Performance Benchmarks

Overview

bun-scikit delivers significant performance improvements over Python’s scikit-learn while maintaining numerical accuracy. Our benchmarks measure real-world workloads using the Heart Disease dataset with 1,025 samples and 13 features.

All benchmarks are automated in CI and the results shown below are from actual benchmark runs. Dataset: test_data/heart.csv with 80/20 train/test split.

Performance Summary

Based on the latest CI benchmark snapshot (2026-02-25):

Regression: 2.2x faster fit, 2.4x faster predict
Classification: 2.5x faster fit, 2.6x faster predict
DecisionTree: 1.6x faster fit, 4.4x faster predict
RandomForest: 6.4x faster fit, 3.9x faster predict

Performance gains are hardware and dataset dependent. Your results may vary based on CPU architecture, data size, and workload characteristics.

Regression Benchmarks

LinearRegression Performance

Pipeline: StandardScaler + LinearRegression(normal)

Metric	bun-scikit	scikit-learn	Speedup
Fit time (median)	0.176 ms	0.389 ms	2.20x
Predict time (median)	0.019 ms	0.045 ms	2.43x
MSE	0.117545	0.117545	identical
R² Score	0.529539	0.529539	identical

Key Observations:

Over 2x faster training with native Zig acceleration
Numerically identical results (MSE delta: 6.4e-14)
Faster predictions with optimized TypeScript runtime

Classification Benchmarks

LogisticRegression Performance

Pipeline: StandardScaler + LogisticRegression(gd,zig)

Metric	bun-scikit	scikit-learn	Speedup
Fit time (median)	0.528 ms	1.293 ms	2.45x
Predict time (median)	0.032 ms	0.083 ms	2.60x
Accuracy	0.8634	0.8634	identical
F1 Score	0.8761	0.8750	+0.001

Key Observations:

Native Zig gradient descent implementation provides 2.5x speedup
Equivalent accuracy with minimal F1 score variance
Both fit and predict operations are significantly faster

Tree-Based Model Benchmarks

DecisionTreeClassifier Performance

Configuration: maxDepth=8

Implementation	Fit time	Predict time	Accuracy	F1 Score
bun-scikit (js-fast)	0.834 ms	0.021 ms	0.9463	0.9488
scikit-learn	1.371 ms	0.093 ms	0.9317	0.9340
Speedup	1.64x	4.44x	-	-

RandomForestClassifier Performance

Configuration: nEstimators=80, maxDepth=8

Implementation	Fit time	Predict time	Accuracy	F1 Score
bun-scikit (js-fast)	31.22 ms	1.76 ms	0.9902	0.9906
scikit-learn	199.63 ms	6.93 ms	0.9951	0.9953
Speedup	6.40x	3.92x	-	-

Key Observations:

Random forests show the largest speedup (6.4x for training)
Highly competitive accuracy with minimal delta
Prediction is consistently 4-5x faster across tree models

Zig vs JavaScript Backend Comparison

bun-scikit supports both native Zig and optimized JavaScript backends for tree models.

DecisionTree Backend Performance

Backend	Fit time	Predict time	Accuracy	F1 Score
js-fast	0.834 ms	0.021 ms	0.9463	0.9488
zig-tree	0.458 ms	0.034 ms	0.8927	0.8991
scikit-learn	1.371 ms	0.093 ms	0.9317	0.9340

Zig vs JS speedup:

Fit: 1.82x faster
Predict: 0.62x (JS is faster for small datasets)

RandomForest Backend Performance

Backend	Fit time	Predict time	Accuracy	F1 Score
js-fast	31.22 ms	1.76 ms	0.9902	0.9906
zig-tree	11.78 ms	0.78 ms	0.9951	0.9953
scikit-learn	199.63 ms	6.93 ms	0.9951	0.9953

Zig vs JS speedup:

Fit: 2.65x faster
Predict: 2.26x faster

The Zig backend (BUN_SCIKIT_TREE_BACKEND=zig) is the default for tree models. For small datasets, the JS backend may be faster for predictions due to lower overhead.

Running Benchmarks

Local Benchmarks

# Run complete benchmark suite
bun run bench

# Generate CI-style snapshot
bun run bench:ci

# Generate snapshot with native Zig kernels
bun run bench:ci:native

# Classification benchmarks only
bun run bench:heart:classification

# Tree model benchmarks
bun run bench:heart:tree

Hot-Path Benchmarks

Compare JS vs Zig backends on synthetic data:

# Run hot-path benchmark
bun run bench:hotpaths

# Verify against baseline (CI regression check)
bun run bench:hotpaths:check

Python Dependencies

To run comparison benchmarks against scikit-learn:

python -m pip install -r bench/python/requirements.txt

Benchmark Methodology

Dataset Details

Source: test_data/heart.csv
Samples: 1,025
Features: 13
Split: 80/20 train/test (deterministic, randomState=42)
Target: Binary classification for heart disease presence

Measurement Process

Both implementations use identical data preprocessing
Timing measurements use median of multiple runs
Fit time includes model training only (excludes data loading)
Predict time measures inference on test set
Metrics calculated using identical test samples

CI Automation

Benchmarks run automatically in GitHub Actions:

CI workflow: Runs on every push/PR
Benchmark Snapshot workflow: Scheduled updates
Results published to bench/results/heart-ci-latest.json
README table auto-updated with latest results

Numerical Accuracy

All benchmarks verify numerical parity with scikit-learn:

Regression: MSE delta < 1e-13, R² delta < 1e-12
Classification: Accuracy matches exactly, F1 delta < 0.002
Tree models: Competitive accuracy with deterministic splits

Minor accuracy differences in tree models are expected due to different tie-breaking strategies and floating-point precision handling.

Get Started

Core Concepts

Guides

Performance

Performance Benchmarks

Overview

Performance Summary

Regression Benchmarks

LinearRegression Performance

Classification Benchmarks

LogisticRegression Performance

Tree-Based Model Benchmarks

DecisionTreeClassifier Performance

RandomForestClassifier Performance

Zig vs JavaScript Backend Comparison

Running Benchmarks

Local Benchmarks

Hot-Path Benchmarks

Python Dependencies

Benchmark Methodology

Numerical Accuracy

Next Steps

Native Runtime

Optimization Tips

Build docs developers (and LLMs) love

Get Started

Core Concepts

Guides

Performance

​Overview

​Performance Summary

​Regression Benchmarks

​LinearRegression Performance

​Classification Benchmarks

​LogisticRegression Performance

​Tree-Based Model Benchmarks

​DecisionTreeClassifier Performance

​RandomForestClassifier Performance

​Zig vs JavaScript Backend Comparison

​Running Benchmarks

​Local Benchmarks

​Hot-Path Benchmarks

​Python Dependencies

​Benchmark Methodology

​Numerical Accuracy

​Next Steps

Native Runtime

Optimization Tips

Build docs developers (and LLMs) love

Overview

Performance Summary

Regression Benchmarks

LinearRegression Performance

Classification Benchmarks

LogisticRegression Performance

Tree-Based Model Benchmarks

DecisionTreeClassifier Performance

RandomForestClassifier Performance

Zig vs JavaScript Backend Comparison

Running Benchmarks

Local Benchmarks

Hot-Path Benchmarks

Python Dependencies

Benchmark Methodology

Numerical Accuracy

Next Steps