Fit ClassSchema on training labels only, then encode both training and validation labels using it. Schemas sort homogeneous string/integer labels, reject missing labels and require at least two classes. Validation may contain one known class; unknown labels are rejected. Binary training requires both observed codes.
from openboost import ClassSchema, MixedData, Problem, RunContext
from openboost.recipes import binary
schema = ClassSchema.fit(["no", "yes", "no", "yes"])
x = MixedData([[0, "a"], [1, "b"], [2, "a"], [3, None]], [1, 2, 3, 4],
("x", "group"), ("numeric", "categorical"))
v = MixedData([[0.5, "a"], [2.5, "unknown"]], [10, 11], x.feature_names, x.feature_kinds)
train = Problem(x, schema.encode(["no", "yes", "no", "yes"]), x.row_ids, classes=schema)
valid = Problem(v, schema.encode(["no", "yes"]), v.row_ids, classes=schema)
fit = binary(train, valid, context=RunContext("binary-demo", 7), rounds=3)
model = fit.state.best_model
probabilities = model.predict_proba(v) # columns match model.classes.values
labels = model.predict_label(v)
assert probabilities.shape == (2, 2)
assert all(label in schema.values for label in labels)Raw state remains [N, 1] logits. Binary.geometry exposes weighted mean logistic
loss and unweighted gradient/curvature. Signed-margin logaddexp loss and separate
sigmoid tails preserve small gradients for confident correct predictions. There
is no hidden curvature floor. The shared Newton adapter applies original weights
once; categorical/numeric trees and fixed/backtracking transactions are unchanged.
Initialization is the logit of clipped training weighted prevalence (clip=1e-6), minus weighted mean offset. With heterogeneous offsets this is a deterministic initializer, not an optimized intercept. Clip must be positive, less than 0.5, and represent an upper probability below 1. Offsets affect objective geometry once and remain outside accepted raw caches.
model.predict returns raw logits. predict_proba returns columns for the two
persisted class labels; predict_label decodes argmax, with exact ties selecting
the first sorted label. Supply inference offsets explicitly to these methods.
Probability inference uses an inference-only transform and needs no training
objective/data. Outputs are probabilities, not a calibration guarantee.
Class order is part of Problem/model identity and follows every accepted/best snapshot. Mismatched train/validation schemas fail. The versioned ensemble format now includes class labels; malformed/unsupported schemas and earlier formats fail loading. Regression recipes reject classification-tagged problems rather than silently treating class codes as numeric regression targets.
This slice verifies binary geometry and complete mixed-feature inference. The binary recipe accepts exactly two classes. Multiclass provides a separate joint vector-tree recipe. Broader classification quality evaluation remains open. No XGBoost/LightGBM/CatBoost parity or CUDA claim.