How to beginner · 3 min read

Zero-shot classification with transformers pipeline

Quick answer

Use the Hugging Face transformers library's pipeline with the zero-shot-classification task to classify text into labels without prior training. Initialize the pipeline with a model like facebook/bart-large-mnli and call it with your text and candidate labels.

PREREQUISITES

Python 3.8+
pip install transformers>=4.30.0
pip install torch (or tensorflow)
Internet connection to download model weights

Setup

Install the transformers library and a backend like torch or tensorflow. The pipeline API simplifies zero-shot classification with pretrained models.

bash

pip install transformers torch

output

Collecting transformers
Collecting torch
Successfully installed transformers-4.30.0 torch-2.0.1

Step by step

Initialize the zero-shot classification pipeline with a suitable model, then classify your input text against candidate labels.

python

from transformers import pipeline

# Initialize zero-shot classification pipeline
classifier = pipeline('zero-shot-classification', model='facebook/bart-large-mnli')

# Input text to classify
sequence_to_classify = "I love programming in Python and exploring AI models."

# Candidate labels
candidate_labels = ['technology', 'sports', 'politics', 'education']

# Perform classification
result = classifier(sequence_to_classify, candidate_labels)

print('Labels:', result['labels'])
print('Scores:', result['scores'])

output

Labels: ['technology', 'education', 'sports', 'politics']
Scores: [0.987, 0.012, 0.0005, 0.0003]

Common variations

Use multi_label=True to allow multiple labels per input.
Switch models to roberta-large-mnli or facebook/bart-large-mnli depending on accuracy and speed needs.
Use async pipelines with transformers>=4.30.0 for concurrency.

python

from transformers import pipeline

# Multi-label classification example
classifier = pipeline('zero-shot-classification', model='facebook/bart-large-mnli')

sequence = "The government passed a new education reform bill."
labels = ['politics', 'education', 'technology']

result = classifier(sequence, labels, multi_label=True)

print(result)

output

{'sequence': 'The government passed a new education reform bill.', 'labels': ['education', 'politics', 'technology'], 'scores': [0.85, 0.75, 0.05]}

Troubleshooting

If you get OSError: Model name 'facebook/bart-large-mnli' was not found, ensure you have internet access to download the model or pre-download it manually.
For slow inference, consider using a smaller model or running on GPU.
If pipeline import fails, verify transformers is installed and updated.

✅

Key Takeaways

Use Hugging Face transformers pipeline with 'zero-shot-classification' for quick text classification without training.
The 'facebook/bart-large-mnli' model is a reliable default for zero-shot tasks.
Enable multi-label classification with the 'multi_label=True' parameter when needed.
Install 'transformers' and a backend like 'torch' to run the pipeline.
Troubleshoot model download and environment issues by checking internet and package versions.

Verified 2026-04 · facebook/bart-large-mnli, roberta-large-mnli

Verify ↗