OpenAI CLIP

Organisation: OpenAI
Family: CLIP
Released: 2021-02-26 (arXiv v1 submission date (https://arxiv.org/abs/2103.00020v1). The model card gives 'Model Date: January 2021' and later checkpoint releases up to April 2022.)
Checkpoints: RN50, RN101, RN50x4, RN50x16, RN50x64, ViT-B/32, ViT-B/16, ViT-L/14, ViT-L/14@336px
Weights: Open · MIT (code and weights)
Links: Paper Model card Code

Training data disclosure: Partial list

Level determined for 0% of benchmark cells.

TestTest data seen0
SeenSeen, split unclear0
TrainTraining data seen0
Rel.Related data seen0
NoNot seen0
?Unknown8

“The model was trained on publicly available image-caption data.”

raw.githubusercontent.com · model card · Data · retrieved 2026-10-15

“We do not intend for this dataset to be used as the basis for any commercial or deployed model and will not be releasing the dataset.”

raw.githubusercontent.com · model card · Data Mission Statement · retrieved 2026-10-15

Training corpora

WIT (WebImageText), OpenAI CLIP pre-training set · 2 sources · pre-training
Evidence that OpenAI CLIP uses it

“We refer to this dataset as WIT for WebImageText.”

arxiv.org · paper · Section 2.2 · retrieved 2026-10-15
Contents of this corpus
- publicly available sources on the Internet (web-crawled image-text pairs) (broad web data: may contain benchmark test sets) · split: unknown
  Collection-level entry. The crawl dates and the websites are not disclosed. The model card says OpenAI crawled 'a handful of websites'.
  
  Evidence
  
  “we constructed a new dataset of 400 million (image, text) pairs collected form a variety of publicly available sources on the Internet.”
  
  arxiv.org · paper · Section 2.2 · retrieved 2026-10-15
  
  “A large portion of the data comes from our crawling of the internet.”
  
  raw.githubusercontent.com · model card · Data · retrieved 2026-10-15
- YFCC100M (filtered subset) · split: unknown
  The model card says 'pre-existing image datasets such as YFCC100M', so other pre-existing datasets may be included but are not named.
  
  Evidence
  
  “This is due to it being constructed out of YFCC100M, which our pre-training dataset contains a filtered subset of.”
  
  arxiv.org · paper · Section 5 (Data Overlap Analysis) · retrieved 2026-10-15
  
  “This was done through a combination of crawling a handful of websites and using commonly-used pre-existing image datasets such as”
  
  raw.githubusercontent.com · model card · Data · retrieved 2026-10-15

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

There was no pre-training deduplication. OpenAI chose to measure overlap afterwards instead (Section 5). The quote 'Instead, we document how much overlap occurs and how performance changes due to these overlaps.' is in Section 5. The checkpoint list comes from the model card's 'Model Versions'. All sizes share WIT.

Sources

Learning Transferable Visual Models From Natural Language Supervision · paper · retrieved 2026-10-15
Model Card: CLIP · model card · retrieved 2026-10-15