v0.1.0 — Open the Hood

Model Garage v0.1.0 — Open the Hood

Component-level neural network surgery, analysis, and composition.

Highlights

• Extract real nn.Module components from 70+ models across 18 vendors
• Inject custom processing between any layers (scaling, noise, blades, debate chambers)
• Analyze activations, entropy, and sparsity layer by layer
• Compose hybrid architectures from mixed-source components
• Registry with full decomposition for GPT-2, Llama, Phi, Mistral, Gemma, Qwen, BERT
• Zero-training Rust expert router with sub-2ms DAG execution

Research

Three papers included with full LaTeX source:

• Blades: +14.2% accuracy via hidden state injection
• MoE Router Miscalibration: 482/896 significant but router ignores them (rho=0.069)
• Sparse Pathways: 2-4x compute savings via domain-aware neuron routing

Install

pip install model-garage

Quick Start

garage open gpt2
garage extract gpt2 --layer 6 --component self_attention
garage analyze gpt2 --prompt "The meaning of life"
garage compare gpt2 distilgpt2

What's Included

• 5 pretrained SAE blades (layers 0, 3, 6, 9, 11)
• Extracted component samples with weights
• 4 decomposition specs (GPT-2, Phi-3.5, Phi-4-reasoning, ESM2)
• Expert router manifests (MedGemma, OCR, Phi-4)
• 77 tests passing
• GitHub Actions CI + automated PyPI publishing