Generating Effective Prompts for Large Language Models: A Guide

With the launch of LLMs like ChatGPT, we've witnessed a wave of chatbots that can really connect with people and deliver some surprising results. These AI tools have been a boon for many, whether it's entrepreneurs venturing into AI-based ventures or students getting a helping hand with their assignments.

This technology is undoubtedly impressive, yet it often produces false positives, leading to the generation of fake or meaningless data. To avoid such issues and ensure the generation of relevant and accurate results, several strategies can be implemented when utilizing the Large Language Models (LLMs).

1. Study the model

When working with a specific model, it is important to verify its specifications, such as determining whether it is designed for single or multi-churn tasks, and assessing its token capacity per prompt. Additionally, referring the model's guidelines like of Gemini and ChatGPT can streamline the process of achieving the intended outcomes.

2. Structure, Clarity and Detail

Making the instructions clear and detailed can help boost the AI’s performance. For example, rather than just writing “ Help me edit the below resume”. You could give more context and be more specific like “ You are a hiring manager who is expert in editing resumes. Help me edit the below resume for a Machine Learning Job based on the given job description: (Resume:{}, Job Description:{})”. Try doing the above with?a structure of:

{

Assigning a role to the model, if needed,

Providing clear and detailed instructions,

Providing examples with labels (like paragraph: {} , summary: {}) if needed.

}

3. Provide examples

When engaging in specific tasks such as summarization, domain-specific question answering, or sentiment detection, it's advantageous to include examples in the prompt to clarify the desired outcome. This practice is commonly referred to as one-shot learning or few-shot learning. If this method proves insufficient and more precise results are required, it's advisable to perform PEFT fine-tuning or full fine-tuning ( potential for catastrophic forgetting), depending on the task's requirements.

4. Understand the response

Many AI chatbots, such as Blackbox.ai and ChatGPT, typically retain conversation histories. However, there are instances where this can lead to repetitive incorrect outcomes. In such cases, it's crucial to either offer feedback or stop the conversation and start a fresh one. For example, if the desired results are not achieved, you can provide a feedback, "The above result has a lot of bullet points. Please reduce the bullet points to 3 for each section in the above resume." Alternatively, you could opt to end the current conversation and approach the prompt with a different perspective.

5. Modify parameters

Digging a little deeper, when utilizing an LLM API or Playground (such as the one offered for ChatGPT), we have the option to adjust various parameters to tailor the output to our preferences. These adjustable parameters include temperature, top p, top k, max tokens, stop sequence, frequency penalty, and presence penalty. For instance, temperature is a parameter that influences the randomness of the generated content. Higher temperature values result in greater randomness in the output, and this setting can be adjusted based on the specific task at hand.

You can read about these parameters in the below article: https://www.dhirubhai.net/pulse/understanding-prompt-parameters-enhanced-performance-llms-chhagani-vfv4c/?trackingId=MAOKI0H%2BQcayWLawzM2yDw%3D%3D

Personal experience:

Here's my personal experience dealing with these AI models as I have tried to craft the perfect prompt. I have experimented with generating data, writing reports, organizing schedules, and much more. One thing I have learned is that the AI doesn't always hit the mark, so it's up to us to try different approaches, offering feedback and examples along the way to guide it toward the desired outcome. Sometimes, these models have tendencies of generating a lot of irrelevant data, but including instructions like "Please avoid false positives" can help trim down the excess. Secondly, Providing relevant data from the specific domain can lead the AI to understand the domain better and mold better results for the given task.

(Illegal tip: Ask the AI to "make the above text more humanlike," and it will polish up the text.)

Conclusion:

The quality of data and the effectiveness of prompts play crucial roles in model training and performance. A well-crafted prompt can steer the model towards desired outputs, while a poorly constructed one may lead to ambiguous or incorrect results. This article aims to provide insights and best practices for generating effective prompts to enhance the performance of the models.

By following guidelines such as clearly defining tasks, providing context, using descriptive language, incorporating examples, experimenting with different formats, and incorporating feedback loops, we can create better prompts.

The upcoming series of articles will cover cost reduction, optimizing prompt engineering, fine-tuning and much more. Stay tuned for further updates :)

Happy reading ?? . For more such articles subscribe to my newsletter: https://lnkd.in/guERC6Qw

I would love to connect with you on Twitter: @MahimaChhagani. Feel free to contact me via email at [email protected] for any inquiries or collaboration opportunities.