moon-online-models

phjphj676/moon-online-models is a native MoonBit toolkit for applications that learn from events continuously: CTR/risk scoring, device telemetry, incremental forecasting, ranking, feature streams, and model monitoring. It is the acceptance version of the August 2026 MoonBit Hackathon project.

What is included

• Dense online learners: RLS, SGD/FTRL logistic regression, Adagrad, ridge, Huber, quantile, softmax, Naive Bayes, k-means, kernels, PCA, factorization, tree stumps, ensembles, and time-series models.
• Sparse and streaming data: sparse vectors, hashing, CSV parsing, categorical encoding, feature crosses, online joins, reservoirs, bootstrapping, and schema-checked feature storage.
• Evaluation and safety: exact and histogram AUC, calibration, regression and ranking metrics, conformal intervals, drift detectors, fairness gaps, cost-sensitive thresholds, validation, gradient/prediction guards.
• Operations: model snapshots and checksums, registry and deployment state transitions, canary experiments, serving batchers, SLO/error budgets, alerts, audit trails, data lineage, privacy budgets, and reproducibility manifests.

The implementation is deliberately dependency-light: the root package only uses moonbitlang/core/math and moonbitlang/core/json. Public state is bounded where a stream can grow without limit; callers can inspect counters, reset state, and reject malformed dimensions at the boundary.

Install

moon add phjphj676/moon-online-models

Minimal example

let model = @moon-online-models.AdagradLogisticRegression::new(
  3,
  learning_rate=0.1,
)
model.update([1.0, 0.0, 0.2], 1.0)
model.update([0.0, 1.0, -0.2], 0.0)
let probability = model.predict([1.0, 0.0, 0.2])

For sparse CTR-style features:

let hasher = @moon-online-models.FeatureHasher::new(1_000_000)
let features = hasher.encode(["country=CN", "device=mobile", "slot=home"])
let model = @moon-online-models.SparseAdagradClassifier::new(1_000_000)
model.update(features, 1.0)
let probability = model.predict(features)

Verification

Run the same checks locally that the repository CI runs:

moon version --all
moon update
moon check --target all
moon test --target all
moon fmt && git diff --exit-code
moon info && git diff --exit-code

The local wasm-gc test suite contains 16 boundary and behavior scenarios; the native benchmark trains 20,000 deterministic events and records its measured output in benchmarks/RESULTS.md. Re-run it with benchmarks/run.ps1 on the acceptance machine.

Package and repository

• Package: phjphj676/moon-online-models@0.2.0
• GitHub: https://github.com/phjphj676/moon-online-models
• GitLink mirror: https://www.gitlink.org.cn/phjphj676/moon-online-models
• Default branch: master
• License: Apache-2.0; see LICENSE.
• The project proposal and acceptance notes are retained in 申报书.md.

The code is original MoonBit implementation authored by the repository owner. No generated build directory or credentials are part of the package.