Cheatsheet: Run a model in a few lines

The one-liner

from transformers import pipeline
classifier = pipeline("sentiment-analysis")
classifier("I love this!")

Picks a default model, downloads and caches it, runs all three steps. Pass model="checkpoint-name" to override the default.

Common task strings for pipeline()

Task string	What it does
sentiment-analysis	Classify text as positive or negative
zero-shot-classification	Classify into labels you supply at call time
text-generation	Continue a prompt (decoder-only)
fill-mask	Fill in a blanked token (encoder-only)
ner	Named-entity recognition (label tokens)
question-answering	Extract an answer span from a context
summarization	Shorten a long text
translation	Translate between languages

The three steps a pipeline hides

Step	What happens	Class
1. Preprocess	Text becomes input_ids + attention_mask	AutoTokenizer
2. Model	Numbers run through the network, output logits	AutoModelFor<Task>
3. Postprocess	Logits become probabilities and labels	softmax + id2label

The tokenizer step

from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
inputs = tokenizer(texts, padding=True, truncation=True, return_tensors="pt")
# inputs -> {"input_ids": ..., "attention_mask": ...}

• padding=True: pad a batch to equal length
• truncation=True: cut over-long inputs
• return_tensors="pt": return PyTorch tensors

The model step: pick the head by class name

Class	Head / output
AutoModel	None; raw hidden states (a vector per token)
AutoModelForSequenceClassification	Classify the whole input
AutoModelForTokenClassification	Label each token (NER)
AutoModelForQuestionAnswering	Answer-span start/end
AutoModelForCausalLM	Next-token generation (decoder-only)
AutoModelForMaskedLM	Fill blanks (encoder-only)

from transformers import AutoModelForSequenceClassification
model = AutoModelForSequenceClassification.from_pretrained(checkpoint)
outputs = model(**inputs)   # outputs.logits

The postprocessing recipe

import torch
probs = torch.nn.functional.softmax(outputs.logits, dim=-1)
labels = model.config.id2label   # {0: 'NEGATIVE', 1: 'POSITIVE'}

Models output logits (raw scores), not probabilities. Softmax normalizes; id2label names the columns.

The one idiom

SomeAutoClass.from_pretrained(checkpoint)

Loads tokenizers, base models, and task heads alike. Change the checkpoint string to change the model; change the head class to change the task.

Words to use precisely

• Checkpoint: a model’s name on the Hub (e.g. distilbert-base-uncased-finetuned-sst-2-english), passed to from_pretrained.
• Hidden states / features: the base transformer’s per-token output vectors, before any task head.
• Head: the small layer(s) on top of the base model that produce task-specific output.
• Logits: raw unnormalized model scores; softmax turns them into probabilities.

Recommended further study

• Hugging Face LLM Course, Chapter 2: “Using Transformers.” huggingface.co/learn/llm-course/chapter2. Released under Apache 2.0; this lesson mirrors its structure with original prose.