Skip to main content

limen.sfd.foundational_sfd

The catalog of packaged, ready-to-run Single-File Decoders.

Canonical docs​

What this package owns​

Owns the production-oriented, packaged SFD modules — each defining an experiment shape via params() and a manifest() that wires the data pipeline together. Does not own model training logic (owned by the matching limen.sfd.reference_architecture model) or experiment execution.

Key entry points​

Entry pointUse caseNotes
logreg_binaryStandard binary-classification SFDManifest-driven
dlinear_regressorLinear time-series regression SFDManifest-driven
lightgbm_binaryGradient-boosted binary classifier SFDManifest-driven
xgboost_regressorRegression-style SFDManifest-driven
random_binaryBaseline classifier for comparisonManifest-driven
rule_basedPredicate-driven, non-learned SFDUses limen.sfd.rule_based
tabpfn_binaryTabPFN-based SFDModule is lazy; model use requires the tabpfn extra

Adjacent modules​

  • limen.sfd.reference_architecture supplies the model function each foundational SFD delegates training to.
  • limen.experiment runs these SFDs, consuming their params() and manifest().
  • limen.data, limen.indicators, limen.features, limen.transforms, and limen.scalers are the building blocks the manifests wire together.

Quick orientation​

foundational_sfd/
├── logreg_binary.py
├── dlinear_regressor.py
├── lightgbm_binary.py
├── xgboost_regressor.py
├── random_binary.py
├── rule_based.py
└── tabpfn_binary.py     # Optional

Things to know​

  • Exports are lazy: the package uses module-level __getattr__, so importing foundational_sfd does not import heavy model dependencies until a specific SFD is accessed. This is why tabpfn_binary can be listed without tabpfn installed.
  • A foundational SFD owns the experiment shape; the matching reference architecture owns the training logic.
  • The simplest path to a new SFD is to copy an existing module, adjust params(), and modify the manifest chain.