| English | 中文 |
Public source moved from skills/model/shared/ml_regression/ to skills/model/shared/machine_learning_model/. The former empty skill shell and stale cache have been removed. This source consolidation does not move result artifacts.
Model navigation now groups models/linear_regression/, models/tree_based/{decision_tree,random_forest}/ and models/time_series/timexer/. The mean baseline belongs to evaluation/. The former shared mlr3/ guidance is retained under frameworks/r/mlr3/; its reference tool and the TimeXer reference keep their identifiers and remain non-executable. Their source tracking, examples and metadata are preserved. Human-facing model families mirror these under methods/shared/machine_learning_model/models/.
Executable code retains scripts/r/ and scripts/python/ with shared planning under scripts/common/. Model-family folders provide navigation and contracts rather than duplicate training pipelines. R still adds the mean baseline automatically; Python still requires its explicit candidate. Selection and calculation logic are unchanged.
| Former relative path | Current relative path |
|---|---|
scripts/run_regression.R |
scripts/r/run_regression.R |
scripts/lib/io.R |
scripts/r/common/io.R |
scripts/lib/target.R |
scripts/r/common/target.R |
scripts/lib/prepare.R |
scripts/r/common/prepare.R |
scripts/lib/select.R |
scripts/r/common/select.R |
scripts/lib/evaluate.R |
scripts/r/common/evaluate.R |
scripts/lib/fit.R |
scripts/r/models/fit.R |
scripts/lib/reproduction.R |
scripts/r/models/reproduction.R |
scripts/lib/explain.R |
scripts/r/contribution/explain.R |
scripts/lib/causal.R |
scripts/r/causal/causal.R |
scripts/review_plan.py |
scripts/common/review_plan.py |
The R module algorithm bodies are preserved. Only the entrypoint source vector changes to resolve the new module locations, in the same order: I/O, target, selection, preparation, fitting, evaluation, contribution and causal. The separate reproduction library remains an explicit caller dependency. The runner sets AIRMOSAIC_R_ENTRYPOINT; source locations resolve from that entrypoint directory. R direct invocation is not a substitute for the runner’s task validation and output protections.
The registered R ID remains ml_regression_workflow, adapter mlr3_regression; its task schema, default learners and PLR restrictions remain unchanged. Python lives in scripts/python/ with its own contract/dependencies and tool ml_python_regression_workflow; see its actual supported behavior. No R library is automatically required by a validated Python-only path. The common planning helper is not a numeric training runtime.
The former methods/shared/ml_regression.md scientific body is retained in common design (sections 1–4 and 8 plus planning), models (section 5), contribution (section 6) and causal (section 7), with an index. General planning and reproduction methods retain their existing homes.