Training Large Language Models for Chatbots


ChatGPT seems to gain a lot of interests and attentions since last November. The core of ChatGPT is GPT (Generative Pre-trained Transformer) which is a large language model trained with huge datasets. The objective of GPT is actually not really about intelligent but generate words that make users be convinced they were written by a human being. While a GPT-powered chatbot (e.g., ChatGPT is using GPT-3 and our research group's vip-bot in Discord is using GPT 3.5) can provide general factual knowledge to users according to many experiements and tests done by others and our research group, the use of Open AI's GPTs is not free and also the GPT is not trained specifically with the course content a teacher really wants his or her students to know and learn. GPT-powered chatbots can only generate something that looks real but in fact is not real for the students when they ask the chatbots.

Imagine if a teacher believes concept #A is more important than concept #B, then he or she will provide all the course materials and supplemental readings to the students. However, if the students use a GPT-powered chatbot like ChatGPT or our vip-bot in Discord who by accident thinks concept #B is more important than concept #A, then what would happen to the students? Perhaps they will fail the course in the final exam?

This research is under the supervision of Dr. Maiga Chang, Professor at the School of Computing and Information Systems, Athabasca University.

Management Dashboard
Web App
Model Testing Dashboard
Web Crawler Dashboard
License Management

About the Current Project

Dr. Chang's research group proposed, developed, piloted, and published an Automatic Text Summarization service named Ask4Summary in 2021 based on Natural Language Processing techniques (Pal, Chang, & Iriarte, 2021). It answered user questions by identifying relevant content and generating summaries only based on the provided materials (Saleh, Chang, & Iriarte, 2022; Saleh, Iriarte, & Chang, 2022) and reached 82.69% success rate for providing quick (0.766 seconds in average) responses of course relevant questions and is open access for users including teachers and students (Kuo, Iriarte, Zou, & Chang, 2023). Following Ask4Summary, an Alberta Innovates/Athabasca University Summer Research Studentship project was proposed in 2024 to design and develop the ChatbotLLM, an open-access system to train educational chatbots [C1]. The ChatbotLLM allows teachers to create courses, choose the desired configuration options, upload the materials, and start to train an educational chatbot on the uploaded course materials. ChatbotLLM's lead developer, AU BScCIS graduate, Supun De Silva, won IEEE Northern Canada Section's Capstone Project Award on October 18, 2024.

This research has the following tasks:

  1. has web-based dashboard (in PHP) for teachers to create courses and upload any kind of course materials (in PDF, DOCX, DOC, PPT, PPTX, TXT, and even MP3) to the server;
  2. the machine will then read and digest (with NLTK and other NLP techniques in Python) uploaded materials of courses for training the correspondent large language models (with TensorFlow or TensorFlow Lite in Python) -- i.e., each course will have a trained large language model;
  3. has a chatbot who can load a trained large language model dynamically so it can respond to students' messages according to the course(s) students are taking -- all the communications will be saved anonymously so we can have domain experts like the teachers to rate the chatbot's performance;
  4. teachers can give their students a course invite link so students can join the system and then interact with the chatbot;
  5. the chatbot will also, from time to time, actively ask the students feedback regarding how they feel about their satisfactions, perceived interactions, and the content it provides, etc.

Terms of Use

The VIP Research Group is a research group led by Prof. Maiga Chang at School of Computing and Information Systems, Athabasca University. This "large language models for chatbots" (chatbot.vipresearch.ca) is one of the research group's works. The research group does have follow-up research plan to improve it and further use it in other research projects.

Almost all of Prof. Chang's works are open access (or open source). The web service (chatbot.vipresearch.ca) is now open access and there is no plan to make it open source. The web service is open access and running on a self-sponsored server, as all of other research projects (see Prof. Chang's advanced projects) they will be always online, improving, and accessible as long as the cost can be affordable and covered by Prof. Chang.

Of course in the case that access volume of the web service becomes too high or any business or commerical institution takes advantage of using it to make money, then the terms of using the web service may look for changes; for examples, donations, personal/academic/business license and subscription models, etc. However, it is really too early to say that.

How to Access

Click on the Dashboard button to create courses and upload course material for a created course. Click on the Try It button and provide some text to see our language model in action!

About Us

...
Our Mission

Our research aims to training large language models, building chatbot, conducting experiments in real courses with teachers, collecting and analyzing collected data, and discussing with teachers and summarizing findings.

...
Our Supervisor

Dr. Maiga Chang is a Full Professor in the School of Computing and Information Systems at Athabasca University, Canada.

...
Research Goal

The research focuses on creating a chatbots that use large language models trained on specific course content to answer student questions.

Our Team

Profile picture of Supun
Supun DE SILVA
2023

Supun De Silva is an undergraduate student at Athabasca University pursuing a Bachelor of Science majoring in Computing Information Systems with a minor in Game Programming. My current research interest is in using generative AI for education, specifically training chatbots using large language models created by content uploaded by instructors. Other areas of interest include cloud computing, system administration, full stack web development, network programming, database administration, game programming, and system analysis and design. Outside of schoolwork, I like reading books and weightlifting.

Profile picture of Zee
Zibusiso MAFAITI
2025

Zibusiso Mafaiti is a 2nd year student pursuing a BSc in computing and information Systems. He is passionate about research particularly it's impact on modern learning.

Profile picture of Owen
Owen LI
2025

Owen is a Software devloper in the school Applied Science at Queen's University, Canada

Profile picture of Tina
Yi-Ting LEE (Tina)
2025

Tina is an undergraduate student majoring in Digital Multimedia Design at Cheng Shiu University, Taiwan. She contributes to the project mainly through flyer design and basic visual layout work. She is passionate about design, especially in visual communication and graphic design.

Profile picture of Victoria
Ting-Yu CHOU (Victoria)
2025

Victoria is a 4th year university student majoring information management at National Central University in Taiwan

Videos

ChatbotLLM Research

In this video recording, one of the research projects we have done, ChatbotLLM, is introduced. ChatbotLLM is an IEEE Northern Canada Section Capstone Project winner. As detailed in our accepted paper, ChatbotLLM is not only capable of learning from material in traditional formats like PDF, PPT(X), DOC(X), and TXT, but also reading XLS(X), image-scanned based PDF, images in PDF, DOC(X), and PPTX, and even MP3 with the help of Google's Speech Recognition library.

Using the ChatbotLLM Web App

This guide shows how to use ChatbotLLM directly in your browser.

Using ChatbotLLM on Discord

This guide shows how to use ChatbotLLM on Discord

Using ChatbotLLM on LINE

This guide shows how to use ChatbotLLM through the LINE messaging service.

Comparing MASA and Ask AU Chatbot

This lecture compares the MASA chatbot and the Ask AU chatbot.

Publications

  • Fuzheng ZHAO, Zibusiso MAFAITI, Yongjin LI, and Maiga CHANG. (2026). Mobile-Accessible Seamless Agent (MASA): Using ChatbotLLM for Grounding Chatbots in Teacher-Provided Materials. In: APSCE 4th International Conference on Artificial Intelligence, Metaverse and Artificial Companions in Education and Society (AI-MetaACES 2026), Kumamoto, Japan, June 25-27, 2026 (accepted)
    PDF
  • Supun DE SILVA and Maiga CHANG. (2025). ChatbotLLM - Training Educational Chatbots on the Materials Uploaded by Teachers. In: 13th International Conference on Information and Education Technology (ICIET 2025), Fukuyama, Japan, April 18-20, 2025. (IEEE)
    DOI PDF

Frequently Asked Questions

  • ChatbotLLM is an open-access educational chatbot service for training course-specific chatbots from materials uploaded by teachers. The ICIET 2025 paper describes the workflow: teachers create a subject, choose configuration options, upload materials, and students can then ask questions through the browser or connected client applications.

  • The published ChatbotLLM and MASA papers emphasize teacher-grounded retrieval. ChatbotLLM searches the teacher-provided course materials with sentence embeddings and cosine similarity, returns the top matching course paragraphs, and can say it does not know when no relevant course content is found instead of inventing an answer.

  • Teachers can upload common teaching materials including PDF, PPT/PPTX, DOC/DOCX, XLS/XLSX, TXT, ZIP packages, MP3, MP4, and WAV. OCR is supported for image-based PDFs, speech recognition for MP3, MP4, and WAV transcription, URL extraction, web crawling, enrichment, and data preparation before chatbot training.

  • Rephrase mode sends retrieved course paragraphs to a generative model to make the answer more readable while preserving the original meaning and key terms. RAG mode sends the question, retrieved context, and prompt to a generative model so the final answer is generated from the selected course context. Teachers can configure these options per subject.

  • Client apps can use the endpoints in the includes folder. Common endpoints include get_trained_models.inc.php for available trained models, get_course_details.inc.php for subject settings, get-model-names-for-course.inc.php for model choices, get-model-output.inc.php for chatbot responses, and update_feedback.inc.php for saving user ratings.

  • Yes. There is a browser-based mobile-accessible client, and the MASA paper shows endpoint documentation on how external clients can request ChatbotLLM responses. The service can also be connected to environments such as Discord, LINE, and VEE by passing the subject, model, message, and optional threshold to the response endpoint.

  • A subject can be public or protected with an access key. Client apps can check and submit access keys through endpoints such as verify_access_key.inc.php and get_course_details.inc.php. If a private subject receives an invalid key, the response endpoint rejects the request and records the access-denied event.

  • Each chatbot response can be stored with a feedback ID. Client apps can then call update_feedback.inc.php with the feedback type, rating, and wording. This supports the research goal of comparing learner perceptions of ChatbotLLM responses with other generative AI chatbot responses.

  • The Publications section links to the ChatbotLLM ICIET 2025 paper and the MASA AI-MetaACES 2026 paper. These papers explain the architecture, teacher-uploaded material pipeline, cosine-similarity retrieval, optional GenAI rephrasing/RAG, and the motivation for reducing hallucinations in education.