Google

Few-shot learning is a method of context scaling that augments a prompt with multiple input-output examples, known as demonstrations. By explicitly providing these examples, the language model can implicitly learn task behavior from them and condition its predictions on this prior information, all without requiring any updates to its internal parameters.

Few-Shot Learning in Prompting

A practical application of in-context learning involves providing a Large Language Model with several examples of text classified by sentiment directly in the prompt. For instance, a prompt might first provide examples like 'I had an amazing day at the park!' (Positive), 'The service at the restaurant was terrible.' (Negative), and 'I think it's going to rain today.' (Neutral). Based on these demonstrations, the model can then successfully classify a new text, such as 'This movie was a fantastic journey through imagination.', as 'Positive'.

Example of In-Context Learning for Sentiment Classification

The following instruction is an example of the initial part of a few-shot prompt, specifically a 2-shot prompt, used for sub-problem generation in the least-to-most prompting method: 'Your task is to decompose a problem into several sub-problems. You will be given a few examples to illustrate how to achieve this.' This instruction sets the stage for the demonstrations (shots) that will follow, guiding the LLM on how to perform the decomposition task.

Example of an Instructional Prompt in a Few-Shot Setting for Sub-Problem Decomposition

A developer is trying to get a large language model to extract the main subject from a user's question and format it as a single keyword. They provide the model with several examples within the prompt before giving it a new, unseen question. However, the model's output for the new question is unreliable. Analyze the examples provided and identify the most significant flaw that prevents the model from learning the desired task. Explain your reasoning.

Troubleshooting a Prompting Strategy

In the context of few-shot learning, demonstrations are example input-output pairs included in a prompt. These samples serve to teach a large language model how to perform a specific task by showing it concrete examples of the desired behavior.

Demonstrations in In-Context Learning

A developer wants a language model to consistently translate informal text messages into a formal, professional tone. The goal is to guide the model's output by showing it examples of the desired transformation directly within the query, without altering the model's permanent parameters. Which of the following inputs best applies this in-context learning method?

A developer provides the following complete text to a large language model to get it to perform a specific information extraction task:

"""
Extract the product name and price from the following sentences into a structured format.

Sentence: The new SuperWidget X1 costs $99.99.
Extraction: {"product": "SuperWidget X1", "price": "$99.99"}

Sentence: For just $45, you can get the BasicModel 2.
Extraction: {"product": "BasicModel 2", "price": "$45"}

Sentence: The premium UltraGadget Pro is on sale for $199.
Extraction:
"""

Analyze the structure of this input. Explain the purpose of each distinct part of the text and describe how this overall structure helps guide the model to produce the desired output without being explicitly retrained.

Analyzing a Prompt's Structure for In-Context Task Learning

A developer is constructing a prompt to teach a language model a new task by providing examples directly in the input. Match each component of the prompt to its specific role in this in-context learning process.

Providing a Large Language Model with a standard few-shot prompt that only includes question-answer pairs without detailed problem-solving steps is often insufficient for complex reasoning tasks. For example, demonstrating that the average of $${}1$$, $${}3$$, $${}5$$, and $${}7$$ is $${}4$$ does not effectively teach the model how to calculate averages. When subsequently asked to calculate the average of $${}2$$, $${}4$$, and $${}9$$, the model remains unable to derive the correct calculation path and may incorrectly output $${}7$$ instead of the correct answer.

Learn Before

Related