Developing of Mini-Scientific Chatbot based on Transformers and LLMs

Abstract

This thesis presents MedSci-GPT, a lightweight multilingual scientific chatbot using a hybrid architecture: GPT-4o API (primary) and fine-tuned Phi-2 (fallback). LoRA (rank 8, alpha 16) was applied on scientific QA datasets. A web app was built with PHP, MySQL, HTML/CSS/JS supporting authentication, conversation management, and Arabic/English/French. Evaluation shows Phi-2 achieves 76% accuracy (0.8s) and GPT-4o 82% (1.8s), proving lightweight scientific assistants run on standard hardware.