Key Words: Large language models; management information systems; artificial intelligence; machine learning; local fine-tuning.
Abstract. The study presents a platform designed for local training and inference of popular large language models (LLMs), addressing the growing need for solutions that preserve the privacy of training data. The platform enables researchers and developers to further fine-tune base models using locally stored and processed data through a modular software architecture. It supports a wide range of models and tasks, offering seamless integration with existing machine learning workflows and tools within enterprise management information systems. Two alternative approaches to contextualization of LLMs’ knowledge bases are implemented within the platform i.e., indirectly through model tuning (parameter-efficient fine-tuning, reinforcement learning with human feedback and tutorial fine-tuning) and directly through extending the original knowledge base via retrieval-augmented generation. The proposed hybrid solution for knowledge contextualization addresses existing challenges within the generative artificial intelligence field, namely accuracy of generated information (hallucinations), as well as transparency and traceability of generated information to source data. Benchmark results from the platform utilization demonstrate competitive performance in terms of latency, context accuracy, and resource utilization compared to cloud-based alternatives. By decentralizing model development and inference, the platform empowers institutions to maintain data sovereignty, reduce operational costs, and promote sustainable practices in the field of artificial intelligence.
