Cheatsheet: The main NLP tasks

The shared loop (every task)

1. Load + clean data (datasets, map, filter)
2. Tokenize (with task-specific alignment)
3. Load AutoModelFor<Task>
4. Pick a data collator + a metric
5. Train with Trainer (or Seq2SeqTrainer)
6. Evaluate + push to Hub

Only the head, the label shape, and the metric change between tasks.

Task -> head -> metric -> shape

Task	Head	Metric	Shape
Sequence classification	AutoModelForSequenceClassification	accuracy, F1	encoder
Token classification (NER)	AutoModelForTokenClassification	seqeval (entity F1)	encoder
Extractive QA	AutoModelForQuestionAnswering	SQuAD EM, F1	encoder
Masked LM	AutoModelForMaskedLM	perplexity	encoder
Causal LM	AutoModelForCausalLM	perplexity	decoder
Summarization	AutoModelForSeq2SeqLM	ROUGE	encoder-decoder
Translation	AutoModelForSeq2SeqLM	BLEU / SacreBLEU	encoder-decoder

Data collators by task

Collator	Used for
DataCollatorWithPadding	Sequence classification (dynamic padding)
DataCollatorForTokenClassification	Token classification (pads labels too)
DataCollatorForLanguageModeling	Masked LM (and causal LM with mlm=False)
DataCollatorForSeq2Seq	Summarization, translation

Wrinkle 1: token-level alignment

Token classification and QA work in token positions, but labels/answers live at word/character level. Use the fast tokenizer:

• word IDs: spread a word’s label across its tokens (NER)
• offsets: map predicted start/end token positions back to characters (QA)

This is why fast tokenizers matter (lesson 6).

Wrinkle 2: sequence-to-sequence

Summarization and translation need:

• AutoModelForSeq2SeqLM
• Seq2SeqTrainingArguments + Seq2SeqTrainer
• DataCollatorForSeq2Seq (targets are sequences, padded + shifted)
• generation-based metrics: ROUGE (summarize), BLEU (translate)

Language modeling: no labels needed

The text is its own supervision; DataCollatorForLanguageModeling builds the targets:

• Masked LM (BERT): random masking, mlm=True (default)
• Causal LM (GPT): next-token, mlm=False
• Metric: perplexity (lower is better)

Words to use precisely

• Token classification: one label per token (NER, POS tagging).
• Extractive QA: return a span of the source; vs generative QA, which writes a new answer (and can hallucinate).
• Perplexity: a language-model quality metric; lower is better.
• seqeval / ROUGE / BLEU / SQuAD: standard metrics for NER / summarization / translation / QA.

Recommended further study

• Hugging Face LLM Course, Chapter 7: “Main NLP tasks.” huggingface.co/learn/llm-course/chapter7. Released under Apache 2.0; this lesson mirrors its structure with original prose.