Validation Loss Curve Calculator

Result summary

Load data and run the analysis.

Waiting

Best run

—

Lowest minimum validation loss

Minimum validation loss

—

Best epoch

Final validation loss

—

Selected analysis range

Generalisation gap

—

At the best validation epoch

Early stopping epoch

—

Based on patience and minimum delta

Overfitting signal

—

Detection status

Convergence status

—

Recent slope and variance

Estimated epochs saved

—

Compared with the final epoch

Interactive loss chart

Hover, zoom, pan, isolate runs, and inspect marked epochs.

Diagnostic interpretation

Diagnostics will appear after analysis.

Input and analysis settings

Validation loss is required. Training and test loss are optional.

Model name

Dataset name

Loss function

Optimizer

Learning rate

Batch size

Validation split

Regularisation strength

Dropout rate

Random seed

Experiment notes

Training runs

Upload CSV

Enter comma, space, semicolon, or line-separated values. Epochs are generated automatically when left empty.

Smoothing method

Smoothing window

EMA strength

Show raw and smoothed curves

Early stopping and overfitting

Patience

Minimum delta

Improvement mode

Required worsening epochs

Sensitivity

Plateau window

Plateau tolerance

Convergence window

Slope tolerance

Variance tolerance

Enable overfitting detection

Recommend restoring best weights

Chart and range options

Chart mode

Loss scale

Line width

Marker size

Chart height

Legend position

Chart title

Epoch-axis label

Loss-axis label

Epoch range

Range start

Range end or N

Target-loss line

Custom epoch marker

Show training loss

Show validation loss

Show test loss

Show gridlines

Show cross-run deviation band

Shade possible overfitting region

Detailed statistics

Run	Minimum	Best epoch	Final	Average	Median	Std. deviation	Range	Final gap	Correlation	Volatility	Status
No analysis yet.

Multiple-run comparison

Run	Lowest loss	Fastest convergence	Stability score	Early stop	Epochs saved	Classification
Add one or more runs to compare.

Epoch data table

The table shows the primary run in the selected range.

Epoch	Training loss	Validation loss	Test loss	Gap	Loss change	Percentage change	Smoothed loss	Status
No epoch data available.

Page 0 of 0

Export and utility tools

Formula used

Generalisation gap at epoch e = Validation loss(e) − Training loss(e)

Absolute reduction = Initial validation loss − Final validation loss

Percentage reduction = (Absolute reduction ÷ Initial validation loss) × 100

Improvement occurs when current loss < best loss − minimum delta

Early stopping occurs after the selected patience expires without a qualifying improvement.

How to use

Enter validation losses for each epoch. Training and test losses are optional. Epoch numbers can be generated automatically.

Select smoothing, early stopping, and chart options. Add more runs when comparisons are needed. Then choose Analyze curves.

Review the best epoch and diagnostic warnings. Inspect the graph and epoch table. Export the final report when finished.

Example data

Epoch	Training loss	Validation loss	Interpretation
1	0.82	0.88	Initial model
5	0.45	0.51	Healthy improvement
10	0.26	0.39	Best validation region
15	0.15	0.48	Possible overfitting

Interpretation guide

A steadily decreasing validation curve suggests useful learning. A flat curve may indicate a plateau. A rising curve can signal deterioration.

Overfitting is more likely when training loss falls. Validation loss may rise during the same period. The growing gap supports that conclusion.

Short fluctuations can be normal. Repeated validation runs provide stronger evidence. Compare several seeds before making final decisions.

Frequently asked questions

What is validation loss?

Validation loss measures model error on held-out data. It estimates performance beyond training samples.

Why can validation loss increase?

It may increase because of overfitting, unstable optimisation, distribution shifts, or noisy validation samples.

Which epoch should I save?

The best checkpoint usually matches the lowest validation loss. Domain metrics should also be reviewed.

What does patience mean?

Patience controls how many non-improving epochs are allowed before early stopping is recommended.

Should I smooth the curve?

Smoothing can reveal trends. Raw values should remain visible when small fluctuations matter.

Can different runs be compared?

Yes. Add runs with different seeds, settings, or architectures and compare their minima and stability.

What is a generalisation gap?

It is validation loss minus training loss. A growing positive gap may indicate overfitting.

Why use a logarithmic scale?

A logarithmic scale helps when loss values cover several orders of magnitude. Values must remain positive.

Does low validation loss guarantee success?

No. Data quality, calibration, fairness, leakage, and task metrics still require separate evaluation.