Few-Shot Learning

Show the Model a Few Examples

Sometimes, no amount of describing your desired output format beats simply showing the model a few examples.

This is Few-shot Learning — providing input-output examples in the prompt so the model "learns" the pattern you expect, then applies it to new inputs.

Zero-shot vs One-shot vs Few-shot

Zero-shot

No examples, just ask directly:

Classify the sentiment of the following sentence as "positive" or "negative":

"The food at this restaurant was really disappointing."

The model usually answers correctly, but the format may be inconsistent (sometimes "negative", sometimes "This is a negative sentiment").

One-shot

One example:

Classify the sentiment as "positive" or "negative".

Sentence: "The weather is beautiful today, I'm feeling great."
Classification: positive

Sentence: "The food at this restaurant was really disappointing."
Classification:

Few-shot

Multiple examples:

Classify the sentiment as "positive" or "negative".

Sentence: "The weather is beautiful today, I'm feeling great."
Classification: positive

Sentence: "The delivery is late again, so annoying."
Classification: negative

Sentence: "This book is brilliantly written, highly recommend."
Classification: positive

Sentence: "The food at this restaurant was really disappointing."
Classification:

With multiple examples, the model more accurately understands:

  1. What the task is — sentiment classification
  2. Output format — only answer "positive" or "negative"
  3. Decision criteria — what counts as positive or negative

How to Choose Good Examples

Example quality directly affects results. Key principles:

1. Cover Different Cases

✅ Include both positive and negative examples
✅ Include simple and complex cases
✅ Include edge cases

If all your examples are positive, the model may lean toward classifying everything as positive.

2. Examples Should Match Target Input

If you're processing technical documents, examples should be technical documents too, not social media posts. Domain matching matters.

3. Consistent Formatting

All examples should follow identical formatting:

❌
Input: apple → fruit
Input: dog → this is an animal
cat → animal

✅
Input: apple
Category: fruit

Input: dog
Category: animal

Input: cat
Category:

4. More Isn't Always Better

Usually 3–5 examples suffice. Too many examples:

  • Consume precious context window
  • Introduce noise
  • Increase token costs

Example Order Matters

Research shows that the ordering of few-shot examples affects model output. Practical tips:

  • Place the most relevant examples closest to the target input (models are more sensitive to recent context)
  • Diversify ordering — don't group same-category examples together
  • If results are poor, try shuffling the order — sometimes that alone fixes things

Practical Applications

Data Transformation

Convert natural language dates to ISO format.

Input: March 15th next year
Output: 2027-03-15

Input: Last day of last month
Output: 2026-02-28

Input: Next Wednesday
Output:

Code Style Conversion

Convert the following JavaScript to TypeScript with type annotations.

JavaScript:
function add(a, b) { return a + b; }

TypeScript:
function add(a: number, b: number): number { return a + b; }

JavaScript:
function greet(name) { return "Hello, " + name; }

TypeScript:

Text Extraction

Extract company name and funding amount from the text.

Text: "ByteDance raised $5 billion in its latest funding round."
Result: {"company": "ByteDance", "amount": "$5 billion"}

Text: "SpaceX completed a new $2.1 billion funding round."
Result: {"company": "SpaceX", "amount": "$2.1 billion"}

Text: "Stripe announced a $6.5 billion strategic investment."
Result:

When Few-shot Helps Most

Few-shot is especially useful for:

  1. Specific output formats — examples are more intuitive than descriptions
  2. Classification tasks — demonstrating each category
  3. Data transformation — input-to-output mapping patterns
  4. Tasks the model struggles with — use examples to "teach" the model your intent

When NOT to Use Few-shot

  • The task description is already clear enough — if the model performs well without examples, adding them just wastes tokens
  • Each example is very long — quickly fills the context window
  • The task is too complex and requires reasoning — Chain of Thought (next chapter) is more appropriate

Key Takeaways

  1. Few-shot = showing the model examples in the prompt. It's one of the most intuitive and effective ways to convey your intent.
  2. Example quality > quantity. 3–5 high-quality, diverse, consistently formatted examples usually suffice.
  3. Example order affects results. Place the most relevant examples closest to the target input.
  4. Few-shot works best for formatted output, classification, and data transformation. For complex reasoning tasks, consider Chain of Thought.