Overview
ðããã³ãããšã³ãžãã¢ãªã³ã°ã®ðšPATTERN.
ð¥Glossary
Shot/Zero-shot/Few-shot
Shotã¯äŸç€º, æšè«éçš, ãã¢ã³ã¹ãã¬ãŒã·ã§ã³. ã¢ãŠããããã®å ·äœäŸ. ãããªãã®ãã»ããã¿ãããª.
Zero-shotã¯äŸç€ºãªã. Few-shotã¯å°æ°ã®äŸç€ºãã.
Input Semantics Pattern
LLMãžã®å ¥åãšããã倿ããŠåºåãçæãããã¿ãŒã³
Output Customization Pattern
ð£Persona Pattern
ããªã㯠xx ã§ã, å°éå®¶ãšããŠâŠ
ð£Template Pattern
åºåãã©ãŒãããæå®.
cf. ðšTemplate Method: ç¶æ¿ã«ããIFã®èŠå®.
Prompt Improvement Pattern
åºåã®è³ªãæ¹å.
Interaction Pattern
UserãšLLMã®ãããšããã¿ãŒã³.
Flipped Interaction Pattern
LLMã«é質åããŠããã.
Context Control Pattern
LLMãæ±ãã³ã³ããã¹ãæ å ±ã«é¢ããã«ããŽãªãŒ.
ããã³ãããã¯ããã¯ãã¿ãŒã³
ð£Chain-of-Thought(CoT)
æèé£é.
è€éãªã¿ã¹ã¯ãæçµçãªè§£æ±ºã«åããè«ççãªã¹ãããã®é£ç¶ã«åºåãããšã§ã人éã®ãããªæšè«ããã»ã¹ãã·ãã¥ã¬ãŒããããã®.
ref. https://arxiv.org/abs/2201.11903
ð£Zero-shot CoT
ãããã, ð€æ®µéçã«èããŠ(Step by Step)ãšãã, ãã®æãå ããã ãã§æšè«èœåãäžãããšããéæ³ã®èšèãšããŠç޹ä»ããããã€ã®çè«ããã.
äžéçãªæšè«ã¹ããããä»ããããšã§è€éãªæšè«ãå¯èœã«ãªã.
ð£Self-Consistency
QAã®æ£çäŸãããã€ãäžããããšã§äºåã«åŠç¿ãããŠããåé¡ãè§£ããããã¯ããã¯.
ðGenerate Knowledge Prompting
æç€ºã®åã«ç¥èãçæãããŠããåé¡ãè§£ããããã¯ããã¯.
- Aã«ã€ããŠèª¬æããŠãã ãã.
- Bã«ã€ããŠèª¬æããŠãã ãã.
- AãšBãã€ãã£ãŠâŠ
Least-to-Most Prompting
decomposite and recomposite.
ããã³ãããã¹ããã©ã¯ãã£ã¹
ref. ðBest practice for prompt engineering with OpenAI API
ðããã³ããã¯ã·ã³ãã«ããã¯ãããŠæ¹åãã
ããã³ããã¯æšè«ãç¹°ãè¿ãããšã§è€éãªãã®ãæšè«ã§ãã.
Start with zero-shot, then few-shot (example), neither of them worked, then fine-tune.
ðZero-shot Prompting(äŸç€ºãªãã®ããã³ãã)ããã¯ãããŠ, ðFew-shot Prompting(ããã€ãäŸãäžããããã³ãã)ããã£ãŠ, ããã§ãããŸãè¡ããªããã°æ¹åããŠãã.
ðæç€ºãšã³ã³ããã¹ããåºåã
æç€ºã¯ããã³ããã®æåã«çœ®ã, æèãšåºåããããããšç²ŸåºŠãäžãã.
ã"""ããæšå¥šãããŠãã. """ã§åºåãããšã¯äžã¯Markdownã®æžåŒã§ããã.
Summarize the text below as a bullet point list of the most important points.
Text: """ {text input here} """
ðBest practice for prompt engineering with OpenAI API
ðãããªãããšããããããããšã
xxããŠã¯ãããªããšããå¶çŽæ¡ä»¶ã§ã¯ãªã, å ·äœçãªæç€ºãäžãã.
ãã£ãŠã»ãããªãããšäŒããã ãã§ãªã, 代ããã«äœããã¹ãããäŒãã.
ããã³ããäºäŸ
ð€ããã³ããäºäŸéãž.
ðReferences
- ãChatGPTãããã³ãããã¿ãŒã³ãŸãšã - Qiita
- ããã³ãããšã³ãžãã¢ãªã³ã°å¿çšç·š
- 500+ Best ChatGPT Prompts
- Prompt Engineering ååŒ·äŒ / 2023.03.21 GPT-4 Prompt å ±åäŒ - Speaker Deck
ðBest practice for prompt engineering with OpenAI API
OpenAI APIã®ããã³ãããšã³ãžãã¢ãªã³ã°å ¬åŒãã¹ããã©ã¯ãã£ã¹.
- Best practices for prompt engineering with OpenAI API | OpenAI Help Center
- ChatGPTããã³ãããšã³ãžãã¢ãªã³ã°ã®ã³ã8ç®æ¡ïœOpenAIå ¬åŒã®ãã¹ããã©ã¯ãã£ã¹ããåŠã¶ïœ - Qiita
ðPrompt Engineering Guide - DAIR.AI
- https://github.com/dair-ai/Prompt-Engineering-Guide
- elvis(@omarsar0)ãšãã人
- æ¥æ¬èª: Prompt Engineering Guide
ãã®DAIR.AIãðfast.aiã®ãããªç念ãæã£ãŠããã®ããã.
Democratizing Artificial Intelligence Research, Education, and Technologies.
https://github.com/dair-ai/ML-YouTube-Courses
ðPrompt Engineering for ChatGPT - Jules White, Coursera
ããã³ãããšã³ãžãã¢ãªã³ã°ãã¿ãŒã³ã®è«æè§£èª¬.
- https://arxiv.org/abs/2302.11382
- Prompt Engineering for ChatGPT Course (Vanderbilt) | Coursera
- ããã³ãããã¿ãŒã³ãåŠã¶ â 西éç«å€ªéãŠã§ããµã€ã
å人çã«ã¯ã¢ãŒããã¯ãã£ãã¿ãŒã³POSAã®Douglas C. Schmidtãããè«æã®æ«å°Ÿã«ååãé£ããŠããã®ãç±ã. ããã³ãããšã³ãžãã¢ãªã³ã°ãGoFãã¶ã€ã³ãã¿ãŒã³ãšçµã³ã€ãã. ðã¢ãŒããã¯ãã£ãã¿ãŒã³(POSA)
ðAccelerate Your Learning with ChatGPT - Dr. Jules White/Barbara Oakley, coursera, å匷ã«å¿çš.
ðRelated
ãªããðãã¶ã€ã³ãã¿ãŒã³ã¿ãã -> ðã¢ãŒããã¯ãã£ãã¿ãŒã³(POSA)ã®äººãé 匵ã£ãŠãã¿ãŒã³ã«ãŸãšããŠã. ðPrompt Engineering for ChatGPT - Jules White, Coursera