davidchiii/toxic-tweets
link to hugging face model
Toxic Media - language-models
A introduction to natural language processing and sentimental analysis through the scope of analyzing negative online behaviors.
Requirements
WSL 2
For Windows users, first install WSL2. Install link here
Powershell
wsl --installContainers are lightweight, abstractions at the application layer that packages code and depencies together. Sharing the same operating system kernel, containers take up less space and run as isolated namespaces.
Docker
This project uses Docker.
Docker Desktop is an alternative for Windows users that provides a GUI alternative to the WSL subsystem. Understand the tradeoffs for using Docker Desktop.
Update repository
sudo apt-get update
sudo apt-get upgradeInstall Docker packages
sudo apt install docker-ce docker-ce-cli containerd.ioInstall VSCode
Located here.
Install DevContainers Extension. Through this extension, you can now open any directory in a development container through the bottom left icon.
Accessing the container
In Docker Desktop, you can find the container ID under the Containers tab.
In WSL2, execute
docker psto show all running containers on the system.
Copy the ID and run the command
docker exec -it [container-id] /bin/shto start a shell in container. 
Writeup
The model was trained on the Toxic Comment Classification Dataset found here.
The challenge aims to classify toxicity into the following categories:
- toxic
- severe_toxic
- obscene
- threat
- insult
- identity_hate
The model was trained using Google Colab and exported to HuggingFace Models. The base model, "distilbert-base-uncased", served as the original model in which the dataset helped fine tune.
The output of the model is a set of values corresponding to how much of each category the input tweet is.
Loss
I calculated loss using binary-cross entropy. For every batch, we can call the model to measure its loss. It hovers around 0.03 to 0.06, dropping sometimes to 0.01.