Few-shot prompting

Ingénierie des prompts avec l'API OpenAI

Fouad Trad

Machine Learning Engineer

Few-shot prompting

Modèle fourni avec des exemples (paires question-réponse)

A visual representation of a few-shot prompt showing its structure with example questions and answers, and finally the question we want the model to answer.

Few-shot prompting

Modèle fourni avec des exemples (paires question-réponse)

A visual diagram showing how we send a few-shot prompt to the LLM.

Few-shot prompting

Modèle fourni avec des exemples (paires question-réponse)

A visual diagram showing how we send a few-shot prompt to the LLM and get an answer in return.

Few-shot prompting

Modèle fourni avec des exemples (paires question-réponse)

A visual diagram showing how we send a few-shot prompt to the LLM and get an answer in return.

Nombre d'exemples :
- Zero -> zero-shot prompting
- One -> one-shot prompting
- Plus d’un -> few-shot prompting

Zero-shot prompting

Fournir un prompt sans exemple
Le modèle génère des réponses en fonction de ses connaissances
Idéal pour les tâches rapides et simples

prompt = "What is prompt engineering?"
print(get_response(prompt))

Prompt engineering refers to designing and refining prompts or instructions given 
to a language model like ChatGPT to elicit desired responses or behaviors. It 
involves formulating specific guidelines or hints to guide the model's output 
towards a desired outcome.

One-shot prompting

Fournir un exemple unique au modèle
Utile pour assurer la cohérence du formatage ou du style

prompt =  """ 
Q: Sum the numbers 3, 5, and 6. A: 3+5+6=14
Q: Sum the numbers 2, 4, and 7. A: 
"""
print(get_response(prompt))

2+4+7=13

One-shot prompting

prompt = """
Q: Sum the numbers 3, 5, and 6. A: The sum of 3, 5, and 6 is 14
Q: Sum the numbers 2, 4, and 7. A: 
"""
print(get_response(prompt))

The sum of 2, 4, and 7 is 13

Few-shot prompting

Fournir plusieurs exemples
Puissant pour les tâches contextuelles

prompt = """
Text: Today the weather is fantastic -> Classification: positive
Text: The furniture is small -> Classification: neutral
Text: I don't like your attitude -> Classification: negative

"""

Few-shot prompting

Fournir plusieurs exemples
Puissant pour les tâches contextuelles

prompt = """
Text: Today the weather is fantastic -> Classification: positive
Text: The furniture is small -> Classification: neutral
Text: I don't like your attitude -> Classification: negative
Text: That shot selection was awful -> Classification: 
"""
print(get_response(prompt))

negative

Prompting few-shot avec un modèle de chat

response = client.chat.completions.create(
  model = "gpt-3.5-turbo",

  messages = [{"role": "user",
             "content": "Today the weather is fantastic"},

             {"role": "assistant",
             "content": "positive"},

              {"role": "user",
             "content": "I don't like your attitude"},
              {"role": "assistant",
             "content": "negative"},

              {"role": "user",
             "content": "That shot selection was awful"}
             ],
  temperature = 0
)

print(response.choices[0].message.content)

negative

Points importants

Sélectionner le nombre de shots en fonction de la complexité de la tâche
- Fewer shots -> tâches basiques
- Diverse shots -> tâches complexes

An image showing a person questioning themselves.

Passons à la pratique !

Ingénierie des prompts avec l'API OpenAI