Spaces:
Running on Zero
Running on Zero
| title: NTK Trainer Demo | |
| emoji: 🎛️ | |
| colorFrom: green | |
| colorTo: yellow | |
| sdk: gradio | |
| sdk_version: 5.49.1 | |
| app_file: app.py | |
| pinned: false | |
| license: mit | |
| models: | |
| - OpenPeerAI/OpenPeer-AI-NTK-Trainer | |
| - OpenPeerAI/OpenPeerLLM | |
| short_description: Fit NTK gate controllers on a frozen LLM live on ZeroGPU | |
| # NTK Trainer Demo | |
| A live demonstration of the | |
| [OpenPeer-AI-NTK-Trainer](https://huggingface.co/OpenPeerAI/OpenPeer-AI-NTK-Trainer): | |
| fine-tuning a frozen causal 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. | |
| Paste tab-separated prompt and completion pairs, choose a gate budget and | |
| step count, and the space loads | |
| [OpenPeerLLM](https://huggingface.co/OpenPeerAI/OpenPeerLLM) on ZeroGPU, | |
| scores a held-out slice, selects and fits the gates with | |
| [ntkmirror](https://github.com/leochlon/ntkmirror), scores again, shows | |
| greedy continuations before and after, plots the learned gates, and offers | |
| the controller file for download. | |
| The same method was used at every rung of the | |
| [My Name is Riemann](https://huggingface.co/OpenPeerAI/My-Name-is-Riemann-Prerelease) | |
| growth ladder. | |
| Maintained by OpenPeer AI. The trainer is released under MIT. | |