It's important to understand a few fundamentals of zero-shot to get the most out of it.
Zero-shot attempts to understand the meaning of your label using what you input into the Natural Language Definition field. What you input here must have a meaningful relationship to the text you wish to classify!
We highly recommend using more specific labels that aren't ambiguous. If you notice mixed results with your classifications, consider getting more specific in your Natural Language Definition.
The best part about zero shot is you can quickly try different variations of your definitions and test the results without having to retrain. You can quickly edit your definition and test it out in the preview box.
You can adjust the confidence levels too to see how that impacts the range of your classification.
This is where ensembles become so handy. If the results from zero-shot aren't giving you the precision you need, you can use Rules & Conditions to get the precision you desire.
For example, you can use a prebuilt model as a condition for zero-shot. In this instance, zero-shot classifies within the boundary of the prebuilt model's results.
Or, you can use zero-shot to define the boundary Rules & Conditions to narrow in on topics or intents within the zero-shot classification.