Optimizing Large Language Models for Domain-Specific Tasks

This research explores the integration of Large Language Models (LLMs) with domain-specific datasets, focusing on optimizing model training in centralized
environments. Using the TSpec-LLM dataset, a collection of telecommunications standards documentation, we developed a modular preprocessing
pipeline to clean and structure the data. We fine-tuned pre-trained T5 models for domain-specific question answering and compared their performance
against retrieval-augmented generation (RAG) approaches. Evaluation metrics, including BLEU and cosine similarity, demonstrated that while RAG excels in knowledge-intensive tasks, fine-tuned models provide efficient solutions for tasks with well-defined datasets. This study highlights the potential of centralized LLM systems for advancing domain-specific AI applications in telecommunications, leveraging methodologies from prior works.