Zero-shot classification

createZeroShotClassifier() scores labels you choose at call time against a text, with no fine-tuning: a natural-language-inference model judges whether "This example is {label}." follows from the text. The default checkpoint is MobileBERT MNLI (~26 MB).

import { createZeroShotClassifier } from 'ngx-transformers';

readonly classifier = createZeroShotClassifier(); // Xenova/mobilebert-uncased-mnli

const scored = await this.classifier.classify(ticket, ['billing', 'bug report', 'feature request']);
// [{ label: 'bug report', score: 0.91 }, { label: 'billing', score: 0.06 }, { label: 'feature request', score: 0.03 }]

Try it: zero-shot story.

API

classify(text, labels, options?) resolves to ClassificationResult[] sorted best first. With no labels it resolves to [] without loading the model.

Option	Default	Effect
multiLabel	false	Score each label on its own, so several can be high, instead of normalising across labels so they sum to 1.
hypothesisTemplate	'This example is {}.'	The sentence the label is inserted into. Match it to your domain.

const tags = await this.classifier.classify(text, ['food', 'repair', 'politics'], {
  multiLabel: true,
  hypothesisTemplate: 'This text is about {}.',
});

Tips

• Label wording matters: the model reads each label as English text. 'refund request' works better than 'REFUND'.
• Cost grows with the number of labels, one inference per label, so keep the list short.
• Xenova/distilbert-base-uncased-mnli (~80 MB) is the more accurate drop-in: createZeroShotClassifier({ model: 'Xenova/distilbert-base-uncased-mnli' }).