Mission 7: Few-Shot Learning

🎯 Your Learning Goals

You understand why examples are more powerful than descriptions

You create effective input-output examples

You apply Few-Shot Learning to text, code, and data

💡

The Power of Examples

Imagine teaching someone what a Haiku is:

Just description:
"A haiku is a Japanese poem with 3 lines and specific syllable counts."

With example:
"A haiku has 5-7-5 syllables:
An old silent pond
A frog jumps into the pond—
Splash! Silence again."

Which explanation helps more? The example! The same applies to AI. A good example says more than a thousand words.

⚡ Zero-Shot vs. Few-Shot: The Difference

❌ Zero-Shot (Instruction Only)

✅ Few-Shot (With Examples)

Task: Classify product reviews as positive/negative

Prompt:
"Classify these reviews as positive or negative."

Problem: The AI doesn't understand exactly what you mean. Stars? Tone? Keywords?

Task: Classify product reviews as positive/negative

Prompt:

Review: "Delivery was fast, but quality mediocre."
Sentiment: neutral


Review: "Absolutely fantastic, would buy again!"
Sentiment: positive


Review: "Totally disappointed, want refund!"
Sentiment: negative


Review: "[YOUR REVIEW]"
Sentiment:

Result: The AI understands the pattern and applies it!

🧬 The Anatomy of Good Examples

1

Clear Input

What goes in? A sentence? Text? Data?

Input: "The product is okay."

2

Clear Output

What should come out? Format, style, length?

Output: {"sentiment": "neutral", "confidence": 0.7}

3

Consistency

All examples in the same format!

✓ SAME FORMAT

4

Variety

Show different cases (edge cases!)

positive / negative / neutral / sarcasm

📋 The Golden Rule:

Input: [Your input here]
Output: [Desired output here]


Input: [Your input here]
Output: [Desired output here]


Input: [ACTUAL TASK]
Output:

The AI will automatically complete the last Output!

🔨 The Few-Shot Builder

Select a task and see how the prompt improves with examples:

Select a task:

Number of examples:

Your Few-Shot Prompt:

AI Response:

📚 Proven Few-Shot Patterns

🎭

Style Transfer

Teach the AI a specific writing style:

Input: Welcome to our store.

Output: Hey you! 🎉 Welcome to the club – so glad you're here!

Marketing Tone of Voice

📊

Structure Definition

Define the desired output format:

Input: I have a doctor's appointment on Monday.

Output: {"what": "Doctor appointment", "when": "Monday", "where": "", "priority": "normal"}

JSON Data Extraction

🔄

Reformulation

Show how text should be transformed:

Input: The product is not bad.

Output: The product is good.

Simplification Clarity

🧩

Classification

Sort content into categories:

Input: My computer won't start anymore.

Output: Category: Hardware → Sub: Boot Problems

Support Routing

✨ Quality Tips for Examples

✅

Dos

Start with simple examples, then increase difficulty
Show edge cases (borderline cases)
Keep examples short and concise
Use 2-5 examples (usually optimal)
End with the real input (without output)

❌

Don'ts

Give contradictory examples
Use too many examples (context limit!)
Only show similar cases
Start with overly complex examples
Change format between examples

🚀 Real-World Applications

📧

Email Categorization

Automatically sort incoming emails into "Important", "Promotions", "Support"...


Input: "50% off all shoes!"

Category: Promotion


Input: "Project deadline tomorrow"

Category: Important

💬

Chatbot Training

Teach your bot how to respond to different requests:


Customer: "When will my delivery arrive?"

Response: "Let me check that for you..."

📱

Content Moderation

Automatically identify problematic comments:


Comment: "You're so stupid!"

Action: Issue warning


Comment: "Good point, but..."

Action: OK

📄

Data Extraction

Extract structured data from unstructured text:


Text: "John Doe, 03/25/1990"

JSON: {"name": "John Doe", "dob": "03/25/1990"}

🔥 Pro Tips for Few-Shot

1

Chain-of-Thought + Few-Shot Combined

Show not just the result in examples, but also the thought process:

"Input: 2+3×4
Thought: First multiplication (3×4=12), then addition (2+12=14)
Output: 14"

2

Use Negative Examples

Also show what is NOT desired to avoid misunderstandings.

3

Real-World Examples

Use actual data from your project as examples – then the output fits perfectly!

4

Dynamic Few-Shot Templates

Create reusable templates with placeholders for different tasks.

✅ Quick Check

1. What is the main principle of Few-Shot Learning?

Send more prompts

Show examples instead of describing

longerWrite longer prompts

2. How many examples are usually optimal?

10+ examples

two-five2-5 examples

Only 1 example

3. What is important about the examples?

They must be complex

similarThey should all be similar

consistentConsistent format

📊 Your Mission Progress

0/3 Objectives Completed

0% Total Progress

🎯 Check all learning objectives to complete the mission!

Few-Shot Learning