OpenPeerLLM NTK Trainer
Model Overview
This package provides a LoRA-free training workflow for OpenPeerLLM-style causal language models by fitting signed log-gate controllers with ntkmirror. It also includes a tinygrad-based gate-controller smoke demo and a benchmark suite that generates charts for quick inspection.
Authors
- Andrew Magdy Kamal Nassief
- Riemann Computing Inc.
- OpenPeer AI
Intended Use
- Fit sparse forward-pass controllers on top of frozen Hugging Face causal language models.
- Run a low-cost local demo that validates gate training logic with tinygrad.
- Generate benchmark artifacts and charts for performance comparisons.
- Stop the demo run early once the requested accuracy target is reached.
Dependencies
- ntkmirror: https://github.com/leochlon/ntkmirror
- Tinygrad: https://github.com/tinygrad/tinygrad
- Optional charting: OpenBB
Benchmark Outputs
The benchmark runner records:
- epoch
- training steps
- wall-clock time
- memory usage
- process and thread counts
- samples per second
- initial and final accuracy
- final loss
- predictability score
- learned gate scales
Charts are written as HTML. The benchmark command writes a combined dashboard HTML plus companion charts, prefers OpenBB chart rendering when the optional chart extra is installed, and otherwise falls back to Plotly so the workflow stays runnable.