CTBoost 0.1.60

CTBoost 0.1.60 corrects categorical target statistics and early-stopping model selection. Conditional-inference feature selection, cut-point optimization, and existing learning defaults remain unchanged.

Corrections

• Fractional CTR priors: positive ctr_prior_strength values below one no longer scale the target prior down at first occurrences or unseen categories. Ordered frequency CTRs also use the intended prior at their first row. The default strength of one and zero-prior cold-start behavior are unchanged.
• Best-model selection at the iteration limit: native training with early stopping now retains the best validation checkpoint when the tree budget expires before patience does. DART restores the recorded ensemble, including its historical tree weights.
• DART warm starts: when a supplied untrimmed ensemble cannot recover its historical best round, early stopping evaluates that full ensemble on the current validation data and selects between it and subsequent improvements. It no longer treats a tree prefix as the historical DART checkpoint. Native and callback training paths keep selected-model state consistent.
• Benchmark helper loading: the sealed adapter loader includes its existing learning-option dependency. Frozen protocols and prior benchmark results are preserved.

Saved-model compatibility

Newly fitted or explicitly refitted preprocessing pipelines use format 4 to identify corrected CTR smoothing. Pipelines loaded from formats 1–3 retain their historical smoothing for prediction and warm starts, and reserialize as format 3. Categorical key encoding remains unchanged.

The Python raw-feature export reader accepts pipeline formats 3 and 4 with categorical key encoding 2. Older runtimes reject format 4 rather than silently using different preprocessing. Existing prepared-feature export boundaries and scalar/vector predictor formats remain unchanged. Applications linking the native headers or static core must rebuild; wheels include a matching extension. See deployment for DART resume semantics and preprocessing compatibility.

Validation and evidence

Pre-release Windows/Python 3.12 implementation validation passed 1,345 distinct tests, with 30 skips for unavailable optional capabilities. This includes 33 CTR regression cases, 22 early-stopping cases, and compiled C++ export checks. Genuine regression, binary and multiclass models saved by the public 0.1.59 wheel retain bitwise predictions after loading and resaving; their raw JSON predictors retain identical predictions as well.

The local development studies used training-only inner validation splits. CTR smoothing at strength 0.2 had 6 wins and 4 losses across 10 datasets, with a +0.0818% median relative error improvement and a −12.1872% regression on splice. All 20 strength-one control pairs had identical predictions. A separate 64-versus-256-bin comparison had 5 wins and 9 losses across 14 datasets, with −0.5381% median improvement. These mixed results do not establish a broad score gain; max_bins=256 and other learning defaults remain unchanged.

The audit report and original artifacts document the study binaries, local freezing before fits, reused baselines, and every regression. The studies were not publicly preregistered or independent holdouts and do not constitute HPO admission or leaderboard results. Published TabArena scores still measure 0.1.58, not this release; see benchmark status.