A Tour of Pipeline Tasks

Common tasks and what they return

Each task name maps to a pipeline class with its own input and output shape. text-classification (alias sentiment-analysis) returns a label and score. token-classification (alias ner) tags each token; with aggregation_strategy="simple" it merges word pieces back into whole entities such as a person or a city. zero-shot-classification lets you supply your own candidate labels at run time; it uses a model trained on natural-language inference to score how well each label fits, so you need no training data. question-answering extracts an answer span from a context paragraph. summarization and translation use encoder-decoder models to write new text, and text-generation continues a prompt or a chat. Vision and audio tasks such as image-classification and automatic-speech-recognition follow the same pattern. The model you choose must have been trained for the task; loading a plain base model into a classification pipeline gives random predictions.

Zero-shot labels and named entities

Zero-shot scores your own labels; NER with aggregation returns whole entities with character offsets.

from transformers import pipeline

zero_shot = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
out = zero_shot("The RBI kept the repo rate unchanged this quarter.",
                candidate_labels=["economy", "sports", "politics"])
print(out["labels"][0])  # highest-scoring label first, likely "economy"

ner = pipeline("ner", model="dslim/bert-base-NER", aggregation_strategy="simple")
for ent in ner("Sundar Pichai visited Bengaluru."):
    print(ent["entity_group"], ent["word"], ent["start"], ent["end"])

Read the task tag on the model card

Every Hub model page shows a task tag such as Text Classification or Token Classification. Pick models from that filtered list. A classic bug is loading a summarisation checkpoint into a classification pipeline and getting nonsense with no error.

Quick check: You must sort support tickets into labels that change every month and you have no labelled data. Which pipeline fits best?

• text-generation
• token-classification
• summarization
• zero-shot-classification