Zero-, one-, and few-shot learning: Quick Reference — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)

Quick Reference: Zero-, One-, and Few-Shot Learning in Prompt Engineering

In the context of prompt engineering for large language models (LLMs), zero-, one-, and few-shot learning techniques enable flexible adaptation of models to new tasks without retraining. These approaches are essential for efficient use of generative AI systems and form a core part of the NVIDIA-Certified Professional: Generative AI LLMs exam.

Key Definitions

• Zero-Shot Learning: The model performs a task based solely on the prompt, without any example demonstrations. It relies entirely on the model's pre-trained knowledge and the instructions provided.
• One-Shot Learning: The prompt includes exactly one example of the task to guide the model's output.
• Few-Shot Learning: The prompt contains a small number (typically 2–10) of examples demonstrating the task, improving model understanding and output quality.

Core Principles

• Prompt Design: Clear, concise instructions are critical. For zero-shot, the prompt must explicitly describe the task. For one- and few-shot, examples should be representative and well-formatted.
• Example Selection: Examples should cover typical cases and edge cases to guide the model effectively.
• Context Length: The number of examples is limited by the model's maximum token context window.
• Output Control: Prompt engineering can influence style, format, and content specificity by including explicit constraints or demonstrations.

When to Use Each Approach

• Zero-Shot: Rapid prototyping or when no labeled examples are available.
• One-Shot: When a single example can clarify ambiguous instructions or task format.
• Few-Shot: When higher accuracy or nuanced understanding is required and a few examples can be provided.

Best Practices

• Use natural language instructions that the model can easily interpret.
• Keep examples consistent in style and format.
• Test prompts iteratively to optimize model responses.
• Combine with chain-of-thought prompting for complex reasoning tasks.

Worked Example: Few-Shot Prompt for Sentiment Classification

Task: Classify the sentiment of a given sentence as Positive or Negative.

Prompt:

Review: "I love this product!" Sentiment: Positive Review: "The service was terrible." Sentiment: Negative Review: "The movie was okay, not great." Sentiment:

Expected Model Output: Neutral or Negative (depending on prompt design)

Explanation: The prompt provides two examples (few-shot) demonstrating the task format, guiding the model to classify the sentiment of the third review.

More in this topic

Prompt Engineering — NVIDIA-Certified Professional: Generative AI LLMsZero-, one-, and few-shot learning: Common Mistakes — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Controlling model output — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Zero-, one-, and few-shot learning: Worked Example — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Chain-of-thought prompting — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Domain adaptation via prompting — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Zero-, one-, and few-shot learning — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)Zero-, one-, and few-shot learning: Practice Questions — Prompt Engineering (NVIDIA-Certified Professional: Generative AI LLMs)