Love? | Devpost

Love?

Inspiration

I copied and pasted a piece of text from me and my mothers text conversations and I saw that it shows who sent what for each line. This gave me the idea to create some sort of Natural Language Processing Model and I thought a flirt checker would be a fun implementation idea

What it does

A user inputs a piece of text and it gives a flirt percentage back to the user.

How we built it

I built it using bert-base-uncased and a dataset named ieuniversity/flirty_or_not from huggingface

Challenges we ran into

My first training set was over one epoch and it did not provide low loss or high accuracy (50%). So I increased the the epoch to 10. As it was training and saving each epoch I noticed that after epoch 6 there were diminishing returns which I assumed was due to overfitting so I stopped the program at epoch 9. Then I decided to reduce the learning rate from default 5-e5 to 2-e5 to hopefully improve performance (accuracy) and add a decaying weight of 0.01 in order to regularize the model and help avoid overfitting. However it actually reduced the eval_accuracy and precision.

Additionally, the epoch previously mentioned overwrote the results from the previous training case. From this I learned that I need to save the data outside the project before running the test again (lol).

Accomplishments that we're proud of

I am proud of my use case of hugging face as well as implementing flask to create a web project. I have never used flask or hugging face. I will not say I did this on my own though. ChatGPT and Claude AI were very helpful and provided great insight into explaining how tokenization works, flask, and more. Using propelAuth

What we learned

I learned buzz words like Natural Language Processing, Tokenization, Flask, how to interpret metrics of AI models, and more. I also implemented PropelAuth by the end of the project