From a spreadsheet to a labelling screen in five minutes

Start with the data you have

Labelling projects often begin with a spreadsheet: a column of sentences, a column of image links, maybe a column of labels from an older run. Azmaval accepts CSV, TSV, JSON, JSON Lines and Label Studio exports, up to 25 MB and 20,000 rows.

What gets guessed

• Each column is typed: short text, long text, number, yes or no, image link, audio link, video link, list of options, existing label or ID.
• Text and media columns are shown to reviewers. IDs and existing labels are kept with the sample but hidden.
• An existing label column becomes the options of a pick-one question, and a list column becomes per-row options.
• Everything is editable in a form, and a preview shows the real reviewer screen at desktop and phone width.

Five minutes

1. Paste your table or drop the file.
2. Read the preview and fix any wrong guess: rename a label, hide a column, change the question.
3. Run the data check, which tests every row the way the import will.
4. Choose private (only you) or crowd, the number of reviews per sample and the pay.
5. Create the campaign and review, or invite others.

What to check before you create it

• The question says exactly what to do in the edge cases.
• Options do not overlap.
• Right-to-left text reads correctly in the preview.
• Links open from a browser that is not signed in to your systems.

When you want more than one question

A sample can carry up to four questions on one screen: for example a pick-one label, a free-text reason, and boxes on an image. Each question has its own result and agreement in the export.