Projects/Pythia ‐ Hypixel

A player report classification system powered by machine learning, designed to assist the Hypixel moderation team in handling player reports more efficiently.

Hypixel's chat carries several million messages a day, which turns into thousands of chat reports for staff to work through. Pythia trains text classification models on past reports - the reported messages, paired with whether staff accepted or rejected them - and uses those to predict both whether a new report is likely to end in a punishment, and which punishment that would be. It runs on TensorFlow and Keras, with the training data exported out of production so the models can be trained on a GPU.

In practice that means reports likely to result in a punishment get prioritized, a confident prediction can issue a temporary mute while the report waits its turn, and the most clear-cut cases are handled automatically. It was never meant to replace the people working through reports - anything Pythia can't reach a clear conclusion on still goes to a staff member, and appeals stayed entirely human-managed.

See the Dev Blog linked below for a detailed overview of the project, its goals, and how it works.