Development of a Generative Language Model Specialized in CNC Machine Tool Manuals

Due to the advance of machine tool technology, operation manuals have been becoming more complex and diverse than before. Accordingly, if machine technicians solve individual troubles happened in own machines, the losses such as human error, increased idle time, and delays still occur because of fast retirements and the lack of expertise in manufacturing companies. In this paper, we propose localized Manual LLM, an AI-based language model that enables real-time problem solving by answering operators' questions on-site. The training data set was developed through rapid engineering using GPT-3.5 Turbo and is divided into two main categories: a question collection manual for general manual topics and an alarm manual for CNC warnings. Llama2 was selected as the pre-trained model and hyperparameters were optimized through fine-tuning through multiple tests, and the final evaluation was based on the training loss and response accuracy. The model can be accessed through a mobile app, allowing operators to receive immediate support.