Improve model accuracy automatically

Improve accuracy asks Tekst to look for better label or entity descriptions, using the examples your team has confirmed or corrected as the yardstick. Anything it finds lands in a draft, so you read the changes and decide whether to publish them.

This article covers improvement runs. For the manual side of accuracy, see Evaluate a classification model.

Before you start

An improvement run needs something to measure against, so make sure:

• The model has examples in its example set.
• Accuracy has been calculated at least once and is not currently recalculating.
• No examples are waiting for a correction.

Start a run

1. Open the model.
2. At the bottom of the label or entity table, select Improve accuracy.

The same control sits at the bottom of the Model accuracy panel, so you can start a run straight after reading your results.

A run can take up to 30 minutes. It carries on in the background, so you can close the model and come back later.

While a run is going

The model is read-only until the run finishes. Publishing, discarding, editing a row, and changing the example set are all unavailable, each with a tooltip saying Not available while an improvement is running. Reading is unaffected: you can still open the settings, the accuracy panel, and the version history.

Cancel stops the run immediately. Nothing is changed on the model and Tekst confirms with Improvement cancelled.

Read the result

When the run ends, the table footer reports what happened.

There are three outcomes:

• Accuracy improved by N% on a green bar. The rewritten descriptions are now in your draft. Close the bar with the X to return to the table.
• No accuracy improvements found on a red bar. Nothing changed. Correcting more examples gives the next run more to work from.
• Improvement run failed on a red bar. Nothing changed, and you can start another run.

On the red bars, Continue editing dismisses the bar and returns the footer to normal.

Review and publish an improvement

An improvement that found something writes a new draft, exactly like an edit you made by hand. It is not live until you publish it.

1. Open the Model accuracy panel and compare the Published model and Draft model cards.
2. Select a label or entity to read its new description in the panel on the right.
3. Check that no individual label got worse, not only that the total went up.
4. If you are happy, select Publish. If you are not, select the revert arrow to discard the draft.

The improvement also appears in the version history as a new version, with a change card for each description that was updated.

When Improve accuracy is unavailable

Hovering the disabled button tells you what is missing:

• Model needs to be evaluated before improvement is possible - the example set is empty.
• Calculate accuracy first. It builds the examples to improve against. - the model has never been evaluated.
• Wait for the accuracy calculation to finish - an evaluation is still running.
• Review N examples before improving. - examples are waiting for a correction after a configuration change.

Related articles

• To understand drafts, see the "What is a model draft?" article.
• To put the result live, see the "Publish a model draft" article.
• Evaluate a classification model
• Evaluate an extraction model