Models

eigenhelm models are .npz files containing a trained PCA eigenspace, exemplar vectors, and calibration data.

Bundled models

Model	Languages	PCs	Vectors	Use case
general-polyglot-v1.npz	Python, JS, TS, Go, Rust	36	8,228	General-purpose (default)
lang-python.npz	Python	—	—	Python-specific
lang-javascript.npz	JavaScript	—	—	JavaScript-specific
lang-typescript.npz	TypeScript	—	—	TypeScript-specific
lang-go.npz	Go	—	—	Go-specific
lang-rust.npz	Rust	—	—	Rust-specific
pattern-cli.npz	Mixed	—	—	CLI tool patterns
baseline.npz	Python	—	—	Baseline reference

When no model is specified, eigenhelm uses the bundled polyglot model.

Choosing a model

• Language-specific models are more discriminating for their target language
• The polyglot model works across all supported languages and is the best default
• Custom models can be trained on your own curated corpora for domain-specific evaluation

Training a custom model

# Prepare a corpus manifest
cat > my-corpus.toml << 'EOF'
[corpus]
name = "my-team-best"
language = "python"
class = "A"

[[target]]
name = "my-code"
url = "https://github.com/my-org/my-repo"
ref = "main"
include = ["src/**/*.py"]
description = "Our best production code"
EOF

# Sync the corpus to a local directory
eh corpus sync my-corpus.toml corpus/my-team-best

# Train
eh train corpus/my-team-best --language python -o models/my-model.npz

See eh train for the full training reference.

Model contents

Each .npz model contains:

• PCA components: The principal component matrix defining the quality manifold
• Corpus statistics: Mean and standard deviation of the training feature vectors
• Calibration data: Empirical score distribution (p25/p75 thresholds) from the training corpus
• Exemplar refs: Compressed code content from selected high-quality samples, used for NCD comparison
• Metadata: Language, corpus class, training date, file count