Log loss (binary cross-entropy) is a metric for classification models that penalizes confident wrong predictions heavily. It is computed as the negative mean of y*log(p) + (1-y)*log(1-p) over all samples, where y is the true label and p is the predicted probability.

What is a good log loss value?

A perfect model scores 0. A random classifier that always predicts 0.5 scores about 0.693 (ln 2). Values below 0.3 are generally considered good; values above 1.0 suggest poor calibration.

Why does log loss go to infinity for extreme predictions?

If a model predicts probability 0 for a positive label, log(0) is undefined (negative infinity). This tool clips probabilities to [1e-15, 1-1e-15] to avoid numerical blow-up, matching the behavior of sklearn's log_loss function.

Can I use this for multi-class problems?

This tool handles binary classification (labels 0 and 1). For multi-class log loss you would need one column of probabilities per class. A binary tool covers the most common use case and is what most Kaggle competitions measure.

How many samples can I paste in?

The tool runs entirely in your browser with no server round-trip, so it handles thousands of samples comfortably. Very large inputs (100k+) may take a moment to render the per-sample table.

Log Loss Calculator

Name: Log Loss Calculator
Availability: InStock
Author: Nham Vu

Paste actual labels and predicted probabilities to instantly compute log loss (binary cross-entropy) for your classification model.

Input Data

Actual Labels (0 or 1, comma-separated)

Predicted Probabilities (0 to 1, comma-separated)

Formula Reference

Log Loss = −(1/N) ∑ [ y·log(p) + (1−y)·log(1−p) ]

N — number of samples

y — actual label (0 or 1)

p — predicted probability [0, 1]

Probabilities are clipped to [1e-15, 1−1e-15] to prevent −∞.

Enter labels and probabilities, then click Calculate

Log Loss

—

Samples

—

Rating

—

Perfect (0) Random (0.693) Poor (2+)

Per-Sample Breakdown

#	Actual	Predicted	Sample Loss	Status

Copied to clipboard

Summary