Search

Most of Large Language Model (LLM) text comprehension capabilities come from generative pre-training on large corpora which includes texts of different domains, languages and tasks. As a consequence the LLM performance in a specific language depends on its representation in the training data which for most state-of-the-art models was biased towards English language. The issue is commonly alleviated by further pre-training on the target language, however, due to limited model capacity this often results in knowledge forgetting and text understanding degradation. We argue that the performance drop can be avoided by employing parameter-efficient tuning methods that preserve the integrity of the original model. In this work, we investigate the effectiveness of different vocabulary optimization and adapter tuning schemes for LLM Russian adaptation. Our experimental results with Solar-10.7B LLM show that language adaptation process can be substantially accelerated by transferring the embeddings from smaller language-tuned counterparts. Moreover, we find that preliminary vocabulary optimization stabilizes further adapter-tuning thus improving target language generalization. By applying our two-stage language adaptation approach we obtain state-of-the-art results on Russian Super Glue and MMLU-RU language understanding datasets for sub-30B parameter open-source LLMs. [ABSTRACT FROM AUTHOR]

This work presents an approach to creating interpreted vector representations of words in which each component of the vector corresponds to a certain interpreted semantic category. To obtain such categories, a lexical and semantic resource in the form of the semantic network RuWordNet is used, as well as a representative corpus of Russian-language texts for generating vector representations. The resulting interpreted vector representations were tested for the ability to display different models in the same vector space. [ABSTRACT FROM AUTHOR]

The NICA accelerator complex includes beam transfer lines and stations for applied research. The first commissioning of the Station of Chip Irradiation (SOCHI) was performed at the end of 2021with С4+ heavy ions extracted from the linear accelerator (HILAC) at an energy of 3.2 MeV/n. The new SOCHI beam transfer line is integrated in the existing HILAC-Booster beamline. The Irradiation Setup for Components of Radioelectronic Apparatus (ISCRA) with ion energy ranging from 150 to 500 MeV/n and the Setup for Investigation of Medical Biological Objects (SIMBO) with the ion energy ranging from 400 to1100 MeV/n are based on the beams extracted from Nuclotron. The equipment of ISCRA and SIMBO stations has been manufactured and is planned to mount in the end of 2022. The beamlines are being designed now. The technical parameters of the beamlines and stations and the results of the first run of the SOCHI station are presented in this study. [ABSTRACT FROM AUTHOR]

This paper presents an approach to creating interpretable word embeddings, in which each component of the vector corresponds to some interpretable semantic category. To obtain such categories, a lexico-semantic resource is used in the form of the RuWordNet semantic network, as well as a representative corpus of Russian-language texts to train vector representations. The resulting interpretable embeddings were evaluated on semantic similarity tasks. [ABSTRACT FROM AUTHOR]

Some fundamental ideas and the results of approximation theory based on them are reviewed. The review covers the development of this branch of mathematics from its origins to the end of the 20th century. [ABSTRACT FROM AUTHOR]

Wide applications of artificial intelligence technologies have led to new threats that cannot be effectively addressed using current tools for secure software development. To meet the challenge, the Research Center for Trusted Artificial Intelligence based on the Institute for Systems Analysis of the Russian Academy of Sciences was founded within the federal project "Artificial Intelligence" in 2021. Its objectives are the creation of scientific and technological basis for building trust in AI technologies. This paper discusses the risks and threats of applying artificial intelligence technologies, and describes the research directions and intermediate results produced at the Research Center for Trusted Artificial Intelligence. [ABSTRACT FROM AUTHOR]

The molecular weight characteristics and physicochemical properties of poly-N-vinyl succinimide synthesized under conditions of reversible addition–fragmentation chain transfer mechanism at different concentrations of the chain transfer agent, dibenzyl trithiocarbonate, were determined. The possibilities of N-vinyl succinimide and N-vinylpyrrolidone copolymers with specified molecular weight characteristics of block and gradient microstructures synthesis are considered A method is shown for producing water-soluble polymer matrices based on copolymers of N-vinylsuccinamic acid and N-vinylpyrrolidone, intended for immobilization of low molecular weight compounds exhibiting bioactive properties. Oil-soluble rimantadine and trimecaine were converted into a water-soluble form by interaction with the produced copolymer. [ABSTRACT FROM AUTHOR]

We give a survey of applications of the ideas and methods of Leonid Kantorovich (1912–1986) to the classical problems of approximation theory. [ABSTRACT FROM AUTHOR]

In this paper we discuss the usefulness of applying semi-automatic checking procedures to existing thesauri for natural language processing—large manually-created lexical-semantic resources. The procedure is based on computation of word vector representations and word semantic similarities on large text collections. The first procedure analyses discrepancies between corpus-based and thesaurus-based word similarities. The second procedure compares the hypernyms (more general words) described in a resource and predicted ones from the relevant collection. We applied the procedures to verification of Russian wordnet RuWordNet. Both procedures helped to find some significant mistakes or inconsistencies in word sense description in RuWordNet, which were difficult to reveal in the resource due to its large volume. The proposed procedures also demonstrate the possibility of fast adaptation of an existing semantic resource to a new domain. [ABSTRACT FROM AUTHOR]

The paper presents the results of applying the BERT representation model in the named entity recognition task (NER) for the cybersecurity domain in Russian. We compare several approaches to domain-specific NER combining BERT fine-tuning on a domain-specific text collection, general labeled data, domain-specific data augmentation, and a domain-specific annotated dataset. We showed that using a BERT model fine-tuned on a domain text collection and pre-trained on the combination of a general dataset and augmented data achieves the best results of named entity recognition. We also studied computational performance of the BERT model in so-called mixed precision regime. [ABSTRACT FROM AUTHOR]

Cryopreservation of fish and amphibian eggs is still an unsolved problem. The formation of ice crystals inside and outside cells acts as a main detrimental factor during a deep freezing of fish eggs, as well as crystal growth (recrystallization and repeated crystallization). Designing efficient cryoprotective media is necessary in order to avoid egg injury from freezing. Additional components that are present in a cryoprotective medium and reduce the thermomechanical stress and cracks of frozen tissues might increase oocyte survival after freezing–thawing. Natural components of eggs and the ovarian fluid are promising as such additives. The formation of ice microparticles was studied in thin layers (0.2 mm) of the ovarian fluid and components of Russian sturgeon egg homogenates upon their cooling to a liquid nitrogen temperature (–196°C). The processes of freezing, ice cracking, and microparticle formation were observed as the temperature was decreased gradually. The shape and size of ice microparticles were found to depend on the composition of the freezing solution. Certain fractions of egg homogenate were assumed to be suitable as components of a cryoprotective medium. [ABSTRACT FROM AUTHOR]

Algorithms for determining the stress intensity factor are developed on the basis of analyzing asymptotic solutions to the problems of deformation of a homogeneous solid with a cut and a plate with a crack located at an interface between two media. The proposed algorithms are used to calculate stress intensity factors using the results of a numerical solution to the problems of loading various flat and spatial homogeneous cracked solids, as well as a plate with a crack located at the interface between two elastic media. It is shown that the calculation results are in good agreement with the data obtained by other methods. [ABSTRACT FROM AUTHOR]

The problem of the greatest possible absolute value of the kth derivative of an algebraic polynomial of order n > k with real coefficients at a given point of the complex plane is considered. It is assumed that the polynomial is bounded by 1 on the interval [-1,1]. It is shown that the solution is attained for the polynomial κ · Tσ, where Tσ is one of the Zolotarev or Chebyshev polynomials and κ is a number. [ABSTRACT FROM AUTHOR]

This paper presents the progress of development of the NICA project vacuum system. The main vacuum requirements in the chain of accelerators and the current status of work are described. Special attention is paid to the problems of achieving ultra-high vacuum in the booster. The test bench for the study of most efficient pumping tools for the beam chamber of the superconducting fast cycling synchrotron has been designed and built for this purpose in cooperation with the VakuumPraha Company (Prague, Czech Republic). The vacuum distribution in the superconducting accelerators with "warm" chamber parts at room temperature was simulated using specialized software. The simulation showed that it is necessary to install additional pumping equipment along the booster perimeter. To solve this problem, the original design of a titanium sublimation pump operating at cryogenic temperatures is under development in collaboration with the Budker Institute of Nuclear Physics (Novosibirsk, Russia). [ABSTRACT FROM AUTHOR]

This paper describes the study of deformation of reinforced concrete. A mathematical model for the interaction of reinforcement with concrete, based on the introduction of a contact layer, whose mechanical characteristics are determined from the experimental data, is developed. The limiting state of concrete is described using the Drucker-Prager theory and the fracture criterion with respect to maximum plastic deformations. A series of problems of the theory of reinforced concrete are solved: stretching of concrete from a central-reinforced prism and pre-stressing of concrete. It is shown that the results of the calculations are in good agreement with the experimental data. [ABSTRACT FROM AUTHOR]

Crack propagation under alternating loading is investigated. Relations between the growth rate of a fatigue defect and loading parameters and the expression for the stress intensity factor are derived for compression of a cracked solid taking into account the possible contact of the crack faces. A model for the deformation of a small region near the crack tip is proposed which allows one to formulate the conditions of residual opening of a growing fatigue crack. The experimental data obtained in tests of steel samples are compared with the results of calculation using the developed procedure. [ABSTRACT FROM AUTHOR]

Experimental studies of the extension of mixed‐mode fatigue cracks were performed on rail steel samples. The direction of growth of fatigue cracks was studied for three types of loading: alternate transverse shear and transverse and longitudinal shears under compressing stresses. For all tested samples, the following common feature in the extension of fatigue cracks was established: the direction of crack growth coincides with the propagation direction of the principal stresses having minimum absolute values. [ABSTRACT FROM AUTHOR]