Skip to content

Latest commit

 

History

History
39 lines (34 loc) · 2.79 KB

SCOPE.md

File metadata and controls

39 lines (34 loc) · 2.79 KB

Parity Scope (Active)

Core Problem

Deliver practical scikit-learn parity for Bun/TypeScript without stalling releases on full long-tail coverage.

Success Criteria

  • Users can build end-to-end classic ML workflows (preprocess, split, train, tune, evaluate) fully in bun-scikit.
  • Each new parity area ships with tests, docs, and deterministic behavior.
  • Releases continue on small increments instead of a single "big bang" parity attempt.

In Scope (Current Wave)

  • Maintain shipped clustering parity: KMeans, DBSCAN, AgglomerativeClustering.
  • Maintain shipped decomposition parity: PCA, TruncatedSVD, FastICA, NMF, KernelPCA.
  • Maintain shipped calibration/meta-estimator parity: CalibratedClassifierCV, VotingClassifier, VotingRegressor, StackingClassifier, StackingRegressor, BaggingClassifier.
  • Maintain shipped boosting parity baseline: AdaBoostClassifier, GradientBoostingClassifier, GradientBoostingRegressor, HistGradientBoostingClassifier, HistGradientBoostingRegressor.
  • Expand linear-model classifier parity with the RidgeClassifier family.
  • Expand classification metric parity with curve/ranking helpers and likelihood-ratio reporting.
  • Expand preprocessing parity with sklearn-style functional helpers layered over shipped transformers.
  • Expand neighbor graph parity with radius-based graph transformers.
  • Expand robust linear-model parity with consensus-fit regression.
  • Expand generalized linear-model parity with Tweedie-family regression.
  • Maintain multiclass support baseline across linear, probabilistic, neighbor, tree/forest, and meta-ensemble classifiers.
  • Maintain multiclass native Zig tree/forest backend parity (no binary-only native restriction).
  • Expand model-selection splitter parity with classical CV splitters (ShuffleSplit, leave-one/leave-p-out variants, PredefinedSplit, TimeSeriesSplit).
  • Keep API style aligned with existing estimator conventions (fit/predict/transform, learned attrs).
  • Maintain estimator introspection baseline via learned feature-importance attributes for tree/forest/boosting families.
  • Maintain parity guardrails with reproducible fixture generation and multi-seed thresholded drift checks.
  • Maintain machine-readable parity matrix coverage (docs/parity-matrix.json) enforced in CI/release gates.

Explicitly Out of Scope (Current Wave)

  • Full scikit module parity in one release.
  • Advanced clustering families (spectral/OPTICS/HDBSCAN-like parity).
  • Advanced decomposition beyond current baseline (sparse coding, incremental decomposition).
  • Boosting/manifold/inspection module breadth (extra-trees, XGBoost-style APIs, partial-dependence/SHAP).

Non-Negotiables

  • New APIs must have unit tests.
  • Existing tests and typecheck must stay green.
  • Scope additions require explicit tracking in this file.