CTBoost 0.1.58¶
CTBoost 0.1.58 includes the optional compact multiclass vector leaves and correctness fixes documented in 0.1.57, plus two fixes found during cross-platform CI. Publication of 0.1.57 to PyPI and GitHub Releases was withheld; its Git tag remains available for provenance.
- Vector updates now preserve scalar accumulation rounding on platforms that contract floating-point multiplication and addition. This prevents small prediction differences from changing later boosting rounds. Conditional feature tests, split selection, and the default scalar strategy remain unchanged.
- Windows C++ export tests explicitly expose the selected compiler's runtime directory while loading the generated DLL. They still compile exported code and compare its predictions with the native model.
Model and predictor formats are unchanged from 0.1.57. See the vector-leaf guide for supported workflows and boundaries.
The completed 0.1.58 TabArena-Lite evaluation
covers all 51 datasets at r0f0, with the default plus 25 frozen HPO
configurations and eight-fold bagging: 1,326 parent results and 10,608 child
fits, with zero imputed CTBoost tasks.
| Evaluation | Lite Elo |
|---|---|
| Default | 1161.9 |
| Tuned | 1262.8 |
| Tuned + ensemble | 1296.9 |
These scores use an 87-row comparison roster. This is a author-run Lite HPO25 evaluation, not a full 200-configuration run, TabArena-Full result, or official leaderboard entry. The workers' 4 CPUs and 28 GB RAM make their timings non-comparable to canonical TabArena runtimes. See benchmark status for provenance and the earlier 0.1.56 result, which used a different comparison roster.