Contamination MapVision / Model
DINOv2
- Organisation
- Meta AI
- Family
- DINO · version 2
- Released
- 2023-04-14 (arXiv v1 submission date (https://arxiv.org/abs/2304.07193v1))
- Checkpoints
- dinov2_vits14, dinov2_vitb14, dinov2_vitl14, dinov2_vitg14, dinov2_vits14_reg, dinov2_vitb14_reg, dinov2_vitl14_reg, dinov2_vitg14_reg
- Weights
- Open · Apache-2.0
- Links
- PaperModel cardCode
Training data disclosure: Partial list
Level determined for 0% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen0
- ?Unknown8
“We chose to include as many datasets as possible in the pretraining data in order to cover as many domains as possible.”
“we collect a raw unfiltered dataset of images from a publicly available repository of crawled web data.”
Training corpora
-
LVD-142M
· 4 sources (complete list) · pre-training
Evidence that DINOv2 uses it
“LVD-142M (see paper)”
raw.githubusercontent.com · model card · Training Data · retrieved 2026-10-15 “We assemble our curated LVD-142M dataset by retrieving, from a large pool of uncurated data, images that are close to those in several curated datasets.”
arxiv.org · paper · Section 3 · retrieved 2026-10-15 Contents of this corpus
-
ImageNet-22k / – · split: unknown
Included 'as is'. The '–' in Table 15 means no split is named.
Evidence
“ImageNet-22k / –”
arxiv.org · paper · Table 15 (inclusion: as is, 14,197,086 images; also used as a sample-based retrieval seed) · retrieved 2026-10-15 “how they were included (as is without retrieval or via sample-based or cluster-based retrieval)”
arxiv.org · paper · Table 15 caption · retrieved 2026-10-15 -
Google Landmarks v2 / train (clean) · split: train
Evidence
“Google Landmarks v2 / train (clean)”
arxiv.org · paper · Table 15 (inclusion: as is, 1,580,470 images; also used as a sample-based retrieval seed) · retrieved 2026-10-15 “how they were included (as is without retrieval or via sample-based or cluster-based retrieval)”
arxiv.org · paper · Table 15 caption · retrieved 2026-10-15 - Mapillary SLS / train · split: train
-
uncurated web images (1.2B unique images from a publicly available repository of crawled web data), retrieved by similarity to curated seed datasets (broad web data: may contain benchmark test sets) · split: unknown
About 124M of the 142M images are web images retrieved as neighbours of seed images. Any public web image, including copies of benchmark images, could be retrieved unless it was removed by the deduplication step. The crawl repository and its dates are not named.
Evidence
“we collect a raw unfiltered dataset of images from a publicly available repository of crawled web data.”
arxiv.org · paper · Section 3 · retrieved 2026-10-15 “This results in 1.2B unique images.”
arxiv.org · paper · Section 3 · retrieved 2026-10-15 “We assemble our curated LVD-142M dataset by retrieving, from a large pool of uncurated data, images that are close to those in several curated datasets.”
arxiv.org · paper · Section 3 · retrieved 2026-10-15
-
ImageNet-22k / – · split: unknown
Benchmark cells
Industrial anomaly detection
Unknown MVTec AD · training data not disclosed in enough detail
Unknown MVTec AD 2 · training data not disclosed in enough detail
Unknown MVTec LOCO AD · training data not disclosed in enough detail
Unknown VisA · training data not disclosed in enough detail
Unknown Real-IAD · training data not disclosed in enough detail
Unknown MPDD · training data not disclosed in enough detail
Unknown BTAD · training data not disclosed in enough detail
Unknown KolektorSDD2 · training data not disclosed in enough detail
Notes
disclosure is partial_list because the curated seeds and the 'as is' datasets are fully listed in Table 15, but the web crawl repository is unnamed and only described in general terms. The checkpoint names are the torch.hub names. The model card covers 8 models, with and without registers, all trained on LVD-142M. Deduplication stated by its authors (not a tracked benchmark): test or validation sets of every benchmark used in the DINOv2 paper.
Sources
- DINOv2: Learning Robust Visual Features without Supervision · paper · retrieved 2026-10-15
- DINOv2 MODEL_CARD.md · model card · retrieved 2026-10-15