SberDevices Language Models Ranked The Best In The World

SberBank is a full cycle company engaging on all levels of product developments from creative, engineering, prototyping, design and construction to software development, quality control and hardware testing. Established in 2019, SberDevices is heavily involved with NLP (Natural Language Processing) development, speech technologies, creating computer vision algorithms, developing biometrics systems and neural interfaces, making use of all the latest developments in artificial intelligence and machine learning.

Second only to humanity, SberDevices’ ruRoberta-large fintetune text model has become the leading text comprehension service, after achieving the best score in terms of accuracy for evaluating large text models from the main Russian-language benchmark, Russian SuperGLUE (General Language Understanding Evaluation). If you were wondering who the next best alternatives were, they were also designed by SberDevices, typifying the tremendous strength that Sberbank has achieved in this field.

David Rafalovsky, the Executive Vice President and Head of Technology at Sberbank, highlighted how Sber’s leading experts ‘have spent several years perfecting Russian-language neural networks’, with the ultimate goal of creating ‘reliable intelligent systems that will solve a variety of Russian-language tasks and will serve as the predecessors of robust, homegrown artificial intelligence.’ This is a huge success. Whilst English is the predominant mode of language amongst many global communications, Russian remains one of the hardest languages to learn as an outsider due to different word orders and inflections. This technology helps ensure accurate understanding of the language to a very high degree of accuracy.

Sber developed the Russian SuperGLUE leaderboard for objective evaluation purposes. Being the first of its kind, the leaderboard ranking is based upon how well a neural network company completes tasks on common sense, goal-setting and logic and meaning used by all data researchers working with Russian-language neural networks. Each model is evaluated through several tasks, including DaNetQA, a set of questions on common sense and knowledge with yes/no answers, RCB (Russian Commitment Bank), a classification of causal linkages between a text and a hypothesis, and PARus (Plausible Alternatives for Russian), which assesses goal-setting, choice of alternative options based on common sense, and more. Such a model is not only beneficial in memorizing tasks or predicting the right result, but also learning the particularities and multitude of phenomena of the Russian language.

This success builds upon Sperbank’s recent strengthening of the ties between Russia and the Middle East in the past year, previously discussed in our article here. It exemplifies how Russia has made steps between Russian businesses and investors from foreign regions to further their macroeconomic stability.