jAIce : The Empathetic Robot

AI Empathetic Response Generator

The Empathetic Robot is a specialized LLM Assistant designed to provide emotionally resonant and context-aware responses. Built using the EmpatheticDialogues dataset—comprising 25,000 conversations grounded in specific emotional situations—the model was trained to prioritize supportive, declarative interactions.

---

How to Use

cd into the /app folder...

Inside main.py file, edit the URL to match the server your are running/

Then, run python3 main.py to start the server.

Next, open the link in the terminal to access the website.

Finally, locate the model section in our website to use jAIce.

Technical Stack

• The datasets containing the prompts and empathetic responses (EmpatheticDialogues) https://github.com/facebookresearch/EmpatheticDialogues
• Python, pandas, and excel to preprocess the data.
• Python and aitextgen for the backend of the product demo
• GPT-2, the model to train on
• Google Colab servers to train the model
• HTML, CSS, and JavaScript to create the front end of the product demo
• Flask to connect the front end and the back end
• AI Camp servers to host the project on a domain

Dataset

We used a novel dataset of ~25,000 conversations grounded in Emotional situations. We trained our model to imitate responses in the dataset. To make sure only replies and not questions were generated, we removed all lines with questions as a response. We also concatenated the speaker’s sentences for each response in a conversation so the model has the full situation for each response.

Type of Model

We used the 124M (small) version of GPT-2, an open-source transformer model developed by OpenAI. It is especially useful for text generation, hence why we were able to utilize it. We fine-tuned the pre-trained model using the Empathetic Response dataset. We used 10,000 steps to train the model.

We used aitextgen to train the model on Google Colab servers. Each model took ~30 minutes to train, which is 6 minutes/1000 steps.

Libraries Used

We used pandas to preprocess the data.
To train the model, we used aitextgen.
To connect the frontend to the backend, we used flask.

About the Team

We are a group of awesome people at AI camp to make cool natural language processing products.

Team Members:

• Phakawat Wangkriangkri
  Project Lead
• Chandrark Muddana
  Product Manager
• Chris Sanrow
  Lead Back End Developer
• Jayce Chanas
  Lead Front End Developer
• Victoria You
  Front End Developer
• Riya Kumar
  Back End Developer