Skip to main content
Advertisement
Live broadcast

The key condition for the competitiveness of Russian AI models is named

0
Photo: IZVESTIA/Eduard Kornienko
Озвучить текст
Select important
On
Off

Russian language models can remain competitive and become universal if their developers have access to a wide range of data, including foreign scientific and educational materials. This is the conclusion reached by the authors of the ITMO Institute of AI LLM market research.

The researchers note that domestic models are better able to cope with tasks related to the history, law and culture of Russia than foreign models. In these areas, the Russian language and the cultural context in which Russian-language data help neural networks navigate are important. According to experts, the optimal share of Russian-language data in the training array of the universal model is 5-30%. At the same time, the Russian—speaking segment of the Internet is several times smaller than the global one. If access to international sources is blocked, the model will be worse able to solve problems unrelated to the Russian context.

"When creating Russian LLMs, Russian-language data is necessary, but isolation from international corpora will lead to the fact that the model will lag further behind its foreign counterparts and cope worse with tasks beyond the local linguistic context. It is important to give Russian LLMs the opportunity to learn from all publicly available content — on the Russian Internet and beyond — so that the models, on the one hand, understand the cultural code of our country, and on the other, develop their cognitive abilities using all available knowledge," the authors of the study emphasized.

They see a separate reserve in Russian materials that are not available on the open Internet. These include the collections of libraries and archives, scientific, educational and professional collections. Such data has almost no effect on the general level of models, but it can significantly improve their performance in narrow professional fields, such as medicine, law, construction, geology, and other fields.

The paper also says that copyright regulations should not narrow the range of data sources for training models and slow down development, but, on the contrary, should remove unnecessary barriers.

Переведено сервисом «Яндекс Переводчик»

Live broadcast