How cisco-ai-defense/model-provenance-kit works
Model Provenance Kit is a Python toolkit and CLI that compares two models or scans one model against a bundled reference database. The main flow accepts model identifiers or local paths, extracts metadata, tokenizer, and weight fingerprints, then scores a comparison or performs a staged database lookup and formats ranked results. The README also documents deep-signal fingerprint downloads, caching, and streaming for large models. Source excerpts show the scanner’s cache use and parts of weight extraction; some internal orchestration details are inferred only where the SOURCE INDEX or README supports the relationship.
Main entry points: scanner.py Scanner · cli.py CLI commands · weight_signals.py Weight signals · lookup.py Three-stage lookup
Components 18
CLI and output
- CLI commands
src/provenancekit/cli.py - Result formatter
src/provenancekit/core/results/formatters.py
Fingerprint analysis
- Scanner
src/provenancekit/core/scanner.py - Metadata signal
src/provenancekit/core/signals/metadata.py - Tokenizer signals
src/provenancekit/core/signals/tokenizer.py - Weight signals
src/provenancekit/core/signals/weight_signals.py - Anchor selection
src/provenancekit/core/signals/anchors.py - Weight streaming
src/provenancekit/core/signals/streamers.py
Matching and scoring
- Three-stage lookup
src/provenancekit/core/lookup.py - Pipeline scoring
src/provenancekit/core/scoring.py
Data and services
- Model loading
src/provenancekit/services/model_loader.py - Tensor utilities
src/provenancekit/utils/tensor.py - Feature cache
src/provenancekit/services/cache.py - Reference database
src/provenancekit/services/database.py - Fingerprint models
src/provenancekit/models/signals.py - Fingerprint download
src/provenancekit/services/download.py
Other components
- User
- Hugging Face
Connections 23
- User → CLI commands: runs command
README.md - CLI commands → Scanner: starts scan or compare
src/provenancekit/cli.py - CLI commands → Fingerprint download: manages fingerprints
src/provenancekit/cli.py - CLI commands → Feature cache: configures cache
src/provenancekit/cli.py - Scanner → Feature cache: reads and writes
src/provenancekit/core/scanner.py - Scanner → Metadata signal: extracts metadata
src/provenancekit/core/scanner.py - Scanner → Tokenizer signals: extracts tokenizer features
src/provenancekit/core/scanner.py - Scanner → Weight signals: extracts weight features
src/provenancekit/core/scanner.py - Weight signals → Anchor selection: uses anchors
src/provenancekit/core/signals/weight_signals.py - Weight signals → Weight streaming: streams weights
src/provenancekit/core/signals/weight_signals.py - Weight signals → Model loading: loads model weights
src/provenancekit/core/signals/weight_signals.py - Weight signals → Tensor utilities: uses tensor helpers
src/provenancekit/core/signals/weight_signals.py - Scanner → Model loading: loads model
src/provenancekit/core/scanner.py - Scanner → Three-stage lookup: runs database scan
src/provenancekit/core/scanner.py - Three-stage lookup → Reference database: loads candidates
src/provenancekit/core/lookup.py - Three-stage lookup → Pipeline scoring: scores candidates
src/provenancekit/core/lookup.py - Scanner → Pipeline scoring: scores model pair
src/provenancekit/core/scanner.py - Three-stage lookup → Fingerprint models: validates fingerprints
src/provenancekit/core/lookup.py - Scanner → Fingerprint models: builds fingerprint results
src/provenancekit/core/scanner.py - CLI commands → Result formatter: formats results
src/provenancekit/cli.py - Fingerprint download → Hugging Face: downloads fingerprints
README.md - Hugging Face → Fingerprint download: provides fingerprint files
README.md - User → Hugging Face: hosts dataset