Prompt Engineering

Two types of LLMs

• Base LLM
  • predicts next word based on text training data
• Instruction tuned LLM
  • fine-tune on instructions and good attempts at following instructions
  • e.g.Reinforcement learning with human feedback (RLHF)

Prompting Principles

Principle 1: Write clear and specific instructions

• Tactic 1: Use delimiters to clearly indicate distinct parts of the input
  • Delimiters can be anything like: ```, """, < >, , :
• Tactic 2: Ask for a structured output (JSON, HTML, etc)
• Tactic 3: Ask the model to check whether conditions are satisfied
• Tactic 4: "Few-shot" prompting
  • provide desired output before giving the prompt

Principle 2: Give the model time to “think”

• Tactic 1: Specify the steps required to complete a task
  • step 1: ...
  • step 2: ...
• Tactic 2: Instruct the model to work out its own solution before rushing to a conclusion

In-context learning/prompting methods

• zero-shot prompting: creating an initial prompt that states the task to be completed but not includes examples
• one/few shot prompting: creating an initial prompt that states the task to be completed and includes a single/few of example question with answer
• chain-of-thought prompting: creating a prompt that demonstrates how to solve similar problems using step-by-step reasoning

Iterative Prompt Development

Capabilities of LLM

Summarizing

• Summarizing
• Extracting

Inferring

examples:

• “What is the sentiment of the following …”
• “Determine five topics that are being discussed in the following test…”

Transforming

• Translation
• Tone Transformation
• File Format Conversion
• Spellcheck/Grammar Check

Expanding

= generate long piece of text from short piece of text

Temperature

= degree of randomness of the model (see Generative Model Configuration#^1ce105)

• use temperature 0 for tasks that require reliability, predictability
• use temperature >0 for tasks that require variety