Architecture Overview

LLM Ripper is organized as a modular framework around extraction, analysis, transplantation, and validation of knowledge from Transformer-based language models.

High-level flow

flowchart LR
  A[Models] -->|activation capture| B[Core: activation_capture]
  B --> C[Core: analysis]
  C --> D[Knowledge Bank]
  D --> E[Core: transplant]
  E --> F[Core: validation]
  C --> G[Causal tracing]
  E --> H[Studio viewer]

Packages

• llm_ripper.core
• activation_capture.py: capture activations to HDF5/NP arrays
• analysis.py: component/feature analysis, head catalogs
• extraction.py: build Knowledge Bank assets
• transplant.py: modular transplant strategies
• validation.py: quantitative checks, diagnostics
• llm_ripper.utils
• model_loader.py: safe/controlled model loading
• run.py: run directory and artifact writers
• config.py: configuration helpers
• llm_ripper.causal: tracing utilities
• llm_ripper.interop: adapters and merging
• llm_ripper.safety: provenance and reports
• llm_ripper.studio: lightweight static viewer

Artifacts structure (RunContext)

Runs are created under runs// with subdirectories: - knowledge_bank, activations, analysis, transplants, validation, causal, traces, counterfactuals, uq, catalog, provenance, reports

See docs/api.md for API reference and README for quickstart.