Casper Engine

Local-first C/C++ training and inference workspace for NIYAH experiments.

What works now

• Core_CPP/niyah.exe: smoke + benchmark binary for NIYAH core.
• niyah_train.exe: standalone trainer built from Core_CPP/niyah_train.c.
• Core_CPP/trainer.exe: C++ training simulation pipeline.
• scripts/build_msvc.ps1: MSVC build script for core binaries.
• scripts/build_gcc.sh: GCC/Clang build script with optional lint/smoke.

Quick start (Windows, PowerShell)

cd C:\Users\Iqd20\Casper_Engine

# one command for the full pipeline: build + train + smoke
.\scripts\niyah.ps1 all

Single actions (unified entrypoint)

.\scripts\niyah.ps1 build
.\scripts\niyah.ps1 corpus
.\scripts\niyah.ps1 train Data_Training/sovereign_knowledge.txt 3 0.001
.\scripts\niyah.ps1 smoke
.\scripts\niyah.ps1 bench -Steps 300 -Size tiny
.\scripts\niyah.ps1 save  -Model niyah_tiny.bin -Size tiny
.\scripts\niyah.ps1 run   -Model niyah_tiny.bin -Prompt "بِسْمِ اللَّهِ" -Tokens 64

Notes

• scripts/niyah.exe and Core_CPP/niyah.exe are different binaries with different --mode support.
• scripts/niyah.ps1 is the unified wrapper to avoid mode/path confusion.
• Training data is expected in Data_Training/sovereign_knowledge.txt.
• niyah_train.exe supports arguments:
  • niyah_train.exe [data_path] [epochs] [lr] [min_lr]
• Prefer relative paths; avoid hardcoded absolute user paths in new code.

Project layout

• Core_CPP/: NIYAH core, smoke driver, training sources.
• Data_Training/: datasets.
• Math_ASM/: assembly experiments.
• UI_CSharp/: optional manager app.
• scripts/: build and run automation.