کتاب مهندسی پرامپت برای مدل‌های زبانی بزرگ

مقدمه‌ای بر مهندسی پرامپت

LLMs Are Magic • Language Models: How Did We Get Here? • Early Language Models • GPT Enters the Scene • Prompt Engineering • Conclusion

درک مدل‌های زبانی بزرگ

What Are LLMs? • Completing a Document • Human Thought Versus LLM Processing • Hallucinations • How LLMs See the World • Difference 1: LLMs Use Deterministic Tokenizers • Difference 2: LLMs Can’t Slow Down and Examine Letters • Difference 3: LLMs See Text Differently • Counting Tokens • One Token at a Time • Auto-Regressive Models • Patterns and Repetitions • Temperature and Probabilities • The Transformer Architecture • Conclusion

حرکت به سمت چت

Reinforcement Learning from Human Feedback • The Process of Building an RLHF Model • Keeping LLMs Honest • Avoiding Idiosyncratic Behavior • RLHF Packs a Lot of Bang for the Buck • Beware of the Alignment Tax • Moving from Instruct to Chat • Instruct Models • Chat Models • The Changing API • Chat Completion API • Comparing Chat with Completion • Moving Beyond Chat to Tools • Prompt Engineering as Playwriting • Conclusion

طراحی برنامه‌های کاربردی مبتنی بر مدل‌های زبانی

The Anatomy of the Loop • The User’s Problem • Converting the User’s Problem to the Model Domain • Using the LLM to Complete the Prompt • Transforming Back to User Domain • Zooming In to the Feedforward Pass • Building the Basic Feedforward Pass • Exploring the Complexity of the Loop • Evaluating LLM Application Quality • Offline Evaluation • Online Evaluation • Conclusion

محتوای پرامپت

Sources of Content • Static Content • Clarifying Your Question • Few-Shot Prompting • Dynamic Content • Finding Dynamic Context • Retrieval-Augmented Generation • Summarization • Conclusion

اسمبل کردن پرامپت

Anatomy of the Ideal Prompt • What Kind of Document? • The Advice Conversation • The Analytic Report • The Structured Document • Formatting Snippets • More on Inertness • Formatting Few-Shot Examples • Elastic Snippets • Relationships Among Prompt Elements • Position • Importance • Dependency • Putting It All Together • Conclusion

رام کردن مدل

Anatomy of the Ideal Completion • The Preamble • Recognizable Start and End • Postscript • Beyond the Text: Logprobs • How Good Is the Completion? • LLMs for Classification • Critical Points in the Prompt • Choosing the Model • Conclusion

عامل‌های گفتگو

Tool Usage • LLMs Trained for Tool Usage • Guidelines for Tool Definitions • Reasoning • Chain of Thought • ReAct: Iterative Reasoning and Action • Beyond ReAct • Context for Task-Based Interactions • Sources for Context • Selecting and Organizing Context • Building a Conversational Agent • Managing Conversations • User Experience • Conclusion

گردش‌کارهای مدل‌های زبانی

Would a Conversational Agent Suffice? • Basic LLM Workflows • Tasks • Assembling the Workflow • Example Workflow: Shopify Plug-in Marketing • Advanced LLM Workflows • Allowing an LLM Agent to Drive the Workflow • Stateful Task Agents • Roles and Delegation • Conclusion

ارزیابی برنامه‌های کاربردی مدل‌های زبانی

What Are We Even Testing? • Offline Evaluation • Example Suites • Finding Samples • Evaluating Solutions • SOMA Assessment • Online Evaluation • A/B Testing • Metrics • Conclusion

نگاهی به آینده

Multimodality • User Experience and User Interface • Intelligence • Conclusion