Dataset Alignment for Fine-Tuning Large Language Models for PJOK Educational Chatbot

BERT is one of the methods in LLMs that can be used to build chatbots. The resulting chatbot is like a question-and-answer system. There are three main stages in adapting BERT for a specific domain: dataset collection, creation, and fine-tuning. BERT is one of the models that has been pre-trained on a very large corpus of text. For Indonesian, there is already IndoBert which has been pre-trained with 220 million words. If it can be adapted to a specific domain, then BERT needs to be fine-tuned with a new dataset that must be prepared in advance. In this research, a chatbot was created using the BERT method and collecting a dataset of PJOK subjects. In this research, a dataset was collected to meet the needs of an educational chatbot in the field of sports and health for elementary school students. The dataset collection involved students, teachers, and student assistants to align it with elementary school textbooks. The result of this research is an alignment dataset containing 300 elementary school physical education (PJOK) questions, complete with context sourced from textbooks. This research successfully produced an alignment dataset with an accuracy of 84 percent.