What is a Prompt?

A prompt is essentially the input or query directed at an LLM to generate a specific response. Prompts can range from simple sentences or questions to complex code snippets, depending on the context. They can also be interconnected, meaning that the output from one prompt can serve as the input for another, facilitating dynamic and intricate interactions with the model.

Prompt Engineering, also referred to as Prompt Design, involves techniques for effectively communicating with LLMs to steer their responses toward desired outcomes, all without modifying the model's weights. This is an experimental process, as the efficacy of different prompt design strategies can vary by model, necessitating considerable experimentation and experience.

For instance, instead of a basic prompt like "write about artificial intelligence," enhancing it with more context, such as "compose a blog post highlighting the applications of artificial intelligence in education with specific examples," can yield improved results.

When Should You Employ Prompts?

Prompt Engineering versus Fine-Tuning

When utilizing a language model for specific tasks, one might consider fine-tuning the model using a training dataset. Fine-tuning is a viable approach, but in scenarios with limited data and computational resources, prompt engineering can be the preferred choice for closely managing model responses.

Once sufficient data is gathered, you may opt for fine-tuning the model or explore a combination of both methods for optimal performance.

Effective Prompt Crafting Techniques

Here are several strategies for designing prompts specifically for OpenAI's GPT-3.5 and GPT-4 models. You can employ one or a blend of techniques based on your needs and challenges. Generally, a prompt may encompass the following components (though not all are required):

1. Instruction Prompting

   One straightforward approach is to clearly instruct the LLM on the desired action or stipulate rules and guidelines that need to be followed.

Example:

• Translate the following sentence into Japanese: 'I love programming.'
• Summarize the paragraph below: {INSERT PARAGRAPH}
• Please answer the following question honestly. If uncertain, say 'I don't know.' {INSERT QUESTION}

Modern AI can adhere to complex instructions, particularly with models trained using Reinforcement Learning from Human Feedback (RLHF) like ChatGPT.

• Ensure instructions are clear, specific, and comprehensible.

  • Repeat key instructions to reinforce understanding.
  • Favor affirmative prompts over negative.
  • Utilize markup (brackets, quotes, bullet points, etc.) to delineate requests.
  • Present instructions in a logical, step-by-step manner.
  • Describe the desired output format in detail, including tone, writing style, and length, along with examples.

2. Role Prompting

   Assigning a role to the AI can enhance context and specificity. For instance, starting with "You are a doctor" or "Act as a prominent lawyer" can shape the AI's responses accordingly.

Example:

• You are a helpful assistant chatbot on a mission to provide information about "The Boring Company."

3. Providing Examples (In-Context Learning)

   While LLMs can deliver accurate responses through zero-shot learning, inaccuracies can arise due to task complexity or insufficient context. Providing examples can improve response accuracy.

Few-shot learning involves presenting quality examples of the desired task, with each example clearly illustrating both input and expected output. This method often outperforms zero-shot learning.

• Consider the structure and format of examples carefully.

  • Address potential biases in example labeling.
  • Use diverse examples arranged randomly to enhance model understanding.

4. Chain-of-Thought

   This method involves prompting the model to articulate its thought process step-by-step. For example, guiding the model to think through a problem logically can lead to more accurate outcomes.

5. Parameter Tuning

   Adjusting parameters can fine-tune model output.

• Set temperature=0 for consistent results.

  • Generate multiple outputs using varied parameters and select the best.
  • Control the output length with max_tokens.

Experimentation is key to finding the right parameter settings for your objectives.

Common Pitfalls in LLM Usage and How to Avoid Them

While LLMs, especially ChatGPT, are powerful tools, they also have inherent limitations that can lead to issues if not addressed.

1. Hallucinations: This is when the model fabricates answers or provides incorrect information.
   • Solution: Incorporate instructions like "answer truthfully based on the following information" and encourage the model to critique its responses.
2. Bias and Inappropriate Content: Given the vast training data, models may produce biased outputs.
   • Solution: Include guidelines in prompts to ensure equality and avoid stereotypes.
3. Outdated Information: LLMs may lack current or specialized knowledge.
   • Solution: Integrate databases or online search tools to provide context.