Few-Shot Learning Techniques

Few-shot learning allows you to teach AI models new tasks or patterns by providing a small number of examples directly in the prompt.

Zero-Shot vs Few-Shot vs Many-Shot

Basic Few-Shot Pattern

Convert informal text to formal business English.

Example 1:
Informal: "gonna need those reports asap"
Formal: "I require those reports at your earliest convenience."

Example 2:
Informal: "cant make the meeting, something came up"
Formal: "I regret to inform you that I am unable to attend the meeting due to an unforeseen circumstance."

Example 3:
Informal: "the project is kinda behind schedule ngl"
Formal: [Model completes this]

Effective Example Selection

Diversity Matters

Include examples that cover different scenarios:

Classify the sentiment of product reviews.

Example 1 (Positive):
Review: "Absolutely love this product! Best purchase ever."
Sentiment: Positive

Example 2 (Negative):
Review: "Complete waste of money. Broke after one day."
Sentiment: Negative

Example 3 (Mixed):
Review: "Great features but terrible battery life."
Sentiment: Mixed

Example 4 (Neutral):
Review: "It does what it says. Nothing special."
Sentiment: Neutral

Now classify:
Review: "The design is beautiful but it's overpriced."
Sentiment:

Edge Cases

Include tricky examples:

Extract company names from text.

Example 1:
Text: "I work at Microsoft."
Companies: ["Microsoft"]

Example 2:
Text: "The Apple product works with Google services."
Companies: ["Apple", "Google"]

Example 3:
Text: "She ate an apple while working on her resume."
Companies: []  # Note: "apple" here is the fruit, not the company

Example 4:
Text: "Amazon rainforest is disappearing, but Amazon.com is growing."
Companies: ["Amazon.com"]  # Only the company reference

Format Preservation

Models follow the exact format of your examples:

JSON Output

Parse event details into JSON.

Input: "Team standup tomorrow at 9am in the main conference room"
Output: {
  "event": "Team standup",
  "date": "tomorrow",
  "time": "9am",
  "location": "main conference room"
}

Input: "Lunch with Sarah next Tuesday at noon"
Output: {
  "event": "Lunch with Sarah",
  "date": "next Tuesday",
  "time": "noon",
  "location": null
}

Input: "Board meeting on March 15th, 2pm, Building A Room 401"
Output:

Markdown Output

Convert bug reports to markdown tickets.

Input: "Login button doesn't work on mobile Safari. Users can't access their accounts."

Output:
## Bug: Login Button Non-Functional on Mobile Safari

**Severity:** High
**Platform:** iOS Safari

### Description
Users are unable to access their accounts due to the login button being non-functional.

### Steps to Reproduce
1. Open site on iOS Safari
2. Navigate to login page
3. Tap login button
4. Observe: button does not respond

### Expected Behavior
Login button should authenticate user and redirect to dashboard.

---

Input: "Dark mode has wrong colors in settings panel"
Output:

Few-Shot for Code Tasks

Code Translation

Convert Python to TypeScript.

Python:
```python
def greet(name: str) -> str:
    return f"Hello, {name}!"

TypeScript:

function greet(name: string): string {
    return \`Hello, \${name}!\`;
}

Python:

from typing import List

def sum_evens(numbers: List[int]) -> int:
    return sum(n for n in numbers if n % 2 == 0)

TypeScript:

function sumEvens(numbers: number[]): number {
    return numbers.filter(n => n % 2 === 0).reduce((a, b) => a + b, 0);
}

Python:

async def fetch_data(url: str) -> dict:
    async with aiohttp.ClientSession() as session:
        async with session.get(url) as response:
            return await response.json()

TypeScript:

### Error Message Generation

```text
Generate helpful error messages for validation failures.

Input: email field is empty
Output: "Please enter your email address. We'll use this to send you important account updates."

Input: password too short
Output: "Your password needs at least 8 characters. Try adding numbers or symbols to make it stronger."

Input: phone number invalid format
Output:

Advanced Techniques

Contrastive Examples

Show what NOT to do:

Write clear variable names.

Bad: x = get_data()
Good: user_profile = get_user_profile_from_database()

Bad: temp = calc(a, b)
Good: total_price = calculate_subtotal(quantity, unit_price)

Bad: res = proc(inp)
Good:

Difficulty Progression

Order examples from simple to complex:

Write SQL queries for these natural language requests.

Request: "Get all users"
SQL: SELECT * FROM users;

Request: "Get users who signed up this month"
SQL: SELECT * FROM users WHERE created_at >= DATE_TRUNC('month', CURRENT_DATE);

Request: "Get the top 10 users by order count with their email addresses"
SQL: SELECT u.email, COUNT(o.id) as order_count
     FROM users u
     JOIN orders o ON u.id = o.user_id
     GROUP BY u.id, u.email
     ORDER BY order_count DESC
     LIMIT 10;

Request: "Get monthly revenue trends for the past year, broken down by product category"
SQL:

Optimal Number of Examples

Common Mistakes

1. Inconsistent formatting: Examples use different structures
2. Too similar: All examples cover the same case
3. Too many: Model gets confused or truncated
4. Wrong order: Complex before simple
5. No edge cases: Model fails on unusual inputs