FINE-TUNING INDONESIAN QUESTION ANSWERING MODEL TRAINING WITH TENSOR PROCESSING UNIT

• Main Author: Adi Wicaksono, Baskoro
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/74784
• Institution: Institut Teknologi Bandung

The research focuses on the application of fine tuning method and TPU within the process of training question answering (QA) for Bahasa Indonesia model. This method is chosen due to its effectiveness in speeding up the training time while still retaining high accuracy value. The research started with exploring theories about Transformers, BERT, distilBERT, and previous researches on question answering models. The research continues with preprocessing the dataset and building the model based on the hyperparameter tuning applied. The model is then ready to be trained and evaluated. There are three metrics that has been chosen to evaluate the model, such as accuracy, F1-Score, and exact match (EM). The results of this experiment elaborated that by using fine-tuning and also with the help of TPU, the model can achieve high accuracy within much faster time. It is hoped that this research is able to serve as a stepping stone to future research and development of question answering model, especially Indonesian question answering.