Prompt Engineering

[manisnesan]

Label: Current

• [ ] Introduction
• [ ] Guidelines
• [ ] Iterative

ChatGPT Prompt Engineering - Course

Related

• [ ] Prompt Engineering Guide from dair.ai
• [ ] Anthropic Prompt Engineering Interactive Tutorial - Google Sheets
• [ ] Brexhq - Prompt Engineering
• [x] Cohere Prompt Engineering Guide
• [x] https://hamel.dev/notes/prompt-eng/course/01_guidelines.html

[manisnesan]

Introduction

• Base LLM

• Instruction tuned LLM - focus of the course

• Takeaway: Be clear about not only the instructions but also with what sort of output and tonality of the output

[manisnesan]

Prompting principles

• Write clear and specific instructions
• Give the model time to think

Tactics

• Use delimiters to clearly indicate parts of the input ( this may hopefully prevent Prompt injections)
• ask for structured output in json, HTML
• ask the model to check if the conditions are satisfied. Check assumption required to do the task
• few shot prompting
• specify the steps to complete the task
  • asking for output in specified format from each steps in addition to providing the steps
• instruct the model to work out its own solution before reaching to its conclusion

Model Limitations

• Hallucinations: Makes up responses that sound plausible but not actually correct
• Solution : First find relevant information then answer the question based on relevant information

[manisnesan]

https://platform.openai.com/docs/guides/gpt-best-practices/six-strategies-for-getting-better-results

[manisnesan]

Let’s first understand the problem and devise a plan to solve the problem. Then, let’s carry out the plan and solve the problem step by step”

Feels like plan-to-solve prompting got eclipsed by chain-of-thought, but planning worked better in this paper. arxiv.org/pdf/2305.04091

via Ethan Mollick

[manisnesan]

Via MindBranches

Refer: Pg 5 https://arxiv.org/pdf/2312.16171v2

[manisnesan]

via Mindbranches

[manisnesan]

https://winder.ai/llm-prompt-best-practices-large-context-windows/

Piecewise document summarization: For summarizing long documents, break them into sections and summarize each one separately. Then, combine these summaries for an overall view. This recursive approach can also include a running summary to maintain context throughout the document.

[manisnesan]

Taxonomy of Prompt Engineering

Great paper summarizing the prompt techniques - The Prompt Report.

• 58 text-only prompting techniques including zero-shot, few-shot, thought generation, ensembling, self-criticsm, and decomposition techniques. Few-shot CoT performs the best among the few techniques they benchmarked
• Prompt Engineering: Meta prompting, auto prompt, automatic prompt engineer, gradient free instructional prompt search, prompt optimization with textual gradients, RLPrompt, dialogue-comprised policy-gradient-based discrete prompt optimization
• Answer Engineering: Answer shape, answer space, answer extractor
• Agents: tool use agents, code-based agents, observation-based agents, RAG
• Security: prompt hacking (prompt injection, jailbreaking), risks (data privacy, code generation concerns, customer service), hardening measures ( prompt-based defense, guardrails, detectors).
• Alignment: ambiguity, biases, calibration, prompt sensitivity

🔗arxiv.org/abs/2406.06608

via Sophie Yang - X post