Imitation learning
replay → model.
Gula · gluttony
Game replays hold thousands of hours of expert behavior: split-second decisions, movement paths, strategies nobody ever writes down.
Lun'Gula decodes them, turns them into feature sequences and trains a model to reproduce that behavior. Each game only needs a replay decoder and a state encoder: the dataset, training, model and export are shared.
Architecture
A four-stage pipeline, from raw replay to exported model. The compute backend is detected automatically (CUDA, DirectML, ROCm, MPS or CPU). Adding a game means writing two classes and registering them in the CLI.
Reads the replay and the beatmap and turns them into a sequence of game states
Sliding windows: the context in, the next action out
2 layers, AdamW optimizer, MSE loss, separate validation set
Portable format, usable from Python, Node.js or C++
Where it stands
Checked in October 2026
Nothing yet