Few-shot prompting provides models with a few input-output examples to induce understanding of a given task, in contrast to zero-shot prompting, where no examples are provided. Providing even a few high-quality examples improved model performance on complex tasks compared to no demonstration.
However, few-shot prompting requires additional tokens to include examples, which can become prohibitive for longer text inputs. Additionally, the selection and composition of prompt examples can significantly influence the model's behavior, and biases such as preference for frequent words can still affect few-shot prompting results.
Although few-shot prompting improves capabilities for complex tasks, especially among large pre-trained models like GPT-3, careful prompt engineering is essential to achieve optimal performance and mitigate unintended model biases.
Learn by example
The principle of the few-shot prompt is to provide the LLM with a small set of relevant examples or demonstrations within the prompt itself. These examples guide the model, illustrating how to approach and respond to a particular type of task or question. Demonstrations are generally structured as follows:
- Input-output pairs: Each demonstration typically consists of an input (e.g., a question or a piece of text) and its corresponding output (the desired answer or solution).
- Format consistency: The demonstrations maintain a consistent format, which helps the model recognize the pattern it should follow.
- Relevance of the task: The examples provided are directly relevant to the task at hand, highlighting the specific skill or knowledge required.
When presented with these demonstrations, the language model engages in a process often called “learning in context” or “learning by example.” It works as follows:
- Pattern recognition: The model analyzes the provided examples, identifying patterns in how inputs are transformed into outputs.
- Task inference: From these models, the model deduces the nature of the task it is asked to perform
- Generalization: The model then attempts to generalize from the given examples to new, unseen inputs.
- Application : finally, the model applies this learned model to the new input provided at the end of the prompt.
Help with reasoning
A few tips can greatly improve an LLM's ability to handle tasks involving multi-step reasoning, logical deduction, or domain-specific knowledge.
“Give a possible diagnosis and explain your reasoning:
Symptoms: fever, cough, fatigue
Diagnosis: cold
Explanation: The combination of fever, cough and fatigue is typical of a cold. No serious symptoms are present, suggesting a mild viral infection.
Symptoms: Chest pain, shortness of breath, dizziness
Diagnosis: Possible heart attack
Explanation: The combination of chest pain, shortness of breath, and dizziness are warning signs of a possible heart attack. Immediate medical attention is required.
Symptoms: headache, sensitivity to light, nausea
Diagnosis:
Explanation : "
This prompt guides the model not only to provide a diagnosis, but also to explain the reasoning behind it, demonstrating complex medical reasoning.
“Diagnosis: Migraine
Explanation: These symptoms together are typical of a migraine. Headache combined with sensitivity to light (photophobia) is a strong indicator of migraine, and nausea often accompanies this condition. Although other causes are possible, this combination strongly suggests migraine.
Code generation
Few-shot prompts can be incredibly useful in helping LLMs generate code that adheres to specific conventions, follows best practices, or meets particular requirements. Examples can demonstrate the correct syntax and structure for a particular programming language.
Let's see how to use few-shot prompts to generate Python functions with docstrings and type hints:
This prompt guides the template to generate a function with appropriate type hints and a detailed docstring, following the established pattern.
THE result would be :
How to practice few-shot prompting
The quality and relevance of the examples you choose are essential to successfully encourage the use of a few examples. Here are some tips:
- Make sure the examples are directly related to the task you want the model to perform. Irrelevant examples can confuse the model and lead to poor performance.
- Use a diverse set of examples that cover different aspects of the task. This helps the model generalize better to new inputs.
- Examples should be clear and unambiguous. Avoid complex or convoluted examples that could confuse the model.
One risk of encouraging the use of few examples is that the model may overfit to the examples provided, leading to results that are too similar to the examples or fail to generalize well to new inputs. To avoid this:
- Use a variety of examples that cover different scenarios and edge cases. This helps the model learn to generalize rather than just imitating the examples.
- Avoid using too many examples, as this can lead to overfitting. A few well-chosen examples are often more effective than a large number of similar examples.
- Test the model's performance on a range of new inputs to ensure it generalizes well. Adjust examples and prompts as needed based on these tests.
By following these best practices, you will be able to create more effective few-shot prompts that guide the model in performing the desired task accurately and efficiently.
EN 
FR