AI in practice · Mixed

Data: where it comes from and what to do when there is none

In one line: In most projects the data exists but is scattered, unlabelled and unclean — and that's the stage that eats most of the time.

Three questions before you start

Where the information physically sits and who has access; what state it's in — format, completeness, duplicates; and whether you're allowed to use it for this purpose. The third question halts more projects than the first two.

When there's no labelling

Start small: two hundred well-labelled examples beat ten thousand labelled carelessly. Write labelling guidelines, label a sample twice, and measure agreement between labellers — disagreement indicates a vague definition, not bad workers.

You can also generate initial labelling with a model and correct by hand. That's fast, and requires checking the sample so as not to perpetuate bias.

Cleaning that pays off

Merging duplicate entities, normalizing Hebrew text, fixing dates and numbers, and removing records with no meaning. Most of the jump in quality comes from here, not from a better model.

Going deeper

Keep the data in versions: which sample was used for which evaluation. Without it, comparing two runs done in different weeks is meaningless, and you can't tell whether you improved or the data just changed.

Data