Zero-Shot Classification

Classify text into any set of labels without needing a fine-tuned model. Provide candidate labels at inference time — the model determines which labels best match the input.

For full API reference (classifyZeroShot(), options, result types, and custom providers), see the Core Classification guide.

See it in action

Try Email Classifier for a working demo.

Recommended Models

Model	Size	Speed	Use Case
`Xenova/mobilebert-uncased-mnli`	~21MB	⚡⚡⚡	Fast, browser-friendly (recommended)
`Xenova/nli-deberta-v3-xsmall`	~90MB	⚡⚡	Better accuracy, still browser-friendly
`onnx-community/ModernBERT-large-zeroshot-v2.0-ONNX`	~297MB	⚡⚡	Highest accuracy (large download)

Email Classification Example

Based on the Email Classifier showcase app:

import { transformers } from '@localmode/transformers';
import { classifyZeroShot } from '@localmode/core';

const model = transformers.zeroShot('Xenova/mobilebert-uncased-mnli');

const categories = [
  'Work',
  'Personal',
  'Shopping',
  'Travel',
  'Finance',
  'Social',
  'Spam',
];

const { labels, scores } = await classifyZeroShot({
  model,
  text: 'Your flight to Paris has been confirmed for March 25th.',
  candidateLabels: categories,
});

console.log(`Category: ${labels[0]}`); // 'Travel'
console.log(`Confidence: ${(scores[0] * 100).toFixed(1)}%`); // '95.2%'

Best Practices

Zero-Shot Tips

Use descriptive labels — "Customer complaint" works better than "type_3"
Limit label count — 3-10 labels works best; too many dilutes scores
Check confidence — Low top scores may mean none of the labels fit well
Consider fine-tuned models — If you always use the same labels, a fine-tuned classifier is faster