AlloyGBM Documentation

AlloyGBM is a Rust-first gradient boosting library supporting regression, binary and multi-class classification, and learning-to-rank, with a Python API oriented around native execution, deterministic training, explicit validation, time-aware workflows, and zero-copy artifact-backed prediction.

The project is strongest on panel-style and finance-style workloads, with competitive performance on general tabular benchmarks across all three task types.

Note

AlloyGBM 0.12.8 is a feature release on top of v0.12.7. The GLM ("poisson", "gamma", "tweedie") and "quantile" objectives now work on GBMRanker and MultiLabelGBMRanker (both multi_label_mode="independent" and "joint"), in addition to single-output GBMRegressor. Only the Classifier / multiclass softmax paths still reject these objectives. No artifact format change — v0.12.7 artifacts load and predict identically under v0.12.8. See Release and platform policy for full notes.

Getting started

If you are new to AlloyGBM, start in this order:

User Guide

• Installation
  • Requirements
  • Install from PyPI
  • Build from source
  • Verify the install
  • Platform policy for 0.12.8
• Quickstart
  • Regression
  • Binary classification
  • Learning-to-rank
  • Multi-output ranking
  • Interaction constraints
  • MorphBoost (optional adaptive mode)
  • Validation and early stopping
  • Model persistence
  • What happens during fit(...)
  • GLM regression objectives (v0.11.0+)
  • Quantile regression (v0.11.1+)
  • SHAP interaction values (v0.11.0+)
  • NaN / missing values
  • Dense array-like inputs
• GBMRegressor
  • Core parameters
  • Boosting mode
  • Stopping and policy controls
  • Leaf and split controls
  • Tree growth strategy
  • Constraints
  • Reproducibility
  • Continuous-feature controls
  • Categorical support
  • DRO leaf solver
  • Factor-neutral boosting
  • Piecewise-linear leaves
  • Multi-label ranking
  • MorphBoost (Adaptive Split Criterion)
  • Warm-starting
  • Diagnostics
  • Main methods
  • Post-fit attributes
  • Regression objectives (v0.11.1+)
  • SHAP interaction values (v0.11.0+)
  • Recommended usage pattern
• GBMClassifier
  • Overview
  • Quick example
  • Parameters
  • Methods
  • Post-fit attributes
  • sklearn compatibility
  • Early stopping
  • Current scope
• GBMRanker
  • Overview
  • Quick example
  • Ranking objectives
  • Parameters
  • Methods
  • Evaluation
  • Group format
  • Early stopping
  • Current scope
• MorphBoost (Adaptive Split Criterion)
  • When To Use It
  • How It Works
  • Enabling It
  • Parameters
  • Learning-Rate Schedules
  • Combining with piecewise-linear leaves
  • Combining with DRO leaves
  • Persistence
• Time-aware validation
  • Why they matter
  • Time series splits
  • Panel splits
  • Using the splits with AlloyGBM
  • Explicit validation with fit(...)
  • Passing time_index into fit(...)
• Feature importances and SHAP
  • Local explanations
  • Global importance
  • TreeSHAP implementation
  • Current expectations
  • Supported estimators
  • Leaf model compatibility
  • SHAP interaction values (v0.11.0+)
• Benchmarks
  • Methodology
  • Current results
  • Metrics by task type
  • How to run the suite
  • Stage timing output
  • Interpretation

Technical reference

Technical Reference

• Architecture
  • High-level component layout
  • Objective implementations
  • Training pipeline
  • Tree growth strategies
  • Artifact design
  • Recent design choices
• API Reference
  • Package root
  • Estimators
  • Validation utilities
  • Evaluation helpers
• Release and platform policy
  • What’s new in 0.12.8
  • What’s new in 0.12.7
  • What’s new in 0.12.6
  • What’s new in 0.12.5
  • What’s new in 0.12.4
  • What’s new in 0.12.3
  • What’s new in 0.12.2
  • What’s new in 0.12.1
  • What’s new in 0.12.0
  • What’s new in 0.11.1
  • What’s new in 0.11.0
  • What’s new in 0.10.6
  • What’s new in 0.10.5
  • What’s new in 0.10.4
  • What’s new in 0.10.3
  • What’s new in 0.10.2
  • What’s new in 0.10.1
  • What’s new in 0.10.0
  • What’s new in 0.9.0
  • What’s new in 0.8.0
  • What’s new in 0.7.5
  • What’s new in 0.7.4
  • What’s new in 0.7.3
  • What’s new in 0.7.2
  • What’s new in 0.7.1
  • What’s new in 0.7.0
  • What’s new in 0.6.0
  • What’s new in 0.5.0
  • What’s new in 0.4.0
  • What’s new in 0.3.2
  • What was new in 0.3.1
  • What was new in 0.3.0
  • What was new in 0.2.0
  • Validated release surface
  • Deferred targets
  • Release checklist summary