A web-app that detects toxicity (curse, insult, threat, hate, etc.) in YouTube comments and deletes them.
- notebooks - Jupyter Notebooks containing exploratory data analysis, training and evaluation part of the machine learning model.
- app - Implementation of the web-app which connects to youtube account, displays channel data, analyzes latest three video's and deletes (rejects) toxic comments found in them.
- Performed exploratory data analysis on the data such as finding no. of instances per class, finding null values, determining max len for the comment.
- Divided the data into train and test set by using stratified sampling technique to maintain class ratio in both set.
- Fine-tuned the BERT model using PyTorch and Hugging Face transformers libraries, evaluated its performance and saved the model. Fine Tuned Toxicity Detection Model .pth file
- Developed a web-app that connects with youtube, accesses the comments of videos and deletes (rejects) toxic comments using FastAPI Framework.
- Used Google OAuth 2.0 and YouTube Data API to authorize and get access to youtube channel.
- Created various views (web-pages) that helps user to navigate through the web-app.
- Integrated the trained ML model with the web-app to classify and delete comments of the selected video.