Customization of LLM-specific modules within ChatGPT

The goal of this thesis is to develop two distinct custom versions of ChatGPT tailored for specific modules from scratch, and to evaluate whether these customized versions perform better and are easier to use compared to the standard ChatGPT. For the evaluation and comparison, experts in relevant fields will be invited to test these large language models and provide their insights. The motivation for this thesis stems from the growing interest in and possibilities for creating customized versions of ChatGPT. Key considerations include usability, differences from the standard version, and ease of use. Additionally, there is a personal interest of the author in determining whether the custom models created will remain useful beyond the completion of this thesis and if they will present any future applications. This thesis is divided into two main sections: theoretical and practical. The theoretical part provides a comprehensive overview and introduction to large language models, describing the specifics of Generative Pre-trained Transformers (GPTs). It covers the essential concept of prompts, discussing how they can be effectively utilized by users. Additionally, this section describes the complexities of fine-tuning and the principles of RAG in the context of these models. The practical part of the thesis describes the development of two distinct custom versions of ChatGPT. The first version is designed to assist with creating resumes, while the second is tailored to aid users in preparing for the Japanese Language Proficiency Test (JLPT), as well as helping teachers with lesson planning and school-related tasks. This section not only details the creation process but also examines the specific functionalities and applications of each customized version by the experts. In the conclusion of the thesis, the evaluation results of the customized ChatGPT were summarized. The conclusion also outlined potential applications for these specialized GPTs, suggesting ways they could be effectively integrated into relevant fields to enhance task-specific functionalities and user experiences.