Prompt Engineering Techniques

Overview

In this project, i applied In-Context Learning (ICL) techniques, including few-shot, one-shot, and zero-shot prompting, to answer multiple-choice questions (MCQs) from the medical MCQ dataset. These prompting techniques allow us to leverage a large language model (LLM) to solve complex medical questions without the need for additional task-specific training.

Types of In-Context Learning Applied:

Zero-Shot Prompting

For zero-shot prompting, the model is given a question and its options without any prior examples. The model attempts to select the correct answer based purely on its pre-trained knowledge.

Example:

     Prompt: "What is the recommended treatment for hypertension?\n(A) Lisinopril\n(B) Amoxicillin\n(C) Acetaminophen"
     Model Output: "(A) Lisinopril"

One-Shot Prompting

In one-shot prompting, we provide one example of a question-answer pair from the dataset before asking the model to choose an answer for the next question. This helps the model better understand the task.

Example:

    Prompt: "Example:\nQuestion: What is the treatment for diabetes?\nAnswer: Insulin\n\nNow answer the following:\nWhat is the 
    recommended treatment for hypertension?\n(A)
    Lisinopril\n(B) Amoxicillin\n(C) Acetaminophen"
    Model Output: "(A) Lisinopril"

Few-Shot Prompting

Few-shot prompting involves providing multiple question-answer pairs as examples to further guide the model in making the correct selection for a new question. This method tends to improve performance over zero-shot and one-shot approaches by offering more context.

Example:

   Prompt: "Examples:\n1. Question: What is the treatment for diabetes?\n Answer: Insulin\n2. Question: What is the treatment for 
   asthma?\n Answer: 
   Albuterol\n\nNow answer the following:\nWhat is the recommended treatment for hypertension?\n(A) Lisinopril\n(B) 
   Amoxicillin\n(C) Acetaminophen"
   Model Output: "(A) Lisinopril"

This flexible prompting approach allows us to explore different ways of adapting LLMs to specific tasks without requiring task-specific fine-tuning.

Prompting with Roles

This project demonstrates how to use role-based prompting to guide Large Language Models (LLMs) in generating context-specific responses. The examples highlight how roles can provide additional context, tone, or style to the model's output.

Features

Basic Prompting: Example of a simple prompt without any role specification.
Role-Based Prompting: Introduces a role context to refine the model's responses.
Custom Tone and Style: Ability to customize the response tone, such as speaking like an English pirate.

Code Examples

Basic Prompt A simple prompt asking a question without specifying a role:

   prompt = """
   How can I answer this question from my friend:
   What is the meaning of life?
   """
   inputs = tokenizer(prompt, return_tensors="pt")
   outputs = llm.generate(**inputs, max_length=200)
   response = tokenizer.decode(outputs[0], skip_special_tokens=True)             
   print(response)

Role-Based Prompt

A role-based prompt providing context and a specific tone:

     role = """
     Your role is a life coach \
     who gives advice to people about living a good life.\
     You attempt to provide unbiased advice.
     You respond in the tone of an English pirate.
     """
     prompt_with_role = f"""
     {role}
     How can I answer this question from my friend:
     What is the meaning of life?
     """
     inputs = tokenizer(prompt_with_role, return_tensors="pt")
     outputs = llm.generate(**inputs, max_length=200)
     response = tokenizer.decode(outputs[0], skip_special_tokens=True)
     print(response)

Requirements

Python >= 3.8
Hugging Face Transformers library
PyTorch Install the dependencies using:
```
   pip install transformers torch
```

Chain-of-Thought Prompting

This notebook provides a foundation for exploring chain-of-thought prompting with an LLM. Modify the prompts and settings as needed to suit specific tasks or use cases.

Structure

Prompting Problem Statement "15 of us want to go to a restaurant. Two of them have cars, each car can seat 5 people. Two of us have motorcycles, each motorcycle can fit 2 people. Can we all get to the restaurant by car or motorcycle?"
Chain-of-Thought Prompt Variants

Simple problem statement.
Problem statement with instructions to think step by step.
Problem statement with a request for step-by-step reasoning and final summary answer.

Model Inference Steps

Tokenize the prompt using a tokenizer.
Generate output using the language model.
Decode the output and print the response.

Requirements

Tokenizer: Used to preprocess the prompt.

Language Model (LLM): Used to generate responses.

                       pip install transformers torch

Usage Instructions

Define the Problem Prompt

The problem is described as a natural language prompt. It is structured in three variations to test the language model's reasoning ability:

Basic prompt.
Prompt with explicit instructions to think step by step.
Prompt requiring a single answer (yes/no) followed by an explanation.

Tokenize the Prompt

   inputs = tokenizer(prompt, return_tensors="pt")

This converts the text prompt into input tensors suitable for the LLM. 3. ### Generate Response

  outputs = llm.generate(**inputs, max_length=512)

The LLM generates a response based on the tokenized input. 4. ### Decode and Print Response

  response = tokenizer.decode(outputs[0], skip_special_tokens=True)
  print(response)

This decodes the LLM’s response and prints the result.

Solving Riddles using Few-Shot Prompting.

Overview

This notebook demonstrates the use of LangChain and the Groq LLM to solve riddles using few-shot prompting. The script utilizes LangChain's FewShotPromptTemplate to format examples and queries to the Groq model. It showcases solving riddles by leveraging iterative reasoning steps to arrive at a solution.

Requirements

Python 3.8+
LangChain
LangChain-Groq Integration (langchain_groq) Install the required Python packages using the following command:

pip install langchain langchain-groq

Notebook Structure

1. Imports

The following key modules are imported:

PromptTemplate and FewShotPromptTemplate from langchain.prompts
HumanMessage from langchain.schema
ChatGroq from langchain_groq

These modules enable prompt formatting, interaction with the Groq LLM, and structured messaging.

2. Initialize the Groq Model

llm = ChatGroq(groq_api_key=groq_api_key, model_name="Gemma2-9b-It")

The ChatGroq object is initialized with the API key and model name.

3. Define Riddles

Three riddles are defined:

Riddle 1: A train speed problem.
Riddle 2: A farmer's animal count problem.
Riddle 3: An age difference problem.

Each riddle is passed as a HumanMessage to the model.

4. Few-Shot Examples

The notebook defines two solved examples:

Example 1: A train speed problem with reasoning steps.
Example 2: A farmer's animal count problem with equations and reasoning steps.

These examples include intermediate reasoning steps and final answers, demonstrating the format expected from the LLM.

5. Prompt Template

A FewShotPromptTemplate is created using the examples, a specific example prompt structure, and a suffix for new input:

prompt = FewShotPromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
    suffix="Question: {input}",
    input_variables=["input"]
)

6. Query the Model

For each riddle, a prompt is formatted using the FewShotPromptTemplate, and the query is sent to the Groq LLM:

message = HumanMessage(content=prompt.format(input=riddle_1))
print(llm([message]))

The responses are printed for each riddle.

Example Outputs

Riddle 1: Train Speed Problem

Question: A train leaves a station traveling at 60 km/h. One hour later, another train leaves the same station traveling in the same direction at 90 km/h. How long will it take for the second train to catch up to the first train?

Answer:

We will follow up with some questions to get the answer.
Follow up: How far ahead is the first train when the second train starts?
Intermediate answer: In one hour, the first train travels 60 km.
Follow up: What is the relative speed between the two trains?
Intermediate answer: 90 km/h - 60 km/h = 30 km/h.
Follow up: How long will it take for the second train to cover the 60 km gap?
Intermediate answer: 60 km ÷ 30 km/h = 2 hours.
Final Answer: It will take 2 hours for the second train to catch up.

Riddle 2: Farmer's Animals Problem

Question: A farmer has chickens and cows. There are 20 heads and 56 legs in total. How many chickens and cows does the farmer have?

Answer:

We will follow up with some questions to get the answer.
Follow up: How many heads are there in total?
Intermediate answer: 20 heads, each animal has one head.
Follow up: How many legs does a chicken and a cow have?
Intermediate answer: A chicken has 2 legs and a cow has 4 legs.
Follow up: How can we set up equations to solve this?
Intermediate answer: Let x be the number of chickens and y be the number of cows. We know: x + y = 20 (heads) and 2x + 4y = 56 (legs).
Follow up: Solve the equations.
Intermediate answer: From x + y = 20, we get y = 20 - x. Substitute into the second equation: 2x + 4(20 - x) = 56. Simplify: 2x + 80 - 4x = 56 → -2x = -24 → x = 12. So, y = 20 - 12 = 8.
Final Answer: The farmer has 12 chickens and 8 cows.

Usage Instructions

Set Up API Key: Replace groq_api_key with your actual API key.
Run the Script: Execute the script in a Python environment with the necessary dependencies installed.
Modify Riddles: Add your own riddles by defining them as strings and formatting them into the prompt.
Interpret Results: The model’s response will be printed in the console for each riddle, including reasoning steps and the final answer.

Customization

Add more examples to the examples list to improve the model's performance on similar questions.
Modify the PromptTemplate structure to adjust the reasoning steps or format.

Troubleshooting

Invalid API Key: Ensure you are using a valid Groq API key.
Dependencies Missing: Verify that all required libraries are installed.
Incorrect Answers: Fine-tune examples in the FewShotPromptTemplate to guide the model more effectively.

Name		Name	Last commit message	Last commit date
Latest commit History 64 Commits
Prompt_Engineering_Techniques.ipynb		Prompt_Engineering_Techniques.ipynb
Prompt_Engineering_Techniques_2.ipynb		Prompt_Engineering_Techniques_2.ipynb
README.md		README.md
Solving_riddles_fewshot examples_.ipynb		Solving_riddles_fewshot examples_.ipynb

OlfatSeid/Prompt_Engineering_Techniques

Folders and files

Latest commit

History

Repository files navigation

Prompt Engineering Techniques

Overview

Types of In-Context Learning Applied:

Zero-Shot Prompting

One-Shot Prompting

Few-Shot Prompting

Prompting with Roles

Features

Code Examples

Role-Based Prompt

Requirements

Chain-of-Thought Prompting

Structure

Requirements

Usage Instructions

Define the Problem Prompt

Tokenize the Prompt

Solving Riddles using Few-Shot Prompting.

Overview

Requirements

Notebook Structure

1. Imports

2. Initialize the Groq Model

3. Define Riddles

4. Few-Shot Examples

5. Prompt Template

6. Query the Model

Example Outputs

Riddle 1: Train Speed Problem

Riddle 2: Farmer's Animals Problem

Usage Instructions

Customization

Troubleshooting

Acknowledgements

About

Resources

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages