Quiz 3 prep 534

What are the three domains of using transformer

prompt engineering

retrieval augmentation generation (RAG)

reasoning product

Zero-shot learning (sentiment classification)

no examples provided to model through prompt only relies on internal information

• template is tried for all possible class labels

• in context learning that does not include any exemplars of tasks

few-shot learning (machine translation)

some examples are provided to model w/prompt

• no instruction provided just examples given the template

• in context learning that contains several exemplar tasks

What are different templates called

verbalizers

this is how we construct

emergence

quantitative changes in system result in qualitative changes in behavior

emergent behaviors

abilities that larger models have and smaller don’t

in-context learning

language models “learns” to do task from textual prompt containing natural language instruction for task, several exemplars of tasks or both

prompt engineering

painstaking process of trying out many different prompts until you find one that works well for task

verbalizer

template we wrap an example in inorder to perform tasks

perplexity

how fast you can locate information in data

chain of thought prompting

each of the exemplars in your few-shot prompt contains logic showing how to solve the tasks

zero-shot chain of thought prompting

we don’t need any exemplars. just append the string “lets think step-by-step” to end of prompt

chain of thought

prompting enables LLMs to generate intermediate reasoning steps before inferring an answer

Type paradigm-1 : zeroshot CoT

without a trigger hint

paradigm 2: few shot CoT (manual CoT)

in context learning method by demonstrating step-by-step reasoning exemplars (demonstrations)

Auto-cot

not scalable

eliminates need for manual designed input

maintain strong performance

key challenges of auto cot

how to obtain representative questions to reflect task patterns

how to obtain rationales to construct demonstrations

frequent error cluster

clustering using k-means to partition all test questions into k-clusters

find frequent error clusters

diversity: higher chance to obtain good demos that is not too heavily pertubated

Draw backs of LLms

hallucination

outdated info

low efficiency in parameterizing knowledge

lack of in depth knowledge in specialized domains

weak inferential capabilities

Retrieval augmented generation

• first retrieves relevant information then LLMs generate answers based on the info

• by attaching a external knowledge base there is no need to retrain entire large model for each specific task

• RAG model is especially suitable for knowledge intensive task

Rag knowledge update

directly updates

RAG external knowledge

access documentation and other databases

RAG data processing

minimal data processing and handling

RAG model customization

not fully customize model behavior or writing style

RAG interpretibility

traced back to specific data structure

RAG computational resources

support retrieval strats and tech related to database

RAG latency requirements

higher latency

RAG redcuing halluctination

less prone to hallucination

RAG ethical and privacy issues

based on external database

Naive RAG

Indexing

retrieval

generation

advanced RAG

index —> optimization —> pre retrieval process —> retrieval —> post retrieval process —> generation