Philippine Region-Based Food and Nutrition Cross-Reference Using Fine-Tuned Generative ChatBot

This paper describes the development of a Large Language Model Chatbot for Philippine Food and Nutrition knowledge base. The objective is to associate the region or province of origin of any available food item in the country to the nutrients contained and the nutritional benefits to the human body, considering the build of the knowledge base coming from 30 web sites or sources. Three models have been used to test and evaluate the performance in accomplishing the task. In the training and fine-tuning efficiency, the Meta Llama 2 7B HF model is the most efficient when varying the training epochs from 1 through 5; but based on evaluation metrics of BERTScore, ROUGE Score and GLEU, the Meta Llama 3 7B Instruct model obtained the highest scores, and the highest acceptance score when tested on 100 pairs of question-and-answer with ground truths. Further demonstration of the chatbot revealed that the objective has been achieved in most query attempts during the testing phase of the development.