Bluesky Feed Toxicity Analysis
Demonstrating the importance of context in assessing online toxicity, with an ML pipeline for scoring and analyzing posts from Bluesky curated feeds.
How toxic is a post, really? In Trust & Safety scenarios context is king: the thread it belongs to, the feed it appeared in, the author’s recent history. Often ML models contribute toxicity scores that can be misleading because they’re narrowly focused on the posts. This project builds tooling to demonstrate how important that context is when assessing the true “toxicity” of posts online.
The pipeline pulls posts from Bluesky’s curated feeds, scores them for toxicity and sentiment, and surfaces per-feed statistics alongside the flagged posts themselves. The broader goal is to iterate on the ML pipeline and build better, more context-aware tooling for content analysis.
Built to work with the toxic-cicd project, which handles model training and deployment.