Few-shot Prompt

Definition

A prompting technique that extends the one-shot approach by providing multiple examples (typically 2-5) to the model before asking it to perform a similar task. This method gives the model more context and pattern recognition opportunities.

Characteristics

• Contains multiple examples (usually 2-5)
• Shows various input-output relationships
• Demonstrates pattern consistency
• Provides broader context than one-shot
• Balances guidance with context efficiency

Example

Prompt:

Classify the emotion in the following statements:

Statement: "I just got promoted at work!"
Emotion: Joy

Statement: "I can't believe I lost my keys again."
Emotion: Frustration

Statement: "The movie made me cry."
Emotion: Sadness

Now classify this statement:
Statement: "I'm so excited for the weekend!"
Emotion:

Expected Response: "Joy" or "Excitement"

Use Cases

• Complex classification tasks
• Pattern recognition
• Style transfer with variations
• Multi-faceted tasks requiring diverse examples
• Tasks with subtle distinctions

Benefits

• Higher accuracy than zero-shot and one-shot
• Shows pattern variations and edge cases
• Reduces overfitting to single examples
• Provides robust task understanding
• Better handles task complexity and nuance

Limitations

• Uses more context tokens
• Requires careful example selection
• May still miss rare edge cases
• Can be overkill for simple tasks
• Example quality and diversity critical

Best Practices

• Use diverse, representative examples
• Include edge cases when relevant
• Maintain consistent output format across examples
• Order examples logically (simple to complex)
• Ensure examples cover the task scope
• Balance positive and negative examples when applicable

Example Selection Strategy

• Cover different input types
• Show various difficulty levels
• Include common and uncommon cases
• Demonstrate consistent labeling
• Avoid contradictory examples

When to Use

• Complex or nuanced tasks
• When accuracy is critical
• Tasks with multiple valid patterns
• Classification with subtle distinctions
• When context space allows for multiple examples