XGBoost in your browser, bit for bit
A 200-tree XGBoost model trained on the California housing data, running in this tab through moonboost, a gradient boosting predictor written in MoonBit and compiled to JavaScript. Nothing is sent to a server. The predictor is a single small module.
Estimate a block's median house value
Tick "missing" to leave a feature out: the model then follows the default branch XGBoost learned for missing values, exactly as XGBoost would.
Same answers as XGBoost
2,000 held-out blocks, scored by XGBoost 3.4.1 in Python and stored as raw float32 bits, scored again here.
Race it against onnxruntime-web
Loads onnxruntime-web from a CDN with the same model converted to ONNX, and times both engines on the 2,000 blocks in this browser (one thread each).
Source, tests and benchmarks: github.com/Juwan-Hwang/moonboost.
Model and data are built by tools/gen_demo.py.