Zero-shot and few-shot classification with a language model, without fine-tuning

A language model can classify text without a training run. You describe the task in plain language (zero-shot), or you also show it a few labelled examples in the prompt (few-shot, also called in-context learning). No model is trained or hosted, and changing the labels is a matter of editing a question.

This guide explains both approaches as spinf runs them, and what a few examples change, measured on our public benchmark.

Zero-shot: the labels are the question

In zero-shot classification the model sees only the content and a question that names the possible answers. With spinf, the question is a template that ends where the answer goes, and the answers are the options to score:

{
  "id": "topic",
  "template": "\n\nTopic (billing, technical, account, shipping, or cancellation):{?}",
  "options": [" billing", " technical", " account", " shipping", " cancellation"]
}

spinf returns the probability of each option, and they sum to 1. Nothing is generated, so there is no free text to parse and no label outside your list. It also returns the floor of each answer: the same question asked with empty content. Reading the lift over the floor removes the question's own lean towards one answer, which often helps zero-shot questions with several answers (see Calibration).

Zero-shot works best when the labels are ordinary words with a clear meaning: sentiment, a yes/no fact, a short list of distinct topics.

Few-shot: labelled examples in the prompt

Few-shot classification adds a handful of labelled examples to the content, before the item to score. Each example is written in the same format as the question, with its answer. The examples show the model your labels, where your boundaries are, and the edge cases your team has already decided.

In-context learning is not fine-tuning: the model's weights do not change. The examples are part of the input of each call, so you can change them, version them and test them on your own data at any time. They are also billed with every input, so keep them short and send many items per call. Writing questions shows the format.

What a few examples change, measured

These figures come from our public benchmark, on spinf-12b (Gemma 4 12B, optimized for scoring). "No examples" is zero-shot, and "calibrated" means it was read as the lift over the floor. "With examples" is 4 or 8 labelled examples in the prompt, as a mean over three different example sets. The examples are never the items being scored.

The synthetic sets were written for the benchmark, labelled by a language model and checked blind by a second model. They are cleaner than real data, so expect lower figures on your own content. The test sets, the questions, the example sets and the code to re-run everything are public: spinf-benchmarks on GitHub.

The pattern is consistent. Examples help most where the labels depend on your own definitions: which tickets a bot can resolve, which of seven policy categories a post falls into, how a review treats one aspect among several.

When zero-shot is already enough

Some questions need no examples at all:

• Clear sentiment. On 300 public product reviews, overall polarity was 97% correct with no examples (calibrated).
• Plain facts. On synthetic news, "does the article report a guidance change?" was 98% correct with no examples, and 99% with 8.

Start zero-shot. Add examples only for the questions where the scores do not separate your labels well enough.

A practical workflow

1. Label a sample. 100–300 items of your own data, with the answers your team would give.
2. Score zero-shot, with two or three wordings of each question. Read the lift over the floor for questions with several answers.
3. Add 4, then 8 examples to the questions that need them. Cover every answer at least once, and include the cases people get wrong. Try more than one example set.
4. Keep the best version fixed. Scores compare over time only under the same template, answers, examples and model build.
5. Choose thresholds on the probabilities, for what is automated and what goes to a person. See Probabilistic if statements.

A round on 300 items with 8 examples is about 400,000 tokens: a few cents per variant at $0.09 per million input tokens, and well within the 50M free tokens of a new account.

Try it

• The classification API page shows a live example and the cost at scale.
• Related guides: Probabilities from a language model and Probabilistic if statements. For the example format, see Writing questions.