class_idx int32 0 999 | class_name stringclasses 998
values | sample_idx int32 0 9 | name stringclasses 10
values | data imagewidth (px) 256 256 |
|---|---|---|---|---|
0 | tench | 0 | 0000.png | |
0 | tench | 1 | 0001.png | |
0 | tench | 2 | 0002.png | |
0 | tench | 3 | 0003.png | |
0 | tench | 4 | 0004.png | |
0 | tench | 5 | 0005.png | |
0 | tench | 6 | 0006.png | |
0 | tench | 7 | 0007.png | |
0 | tench | 8 | 0008.png | |
0 | tench | 9 | 0009.png | |
1 | goldfish | 0 | 0000.png | |
1 | goldfish | 1 | 0001.png | |
1 | goldfish | 2 | 0002.png | |
1 | goldfish | 3 | 0003.png | |
1 | goldfish | 4 | 0004.png | |
1 | goldfish | 5 | 0005.png | |
1 | goldfish | 6 | 0006.png | |
1 | goldfish | 7 | 0007.png | |
1 | goldfish | 8 | 0008.png | |
1 | goldfish | 9 | 0009.png | |
2 | great white shark | 0 | 0000.png | |
2 | great white shark | 1 | 0001.png | |
2 | great white shark | 2 | 0002.png | |
2 | great white shark | 3 | 0003.png | |
2 | great white shark | 4 | 0004.png | |
2 | great white shark | 5 | 0005.png | |
2 | great white shark | 6 | 0006.png | |
2 | great white shark | 7 | 0007.png | |
2 | great white shark | 8 | 0008.png | |
2 | great white shark | 9 | 0009.png | |
3 | tiger shark | 0 | 0000.png | |
3 | tiger shark | 1 | 0001.png | |
3 | tiger shark | 2 | 0002.png | |
3 | tiger shark | 3 | 0003.png | |
3 | tiger shark | 4 | 0004.png | |
3 | tiger shark | 5 | 0005.png | |
3 | tiger shark | 6 | 0006.png | |
3 | tiger shark | 7 | 0007.png | |
3 | tiger shark | 8 | 0008.png | |
3 | tiger shark | 9 | 0009.png | |
4 | hammerhead | 0 | 0000.png | |
4 | hammerhead | 1 | 0001.png | |
4 | hammerhead | 2 | 0002.png | |
4 | hammerhead | 3 | 0003.png | |
4 | hammerhead | 4 | 0004.png | |
4 | hammerhead | 5 | 0005.png | |
4 | hammerhead | 6 | 0006.png | |
4 | hammerhead | 7 | 0007.png | |
4 | hammerhead | 8 | 0008.png | |
4 | hammerhead | 9 | 0009.png | |
5 | electric ray | 0 | 0000.png | |
5 | electric ray | 1 | 0001.png | |
5 | electric ray | 2 | 0002.png | |
5 | electric ray | 3 | 0003.png | |
5 | electric ray | 4 | 0004.png | |
5 | electric ray | 5 | 0005.png | |
5 | electric ray | 6 | 0006.png | |
5 | electric ray | 7 | 0007.png | |
5 | electric ray | 8 | 0008.png | |
5 | electric ray | 9 | 0009.png | |
6 | stingray | 0 | 0000.png | |
6 | stingray | 1 | 0001.png | |
6 | stingray | 2 | 0002.png | |
6 | stingray | 3 | 0003.png | |
6 | stingray | 4 | 0004.png | |
6 | stingray | 5 | 0005.png | |
6 | stingray | 6 | 0006.png | |
6 | stingray | 7 | 0007.png | |
6 | stingray | 8 | 0008.png | |
6 | stingray | 9 | 0009.png | |
7 | cock | 0 | 0000.png | |
7 | cock | 1 | 0001.png | |
7 | cock | 2 | 0002.png | |
7 | cock | 3 | 0003.png | |
7 | cock | 4 | 0004.png | |
7 | cock | 5 | 0005.png | |
7 | cock | 6 | 0006.png | |
7 | cock | 7 | 0007.png | |
7 | cock | 8 | 0008.png | |
7 | cock | 9 | 0009.png | |
8 | hen | 0 | 0000.png | |
8 | hen | 1 | 0001.png | |
8 | hen | 2 | 0002.png | |
8 | hen | 3 | 0003.png | |
8 | hen | 4 | 0004.png | |
8 | hen | 5 | 0005.png | |
8 | hen | 6 | 0006.png | |
8 | hen | 7 | 0007.png | |
8 | hen | 8 | 0008.png | |
8 | hen | 9 | 0009.png | |
9 | ostrich | 0 | 0000.png | |
9 | ostrich | 1 | 0001.png | |
9 | ostrich | 2 | 0002.png | |
9 | ostrich | 3 | 0003.png | |
9 | ostrich | 4 | 0004.png | |
9 | ostrich | 5 | 0005.png | |
9 | ostrich | 6 | 0006.png | |
9 | ostrich | 7 | 0007.png | |
9 | ostrich | 8 | 0008.png | |
9 | ostrich | 9 | 0009.png |
Dataset Distillation via Drifting: distilled data
The distilled datasets from Dataset Distillation via Drifting (NeurIPS 2026).
- Browse: drifting-distillation.github.io/data
- Code: github.com/GeorgeCazenavette/drifting-distillation
Usage
Each distilled set is its own config with a single train split:
from datasets import load_dataset
ds = load_dataset("georgecazenavette/drifting-distillation", "imagenet-1k_dinov3_noise_ipc10", split="train")
To save a set as an image folder ({class:04d}/{name}), download its parquet files and write
out the data bytes:
import os, glob
import pyarrow.parquet as pq
from huggingface_hub import snapshot_download
files = "image/imagenet-1k/dinov3_noise/10_ipc*.parquet" # path of the set (see Format)
root = snapshot_download("georgecazenavette/drifting-distillation", repo_type="dataset", allow_patterns=files)
for path in sorted(glob.glob(os.path.join(root, files))):
pf = pq.ParquetFile(path)
for i in range(pf.num_row_groups): # one row group per class
for r in pf.read_row_group(i, columns=["class_idx", "name", "data"]).to_pylist():
os.makedirs(f"dinov3_noise_ipc10/{r['class_idx']:04d}", exist_ok=True)
with open(f"dinov3_noise_ipc10/{r['class_idx']:04d}/{r['name']}", "wb") as f:
f.write(r["data"])
Format
Files are {modality}/{dataset}/{variant}/{ipc}_ipc.parquet; sets larger than 4 GB are split
into _partN.parquet. Each file has one row per sample and one row group per class, so a
single class can be read without downloading the rest. manifest.json lists every set and its files.
| column | type | |
|---|---|---|
class_idx |
int32 | class index (sorted-wnid / sorted-folder order) |
class_name |
string | |
sample_idx |
int32 | index within the class |
name |
string | file name |
data |
image / video / audio | file bytes |
- Images: the 256x256 PNGs evaluated in the paper.
- Video: 16-frame 256x256 clips, stored as MP4.
- Audio: 524,288-sample (32.8 s) clips at 16 kHz, stored as 16-bit WAV.
Sets
noise sets are distilled from noise; real sets are initialized from RDED crops.
ImageNet-1k (12 sets)
| config | backbone | setting | classes | per class | size |
|---|---|---|---|---|---|
imagenet-1k_dinov3_noise_ipc10 |
DINOv3 ViT-B | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3_noise_ipc50 |
DINOv3 ViT-B | noise | 1000 | 50 | 6.6 GB |
imagenet-1k_dinov3_real_ipc10 |
DINOv3 ViT-B | real | 1000 | 10 | 1.2 GB |
imagenet-1k_dinov3_real_ipc50 |
DINOv3 ViT-B | real | 1000 | 50 | 5.8 GB |
imagenet-1k_mocov3_noise_ipc10 |
MoCo-v3 | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_mocov3_noise_ipc50 |
MoCo-v3 | noise | 1000 | 50 | 6.3 GB |
imagenet-1k_mocov3_real_ipc10 |
MoCo-v3 | real | 1000 | 10 | 1.2 GB |
imagenet-1k_mocov3_real_ipc50 |
MoCo-v3 | real | 1000 | 50 | 5.9 GB |
imagenet-1k_resnet18_noise_ipc10 |
ResNet-18 | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_resnet18_noise_ipc50 |
ResNet-18 | noise | 1000 | 50 | 6.4 GB |
imagenet-1k_resnet18_real_ipc10 |
ResNet-18 | real | 1000 | 10 | 1.2 GB |
imagenet-1k_resnet18_real_ipc50 |
ResNet-18 | real | 1000 | 50 | 5.9 GB |
ImageNet-1k (more backbones) (17 sets)
| config | backbone | setting | classes | per class | size |
|---|---|---|---|---|---|
imagenet-1k_clip_noise_ipc10 |
CLIP ViT-B | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_clip_noise_ipc50 |
CLIP ViT-B | noise | 1000 | 50 | 6.5 GB |
imagenet-1k_clip_real_ipc10 |
CLIP ViT-B | real | 1000 | 10 | 1.1 GB |
imagenet-1k_clip_real_ipc50 |
CLIP ViT-B | real | 1000 | 50 | 5.5 GB |
imagenet-1k_eupe_noise_ipc10 |
EuPE ViT-B | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_eva02_noise_ipc10 |
EVA-02 ViT-B | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_eva02_noise_ipc50 |
EVA-02 ViT-B | noise | 1000 | 50 | 6.7 GB |
imagenet-1k_eva02_real_ipc10 |
EVA-02 ViT-B | real | 1000 | 10 | 1.2 GB |
imagenet-1k_eva02_real_ipc50 |
EVA-02 ViT-B | real | 1000 | 50 | 5.9 GB |
imagenet-1k_siglip2_noise_ipc10 |
SigLIP2 ViT-B | noise | 1000 | 10 | 1.4 GB |
imagenet-1k_siglip2_noise_ipc50 |
SigLIP2 ViT-B | noise | 1000 | 50 | 7.0 GB |
imagenet-1k_siglip2_real_ipc10 |
SigLIP2 ViT-B | real | 1000 | 10 | 1.2 GB |
imagenet-1k_siglip2_real_ipc50 |
SigLIP2 ViT-B | real | 1000 | 50 | 6.3 GB |
imagenet-1k_vit_noise_ipc10 |
ViT-B (supervised) | noise | 1000 | 10 | 1.3 GB |
imagenet-1k_vit_noise_ipc50 |
ViT-B (supervised) | noise | 1000 | 50 | 6.7 GB |
imagenet-1k_vit_real_ipc10 |
ViT-B (supervised) | real | 1000 | 10 | 1.1 GB |
imagenet-1k_vit_real_ipc50 |
ViT-B (supervised) | real | 1000 | 50 | 5.7 GB |
Ablations (12 sets)
| config | backbone | setting | classes | per class | size |
|---|---|---|---|---|---|
imagenet-1k_dinov3_noise_augs0_ipc10 |
DINOv3 ViT-B | no augmentation | 1000 | 10 | 1.4 GB |
imagenet-1k_dinov3_noise_augs1_ipc10 |
DINOv3 ViT-B | 1 augmentation | 1000 | 10 | 1.4 GB |
imagenet-1k_dinov3_noise_augs2_ipc10 |
DINOv3 ViT-B | 2 augmentations | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3_noise_augs5_ipc10 |
DINOv3 ViT-B | 5 augmentations | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3_noise_modecls_ipc10 |
DINOv3 ViT-B | CLS token only | 1000 | 10 | 1.4 GB |
imagenet-1k_dinov3_noise_modeclsmean_ipc10 |
DINOv3 ViT-B | CLS + mean patch token | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3_noise_nodecorr_ipc10 |
DINOv3 ViT-B | no color decorrelation | 1000 | 10 | 1.5 GB |
imagenet-1k_dinov3_noise_pixel_ipc10 |
DINOv3 ViT-B | pixel parameterization (no pyramid) | 1000 | 10 | 1.7 GB |
imagenet-1k_dinov3_noise_temp0.05_ipc10 |
DINOv3 ViT-B | temperature 0.05 | 1000 | 10 | 1.4 GB |
imagenet-1k_dinov3_noise_temp5_ipc10 |
DINOv3 ViT-B | temperature 5 | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3large_noise_ipc10 |
DINOv3 ViT-L | model size | 1000 | 10 | 1.3 GB |
imagenet-1k_dinov3small_noise_ipc10 |
DINOv3 ViT-S | model size | 1000 | 10 | 1.4 GB |
ImageNette / ImageWoof (16 sets)
| config | backbone | setting | classes | per class | size |
|---|---|---|---|---|---|
imagenette_dinov3_noise_ipc10 |
DINOv3 ViT-B | noise | 10 | 10 | 14 MB |
imagenette_dinov3_noise_ipc50 |
DINOv3 ViT-B | noise | 10 | 50 | 67 MB |
imagenette_dinov3_real_ipc10 |
DINOv3 ViT-B | real | 10 | 10 | 12 MB |
imagenette_dinov3_real_ipc50 |
DINOv3 ViT-B | real | 10 | 50 | 59 MB |
imagenette_resnet18_noise_ipc10 |
ResNet-18 | noise | 10 | 10 | 13 MB |
imagenette_resnet18_noise_ipc50 |
ResNet-18 | noise | 10 | 50 | 65 MB |
imagenette_resnet18_real_ipc10 |
ResNet-18 | real | 10 | 10 | 12 MB |
imagenette_resnet18_real_ipc50 |
ResNet-18 | real | 10 | 50 | 60 MB |
imagewoof_dinov3_noise_ipc10 |
DINOv3 ViT-B | noise | 10 | 10 | 14 MB |
imagewoof_dinov3_noise_ipc50 |
DINOv3 ViT-B | noise | 10 | 50 | 70 MB |
imagewoof_dinov3_real_ipc10 |
DINOv3 ViT-B | real | 10 | 10 | 13 MB |
imagewoof_dinov3_real_ipc50 |
DINOv3 ViT-B | real | 10 | 50 | 63 MB |
imagewoof_resnet18_noise_ipc10 |
ResNet-18 | noise | 10 | 10 | 14 MB |
imagewoof_resnet18_noise_ipc50 |
ResNet-18 | noise | 10 | 50 | 68 MB |
imagewoof_resnet18_real_ipc10 |
ResNet-18 | real | 10 | 10 | 13 MB |
imagewoof_resnet18_real_ipc50 |
ResNet-18 | real | 10 | 50 | 63 MB |
Audio and video (2 sets)
| config | backbone | setting | classes | per class | size |
|---|---|---|---|---|---|
fma_musicfm-fma_ipc5 |
MusicFM | noise | 8 | 5 | 80 MB |
hmdb51_videomaev2_ipc5 |
VideoMAE-v2 | noise | 51 | 5 | 25 MB |
Teachers
teachers/ holds the 10-class ResNet-18 teachers for ImageNette and ImageWoof that the code
uses for real initialization and knowledge-distillation evaluation (downloaded automatically).
They are RDED's released observers, redistributed under
RDED's Apache-2.0 license, with the ImageWoof classifier rows reordered to sorted-wnid class
order; see teachers/README.md.
Citation
@inproceedings{cazenavette2026drifting,
title={Dataset Distillation via Drifting},
author={Cazenavette, George and Quan, Angelina and Daras, Giannis and Torralba, Antonio and Sitzmann, Vincent},
booktitle={Advances in Neural Information Processing Systems},
year={2026}
}
- Downloads last month
- 520