Skip to content

mlframe documentation index

Conceptual guides and design notes for mlframe. Copy-pasteable per-feature snippets live in docs/examples/; this index covers the deeper material in the docs root.

User-facing guides

File Topic
baseline_diagnostics_guide.md BaselineDiagnostics — ablation table + init_score baseline; the default-ON diagnostic that tells you whether composite targets will help
dummy_baselines_guide.md Parameter-free dummy baselines (median / mean / mode / lag / AR(1)) and the dummy-floor gate
calibration_policy.md pick_best_calibrator — OOF-ECE-with-bootstrap-CI selection between NoCal / Sigmoid / Isotonic
honest_diagnostics_guide.md The honest-diagnostics aggregator: holdout-only verdicts, dummy-floor delta, calibration ECE, DeLong significance
feature_handling_examples.md FeatureHandlingConfig cookbook — text / categorical / per-model handling recipes wired via feature_handling_config=
composite_targets_tutorial.ipynb End-to-end TVT walkthrough notebook for composite-target discovery
sidecar_verification_guide.md What a sha256 "sidecar" file is, why saved models are loaded through one, and what it does/doesn't protect against
composite_targets_guide.md CompositeTargetEstimator and automatic composite-target discovery — transform catalog, config, and when a composite target beats the raw one
composite_config_reference.md CompositeTargetDiscoveryConfig field-by-field reference
MULTI_OUTPUT.md Multilabel / multiclass classification design notes
MULTI_TARGET_REGRESSION.md Multi-target (K-column) regression support matrix and integration roadmap
SELECTION_BIAS.md PU-learning and drift-correction wrappers for biased-sampling problems
DEBUGGING_UPSTREAM_ERRORS.md Decoding cryptic errors raised by sklearn / LightGBM / CatBoost under the suite
visualization.md Reference for the charts train_mlframe_models_suite emits and how to tune them
gallery/index.md Rendered example gallery of every chart type

See also docs/examples/composite_targets.md for the tiered composite-target recipes and docs/examples/fixed_splits.md for exact, reusable train/val/test splits.

Internal / research notes

Design audits, literature surveys and roadmaps (working notes, not user API documentation) Some items have since shipped; each doc carries a status note where relevant. | File | Topic | |---|---| | [`NESTED_PARALLEL.md`](NESTED_PARALLEL.md) | Why two threads must never enter one numba `parallel=True` kernel, and the scanner that finds the places they can | | [`BENCHMARK_PREREGISTRATION.md`](BENCHMARK_PREREGISTRATION.md) | Binding pre-registration for the FS/FE/HPO benchmark — kill criterion, null hypothesis, ROPE, reserved seed ranges, control arms, and the falsifiable hypothesis table | | [`MRMR_RESEARCH.md`](MRMR_RESEARCH.md) | MRMR / feature-selection literature survey + FS competition matrix; most proposals shipped, 4 open estimator ideas | | [`FE_MRMR_IDEAS_BACKLOG.md`](FE_MRMR_IDEAS_BACKLOG.md) | FE gate/operator backlog status note — shipped / rejected / 3 open items | | [`pysr_fe_upgrade_research.md`](pysr_fe_upgrade_research.md) | PySR symbolic-regression FE tuning survey + operator-preset design (shipped) | | [`date_features_kaggle_research.md`](date_features_kaggle_research.md) | Calendar / cyclical date-feature survey behind `create_date_features` | | [`NUMERICAL_STABILITY_REPORT.md`](NUMERICAL_STABILITY_REPORT.md) | Catastrophic-cancellation audit + benchmark of the numba moment kernels | | [`WAVE5_GPU_ROADMAP.md`](WAVE5_GPU_ROADMAP.md) | GPU-acceleration roadmap notes | | [`SHAP_PROXIED_FS_GAME_THEORY.md`](SHAP_PROXIED_FS_GAME_THEORY.md) | `ShapProxiedFS` game-theory extensions (gt_01-gt_09) research verdict |