddings.
flickr30k_all	train	155,070 captions	31,014	Flickr30K train/validation/test merged into one split for full-dataset training.
imagenet1k	validation	50,000 rows	50,000	ImageNet-1K validation features.
laion1m	train	984,340 captions	982,723 images	LAION subset feature dump. DINOv3 image tables store one row per unique image_uri; loaders expand to caption rows when needed.

Main Embedding Sets

Dataset	Folder	Type	Rows	Dim	Notes
flickr30k	text_embeddings/tinyclip_vit_39m_16_text_19m_yfcc15m	text	155,070	512	One embedding per caption across train/validation/test.
flickr30k	image_embeddings/dinov3_vits16_pretrain_lvd1689m	image	31,014	384	DINOv3-Small, one embedding per unique image across train/validation/test.
flickr30k_all	text_embeddings/tinyclip_vit_39m_16_text_19m_yfcc15m	text	155,070	512	Merged single train split.
flickr30k_all	image_embeddings/dinov3_vits16_pretrain_lvd1689m	image	155,070	384	DINOv3-Small, expanded to caption-row order.
flickr30k_all	image_embeddings/dinov3_vitb16_pretrain_lvd1689m	image	155,070	768	DINOv3-Base, expanded to caption-row order.
flickr30k_all	image_embeddings/dinov3_vitl16_pretrain_lvd1689m	image	155,070	1024	DINOv3-Large, expanded to caption-row order.
imagenet1k	image_embeddings/dinov3_vitb16_pretrain_lvd1689m	image	50,000	768	ImageNet validation.
imagenet1k	image_embeddings/dinov3_vitl16_pretrain_lvd1689m	image	50,000	1024	ImageNet validation.
laion1m	image_embeddings/dinov3_vitb16_pretrain_lvd1689m	image	982,723	768	DINOv3-Base, one row per unique image_uri.
laion1m	image_embeddings/dinov3_vitl16_pretrain_lvd1689m	image	982,723	1024	DINOv3-Large, one row per unique image_uri.

All listed DINOv3 image embeddings are L2-normalized CLS-token features.

Loading Examples

Load DINOv3-Small features for merged Flickr30K:

from huggingface_hub import hf_hub_download
import pandas as pd

repo_id = "StanislavLev/tiny-clip-image-encoders-adapter"

path = hf_hub_download(
    repo_id=repo_id,
    repo_type="dataset",
    filename="flickr30k_all/image_embeddings/dinov3_vits16_pretrain_lvd1689m/train-00000.parquet",
)
image_features = pd.read_parquet(path)

Load split-preserving Flickr30K DINOv3-Small validation features:

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="StanislavLev/tiny-clip-image-encoders-adapter",
    repo_type="dataset",
    filename="flickr30k/image_embeddings/dinov3_vits16_pretrain_lvd1689m/validation.parquet",
)
validation_image_features = pd.read_parquet(path)

Load the encoder index:

from huggingface_hub import hf_hub_download
import yaml

path = hf_hub_download(
    repo_id="StanislavLev/tiny-clip-image-encoders-adapter",
    repo_type="dataset",
    filename="manifests/encoders.yaml",
)
with open(path, "r", encoding="utf-8") as f:
    encoders = yaml.safe_load(f)

Common Columns

Metadata tables include image_id, caption_id, caption_index, caption, file_name, image path fields, split, and source fields where available.

Text embedding tables include caption_id, image_id, split, caption_index, encoder metadata, embedding_dim, and embedding.

Image embedding tables include image_id, image_uri where available, split, file_name, image path fields, encoder metadata, embedding_dim, and embedding.

Notes

flickr30k_all is designed for simple row-aligned adapter training: caption embeddings and expanded image embeddings share the same row count and split.

flickr30k keeps the original split structure and stores one row per unique image for each available DINOv3 encoder, which is useful when you want image-level features without caption duplication.

laion1m DINOv3 Base and Large stores are keyed by image_uri and contain one feature row per successfully encoded unique image. Training loaders should join/expand these rows against metadata instead of assuming raw row order alignment.

Check the original dataset and model licenses before redistributing derived models or using these features outside the allowed terms of the source assets.