Contamination MapVision / Model
SigLIP 2
- Organisation
- Google DeepMind
- Family
- SigLIP · version 2
- Released
- 2025-02-20 (arXiv v1 submission date (https://arxiv.org/abs/2502.14786v1). The HF repo google/siglip2-so400m-patch14-384 was created 2025-02-17.)
- Checkpoints
- ViT-B/32, ViT-B/16, ViT-L/16, So400m/14, So400m/16, g-opt/16, B/16 NaFlex, google/siglip2-so400m-patch14-384
- Weights
- Open · Apache-2.0
- Links
- PaperModel cardCode
Training data disclosure: Partial list
Level determined for 25% of benchmark cells.
- TestTest data seen0
- SeenSeen, split unclear0
- TrainTraining data seen0
- Related data seen0
- NoNot seen2
- ?Unknown6
“We use the WebLI dataset [10] containing 10 billion images and 12 billion alt-texts covering 109 languages.”
Training corpora
-
WebLI (Web Language-Image)
· 1 source · pre-training
Evidence that SigLIP 2 uses it
“SigLIP 2 is pre-trained on the WebLI dataset”
huggingface.co · model card · Training data · retrieved 2026-10-15 “we compose the mixture such that 90% of the training image-text pairs is sourced from English web pages, and the remaining 10% from non-English web pages”
arxiv.org · paper · Section 2.1 · retrieved 2026-10-15 Contents of this corpus
-
images and texts from the public web (alt-text, page title, OCR) (broad web data: may contain benchmark test sets) · split: unknown
Collection-level entry. The crawl is not public and the source pages are not documented.
Evidence
“we introduce WebLI, a multilingual imagelanguage dataset built from images and texts available on the public web.”
arxiv.org · paper · Section 3.2 · retrieved 2026-10-15 “The WebLI dataset covers about 10 billion images and 12 billion alt-texts in 109 languages.”
arxiv.org · paper · Appendix B · retrieved 2026-10-15 “The dataset is built from a subset of public web pages.”
arxiv.org · paper · Appendix G, WebLI datasheet · retrieved 2026-10-15
-
images and texts from the public web (alt-text, page title, OCR) (broad web data: may contain benchmark test sets) · split: unknown
Benchmark cells
Industrial anomaly detection
Unknown MVTec AD · training data not disclosed in enough detail
Not seen MVTec AD 2 · its test items were published from 2025-03-27, after the model's stated training-data cutoff (2022-12-31)
Evidence (3)
“Over what timeframe was the data collected? 2021-2022”
“arXiv:2503.21622v1 [cs.CV] 27 Mar 2025”
“All image data is available at https://www.mvtec.com/company/research/datasets/mvtec-ad-2.”
Unknown MVTec LOCO AD · training data not disclosed in enough detail
Unknown VisA · training data not disclosed in enough detail
Not seen Real-IAD · its test items were published from 2024-03-19, after the model's stated training-data cutoff (2022-12-31)
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
The paper says: 'we release model checkpoints at four sizes: ViT-B (86M), L (303M), So400m (400M), and g (1B)'. Several resolutions per size are listed in big_vision README_siglip2.md. The paper does not say whether the WebLI snapshot differs from PaLI's 2021-2022 collection.
Sources
- SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding, Localization, and Dense Features · paper · retrieved 2026-10-15
- google/siglip2-so400m-patch14-384 model card · model card · retrieved 2026-10-15
- big_vision README_siglip2.md · readme · retrieved 2026-10-15