Few-Shot Demonstrations and Prompt Datasets

A few-shot prompt puts a handful of input-output examples in front of the new request. The model reads a local pattern off them — labels, formatting, tone, procedure — and answers accordingly. Nothing in its parameters changes.

That framing is not folklore. It is how GPT-3 was evaluated in the first place. The 175-billion-parameter model was tested at NeurIPS 2020 in zero-, one- and few-shot settings, with the demonstrations reaching it only through the prompt. Brown and 30 co-authors put the condition in the abstract: “For all tasks, GPT-3 is applied without any gradient updates or fine-tuning, with tasks and few-shot demonstrations specified purely via text interaction with the model.”

The same mechanism transmits what you did not intend to teach. A demonstration block can make the model copy an irrelevant detail. It can push the model toward the label it saw last. It can let examples that are too alike stand in for the rule they were meant to illustrate. Demonstration selection is a small dataset-design problem inside the context window. The measurements in this lesson show how small the margin for carelessness is.

Choose examples according to the behavior you need to teach. Format examples show fields, ordering, units, or a valid schema instance. Boundary examples contrast cases that differ on one decisive condition. Procedure examples demonstrate intermediate tool use, evidence handling, or escalation. Style examples set audience, tone, density, or vocabulary under a stable task. Failure examples show abstention, missing evidence, and out-of-scope behavior.

Name which of these a given example is for. It is the cheapest way to notice that a set holds five of one kind and none of the rest.

Nearest examples are not automatically the best teaching set. Both halves of that sentence have been measured.

Retrieving examples that resemble the query does work, and there is a published baseline for how well. A 2022 method called KATE does exactly that: instead of sampling demonstrations at random, fetch the ones semantically closest to the test query. Jiachang Liu and colleagues report that “the retrieval-based prompt selection approach consistently outperforms the random selection baseline”. In the published version it reaches 45.5% on open-domain question answering over Natural Questions, and 44.3% on the ToTTo table-to-text dataset.

Coverage is the answer to where that strategy runs out. Levy and colleagues tested compositional generalization in 2023 and found that “selecting similar demonstrations is insufficient, as often no example will be similar enough to the input”. Selecting instead a diverse set that covers the structures the output has to contain substantially improved performance across three compositional semantic parsing datasets.

The three strategies are goals, not a ranking. A static curated set is easy to version and audit and stays consistent across requests; it can also waste context on irrelevant cases and miss rare input types. A similarity-selected set adapts to domain and terminology, but it depends on embedding quality, can reinforce a wrong neighborhood, and needs diversity and label-balance controls. A coverage-oriented set makes contrasts explicit and supports systematic testing; it is often less lexically similar, and you need a task taxonomy before you can build one at all. The two results above tell you which goal is the one currently failing.

These failures are common even when each example is individually correct. The first and the last of them have been quantified.

Treat examples as a dataset with a measurable purpose. The measurements above set the order of work.

1. Name the behavior. Specify whether the examples teach format, boundary, procedure, or style.

2. Add contrasts. Pair similar inputs with different outcomes driven by one condition. That is the point of the compositional-generalization result: “selecting similar demonstrations is insufficient, as often no example will be similar enough to the input”.

3. Balance labels and slices. Control class frequency, language, domain, and risk coverage. Majority-label bias is one of the three effects that had to be calibrated away for up to 30.0% absolute.

4. Randomize order tests. Measure whether position changes the prediction. This step is not a precaution. Permutation alone spans near state-of-the-art to random guessing, and entropy-based ordering was worth 13% relative across eleven text classification tasks.

5. Perturb surface details. Replace names, values, and wording while preserving the rule, so that template copying shows up as a failure rather than as a passing score.

6. Version and evaluate. Lock examples with the prompt and a protected regression set. Roughly 4.7 million evaluation samples from 263 benchmarks reached GPT-3.5 and GPT-4 in the first year after release, largely through prompts. That is what an unversioned example pool does to an evaluation set over time.

Demonstrations earn their place when they communicate distinctions that instructions describe poorly. The published results say precisely where that value sits. Not in the correctness of each label — Min and colleagues randomized those at little cost. It sits in the label space, the input distribution, the format, the balance, and the order. Demonstrations are costly when they consume context, leak data, or create brittle copying.

The next lesson moves from examples to decomposition. Instead of showing complete answers, a workflow can divide a task into stages with observable intermediate contracts.