Використання локальних LLM-моделей для стандартизації та багатомовного перекладу технічних назв виробів

Abstract

The paper considers an approach to automated standardization of technical names of hardware using local large language models (LLM), which allows to significantly increase the accuracy of processing specialized data. The architecture of interaction of python scripts with the local Lm studio server is described in detail, including the organization of api requests, the structure of input and output data in JSON format, as well as the mechanisms for saving results in the sqlite database (items.db). The paper provides examples of generating requests (prompts) containing raw product names, the expected standard format and product size, as well as examples of checking the correctness of formats after generating results. It shows how the processing of a series of records from the database is implemented, the quality control of the results obtained and the display of information about errors are implemented, which allows to promptly correct both prompts and training data of the model. The feasibility of switching to local solutions for processing technical data instead of using cloud services is substantiated. This approach allows to significantly reduce financial costs for infrastructure, to increase data security and confidentiality, and to ensure stable control over information processing workflows. It is shown that local inference of large language models can effectively solve standardization problems in application systems, ensuring high accuracy of pattern recognition, correct compliance with international standards (DIN, ISO, DSTU) and the possibility of integration into existing enterprise workflows without involving thirdparty services. In addition, the work demonstrates the practical implementation of multilingual translation of technical names, which increases the versatility of the solution and facilitates interaction with international databases and documentation.