Skip to content

ChangeLog

An overview of every SuperModelingFactory release. All 0.3.x releases predate the established release process and are consolidated in the v0.3.x summary; starting with 0.4.0, each version has its own page with the full list of fixes, behavior-change notes, and backward-compatibility impact.

Release Policy

  • Patch versions (0.x.Y): pure hardening releases with zero API removals and zero numerical differences on healthy inputs. Batches are labeled by severity: HIGH (H/N), MEDIUM (M/N), LOW (L/N).
  • Minor versions (0.X.0): include at least one behavior change (a flipped default, a change in return-value semantics, or a changed required field). Any breaking default must first land behind an opt-in parameter in the previous patch release, and only then be flipped in the next minor (see the cross_vars pattern from 0.3.19 → 0.4.0, and the planned flip of the 0.4.2 missing_policy default).
  • Major versions (X.0.0): major architectural changes. None are currently planned.

Version Index

  • v0.9.3 - Closes the silent leak gap of the guarded screening IV: iv_upper_threshold defaults to 2.0 and gates on iv_raw (never the guarded iv), a leak warning and iv_gap_flag mark features whose iv_raw is far above iv, weighted_feature_screen exposes screen_iv_convention, every leak-gate drop warns, and no empty-stage fallback brings a dropped leak back (it keeps the fallback_top_k best soft drops, raises when only leaks are left; on_empty_stage='keep_all' is the old fallback); behavior changes shipped as an exception to the release policy (migration note)

  • v0.9.2 - The screening IV on WOE bins leaves class-pure bins out of the sum (screen_iv_convention='guarded', the default; 'raw' reproduces 0.9.1), a behavior-changing fix shipped as an exception to the release policy. New columns iv_raw, iv_sv_pure and n_sv_pure, a list of class-pure special-value bins, the leakage-gate reason iv_above_upper_sv_pure; scoring unchanged (upgrade guide)

  • v0.9.1 - The first published 0.9 release (0.9.0 was never published; upgrade guide). Class-pure WOE bins are merged by default (small_bin_policy='merge'); behavior-changing fixes shipped as an exception to the release policy: missing values get their own bin on every binning path, RejectInferencePipeline no longer trains unlabelled rows as goods, inverts high_good or leaks the OOT into the pre-score, and ScoreComparisonPipeline compares scores on the same rows; ODPSRunner needs ODPS_PROJECT and ODPS_ENDPOINT

  • v0.9.0 - Not published (its changes ship in 0.9.1). Three defaults flip (unseen_special_policy='neutral', sv_total_basis='all', warm_start_score_scope='full'); every text is English and every public docstring documents its parameters; the defects of the documentation verification and of the end-to-end audit of both pipelines are fixed (batch-mode selection over all batches, pooled tables without the stand-in OOT, consistent weighted and unweighted IV, evaluation and output-directory fixes)

  • v0.8.2 - Seven defect batches: unseen_special_policy for declared-but-unseen special values, warnings visible at import again, SMF's own warning noise and the bugs behind it, single-class evaluation, bool features in binning/WOE/FVP, and no more process-wide pandas option changes or writes into the caller's DataFrame

  • v0.8.1 - By-group WOE chart IV fixed to in-group IV (special bins, fitted SV decisions, unsmoothed single-class cells excluded), SV decisions in Format-A attrs, sparse bin-id alignment, dict-wrapped load_woe_bins, and joblib ≥ 1.6 compatibility for ParallelApplyEngine

  • v0.8.0 - SV bin governance: sv_min_bin_size / sv_small_policy low-share fallback and sv_woe_smoothing / sv_smoothing_alpha bad-rate shrinkage on both WOE engines, defaults bit-for-bit identical to 0.7.2

  • v0.7.2 - 0.7.1 boundary hardening: categorical small-bin fit governance, verified Format-A round-trips, frozen FVP transform audits, weighted VIF, and constant-weight corr parity

  • v0.7.1 - First 0.7.x release (0.7.0 burned pre-release): default-governance flip, categorical VIF support, and the unweighted-corr categorical crash fix with mixed WOE matrix basis

  • v0.6.9 - Categorical transform guards (missing-drift + str-fallback tripwires, coverage stats) and CMP self-fit param split

  • v0.6.8 - Weighted-path audit closure: hard-leak IV upper band on all paths, weighted G03/G04 evidence, floored G05/G06 ranking

  • v0.6.7 - Selection & binning governance: WOE ordering, IV/VIF/group gates, LR p-value elimination, verifiable split materialization

  • v0.6.6 - Evaluation governance: multi-label, special bins, direction, eval weights

  • v0.6.5 - OOT governance: candidate mode, split gates, no silent OOT synthesis

  • v0.6.4 - ODPS proc means, weighted evaluation figures, and wider grouped WOE charts

  • v0.6.3 - Close-out of equal_freq WOE-plot fix: if group: guard hoisted in get_bivar_graph

  • v0.6.2 - Fix equal_freq WOE plots silently producing zero PNGs

  • v0.6.1 - Independent FeatureValidation plot output control

  • v0.6.0 - Vectorized wide-table analysis and bounded-memory execution

  • v0.5.9 - Custom evaluation splits and FVP handoff metadata

  • v0.5.8 — ODPS pull-probe diagnostics

  • v0.5.7 — Categorical screening and RI lr NaN handling

  • v0.5.6 — Pipeline and data IO correctness

  • v0.5.5 — Parallel ODPS row-number pull

  • v0.5.4 — Pipeline GUI schema metadata

  • v0.5.3 — Cross-module hardening

    0.6.0 spec bug batch shipped on the 0.5.x line: UAT aggressive coercion, prediction NaN stats, all-zero weight guards, ODPS partition atomic swap, RI NaN-score policy, evaluator duplicate protection, unified evaluate return type, safer AUC failures, and NaN handling for chi2/VIF.

  • v0.5.1 — Feature / WOE hygiene

    Weighted IV zero-mass guards, PSI one-sided bucket smoothing, safer WOE missing sentinel, clearer WOE empty-dict errors, and per-variable failure diagnostics.

  • v0.5.0 — 2026-07-07

    MEDIUM batch — PSI missing_policy default flipped (breaking), 6 upstream public APIs now expose the kwarg, and MonotoneWOEBinner.apply_woe gains unseen_category_policy and _unseen_category_stats

  • v0.4.2 — 2026-07-06

    HIGH hotfix batch — atomic ODPS upload, split_df exclude_cols semantics fix, HardCutoffInferrer NaN guard, Weighted_Screen missing bin, new PSI_Tool.missing_policy parameter

  • v0.4.1 — 2026-07-06

    LOW hygiene batch — NaN handling for the default reject-inference cutoff, predict_positive shape/length/finite validation, and feature_validation stability for mixed Interval+NaN groupby

  • v0.4.0 — 2026-07-06

    MEDIUM hygiene batch #2 — cross_vars default change (breaking), opt-in numeric coercion for object columns, vectorized correlation computation, sample-analysis dry-run, and two more items

  • v0.3.x Summary — before 2026-06-30

    0.3.4 → 0.3.18, covering the WOE binning engine, unified feature_screen, weighted screen, FeatureScreeningArtifact, Pipeline hardening batches H1-H4, and hygiene batches M2-M4

  • SMF main repo Releases
  • Full tag list
  • API breaking changes and behavior flips are called out explicitly on the corresponding version page; search for "breaking change" or "behavior change".