Fine-tuning Large Language Model (BERT) for Islamic Moral Inquiry and Response
Nurul Aiman Binti Mohd Nazri, A'wathif Binti Omar, Azizah Hussin · International Journal on Perceptive and Cognitive Computing · 2025
The development of Large Language Models (LLM) that are capable of understanding and responding to issues from an Islamic perspective is extremely insightful as it will benefit many people. For an LLM to do so, it is not enough for the model to only understand the language, but it also needs to understand the context and specific doctrines within the Islamic texts due to the complexity of Islamic jurisprudence and moral philosophy. Therefore, in this research, we intend to fine-tune an LLM model which is known as Bidirectional Encoder Representations from Transformers (BERT) for Islamic moral inquiry and response. By incorporating Islamic principles, norms, and teaching into the model, we aim to enhance the pre-trained BERT model’s ability to perform moral-related Question Answering (QA) tasks. The original model that we chose is deepset BERT model which was built based on BERT-large and meticulously pre-trained using the SQuaD 2.0 dataset, specifically for QA tasks. We fine-tune the model using the data extracted from “Islam: Questions and Answers: Character and Morals”, the Volume 13 of a Series of Islamic Books by Muhammad Saed Abdul-Rahman, where the data has been cleaned and pre-processed. The fine-tuning process used supervised learning techniques, to ensure its proficiency in understanding Islamic principles, providing accurate, contextually appropriate, and theologically sound responses. We assessed the model using F1 score and Levenshtein similarity evaluation metrics where F1 score merges precision and recall by computing their harmonic mean, while Levenshtein similarity compares the predicted and actual answers at the character level by normalizing the Levenshtein distance. Our research yielded significant success, evidenced by the remarkable enhancement in the average F1 scores and Levenshtein similarities, soaring from 0.30 and 0.24, to 0.74 and 0.67 respectively.