Master Few-Shot Learning for Better LLM Prompts

1. 1
   Establish a Zero-Shot Baseline

   First, ask the model to perform a task with no examples. This is called 'zero-shot' prompting. It helps you understand the model's default behavior and provides a baseline to measure the improvement from Few-Shot Learning.
2. 2
   Introduce a Single Example (One-Shot)

   Now, add one complete input-output example to your prompt before the final query. This single demonstration, or 'one-shot' learning, shows the model the exact format and type of answer you expect, significantly improving its performance.
3. 3
   Apply Few-Shot Learning with Multiple Examples

   Expand to 2-5 examples to apply Few-Shot Learning. Providing multiple examples helps the model understand nuance and handle more complex cases. This reinforces the desired pattern and increases the reliability of the output.
4. 4
   Select Diverse and High-Quality Examples

   The success of Few-Shot Learning depends on your examples. Choose examples that are clear, accurate, and cover a range of scenarios (e.g., positive, negative, and neutral). The model learns the pattern you provide, so high-quality inputs lead to high-quality outputs.

import os
from openai import OpenAI

# Make sure to set your OPENAI_API_KEY environment variable
# export OPENAI_API_KEY='your-key-here'

client = OpenAI()

def classify_sentiment_with_few_shot(text_to_classify):
    """Uses Few-Shot Learning to classify text sentiment."""

    # 1. Define high-quality examples
    examples = [
        {"text": "This movie was fantastic, a must-see!", "sentiment": "positive"},
        {"text": "I'm not sure how I feel about this product.", "sentiment": "neutral"},
        {"text": "The customer service was incredibly slow and unhelpful.", "sentiment": "negative"}
    ]

    # 2. Construct the prompt with examples
    prompt_parts = []
    for example in examples:
        prompt_parts.append(f"Text: {example['text']}\nSentiment: {example['sentiment']}")
    
    # Add the new text to classify
    prompt_parts.append(f"Text: {text_to_classify}\nSentiment:")

    final_prompt = "\n\n---\n\n".join(prompt_parts)
    
    print(f"--- Generated Prompt ---\n{final_prompt}\n------------------------")

    # 3. Call the API
    try:
        response = client.chat.completions.create(
            model="gpt-3.5-turbo",
            messages=[
                {"role": "system", "content": "You are a helpful assistant that classifies text sentiment as positive, negative, or neutral."},
                {"role": "user", "content": final_prompt}
            ],
            temperature=0,
            max_tokens=10
        )
        return response.choices[0].message.content.strip()
    except Exception as e:
        return f"An error occurred: {e}"

# Text we want to classify
new_text = "The setup was easy but the battery life is disappointing."

# Get the classification
predicted_sentiment = classify_sentiment_with_few_shot(new_text)

print(f"\nInput Text: '{new_text}'")
print(f"Predicted Sentiment: {predicted_sentiment}")