Hugging Face Pipelines

Hugging Face pipelines are the simplest way to run a model with transformers. A pipeline wraps three steps, tokenizing the input, running the model and turning raw numbers into readable results, into one function call.

Choosing a Task and a Model

Pass a task name, and optionally a model ID. Always pinning a model is a good habit: the default for a task can change between versions, and you want results that do not shift under you.

from transformers import pipeline

sentiment = pipeline(
    "text-classification",
    model="distilbert/distilbert-base-uncased-finetuned-sst-2-english",
)
print(sentiment(["The delivery was fast.", "The box arrived crushed."]))
[{'label': 'POSITIVE', 'score': 0.9991}, {'label': 'NEGATIVE', 'score': 0.9993}]

Passing a list processes several inputs at once and returns one result per item.

Common Pipeline Tasks

  • text-classification: sentiment, spam, topic labels
  • token-classification (ner): find names, places and organizations
  • question-answering: extract an answer span from a context paragraph
  • summarization and translation: sequence-to-sequence text tasks
  • zero-shot-classification: classify into labels you choose at run time
  • text-generation: continue text or chat with an LLM
  • image-classification, automatic-speech-recognition and many more

Zero-Shot Classification

Zero-shot models let you invent labels without training anything. This is handy for quick prototypes such as routing support tickets:

from transformers import pipeline

zs = pipeline("zero-shot-classification", model="facebook/bart-large-mnli")
result = zs(
    "My invoice shows the wrong amount for March.",
    candidate_labels=["billing", "technical issue", "account access"],
)
print(result["labels"][0], round(result["scores"][0], 3))
billing 0.942   (sample output)

Named Entity Recognition

from transformers import pipeline

ner = pipeline("ner", model="dslim/bert-base-NER", aggregation_strategy="simple")
for ent in ner("Naeem moved from Lahore to Dubai to work at Emirates."):
    print(ent["entity_group"], ent["word"])
PER Naeem
LOC Lahore
LOC Dubai
ORG Emirates

aggregation_strategy="simple" merges word pieces back into whole words, so you get "Lahore" instead of separate sub-word tokens.

Running a Pipeline on a GPU

Use the device argument: "cuda" for NVIDIA, "mps" for Apple Silicon, or "cpu". For large models, device_map="auto" (which needs accelerate) spreads the weights over all available devices.

pipe = pipeline("summarization", model="facebook/bart-large-cnn", device="cuda")
text = "Hugging Face hosts models, datasets and demo apps. " * 10
print(pipe(text, max_length=40, min_length=10)[0]["summary_text"])
Hugging Face hosts models, datasets and demo apps . ...   (sample output)

Frequently asked questions

What is a pipeline in Hugging Face?

A high-level helper in transformers that bundles a tokenizer, a model and post-processing for a specific task, so you can get predictions with a single call.

How do I make a Hugging Face pipeline use the GPU?

Pass device="cuda" (or device=0) when creating the pipeline, or device_map="auto" for large models with accelerate installed.

Which model does a pipeline use by default?

Each task has a default checkpoint, but it can change between versions and a warning is printed. Pass model= explicitly for reproducible results.