Sberdevices developers conducted testing their own text pattern. According to the results of the test, the RUROBERTA Language model is only inferior to a person, “Lente.ru” in the company.
The model estimate was made using the Benchmark Russian Superglue, which checks the Language data of the AI to logic, common sense, goaling and understanding of the meaning of the text created on the basis of the Glue and Superglue English methodology (General Language Understanding Evaluation – an assessment of a general understanding of the language). According to developers, all the developers of Russian-speaking neural network are used by this test.
RUBERT was taken as the basis of RUROBERTA. Training took place on a large number of Russian-speaking text using Christophary’s supercomputer and took three weeks. The final data set took 250 gigabytes. According to the results of the test, the RUROBERTA model is recognized as the world’s best to recognize the Russian language and not only remembers the tasks, but also mastered the diversity of the phenomena of the language.
“Our ultimate goal is to create reliable intelligent systems to solve diverse tasks in Russian, which can be predecessors of strong artificial intelligence of the domestic assembly,” – commented on the executive vice president of Sberber David Rafalovsky