Skip to content
/ ModelCypher Public template

ModelCypher - Decipher the high dimensional geometry of LLMs. An open source x-ray into LLM representation structure.

License

Notifications You must be signed in to change notification settings

Ethyros-AI/ModelCypher

ModelCypher

Geometric diagnostics for LLM representations. Measures intrinsic dimension, curvature, entropy, and representational similarity to guide model merging, monitor training stability, and detect behavioral drift.

Install

poetry install

Requires Python 3.11+. Primary backend is MLX (macOS/Apple Silicon). JAX backend available for Linux/TPU.

Quick Start

# Probe a model's architecture and geometry
mc model probe /path/to/model

# Merge two models with geometric alignment
mc merge --source /path/to/source --target /path/to/target --output-dir /path/to/output

# Analyze spatial geometry encoding
mc geometry spatial probe-model /path/to/model

# Measure entropy dynamics
mc thermo measure --model /path/to/model "Your prompt here"

Core Capabilities

Command Group Purpose
mc model Probe, fetch, register, validate models
mc merge Cross-architecture model merging pipeline
mc geometry Representational geometry analysis (30+ subcommands)
mc thermo Linguistic thermodynamics and entropy measurement
mc safety Behavioral drift and refusal pattern detection
mc train Training with geometry monitoring
mc infer Entropy-aware inference with security monitoring

Run mc --help for full command list. Run mc <command> --help for subcommands.

MCP Server

poetry run modelcypher-mcp

Tools available. See docs/MCP.md for the full catalog.

Documentation

Doc Purpose
CLAUDE.md AI assistant guidance and architecture
docs/CLI-REFERENCE.md Command reference
docs/GEOMETRY-GUIDE.md Geometry metrics explained
docs/GLOSSARY.md Terminology

License

AGPL-3.0. See LICENSE.