Data Size-versus-Performance Plot Calculator

Analysis results

2 models

Logistic Regression

More data may still help.

Current dataset size5,000

Current ACCURACY0.849

Best observed performance0.849

Dataset size at best5,000

Total change0.169

Estimated maximum0.89

Target dataset size—

Recommended next size7,500

Interactive performance plot

Hover over points for exact values. Use the toolbar to zoom and reset.

Logistic Regression

y = 0.889996 − 1.69046·x^−0.45

Mean performance0.7868

Median performance0.805

Standard deviation0.0658

Size correlation0.739

Fit R-squared0.9929

Fit RMSE0.0051

Average generalisation gap0.086

Detected plateau size—

Current estimated cost$415

Target estimated cost$—

Budget-supported size13,943

Remaining headroom0.041

Random Forest

y = 0.953614 − 2.05288·x^−0.45

Mean performance0.8283

Median performance0.85

Standard deviation0.0799

Size correlation0.7453

Fit R-squared0.9916

Fit RMSE0.0067

Average generalisation gap0.1238

Detected plateau size—

Current estimated cost$415

Target estimated cost$304.01

Budget-supported size13,943

Remaining headroom0.0486

Detailed calculations

Model	Size	Training	Validation	Test	Std.	Primary	Change	% change	Marginal / 1,000	Gap	Fitted	Residual
Logistic Regression	1	0.91	0.68	0.67	0.03	0.68	—	—	—	0.23	0.6772	0.0028
Logistic Regression	25	0.89	0.74	0.73	0.025	0.74	0.06	8.82	0.4	0.15	0.7491	-0.0091
Logistic Regression	5	0.87	0.79	0.78	0.02	0.79	0.05	6.76	0.2	0.08	0.7868	0.0032
Logistic Regression	1,000	0.86	0.82	0.81	0.016	0.82	0.03	3.8	0.06	0.04	0.8145	0.0055
Logistic Regression	2,500	0.855	0.842	0.838	0.012	0.842	0.022	2.68	0.0147	0.013	0.84	0.002
Logistic Regression	5,000	0.852	0.849	0.846	0.009	0.849	0.007	0.83	0.0028	0.003	0.8534	-0.0044
Random Forest	1	0.99	0.7	0.69	0.035	0.7	—	—	—	0.29	0.6952	0.0048
Random Forest	25	0.98	0.77	0.76	0.029	0.77	0.07	10	0.4667	0.21	0.7825	-0.0125
Random Forest	5	0.96	0.83	0.82	0.022	0.83	0.06	7.79	0.24	0.13	0.8283	0.0017
Random Forest	1,000	0.94	0.87	0.86	0.017	0.87	0.04	4.82	0.08	0.07	0.8619	0.0081
Random Forest	2,500	0.925	0.895	0.89	0.012	0.895	0.025	2.87	0.0167	0.03	0.8929	0.0021
Random Forest	5,000	0.918	0.905	0.901	0.01	0.905	0.01	1.12	0.004	0.013	0.9092	-0.0042

Calculator inputs

Observation data

Model, size, training, validation, test, standard deviation

Use CSV, semicolons, or tabs. Two-column size and performance input also works.

Import CSV

Metric and analysis

Performance metric

Custom metric name

Primary analysis series

Metric direction

Curve fitting

Moving-average window

Targets and thresholds

Target performance

Diminishing-return threshold

Marginal-gain sample unit

Confidence level

Cost and budget

Collection cost per sample

Labelling cost per sample

Storage cost per sample

Training cost per experiment

Available budget

Chart appearance

Chart title

X-axis label

Y-axis label

Dataset-size axis

Chart height

Marker size

Line width

Formula used

P(n) = Pmax − a × n^−b

Marginal gain = (P₂ − P₁) ÷ (n₂ − n₁)

Generalisation gap = training performance − validation performance

Confidence band = observed performance ± z × standard deviation

The calculator also tests linear, logarithmic, polynomial, power, and exponential curves. Automatic mode selects the lowest fitting error. Extrapolated targets should still be validated experimentally.

How to use

Enter one observation per line using the displayed column order.
Select the metric, direction, primary series, and fitting method.
Set target, confidence, cost, threshold, and chart controls.
Calculate, inspect the curve, and review diminishing returns.
Export the chart, table, CSV data, or PDF report.

Example data format

Model	Dataset Size	Training	Validation	Test	Std.
Model A	100	0.95	0.70	0.69	0.03
Model A	500	0.91	0.82	0.81	0.02
Model B	1000	0.93	0.88	0.87	0.01

Export and reporting

Frequently asked questions

What does this plot measure?

It measures model performance while the available dataset size increases.

What indicates diminishing returns?

Very small consecutive improvements suggest additional data adds limited value.

Can lower values represent better performance?

Yes. Select lower-is-better for loss, error, or perplexity metrics.

What is the generalisation gap?

It is the difference between training and validation performance values.

Why use confidence bands?

They show uncertainty across repeated runs, folds, or sampled experiments.

Which curve should I select?

Automatic mode compares supported curves and chooses the lowest RMSE.

Are target estimates guaranteed?

No. They are extrapolations and require confirmation using new experiments.

Can multiple models be compared?

Yes. Use different model names in the first input column.

Can the calculator estimate collection costs?

Yes. Enter collection, labelling, storage, training, and budget values.