| English | 中文 |
Browse the model-family directory for linear models, the tree family and reference-only sequence models. Decision trees and random forests share one family; the baseline belongs to common evaluation. Implemented models link full per-language task/runner examples. This reference describes the common fitting interface.
Choose the scientific family with model comparison methods. The R path currently supports a mean baseline plus declared regr.rpart and regr.ranger candidates. Learner choice and bounded parameters belong in the existing task JSON; no per-family pipeline copy is needed. R operations, supported settings records exact bounds, defaults, validation selection and absent tuning/CV/refitting. Python families and dependency checks are declared separately in the Python path.
Common preparation lives in scripts/r/common/. The R fitting module is scripts/r/models/fit.R, followed by held-out evaluate.R. reproduction.R is a separate development-only fixed-fold forest library; it is not called by the registered prediction entry. Source io.R before calling reproduction_forest() or reproduction_search_forest(). Preserve externally fixed folds and keep test rows outside search. See the independent reproduction method.
Use the complete R calls and result checks, including candidate metrics, actual cohorts, saved preprocessing, frozen predictions, independent metric recalculation and reload equality. A successful model fit does not establish scientific validity.
For development-only package-driven parameter search, use the separate automatic-tuning helper. It takes reviewed grouped folds and bounded ranges; the registered fitting modules above do not invoke it or accept its settings. Its full dependency-based execution remains separately recorded.