OpenLVD-200M

200M curated web image URLs with captions. This is an updated version of the one used in the original SigLino paper. Built for vision foundation model training and distillation. This repo holds metadata only. We do not host images.

• Rows: 197,657,490
• Files: 200 parquet files, data/part-00000.parquet to data/part-00199.parquet.
• About 1M rows per file. zstd compression.
• Paper: SigLino, arXiv 2512.20157

Sources

Source	source value	Selected	After NSFW filter
COYO-700M	coyo	117,043,287	115,589,056
DFN-2B (subset of DataComp CommonPool)	dfn	44,296,498	43,984,937
Re-LAION-5B research-safe	relaion	38,660,215	38,083,497
Total		200,000,000	197,657,490

The NSFW filter removed 2,342,510 rows (1.17%): every row with a Falconsai/nsfw_image_detection nsfw probability above 0.2.

How it was built

Pool

• Images from COYO-700M, DFN and Re-LAION.
• Each image embedded with the DINOv3 ViT-L/16 CLS token.

Clustering

• Spherical 4-level hierarchical k-means.
• Method from Vo et al. 2024, Automatic Data Curation for Self-Supervised Learning: A Clustering-Based Approach.

Level	Clusters
1 (finest)	24,799,866
2	992,000
3	99,200
4 (top)	24,800

Sampling

• Balanced hierarchical sampling down to 200M.
• Within-cluster strategy: random (r in the Vo et al. code).
• All 24,800 top-level clusters are represented.
• Images per top-level cluster: median 4,833, max 18,991.

Columns

Column	Type	Description
uid	string	Stable 12-digit id, e.g. 000000000000.
url	string	Original image URL.
sha256	string	SHA-256 of the original downloaded image bytes. From the img2dataset download metadata of the source datasets.
caption	string	Original alt-text from the source dataset. Mostly English, some other languages. A few rows are empty.
width	int32	Original image width in pixels.
height	int32	Original image height in pixels.
source	string	coyo, dfn or relaion.
cluster_level1	int32	Level 1 cluster id (finest, 24,799,866 clusters).
cluster_level2	int32	Level 2 cluster id (992,000 clusters).
cluster_level3	int32	Level 3 cluster id (99,200 clusters).
cluster_level4	int32	Level 4 cluster id (top, 24,800 clusters).
nsfw_score	float32	Probability of the nsfw class from Falconsai/nsfw_image_detection. Rows above 0.2 were removed.

Download the images

Use img2dataset.

Get the metadata:

huggingface-cli download tiiuae/OpenLVD-200m --repo-type dataset --local-dir . --include "data/*"

Download the images:

img2dataset --url_list data/ --input_format parquet \
  --url_col url --caption_col caption \
  --compute_hash sha256 --verify_hash '["sha256","sha256"]' \
  --save_additional_columns '["uid","source","cluster_level1","cluster_level2","cluster_level3","cluster_level4"]' \
  --output_format webdataset --output_folder openlvd200m \
  --processes_count 16 --thread_count 64 \
  --image_size 336 --resize_mode keep_ratio_largest --resize_only_if_bigger True \
  --downscale_interpolation bicubic --encode_quality 95

Load the metadata

from datasets import load_dataset

ds = load_dataset("tiiuae/OpenLVD-200m", split="train", streaming=True)
for row in ds.take(3):
    print(row["uid"], row["source"], row["url"], row["caption"])

License

• Our metadata (uid, cluster ids, nsfw_score, selection): CC-BY-4.0.
• url and caption come from the source datasets and stay under their terms:
  • COYO-700M: CC-BY-4.0, Kakao Brain.
  • DataComp CommonPool (DFN rows): CC-BY-4.0. The DFN-2B selection is released by Apple under the Apple Model License for Research (apple-amlr), research use only. Copy in LICENSE.apple-amlr.
  • Re-LAION-5B: Apache-2.0, LAION, gated research release.
• Images belong to their owners. Downloading and using them is your responsibility.

Citation

@article{chaybouti2025siglino,
  title   = {SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models},
  author  = {Chaybouti, Sofian and Narayan, Sanath and Dahou, Yasser and L{\^e} Khac, Ph{\'u}c H. and Singh, Ankit and Huynh, Ngoc Dung and Para, Wamiq Reyaz and Kuehne, Hilde and Hacid, Hakim},
  journal = {arXiv preprint arXiv:2512.20157},
  year    = {2025}
}