Canonical docs
- Single-File Decoder
- Built-In SFDs
- Reference Architecture
- Experiment Manifest
- Universal Experiment Loop
In the research projects
The Python project supplies params, prep and model as ordinary caller-owned functions. The Compiled project names its pipeline and model in YAML and runs them through the CLI. The Python modules indexed here implement those families. Editing them is optional component development, described in the contributor appendix.
What this package owns
Owns packaged SFD modules and the reference architectures they rely on. Does not own experiment execution, data retrieval, or the lower-level indicator, feature, scaler, and transform libraries it composes.
Key entry points
| Entry point | Use case | Notes |
|---|---|---|
logreg_binary |
Standard manifest-driven binary-classification SFD | Exported at the package root |
random_binary |
Baseline classifier for comparison | Exported at the package root |
ridge_regressor |
Ridge next-return regression SFD | Under foundational_sfd |
xgboost_regressor |
Regression-style SFD | Exported at the package root |
foundational_rule_based |
Rule-based foundational SFD | Exported at the package root |
foundational_sfd |
Full catalog of packaged SFDs | Subpackage with production-oriented SFD modules |
reference_architecture |
Model-function templates without the packaged data pipeline | Starting point for custom SFD work |
Adjacent modules
limen.experimentruns SFDs and consumesparams()plus eithermanifest()or customprep/modelfunctions.limen.data,limen.indicators,limen.features,limen.transforms, andlimen.scalersprovide the building blocks that manifest-driven SFDs wire together.limen.metricsis commonly used inside the reference architecture model functions.
Quick orientation
| Module or family | Responsibility |
|---|---|
foundational_sfd/ |
Packaged SFDs with params + manifest |
foundational_sfd/tabpfn_binary.py |
Optional TabPFN definition |
reference_architecture/ |
Model-function implementations and templates |
sfd/├── README.md├── __init__.py├── foundational_sfd/├── reference_architecture/└── rule_based/Merged source PR · 87894eb8Package layout · Limen 6.0.1 · current package validation · package source
Things to know
- A foundational SFD owns the experiment shape, while the matching reference architecture owns the model training logic.
- The lazy
tabpfn_binarymodule and symbols are importable without the extra; constructing or training the TabPFN model requiresvaquum-limen[tabpfn]. - Component authors can copy an existing foundational SFD, adjust
params(), and modify its manifest chain. Compiled Research users select and configure the corresponding YAML template. - SFD modules are expected to stay cheap and stateless at import and manifest-construction time.