Gradient boosted tree training and inference in TypeScript for Deno.

// Import the XGB class and types
import XGB, { type XgbDump, type XgbObjective, type XgbParams } from '@neabyte/xgb'

import XGB from '@neabyte/xgb'

// Train a binary classifier and predict labels
const model = new XGB({ objective: 'binary:logistic', nEstimators: 50 })
const x = [
  [1, 2],
  [3, 4],
  [5, 6],
  [7, 8]
]
const y = [0, 0, 1, 1]
model.fit(x, y)
const labels = model.predict(x)
const probabilities = model.proba(x)

The constructor accepts either a string objective or a partial config object. When no argument is provided it defaults to an empty object.

If config is a string, it is wrapped as { objective: config }. The provided config is merged with default parameters using a spread, so any fields you omit will take their default values.

The constructor looks up an internal strategy by the resolved objective. If the objective is unknown, it throws a RangeError with a message listing all valid objectives.

Both objective and params are exposed as readonly public properties on the instance.

// Construct with an objective string shorthand
const a = new XGB('reg:squarederror')

// Construct with a partial config object
const b = new XGB({ objective: 'multi:softprob', nEstimators: 200, maxDepth: 4 })

// Construct with defaults (binary:logistic, 100 rounds)
const c = new XGB()

All fields on XgbParams with their types and defaults from the built-in default config.

// Serialize a trained model and restore it later
const dump = model.dump()
const json = JSON.stringify(dump)
const restored = XGB.load(JSON.parse(json))
const predictions = restored.predict(x)

The dump method produces a plain XgbDump object that captures the format version, objective, parameters, base score, base margins, feature and class counts, and all trained trees. Every field is copied rather than shared, so mutating the dump cannot corrupt the model it came from. The XGB.load static method accepts this object, validates the format tag, the required fields, and the shape agreement between numClass, baseMargin, and the tree nesting, and returns a fully restored model that is ready for prediction. Serialization to and from JSON is handled by the caller.