Proposing a minimal-coding methodology to customize large language models to automate services in information organizations

Rarely do any studies propose a minimal-coding methodology to customize a Large Language Model (LLM) as part of the AI-driven automation of information services in libraries. This paper fills in the gap by illustrating a four-stage methodology to customize Llama 3 when automating cataloging, chat support, and recommendation services. Firstly, we collected and processed library data, that is, 197 frequently asked questions from the library website of a leading Research-I university in the US, and metadata of 450,320 books and their one-million-plus user ratings from the Book-Crossing dataset. Secondly, we applied the finetuning technique to customize the LLM using 80 percent of the collected data. Thirdly, we tested the customized LLM using the remaining 20 percent of the collected data. We used BERT F1 and ROUGE scores to evaluate the performance of the customized LLM, which confirmed the quality of its output. Finally, we uploaded the customized LLM to HuggingFace, an open-access repository. We present benefits for non-technical information professionals (e.g., minimal-coding methodology), information organizations (e.g., transportability of the proposed methodology, extending and expanding services), and patrons (e.g., privacy). Theoretical implications at the intersection of AI, information organization, and human-information interaction are discussed at the end.