Large language models (LLMs) are widely deployed in settings where both reliability and efficiency matter. We present a calibrated, seed‑robust empirical comparison of an encoder fine‑tuned model (bidirectional encoder representations from transformers (BERT)‑base) and a decoder in‑context model (generative pre-trained transformer (GPT)‑2 small) across Stanford question answering dataset v2.0 (SQuAD v2.0) and general language understanding evaluation (GLUE)-multi-genre natural language inference (MNLI), Stanford sentiment treebank 2 (SST‑2). Beyond accuracy, we assess reliability (expected calibration error with reliability diagrams and confidence–coverage analysis) and efficiency (latency, memory, throughput) under matched conditions and three fixed seeds. BERT‑base yields higher accuracy and lower calibration error, while GPT‑2 narrows gaps under few‑shot prompting but remains more sensitive to prompt design and context length. Efficiency benchmarks show that decoder‑only prompting incurs near‑linear latency/memory growth with k‑shot exemplars, whereas fine‑tuned encoders maintain stable per‑example cost. These findings offer practical guidance on when to prefer fine‑tuning versus prompting and demonstrate that reliability must be evaluated alongside accuracy for risk‑aware deployment.
31 October 2024 in Eastern-European Journal of Enterprise Technologies
The object of the study is the conceptual prototype of a microprocessor resource-saving relay protection system. Currently, relay protection ensures electrical networks reliable and efficient work, however, the traditional architecture is proprietary, not allowing to fix and replace damaged parts without a company specialist. Therefore, the open-architecture relay protection is a very pressing issue, but the problem lies in meeting the relay protection requirements. The data transmission protocols nRF and ESP-NOW, Hall sensors evaluation for AC current measurement determination and sensor accuracy improvement was implemented. Experimental validation demonstrated that nRF and ESP-NOW protocols meet the delay and reliability requirements, however, the nRF protocol is more suitable due to its flexibility and obstacle penetration. The data demonstrated that the most effective conditions are without obstacles at 15 meters from the modem and with obstacles 5 meters from the modem. The experiment of Hall sensors characteristics determination demonstrated the accuracy of current measurement with the set values of the opening and closing currents. Nevertheless, it is not accurate (12.45 %) for the relay protection application. Therefore, the application of the changing values of the opening and closing currents is more effective and accuracy reaches 6.92 %. As a result, the service life of the Hall sensor was determined, and even after 10 million openings, the open state time remained unchanged. Therefore, the approximation function for current amplitude determination depending on open state time was found. On the other hand, Hall sensors may suffer from temperature drift and require further optimization to be fully reliable. The study limitation is the current range from 0 to 800 A
This article discusses the current problems of forest and steppe fires in the Republic of Kazakhstan, their increasing frequency, impact on ecosystems and the economy, as well as on human safety. The necessity of using modern technologies in fire control and management, including unmanned aerial vehicles (UAVs), is discussed. In computer programming, the mathematical formula of the Stefan-Boltzmann law was used to model a thermal process as a physical phenomenon. As a result of calculations in the Python program, the dependence of heat flow on distance and smoke coefficient was confirmed. In the course of the study, the results of calculating thermal radiation based on the methodology for evaluating optical and thermal indicators for early fire prevention were obtained. It has been shown that the integration of UAVs and telemetry improves control over fire conditions, increases the accuracy of fire forecasting by promptly obtaining information in real time. This approach is proposed as an important way to protect ecosystems and human life.
The search for efficient photocatalysts for sustainable hydrogen production has driven growing interest in barium titanate (BaTiO3)-based materials, particularly through polymorph control, surface engineering, and nonmetal and transition-metal doping. In this work, we provide an atomic-scale understanding of structural modifications in nitrogen-, fluorine-, and rhodium-doped BaTiO3 using Density Functional Theory (DFT), as well as pristine and fluorine-substituted BaTiO3 using reactive force-field molecular dynamics (ReaxFF-MD) simulations. DFT results for pristine and doped tetragonal BaTiO3, as well as pristine hexagonal BaTiO3, reveal that nitrogen and rhodium substitutions enhance the covalent character of Ti-N and Rh-O bonds and promote the redistribution of electron density, as evidenced by noncovalent interaction (NCI) and critical point (QTAIM) analyses, whereas fluorine substitution leads to more ionic Ti-F bonding. ReaxFF-MD simulations of pristine and fluorine-substituted BaTiO3 in contact with water molecules demonstrate that fluorine substitution suppresses interfacial O-H bond formation and promotes ordered molecular hydration layers near titanium sites, as reflected in bond statistics and radial distribution functions. This study provides molecular insights into the role of N, F, and Rh doping in BaTiO3 using DFT, and the role of fluorine doping in BaTiO3 at the water–solid interface using ReaxFF-MD simulations, demonstrating that this integrated computational approach provides a solid basis for the rational design of next-generation materials for energy-related applications. Direct calculations of photocatalytic activity, charge transfer rates, and ferroelectric polarization effects were not performed in this work and remain important directions for future study.
Представлены результаты численного моделирования низкотяговых схем увода космического аппарата с геостационарной орбиты на орбиту захоронения высотой около 280 км. Расчёты проводились для спутника с параметрами, весьма близкими к параметрам казахстанского геостационарного спутника KazSat-2. Расчёты показали, что при использовании стационарного плазменного двигателя спутник достигает орбиты захоронения за двое суток. При использовании газового двигателя – за восемь суток. При этом расход топлива примерно на 60 кг больше, чем у СПТ-70, и составил 61,826 кг. Учёт этого фактора весьма важен при планировании увода спутника на орбиту захоронения. Хотя повышение орбиты спутника приведёт к дрейфу на запад, изменение долготы за весь период увода не превысит 0,3 градуса.
30 September 2025 in Bulletin of Shakarim University. Technical Sciences
In recent years, the use of neural networks has expanded significantly through the development of models such as Generative Pre-trained Transformer (GPT) and various variations of convolutional neural networks (CNN) for various machine vision tasks. One of the key tasks in this area is the detection of objects in images. This article presents a comparative analysis of GPT-based models, pre-trained models and created artificial neural networks in the context of object detection. Object detection is a key task in computer vision, and applications cover various fields such as autonomous driving, surveillance, and medical imaging. The study begins by outlining the basics of object detection and the importance of choosing the right model for effective implementation. The advantages of their extensive pre-training are juxtaposed with the challenges associated with high computing requirements and limited customization.