NTK Trainer Demo

Live demonstration of the OpenPeer-AI-NTK-Trainer: adapting a frozen language model by fitting a small set of signed log-gates on its layer outputs, with the fitting objective derived from neural tangent kernel theory, instead of training LoRA adapters or full weights.

The base model here is the 13.1B My Name is Riemann Prerelease, the model this trainer was actually used on. Pick a target writing style; the space fits a controller on forty sentences in that style, then scores held-out sentences from every style, before and after. A working controller lowers the model's surprise on the target style and leaves the others alone. You also see, token by token, how the gates changed the model's view of any text you paste, and you can download the controller.

Gates re-weight what the model already knows. They do not add facts or teach new output formats, so this demo measures what they change rather than asking them to produce a specific answer.

OpenPeer-AI-NTK-Trainer is released under MIT. Depends on ntkmirror. Maintained by OpenPeer AI.