PassGPT-Demo

A course-specific chat model powered by GPT.

Make you learn efficiently and Pass the course! Let's meet PassGPT!

Project Status

This repository marks the conclusion as a course project. For ongoing developments and future updates on PassGPT, please visit the new repository at PassGPT.

Thank you for your interest and support!

Introduction

We aim to bridge the gap between general chatbots and course-specific chatbots. While general LLMs are proficient in generating human-like text, they lack course-specific knowledge.

Our goal is to combine the strengths of both by building a course-specific chat model powered by GPT, utilizing the dataset from the EE6405: Natural Language Processing course.

Main Functions

1. Course information retrieval. e.g. course materials, concepts related informations
2. Generate course-specific content. e.g. examples, sample practices, quiz
3. Study & review assistance. e.g. summarize, tips

Working Flow

Data Preprocessing

We preprocess the course materials from EE6405: Natural Language Processing, including slides and code materials, using:

• Traditional NLP methods (e.g. TF-IDF, LDA, etc.)
• Multimodel-GPT-4

By extracting structured information from the course materials, and build a database for the course-specific chat model.

Database Construction

The database is constructed using MySQL, by the following schema:

Structure: Courses -> Files -> Concepts -> Subconcepts -> Examples

The concepts table has hierarchical structure by using concept_id and parent_id to fully represent the content structure.

Model (Agent)

We propose a novel Agent model combined from 2 parts:

• Query GPT: process the user input into SQL query, interacting with our database.
• Course GPT: generate the course-specific content based on the query result.

Both parts are based on GPT-4, without any additional training/fine-tuning.

Repository Overview

• data/: Note that we are not able to release the original course matrials due to the policy.
  • json/: extracted structed course .json file from GPT-4 also traditional NLP methods.
• database/:
  • json_to_sql.py: using .json to insert entries into SQL database.
  • pgpt_db.sql: You can construct same database by running this script.
• models/: Models for PassGPT.
  • pass_gpt.py: The main model(agent) for PassGPT.
• main.py: Main script for running PassGPT.
• example_results/: Example results of PassGPT, where some of them are good/bad examples(indicating the requirements of well-structrued user input)
• preprocess/:
  • traditional_nlp/: Traditional NLP methods for data (slides) preprocessing.
• requirements.txt: List of packages required to run the code.
• final_presenation.pdf: Final presentation slides for reference.

Setup

Environment

The packages are listed in requirements.txt. Run the following command for setting up the environment:

conda create --name pgpt --file requirements.txt
conda activate pgpt

Database

You can construct your own MySQL database by running the following command:

mysql -u root -p < database/pass_gpt.sql

Then modify the models/pass_gpt.py database information to your own database.

OpenAI API

You need to have an OpenAI API key to run the code. You can get one from OpenAI. Once you have the key, you can set it as an environment variable:

vi ~/.bashrc # open your bashrc file
export OPENAI_API_KEY="your-api-key" # add this line in your bashrc
# :wq to save and exit
source ~/.bashrc

Run

You can run the code in your terminal by:

python main.py

Then you can input your questions or commands to PassGPT in the terminal as show in the figure below:

Enjoy your learning journey with PassGPT!

Contributors

• Ke Xueyi:

  • Conceptualization and idea development.
  • Scripting of the model demo.
  • Database construction, setup, and management.
  • Scripting and integration of the final agent.
  • Repository management.

• Gu Yuhao:

  • Application of traditional NLP techniques for preprocessing slide materials.
  • Video Editing for final presentation.

• Jia Yunfeng & Lai Bohan:

  • Prompt engineering for quiz and practice generation.
  • Preprocessing of code materials data.

• Zhao Neng:

  • Utilization of GPT-4 for data preprocessing in slide materials.
  • Attempted front-end demo development.

• Zhang Jiali:

  • Application of ANTLR for data preprocessing in code materials.