EPFLearn: AI Student Mentor 📚

Introducing EPFLearn, a chatbot designed to answer students’ questions on specialized course material. The chatbot leverages Google’s pre-trained T5 Transformer as its foundation model, which was further fine-tuned using the StackOverflow and NLP4Education datasets.

Modern NLP: Course project Milestone 3

Team: Antoine Bonnet, Silvia Romanato and Alexander Sternfeld

How to run the code

To run the code, you first need to install the requirements in the requirements.txt file.

pip install -r requirements.txt

To use our trained chatbot, you will need to download the checkpoint from here and place it in the checkpoints/finalChatbot folder. You can then instantiate the chatbot and ask it some questions as follows.

checkpoint = 'path/to/finalChatbotDirectory'
chatbot = Chatbot(checkpoint)

chatbot.ask('What is the difference between genetics and epigenetics?')

Repository structure

Here is a list of relevant scripts used in this project.

• final_report_syntax_sorcerers.pdf : Final report of the project.
• gen_model: Generative model
  • chatbot.py : Chatbot class for interaction
  • gen_script_syntax_sorcerers.py : Generating answers with fine-tuned chatbot
  • finetune.py : Fine-tuning a generative language model on specialized content
  • load_data.py : Data pre-processing
  • milestone3.ipynb : Overview
• reward_model: Reward model
  • model.py : Model class
  • evaluate.py : Evaluation script
  • milestone2.ipynb : Overview
  • m2_data_preparation.ipynb : Data pre-processing for reward model
• checkpoints : Model checkpoints
  • finalChatbot : Final chatbot (fine-tuned on StackOverflow and EPFL datasets)
  • midChatbot : Intermediary chatbot (fine-tuned on StackOverflow dataset)
  • rewardModel : Reward model (trained on StackOverflow and EPFL datasets)
• data : Data files.
  • gen_model : Fine-tuning data for the generative model
    • answers_syntax-sorcerers.json : Sample answers generated by chatbot
    • gen_dataset_syntax-sorcerers_EPFL.json : EPFL dataset
    • gen_dataset_syntax-sorcerers_StackOverflow.json : StackOverflow dataset
    • gen_dataset_syntax-sorcerers.json : Combined EPFL and StackOverflow datasets
  • reward_model : Training data for the reward model
    • reward_dataset_syntax-sorcerers_EPFL.json : EPFL dataset
    • reward_dataset_syntax-sorcerers_StackOverflow.json : StackOverflow dataset
    • reward_dataset_syntax-sorcerers.json : Combined EPFL and StackOverflow datasets
• requirements.txt : Packages required to run the code
• python.txt : Python version used