OpenAI Cookbook - Note

Better Prompts

Reference: Techniques to improve reliability.

• Tricks work: Lets think step-by-step , You are an expert

• Give clearer instructions and context

• Split complex tasks into simpler subtasks

• *Structure the instruction to keep the model on task

• *Ask for justifications of many possible answers, and then synthesize

• *Generate many outputs, and then use the model to pick the best one

• Prompt the model to explain before answering

  • Zero-shot

    Using Let's think step by step. This works well for multi-step arithmetic problems, symbolic reasoning problems, strategy problems, and other reasoning problems.

    (Reference: 2022 Large Language Models are Zero-Shot Reasoners)

  • Few-shot example-based

    Q: xxx A: xxx Q: xxx

    This works for math problems, questions related to sports understanding, coin flip tracking, and last letter concatenation. Examples ≤ 8 shall be enough.

    (Reference: 2022 Chain of Thought Prompting Elicits Reasoning in Large Language Models)

• Fine-tune custom models to maximize performance

  • To meet the requirement of thousands of examples, costly to write, using few-shot prompt to generate candidate explanations — Self-taught Reasoner (STaR).

    (Reference: STaR: Bootstrapping Reasoning With Reasoning.)

Extensions to chain-of-thought prompting

• Selection-inference prompting Reference: 2022 Selection-Inference: Exploiting Large Language Models for Interpretable Logical Reasoning
• Faithful reasoning architecture Reference: 2022 Faithful Reasoning Using Large Language Models
• Least-to-most prompting Reference: 2022 Least-to-most Prompting Enables Complex Reasoning in Large Language Models

Others

• Maieutic prompting Reference: 2022 Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations

More

• Self-consistency Reference: 2022 Self-Consistency Improves Chain of Thought Reasoning in Language Models
• Verifiers Reference: 2022 TrainingVerifiers to Solve Math World Problems

Theories of reliability

• Probabilistic graphical models Reference: 2022 Language Model Cascades

Takeaways

Bibliography from the article

Prompt examples

• Instruction
• Completion
• Scenario
• Demonstration (few-shot learning)
• Fine-tuned. This means no need to add instruction, where special characters like ‘###’ or ‘→’ would tell the model what to do next.

Reference: How to work with large language models.

/Prompt/ /OpenAI/ /GPT/ /Note/