Skip to main content

Few-shot prompt templates

Few-shot learning is a technique in machine learning where models are trained to perform tasks using just a few examples. This technique can be applied to large language models as well by providing a few examples in the prompt to guide the model's response.

Few-shot prompting is useful when you want the language model to perform a very specific task, but don't have enough data to fine-tune the model. By providing just a few examples in natural language, the model can learn to generate similar responses.

Constructing a few-shot prompt template​

There are two main ways to construct a few-shot prompt template:

1. Using a set of examples​

To create a few-shot prompt template from a set of examples:

  1. Create a list of input-output examples as dictionaries:
examples = [
  {"input": "happy", "output": "sad"},
  {"input": "hot", "output": "cold"} 
]
  1. Create a PromptTemplate to format the examples: 
example_prompt = PromptTemplate(
  input_variables=["input", "output"],
  template="Input: {input}\nOutput: {output}"
)
  1. Construct the FewShotPromptTemplate using the examples and formatter:
prompt = FewShotPromptTemplate(
  examples=examples,
  example_prompt=example_prompt,
  # Other parameters like suffix here
)

2. Using an example selector​

To use a more advanced example selector:

  1. Create the examples list like before

  2. Create an ExampleSelector instance - this will select the examples intelligently:

example_selector = SemanticSimilarityExampleSelector(
  examples=examples,
  embedding_model=OpenAIEmbeddings(),
  vector_store=Chroma()  
)
  1. Construct the prompt template using the selector:
prompt = FewShotPromptTemplate(
  example_selector=example_selector,
  example_prompt=example_prompt,
  # Other parameters
)

The example selector allows picking examples based on semantic similarity to the input, which creates higher quality prompts.

Use cases​

Few-shot prompting is useful for tasks like:

  • Classification
  • Translation
  • Question answering
  • Dialogue agents

And more. The key is quickly adapting the model to new tasks from just a few examples.