TalkToMe

• 

  Web Application

Inspiration

We wanted to experiment with an advanced language model and saw this hackathon as an amazing opportunity. We wanted to offer a solution that might help a person discover his/her/their/...own issues by themselves. We do believe that talking to other people is essential for that, but this would give the user a somewhat more comfortable temporary solution.

What it does

It trains an advanced language model, specifically GPT-2, on scraped chat logs to generate responses based on the user's input text. The chats are the user's and no one else's.

How we built it

We used flask, HTML, CSS, and JavaScript for the web application. GPT-2 was trained using gpt-2-simple which is built on top of TensorFlow in Python.

Challenges we ran into

We trained a BERT model in Google Colab for sentiment analysis of the samples produced by GPT-2 but we weren't able to load it locally due to dependency issues.

Accomplishments that we're proud of

Trained a complex language model that predicts reasonable responses.

What we learned

GPT-2 Architecture Flask, HTML, CSS, JavaScript, TensorFlow, Keras, HuggingFace

What's next for TalkToMe

Training GPT-3 and gathering a whole lot more data so that the model can learn more complex patterns.

Built With

• css
• flask
• gpt-2-simple
• html
• javascript
• keras
• python
• tensorflow