Skip to content

Settings Reference

Complete reference for all training and evaluation settings.

TrainSettings

Required Parameters

Parameter Type Description
n int Number of trees to build per target class
out_folder string Directory to save trained models

Optional Parameters

Parameter Default Description
max_depth 10 Maximum tree depth (prevents overfitting)
frac_eval_cat 0.5 Fraction of data for bin evaluation vs weight-based grouping
max_eval_fit 1000 Maximum rows sampled per node
min_eval_fit 10 Minimum samples before stopping recursion
n_dims 3 Feature combinations to evaluate (1=single, 2=pairs, 3=triplets)
n_cat 3 Number of bins per feature
calcs_per_dim 5000 Maximum combinations to evaluate per dimension
neutral_faktor 0.0 Threshold for neutral branch (within [-n, n] goes to Neutral)

Example: TrainSettings

# Minimum required

n: 5

out_folder: my_model

# With all options

n: 10

out_folder: my_model

max_depth: 15

frac_eval_cat: 0.5

max_eval_fit: 1000

min_eval_fit: 50

n_dims: 3

n_cat: 5

calcs_per_dim: 5000

neutral_faktor: 0.0

EvalSettings

Required Parameters

Parameter Type Description
in_folders list Directories containing trained models
out_folder string Directory for evaluation results

Optional Parameters

Parameter Default Description
out_file null Output file path (.csv or .parquet)
keep_cols [] Columns to include in output
max_parallel_where 100 Split SQL into batches if more conditions than this

Example: EvalSettings

# Minimum required

in_folders:

  - model

out_folder: results

# With all options

in_folders:

  - model_v1

  - model_v2

out_folder: results

out_file: predictions.csv

keep_cols:

  - customer_id

  - date

max_parallel_where: 500

Parameter Guide

n: Number of Trees

Value Effect
1 Fast, baseline
3-5 Good balance
10+ More accurate, slower

max_depth

Value Effect
5-10 Shallow, fast, general
10-15 Medium
15+ Deep, may overfit

n_dims

Value Effect
1 Single features only
2 Feature pairs
3+ Complex interactions

n_cat

Value Effect
2-3 Few bins, general
5 Default
10+ Many bins, specific

calcs_per_dim

Value Effect
null No limit
1000 Quick
10000 Thorough
100000+ Exhaustive

Quick Reference Table

flowchart LR subgraph "Simple Problem" S1[n: 1] --> S2[n_dims: 1] --> S3[n_cat: 3] --> S4[max_depth: 10] end subgraph "Normal Problem" N1[n: 3-5] --> N2[n_dims: 2] --> N3[n_cat: 5] --> N4[max_depth: 15] end subgraph "Complex Problem" C1[n: 10+] --> C2[n_dims: 3+] --> C3[n_cat: 8+] --> C4[max_depth: 20] end