Work

Numerical Methods and a Software Framework for the Adaptive Fine-tuning of Uzbek Vision-language Models

Year: 2026
Type: article
Abstract: This article proposes numerical methods and a software framework architecture for the adaptive fine-tuning of Uzbek-language vision-language models under limited computational resources. The mathematical formulation combines contrastive alignment, image-conditioned text likelihood, and regularization criteria. Low-rank parameterization, AdamW iterations, gradient-norm clipping, and validation-based step-size control are described. The functions of the data, computational core, state-persistence, retrieval, and evaluation modules are specified. Analytical estimates of memory consumption are provided for parameter states and the similarity matrix. Two equivalent computational forms of a low-rank operator were examined using a control example, yielding a maximum absolute difference of less than 1.34·10⁻¹⁵. This result demonstrates the algebraic consistency of the operator. A reproducible computational-experiment protocol is presented for determining accuracy and performance on real data.
Language: English
Open Access status: diamond
OpenAlex ID: W7234638650
FWCI:
Cites:
Cited by:
Related to: