Prompt Engineering Techniques

Flashcards based on lecture notes covering various prompt engineering techniques for language models.

Direct Prompting

Simple, clear instructions without role-setting.

Direct Prompting Reference

Original GPT-2 paper, Radford et al., 2019

Few-Shot Prompting

Providing a few examples to guide the model.

Few-Shot Prompting Reference

Brown et al., 2020, Language Models are Few-Shot Learners

Details of Few-Shot Prompting

expected response format, context, and style, adaptability to complex and diverse tasks (without task-specific training data)

Zero-shot Prompting

Asking the model to perform a task with no examples; more subjective task than direct prompting.

Zero-shot Prompting Reference

Brown et al., 2020

Role-Based Prompting

Instructing the model to 'act as' a specific persona.

Role-Based Prompting Reference

Usage began in GPT-3 applications, Brown et al., 2020

Step-by-step Prompting

Breaking down a complex task into simpler, ordered instructions, guiding the model to complete each part sequentially.

Approach of Step-by-step Prompting

Explicitly structure each part of the task to make it easier for the model to handle in smaller pieces, often directing the model to complete one step before moving on to the next.

Typical Use Case of Step-by-step Prompting

Procedural tasks where each step is clearly defined

Chain of Thought (CoT) Prompting

Guiding the model to explain its reasoning step-by-step.

Chain of Thought (CoT) Prompting Reference

Wei et al., 2022

Benefits of CoT

reasoning and math problems

PROGRAM-OF-THOUGHTS

generate python code, delegate computations

CHAIN-OF-SYMBOL

formulas over natural language

STRUCTURED CHAIN-OF-THOUGHT

bullet points, numbered lists, or labeled sections

TREE-OF-THOUGHTS

allow branches, then do evaluation and selection

Instruction Tuning

Fine-tuning on diverse prompts phrased as instructions; reduces need for prompt engineering.

Instruction Tuning Reference

Chung et al., 2022, 'Scaling Instruction-Finetuned Language Models'

Self-Consistency Prompting

Asking the model the same question multiple times for reliability.

Self-Consistency Prompting Reference

Wang et al., 2022, 'Self-Consistency Improves Chain of Thought Reasoning'

ReACT (Reasoning + Acting) Prompting

Combining reasoning and action steps within prompts.

ReACT (Reasoning + Acting) Prompting Reference

Yao et al., 2022, 'ReAct: Synergizing Reasoning and Acting'

Reasoning Step (ReACT)

Logically thinking through the task, breaking it down and analyzing each part to generate a structured thought process.

Action Step (ReACT)

Based on the reasoning, the model performs a specific action, such as providing an answer, making a choice, or executing a command.

Reinforcement Learning from Human Feedback (RLHF)

Using human feedback to refine model responses; important for alignment.

Reinforcement Learning from Human Feedback (RLHF) Reference

Ouyang et al., 2022, 'Training Language Models to Follow Instructions with Human Feedback'

Alignment

Intent/Goal/Purpose Matching, Ethical and Safety Standards, Contextual Appropriateness

Example Prompt w/ Structured Input

Please provide details of the event in the following format: { 'Event' : 'Event Name' , 'Date' : 'Event Date' , 'Location' : 'Event Location' }

Formats for Structured Input

Key-Value CSV, JSON, XML, RDF

Data Augmentation with Synthetic Prompting

Generating synthetic data to improve generalization

Data Augmentation with Synthetic Prompting Reference

Gao et al., 2021, 'Making Pre-trained Language Models Better Few-shot Learners'

Synthetic Prompts Example

List some advantages of regular physical activity, What positive effects does exercise have on health?, Why is exercising good for you?

Active Prompt Adjustment

User dynamically refining prompts based on responses

Error Identification and Correction Prompting

Asking the model to identify and correct errors.

Error Identification and Correction Prompting Reference

PaLM models, Chowdhery et al., 2022

Confidence Scoring

PaLM uses confidence scoring mechanisms to gauge its certainty about a given response

Step-by-Step Verification

For multi-step reasoning tasks, PaLM can evaluate each step’s validity before proceeding to the next.

Redundant Answer Generation

PaLM sometimes generates multiple responses for the same query and evaluates these responses for consistency.