This paper reports a systematic study of low-rank adaptation (LoRA)-based fine-tuning applied to Qwen3 language models (4B, 8B, and 14B parameters) for the task of legal question answering within the jurisdiction of the Republic of Kazakhstan. The bilingual dataset comprises 63,114 question–answer pairs (76.2% Russian, 23.8% Kazakh) covering 11 legal domains. Models are evaluated through both automated metrics (BERTScore, citation accuracy, and hallucination rate) and blind expert assessment by a panel of two practising legal experts. Key findings: (1) all fine-tuned models reach BERTScore F1 close to 90% (89.6–90.2%) versus 82.2–83.1% for untuned base models; (2) fine-tuned models outperform GPT-4o (87.2%) and GPT-4o-mini (86.7%) on semantic similarity while exhibiting far lower hallucination rates (27–29% vs. 83–90%); (3) blind expert assessment confirms the advantage of fine-tuned models, with panel mean completeness scores of 4.28/5 versus 1.95/5 for base models (quadratically weighted Cohen’s κ = 0.80–0.95 across rating dimensions, indicating substantial to almost perfect inter-rater agreement); and (4) we identify a practical scaling plateau: paired Wilcoxon tests (n = 500) detect statistically significant but practically small differences across the 4B, 8B, and 14B fine-tuned variants (largest mean gap 0.67 pp on BERTScore F1; Cohen’s |d_z| ≤ 0.34), gains too small to justify the 3.5× parameter increase. These findings show that parameter-efficient adaptation of compact open-source models can match or exceed commercial LLMs for specialised legal QA in a low-resource bilingual context. We note one scope restriction: three domains (administrative, criminal, and housing law) are represented only in Russian, so the model is not validated for Kazakh language queries in these areas.

Porosity, absolute permeability and diffusion coefficient are important characteristics of the flow of fluids in the pore space of rocks, the determination of which is resource-intensive and time-consuming. With the development of deep machine learning methods over the past 3–4 years, artificial neural networks have begun to be actively used in determining the transport properties of the “liquid-porous medium” system and the geometric characteristics of the pore space of samples based on their images. This method allows you to quickly determine the desired properties with acceptable accuracy. Therefore, the question arises about the effectiveness and adequacy of deep machine learning methods for these purposes. This article provides a scientific review of open literature sources on the determination of absolute permeability, diffusion coefficient and porosity from images obtained by different scanning methods. We also used our own data, namely images for 4 carbonate samples, and presented the results of predicting the connected porosity of these samples based on their X-ray images using the convolutional neural network model we built. The review showed that images of rock samples obtained using various scanning methods make it possible to calculate their transport properties with high reliability in a significantly short time. This means that deep machine learning can be a good alternative tool for calculating the properties of rock samples based on their images. The model we built showed the predictive ability of the porosity of 3 carbonate samples with a reliability coefficient of 0.936–0.976.

Background: Absolute permeability plays an important role in studying the fluids flow in porous media during the development of oil and gas reservoirs, the injection of CO2 into reservoirs for storage, the monitoring of pollutants migration in underground aquifers, and the modeling of catalytic systems. Therefore, an accurate and fast evaluation of its values is an important task. Aim: The purpose of this article is to study the applicability of machine learning methods for predicting the absolute permeability of carbonate samples, as well as ways to improve the prediction of permeability. Materials and methods: The input data is 408 small volumes extracted from four cylindrical carbonate samples composed almost entirely of calcite. Input data includes total and connected porosity, specific surface area, radii of all and only connected pores, coordination number, throat radius and length, tortuosity, and absolute permeability. Permeability prediction is carried out using regression machine learning methods such as random forest, extremely random trees and extended gradient boosting. Parameters (data) of small volumes were determined using pore-scale modeling of water flow in their pore space applying a specialized Avizo software. Results: Data of small volumes extracted from fractured and non-fractured samples were analyzed, and the results showed that there are good relationships between many parameters of small volumes. For example, the connected and total porosity have a second-order polynomial relationship with a high correlation coefficient. Using the above-mentioned regression machine learning methods, absolute permeability values were predicted when input data divided into training and testing data in a ratio of 80/20 and 70/30. Conclusion: Using the logarithm of permeability instead of permeability itself and considering fractured and non-fractured samples separately, can increase the accuracy of absolute permeability prediction using the above-mentioned machine learning methods up to 90%. The extremely random trees method is the most accurate among the three machine learning methods considered for our task.

This study presents the development and experimental evaluation of the Retrieval-Augmented Generation (RAG) system for the Kazakh language with an emphasis on comparative analysis of information retrieval methods. The main purpose of the work was to test the hypothesis of the superiority of a hybrid approach combining statistical (BM25) and semantic (vector search) methods over individual approaches. Based on the corpus of legal documents of the Republic of Kazakhstan, 270 experiments were conducted using three data extraction methods combined with six modern large language models (LLM). The results demonstrate that the hybrid method achieves the highest accuracy (82.2%), statistically significantly surpassing vector search by 3.3% (p < 0.01) and BM25 by 5.5% (p < 0.001). All three methods showed accuracy above 75%, which confirms the high efficiency of RAG systems for the Kazakh language. The analysis also revealed that hybrid search provides the greatest stability of results when working with different language models. This study makes a significant contribution to the development of RAG systems for languages with limited resources, offering an empirically based methodology to improve the accuracy and reliability of response generation.