Notes of ChatGPT Prompt Engineering for Developers

For ease of referring and checking in the future, I summarized this outline for the short course ChatGPT Prompt Engineering for Developers at DeepLearning.AI.

Introduction

Introducing the general background of natural language process (NLP) with deep learning (DL).

• Large language models (LLMs) [I think here the term refers to GPT-like LLMs] were categorized into two types.

• Base LLM

• • Predicts the next word based on the training data

• Instruction Tuned LLM

• • • Tries to follow instructions (Fine-tuning based on instructions and making good attempts at following those instructions)
    • RLHF: Reinforcement Learning with Human Feedback (arXiv paper)

  • Helpful, Honest, Harmless [at least we hope so]

• This short course comes with examples and practices on Jupyter Notebooks. I don’t think we can download them. So I organized and kept them in my GitHub repository.

Guidelines

Introducing two prompting principles and how to write effective prompts.

• Principle 1: Write clear and specific instructions

• • As clear and specific as possible.
  • You don’t have to write short prompts. In many cases, longer prompts can provide more details to general, more relevant outputs.
  • Skill 1: Use delimiters to clearly indicate distinct parts of the input (there is no limitation of delimiters, any punctuations that can separate the sentences will work). This can effectively reduce constructive instructions in the prompts.
  • Skill 2: Ask for structured output (like HTML, JSON, etc.)
  • Skill 3: Check whether conditions are satisfied. Check the assumptions required to do the task. (adding prompts to improve the consistency of outputs)
  • Few shot prompting. Give successful examples of completing tasks. Then ask the model to perform the task.

• Principle 2: Give the model time to “think.”

• • Skill 1: Specify the steps to complete a task. [step by step instructing]
  • Skill 2: Instruct the model to work out its own solution before rushing to a conclusion. [ask the model to do step-by-step outputs.]

• Model Limitations: Hallucination: make statements that sound plausible but are not true. To address it, the suggestion is to find relevant information and then answer the question based on the relevant information. [However, personally, I think this might be too vague sometimes.]

• Iterative: This is originally the third class. However, I think it is more like a trick or concept that can be applied to any use case. In plain language, we can continue updating and refining the prompts/instructions based on the outputs from the model. In the end, we can build prompts to let the model generate outputs closest to our expectations.

Applications

• Summarizing: summarize text (with multiple objects) with a focus on specific topics. Similarly, you can “extract” information (named entity recognition) of multiple objects from the input text.

• Inferring: infer sentiment and topics from product reviews and news articles, such as sentiment of the text, what types of emotions are reflected from the text, and extraction of objects ( similar to “extract” above).

• Transforming: language transformation tasks, such as language translation, spelling and grammar checking, tone adjustment, and format conversion.

• Expanding: generate text based on prompts specifying the requirement.

• Chatbot: utilize the chat format to have extended conversations with chatbots personalized or specialized for specific tasks or behaviors.

• • ‘system’ role (high-level) can send ‘content’ to set the behavior of the assistant (chat model) to provide answers to the user (you) based on the prompts. System messages provide developers with a way to frame the conversation without making the requests a part of the conversation.

Conclusion

LLMs are very powerful. Please use them responsibly. Please only build things that will have a positive impact.