PJSC Sberbank : Sber makes new expanded versions of ruCLIP model available publicly

The ruCLIP line of models for ranking images and tags in Russian and evaluating the semantic similarity of images and text developed by the Sber AI and SberDevices teams is now available on GitHub, complementing ruCLIP Small, which was published previously.

ruCLIP Base exclusive and ruCLIP Large exclusive, the commercial versions with higher quality and more parameters, are available in DataHub on SberCloud ML Space, which features pre-trained models, datasets, and containers. Using a number of datasets, the models successfully overtook the original English-language CLIP model and the Russian-English translator.

The successful training of ruCLIP and the models' open-source availability will facilitate the efficient resolution of a number of computer vision-related tasks in a variety products and services using a zero-shot setup, without requiring expensive additional training.

The release features six ruCLIP models that differ in terms of the sizes of the patch used (14×14, 16×16, 32×32) and the size of the input images (224×224, 336×336 и 384×384). The semantics behind the naming of the models is as follows:

• ruclip-vit-base-patch16-224
• ruclip-vit-base-patch32-224
• ruclip-vit-base-patch32-384 - ruCLIP Base
• ruclip-vit-large-patch14-224 - ruCLIP Large
• ruclip-vit-large-patch14-336 - ruCLIP Large exclusive - SberCloud ML Space DataHub only
• ruclip-vit-base-patch16-384 - ruCLIP Base exclusive - SberCloud ML Space DataHub only

The six new trained models can be compared in detail on GitHub. They were trained based on an independently amassed dataset made of 240 million pairs. The process took 12 full days using 256 Tesla GPU A100s on the SberCloud ML Space platform.

Alexander Vedyakhin, first deputy chairman of the executive board, Sberbank:

"The Sber ecosystem is a leader in ML solutions. We are already offering developers, data scientists, and business representatives even more instruments and services, from ML development platforms like SberCloud ML Space to completed ML Solutions like SmartSpeech. Over the last year, the Sber AI and SberDevices teams joined up to release transformer models, ruGPT-3 & family, which include the popular text-to-image model ruDALL-E. The models top various benchmark rankings and, as opposed to the majority of similar solutions, are publicly available. The exclusive commercial models are available in DataHub on SberCloud ML Space. This helps businesses resolve numerous tasks related to the creation of their own breakthrough products based on ML, accelerate time to market, and reduce development expenses."

Sber AI is Sber's R&D division. It is in charge of developing AI technology and implementing it in various areas of life and business.

SberDevices, a Sber ecosystem company, is a center of expertise on AI-powered solutions in fields such as speech technology, natural language processing, and face and voice biometrics. Established in May 2019 as a department within Sberbank's Technology Unit, the company also focuses on creating smart devices for end users and corporate clients.

SberCloud is the Sber ecosystem's cloud platform. The company offers a wide range of infrastructure and platform cloud solutions, as well as tools for working with artificial intelligence based on the Christofari and Christofari Neo supercomputers. SberCloud's IT platforms and services are at the core of the Sber digital ecosystem and are also available to external clients, including companies, governmental organizations, and individuals.