README.md

trinity-training

HSLM (Hybrid Symbolic Language Model) training infrastructure — Ternary neural networks, Beal conjecture, zeroth-order optimization, Railway deployment.

✨ Features

• 🔢 HSLM Model — ~1.24M ternary parameters, ~248KB compressed
• 🧠 Sacred Attention — φ-weighted mechanism for HSLM
• 📐 Autograd — reverse-mode automatic differentiation
• 🤖 Zeroth-Order — perturb-and-measure optimization (no backprop)
• 🚂 T-JEPA — jigsaw predictive coding self-supervision
• 🌐 Railway Deployment — cloud farm for distributed training
• 📊 Benchmarks — MNIST, CIFAR-10, neural network tests

📦 Installation

# Clone with zig-golden-float submodule
git clone --recursive https://github.com/gHashTag/trinity-training.git
cd trinity-training
git submodule update --init --recursive

🏗️ Modules

src/
├── hslm/          (70+ files)
│   ├── model.zig
│   ├── trainer.zig
│   ├── train.zig
│   ├── autograd.zig
│   ├── attention.zig
│   ├── sacred_attention.zig
│   └── ...
├── bench/          benchmarks
├── data_loaders/  MNIST, CIFAR-10
└── tri/             training orchestration
data/               (208MB)

🌌 Ecosystem

Core dep: zig-golden-float.

Cloud platforms:

• Railway — multi-account farm for distributed training
• Fly.io — multi-region swarm deployment

📜 License

MIT © gHashTag