Sentiment Analysis of Bsky Posts

In this Project, we want to classify the sentiment (positive, negative, neutral) of Bsky posts. To this end, we first train a model on labeled tweets using the gxb912/large-twitter-tweets-sentiment dataset. However, we do not use the provided labels and instead use a teacher model (a LLM) to label the data; we distill the LLM into a smaller, specialized language model for sentiment analysis. This allows us to harness the predictive power of a LLM with significantly lower hardware requirements.

Additionally, we deploy the model as part of a fullstack app which allows a user to provide a Bsky feed (their own or a global hot feed) to infer the general sentiment of the posts in the feed.

Project Structure

• data Contains data for training the sentiment prediction model
• sentiment-analyis Contains notebooks and scripts for analyzing posts, data preparation, model training and evaluation
• models Checkpoints of the sentiment prediction model
• app Front- and Backend for analyzing Bsky feeds