Notes on ChatGPT Prompt Engineering for Developers

Source: https://learn.deeplearning.ai/courses/chatgpt-prompt-eng

Guidelines

There are two types of LLMs:

1. Base LLMs: These predict the next word based on text training data only and cannot provide specific responses. This is for basic tasks.
2. Instruction-Tuned LLMs: These try to follow instructions, fine-tune LLMs, and follow instructions with RLHF (Helpful, Harmless, Honest). They are practically used.

Principles of Prompt Engineering for Instruction-Tuned LLMs:

• Write clear and specific instructions.
• Give the model time to think.

Steps for Prompt Engineering:

• Install libraries such as OpenAI, OS, and load the helper function (the latest being OpenAI library 1.0.0).
• The helper function includes:

def get_completion(prompt, model="gpt-3.5-turbo"): messages = [{"role": "user", "content": prompt}] response = openai.ChatCompletion.create( model=model, messages=messages, temperature=0, # This is the degree of randomness of the model's output ) return response.choices[0].message["content"]

This Python function, get_completion, interacts with OpenAI's ChatGPT model (specifically using the "gpt-3.5-turbo" variant) to generate responses to prompts. Here's a breakdown of how it works:

• Input:prompt: Input text or prompt provided to the function, representing the starting point or context for the conversation.model (optional): Specifies the version of the GPT model to use. Defaults to "gpt-3.5-turbo".
• Formatting the Input:The input prompt is formatted into a list of dictionaries called messages, with each dictionary representing a message in the conversation. The prompt is formatted as a user message.
• Generating Response:The function uses OpenAI's ChatCompletion.create method to generate a response to the provided prompt.It passes the model to use (model), the messages formatted earlier (messages), and additional parameters like temperature.temperature: Controls the randomness of the generated response. Setting it to 0 makes the response deterministic.
• Extracting Response:The function retrieves the generated response from the response object.It accesses the choices attribute of the response object, containing a list of potential completions.Since temperature is set to 0, there's only one choice, so it retrieves the first choice using response.choices[0].Finally, it extracts the content of the message from the first choice using response.choices[0].message["content"].
• Output:The function returns the content of the generated response as a string.

In summary, this function allows easy interaction with OpenAI's ChatGPT model, generating responses to prompts in a conversational manner, and providing a streamlined way to integrate AI-generated text into applications or projects.

Best Practices

• Use delimiters to indicate distinct parts of input. These could be '', "", <>, <tag>, </tag>, or :.
• Ask for structured output, such as JSON or HTML.
• Ask the model to check whether conditions are satisfied.
• Use few-shot prompting.
• Specify the steps required to complete a task.
• Instruct the model to work out its own solution before rushing to a conclusion.
• Instruct the model to test/validate its own results.
• Ensure human-in-the-loop for checking model hallucinations.

Iterative: Prompt development to interact with information-loaded LLMs functions in a similar way to machine learning models: idea -> implement -> iterate.

Best Practices:

• If the provided text is too long, ask it to shorten.
• If the LLM output focuses on the wrong details, get it back on track.
• Ask it to extract information and organize it in a table based on the use case.

Summarizing:

Best Practices:

• Ask the LLM to summarize the text.
• Provide a word or character limit.
• Ask it to focus on a topic within the provided text input.
• If the output introduces an unrelated topic, ask LLMs to extract instead of summarizing.

Inferring:

LLMs can identify sentiment and emotions from a text.

Best Practices:

• Ask LLMs to identify specific sentiments.
• Extract specific information, such as names of persons mentioned in the text.
• Ask LLMs to multitask, accomplishing multiple tasks at the same time.
• Ask LLMs to infer topic titles based on the text.

I used the inferring functionality in my personal podcast app.

Transforming:

Translates from one language to another. LLM can also act as a universal translator at once, translating sentences from several languages into one. LLMs can produce and change tones. They can also change text formats from JSON to HTML, etc.

Expanding:

LLMs can integrate information from two different texts.