Foundation Models and Their Training Data

Anyone who reads about prompting large models quickly picks up a simple picture. The prompt is what you type. The context is something extra you add to it. The examples you include are a kind of context. Under this picture, prompt and context are close to synonyms, and the terms zero-shot and few-shot sound like labels for how much background you supply. Before reading on, make a guess: do you think prompt and context always name the same thing?

The picture is partly right and partly misleading. This article tests it, because a learner who assumes one fixed meaning will misread a good deal of documentation. The goal is to build a precise vocabulary first, then follow what happens inside a model when a prompt is processed.

The GPT-3 paper by Brown and colleagues, published in 2020, used the word context for the entire input given to a model. Under that usage, context and prompt name the same thing. The word is also used to mean contextual information, which is the material a model needs in order to carry out what the prompt asks. A third usage appears in Google's PaLM 2 documentation, which describes context as the text that shapes how the model responds throughout a conversation. That documentation gives examples such as words the model may or may not use, topics to focus on or avoid, and the required response format or style. Taken literally, that definition makes context very close to a task description.

Why does this matter? The mix-up is more than a matter of taste, because one sentence can be read in different ways. Consider the sentence 'the context is too long.' It might mean that the whole input is too long, or that the background material is too long. These two readings lead to different fixes. Shortening the background material does nothing if the real problem is the length of the whole input, and vice versa.

Under the course convention, the context is a part of the prompt, not a separate thing sitting beside it. The diagram below shows one way to arrange the parts of a prompt. The whole prompt contains a context section, which holds the task-relevant information such as instructions and examples, and a query, which is the input the model is asked to answer. The source does not give a fixed layout, so treat this arrangement as one reasonable way to see the nesting.

In-context learning is the name for a model learning from the prompt and then performing a task based on it. The model is a foundation model that has learned from a massive amount of data during training. What is notable is that the prompt itself, rather than any further training, determines which task the model carries out. The same underlying model can answer very different requests depending on what the prompt contains.

Read the diagram as a flow. The prompt and the broad knowledge from training meet inside the model. The prompt does not add new knowledge in the sense of retraining; instead it selects which behavior appears in the output. This is why a single model can translate, summarize, or classify, depending on the examples and instructions it receives.

In the zero-shot column, the model receives the task and the query and nothing else. In the few-shot column, the same task is followed by several worked examples before the query. Each example is a shot. The query stays unanswered in both cases, because the model's job is to produce the answer for it.

Before you move on, trace the state. Start with a prompt that has no examples. The model receives the instruction and the query, and its output is produced from those contents plus training. Now add one example in front of the query. The input has changed, but the model's weights have not. The only thing that changed is what the model reads. Predict what happens if you keep adding examples: which part of the process changes, and which part stays fixed?

There is no universal count of examples. The right number depends on the model and the application. The general pattern is that more examples help a model learn better. The limit is the model's maximum context length, because every example uses space in the prompt. More examples also make the prompt longer, which raises the inference cost of each request. For this reason, you will usually need to experiment to find the number that works for your task.

Benefit is not uniform. For GPT-3, few-shot learning showed a significant improvement over zero-shot learning. In the use cases of a 2023 Microsoft analysis, few-shot learning gave only limited improvement on GPT-4 and a few other models. One reading of that result is that more powerful models are better at understanding and following instructions, so they need fewer examples. The source cautions that such a study may underestimate few-shot examples in domain-specific work. For instance, if a model has seen few examples of the Ibis dataframe API in its training data, adding Ibis examples to the prompt can still make a large difference.