AWS AI Practitioner Flashcards

1
Q

Portion of training training data is labeled and feedback is provided in the form of rewards or penalties. What type of learning

A

Reinforcement learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
2
Q

What are the two types of inferencing?

A

Batch and Real time

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
3
Q

Deep learning is used in which use 2 cases?

A

Computer vision and NLP

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
4
Q

What are FMs in generative AI?

A

Pretrained models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
5
Q

What are Transformer models?

A

Builds encoder decoder concept in genAI. They use self-attention to process input data. Self-attention allows the model to weigh the importance of different words in a sentence when encoding a particular word

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
6
Q

Are FMs pre trained using reinforced learning? True or False?

A

False. FMs are typically pre-trained through self-supervised learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
7
Q

Where is pre-text tasks used for?

A

In self-supervised learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
8
Q

Self-supervised learning makes use of the structure within the data to autogenerate labels. True or False?

A

True

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
9
Q

Optimization of pre trained FMs are done using what?

A

Prompt engineering,
Retrieval-augmented generation (RAG),
Fine-tuning on task-specific data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
10
Q

LLMs, Diffusion and Multiodel models are what?

A

These are FM models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
11
Q

These are numerical representations of tokens, where each token is assigned a vector (a list of numbers) that captures its meaning and relationships with other tokens?

A

Embeddings

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
12
Q

What is a context window?

A

The maximum number of tokens a LLM model can take when generating text

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
13
Q

What is a vector?

A

It is an array of numercial values

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
14
Q

What is the process of vectorization?

A

Text -> [Tokenization]->Tokens -> [Embeddings Model] -> Vectors

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
15
Q

What is the process of vectorization in Bedrock KBs using RAG?

A

Customer KB->[Upload in Amazon S3]->[Select a vector DB]->[Select a Model]->[Sync with customer KB]->Vectorization of Customer KB text

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
16
Q

What is Watermark detection for Amazon Bedrock?

A

Identify images generated by Amazon Titan Image Generator, a foundation model that allows users to create realistic, studio-quality images in large volumes and at low cost, using natural language prompts

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
17
Q

What is continued pretraining in Amazon Bedrock?

A

You provide unlabeled data to pre-train a model by familiarizing it with certain types of inputs

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
18
Q

Which is the models which gradually add more and more meaningful information to this noise until they end up with a clear and coherent output, like an image or a piece of text?

A

Diffusion model

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
19
Q

Which model has generator and discriminator?

A

Generative adversarial networks

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
20
Q

Which model has encoders and decoders?

A

Varional autoencoders

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
21
Q

What are the components of prompt engineering

A

Instructions, Context, Input data and Output indicator

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
22
Q

What are non-determistic LLMs popularly called?

A

Generative Language Models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
23
Q

What is a supervised learning process that involves taking a pre-trained model and adding specific, smaller datasets?

A

Fine tuning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
24
Q

Two types of fine tuning

A

Instruction fine-tuning and Reinforcement learning from human feedback (RLHF)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
25
Q

Fine tuning does it add weight to the data?

A

Yes

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
26
Q

What is Retrieval-augmented generation (RAG)?

A

Supplies domain-relevant data as context to produce responses based on that data.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
27
Q

To create fine tuned models in Bedrock what is the pricing option?

A

Provisioned Throughput only which billed by the hour

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
28
Q

How is RAG different from fine tuning

A

Rather than having to fine-tune an FM with a small set of labeled examples, RAG retrieves a small set of relevant documents and uses that to provide context to answer the user prompt

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
29
Q

What are two types of supervised learning?

A

Classification and Regression

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
30
Q

Predicting continuous or numerical values based on one or more input variable?

A

Regression

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
31
Q

Forcasting uses which supervised learning technique?

A

Regression

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
32
Q

Diagnostic uses which supervised learning technique?

A

Classfication

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
33
Q

What are two types of unsupervised learning?

A

Clustering and Dimensionality reduction

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
34
Q

Examples of RAG vector databases?

A

Amazon OpenSearch Service(KNN capability, vector embeddings),
DynamoDB(high performance,vector embeddings),
Aurora(RDS),
RDS for PostgreqSQL(RDS and open source),
Neptune(GraphQL)

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
35
Q

Grouping of unstructured data is done in which type of unsupervised learning?

A

Clustering

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
36
Q

Reducing the number of features or dimensions in a dataset in which type of unsupervised learning?

A

Dimensionality reduction

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
37
Q

Which learning type continuously improves its model by mining feedback from previous iterations?

A

Reinforcement learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
38
Q

In which learning the reward of a desired outcome is known, but the path to achieving it isn’t?

A

Reinforcement learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
39
Q

How to reduce toxity risk in generative AI?

A

Use guardrail models

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
40
Q

What does guradrail models do?

A

These models will detect and filter out unwanted content

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
41
Q

What is the risk term for when model generates inaccurate responses that are not consistent with the training data?

A

Hellucinations

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
42
Q

What is the risk term when model might generate different outputs for the same input?

A

Nondeterminism

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
43
Q

What is the risk term when the information shared with your model can include personal information and can potentially violate privacy laws?

A

Data security and privacy concerns

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
44
Q

What is the risk term when output generated by model has PII?

A

Regulatory violations

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
45
Q

Which generative AI model used for chatbots?

A

Llama

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
46
Q

Which generative AI model used for code generation?

A

Claude

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
47
Q

Which generative AI model used for code gaming?

A

Stable Diffusion

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
48
Q

Which generative AI has embeddings?

A

Amazon Titan

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
49
Q

Which generative AI has a use case of Healthcare – summarize key ideas from long text?

A

Command

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
50
Q

What are the capabilities of generative AI?

A

SPARCD

Adaptability
Responsiveness
Simplicity
Creativity and exploration
Data efficiency
Personalization

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
51
Q

Recommendation engines, gaming, and voice assistance are examples of which type of AI system?

A

Traditional AI

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
52
Q

Chatbots, code generation, and text and image generation are examples of which type of AI system?

A

Generative AI

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
53
Q

When is a model is underfitted?

A

When a model has a high bias

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
54
Q

Overfitting happens when?

A

When model performs well on the training data but does not perform well on the evaluation data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
55
Q

SageMaker Clarify

A

You can automatically evaluate FMs for your generative AI use case with metrics such as accuracy, robustness, and toxicity to support your responsible AI initiative

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
56
Q

SageMaker Clarify is used for text based models only. True or False?

A

True

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
57
Q

Model evaluation on Amazon Bedrock

A

Evaluate, compare, and select the best foundation model for your use case in just a few clicks

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
58
Q

Can in human evaluation, we can automate it?

A

Yes, using built in task types

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
59
Q

Amazon Bedrock Guardrails

A

Guardrails helps control the interaction between users and FMs by filtering undesirable and harmful content, redacting personally identifiable information (PII), and enhancing content safety and privacy in generative AI applications

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
60
Q

Amazon SageMaker Data Wrangler

A

Offers three balancing operators: random undersampling, random oversampling, and Synthetic Minority Oversampling Technique (SMOTE) to rebalance data in your unbalanced datasets

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
61
Q

Amazon SageMaker Experiments

A

Provide scores detailing which features contributed the most to your model prediction on a particular input for tabular, natural language processing (NLP), and computer vision models

You can use to create, manage, analyze, and compare your machine learning experiments.

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
62
Q

Amazon A2I

A

Human review of ML predictions

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
63
Q

SageMake governance tools

A

Amazon SageMaker Role Manager - define minimum permissions in minutes

Amazon SageMaker Model Cards - capture, retrieve, and share essential model information, such as intended uses, risk ratings, and training details, from conception to deployment

Amazon SageMaker Model Dashboard - You can keep your team informed on model behavior in production, all in one place

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
64
Q

AWS AI Service Cards

A

Responsible AI documentation

Basic concepts to help customers better understand the service or service features
Intended use cases and limitations
Responsible AI design considerations
Guidance on deployment and performance optimization

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
65
Q

What are the four elements of responsible agency in responsible AI?

A

Value alignment
Responsible reasoning skills
Appropriate level of autonomy
Transparency and accountability

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
66
Q

What is Data curation?

A

Curating datasets is the process of labeling, organizing, and preprocessing the data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
67
Q

What are the three steps of data curation?

A

Data preprocessing, augmentation and audit

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
68
Q

Whar are the AWS tools for transparency?

A

AWS AI Service Cards and Amazon SageMaker Model Cards

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
69
Q

Whar are the AWS tools for explainability?

A

SageMaker Clarify and SageMaker Autopilot

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
70
Q

What to use to build your custom ML model in SageMaker with less code?

A

Sagemaker Canvas using AutoML powered by Sagemaker Autopilot

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
71
Q

Which is more detailed interpretability or explainability ?

A

Interpretability

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
72
Q

PDP graphs are used for what?

A

They tell how single feature influence the predicted outcome. Used for interpretibiity and explainability

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
73
Q

What are human centered design concepts?

A

Design for amplified decision making
Design for unbiased decision making
Design for human and AI learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
74
Q

This principle seeks to maximize the benefits of using technology while minimizing potential risks and errors, especially risks and errors that can occur when humans make decisions under stress or in high-pressure environments

A

Design for amplified decision-making

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
75
Q

The design of decision-making processes, systems, and tools is free from biases that can influence the outcomes

A

Design for unbiased decision-making

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
76
Q

This priciple aims to create learning environments and tools that are beneficial and effective for both humans and AI

A

Design for human and AI learning

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
77
Q

Reinforcement learning from human feedback

A

(RLHF) is an ML technique that uses human feedback to optimize ML models to self-learn more efficiently

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
78
Q

Which AWS tool provide RLHF?

A

Amazon SageMaker Ground Truth

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
79
Q

What is Feature engineering?

A

It is the process of creating, transforming, extracting, and selecting variables from data. Convert raw data into meaningful data

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
80
Q

Common ratio for data training, validation and testing?

A

80,10,10 or 70,15,15

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
81
Q

Amazon Sagemake which part does “Collecting, analyzing, and preparing your data”?

A

Amazon SageMaker Data Wrangler

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
82
Q

Amazon Sagemake which part does “Managing Features”?

A

Amazon SageMaker Feature Store

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
83
Q

Amazon Sagemake which part does “Model training and evaluation”

A

Amazon SageMaker Canvas

How well did you know this?
1
Not at all
2
3
4
5
Perfectly
84
Q

Amazon SageMaker Canvas

A

Access to ready models from Bedrock and Jumpstart and No coding is required
It is integrated with Comprehend, Rekognition and Textextract

85
Q

Amazon SageMaker JumpStart?

A

Provides pretrained, open source models that customers can use for a wide range of problem types

86
Q

Amazon Sagemaker which part does “Model evaluation”

A

Amazon SageMaker Experiments

87
Q

Amazon Sagemaker which part does “Hyperparameter tuning”

A

Amazon SageMaker Automatic Model Tuning

88
Q

What is learning rate hyperparameter?

A

So if you have a higher learning rate, that means that your model is going
to have a faster conversions, but there is a risk of you to overshoot the optimal solution because while you’re going too fast for learning.

And if you have a low learning rate, it may be more precise but slower convergence

89
Q

Amazon Sakemaker which part does “Monitoring”

A

Amazon SageMaker Model Monitor

90
Q

Supervised learning Sagemaker built in algorithms

A

(FxKL)
Linera learner
Factorization machines
XGBoost
K-Nearst Neighbours(KNN)

91
Q

Unsupervised learning Sagemaker built in algorithms

A

Clustering - K-means, LDA
Topic Modeling - LDA
Embeddings - Object2Vec
Anomoly detection - Random cut forest, IP insights
Dimensionality reduction - Pricipal component analysis (PCA)

92
Q

Image processing sagemake built in algorithms

A

Image classification - MXNet tensor flow
Object detection - MXNet tensor flow
Semantic segmentation - FCN,PSP,Deeplab V3
Time series - DeepAR

93
Q

Text Analysis sagemake built in algorithms

A

Text classification -Blazing text
Word2Vec - Blazing text
Machine translation - Sequence to sequence
Topic modeling - LDA,NTM
Speech - Sequence to sequence

94
Q

When there is greater gap between predicted and actual value what it means?

95
Q

When the predicted values are too much dispersed what it means?

A

High variance

96
Q

How to reduce high variance?

A

Feature selection for more important features and multiple sets of training and test sets of data

97
Q

What matrix help classify why and how a model gets something wrong?

A

Confusion matrix

98
Q

Confusion matrix?

A

It is used to evaluate the performance of the model that does classfication

99
Q

Formula for accuracy

A

(TP+TN)/(TP+FP+TN+FN)

100
Q

Formula for Precision

A

(TP)/(TP+FP)

101
Q

When to use precision over accuracy?

A

When the cost of false positives are high in your particular business situation

Think about a classification model that identifies emails as spam or not. In this case, you do not want your model labeling a legitimate email as spam and preventing your users from seeing that email.

102
Q

Formula for Recall

A

(TP)/(TP+FN)

103
Q

When to use recall over precision?

A

If it is extremely important and vital to the success of the model that it not give false negative results

Think about a model that needs to predict whether a patient has a terminal illness or not

104
Q

Formula for AUC-ROC

A

ROC is a probability curve, and AUC represents the degree or measure of separability.

In general, AUC-ROC can show what the curve for true positive compared to false positive looks like at various thresholds.

105
Q

What is Mean squared error?

A

You take the difference between the prediction and actual value, square that difference, and then sum up all the squared differences for all the observations and divide by number of predictions

106
Q

What is R squared?

A

R squared explains the fraction of variance accounted for by the model

107
Q

What is the difference between Mean squared error and R squared?

A

MSE focuses the measure of model performance.
R squared provides a measure of the model’s goodness of fit to the data.

108
Q

What is A/B testing or the canary deployments technique?

A

Developers can experiment with two or more variants of a model and help achieve the business goals.

109
Q

For real time interactive workloads with low latency requirements, which deployment model?

110
Q

Requests with large payload sizes(upto 1GB), high processing time and near real time latency requirement, which deployment model?

A

Asynchronous

111
Q

Large dateset and don’t need a persistent endpoint, which deployment model?

A

Batch transform

112
Q

Workloads that have idle periods and can tolerate cold starts, which deployment model?

A

Serverless

113
Q

Providing self-service enviornments and curated data sets, which benfit of MLOps?

A

Productivity

114
Q

Incoporating CI/CD practices, which benfit of MLOps?

A

Reliability

115
Q

MlFlow

A

Manage entire ML lifecycle in Sagemaker Studio

116
Q

Automating all steps in ML development lifecycle, which benfit of MLOps?

A

Repeatibility

117
Q

Versioning all inputs and outputs , which benfit of MLOps?

A

Auditibility

118
Q

Policies to guard against model bias and track changes to data statistical properties, which benfit of MLOps?

119
Q

Sagemaker pipeline

A

Prepare data:
Sagemake data wrangle
Sagemake processing job

Curate feature:
Sagemake feature store

Experiment tracking:
Sagemaker experiments

Train model:
Sagemaker training job

Evaluate model:
Sagemaker processing job

Register model:
Sagemaker model registry

Deploy model:
Deployments

Manage model:
Sagemaker model monitor

120
Q

GenerativeAI pretrained models - J2

A

LLM - text generation, contextual question answering, summarization, and classification

121
Q

GenerativeAI pretrained models - Amazon Titan

A

Embeddings, text generation, and image generation

122
Q

GenerativeAI pretrained model - Claude

A

Art vision and text AI models

123
Q

GenerativeAI pretrained model - Command XL

A

Text-based responses based on prompts

124
Q

GenerativeAI pretrained models -Llama 3

A

LLM - generate coherent and contextually relevant text

125
Q

GenerativeAI pretrained models -Mistral Large

A

large reasoning capabilities or are highly specialized, like synthetic text generation, code generation, RAG, or agents

126
Q

GenerativeAI pretrained models -Stable Diffusion

A

Can generate images of from text input.

127
Q

How to increase the performace of the model?

A

Using prompt engineering, RAG, fine-tuning, or automation agents

128
Q

Prompt engineer works for which models?

129
Q

What is demonstrations, or task-specific instructions called in prompt engineering?

A

Augmentation

130
Q

What is Iteratively refining and adjusting the prompts called?

131
Q

What is combining multiple prompts or generation strategies?

A

Ensembling

132
Q

What is prompt searching, prompt generation, or prompt retrieval from large prompt libraries?

133
Q

Three use cases of RAG

A
  1. Customer support, virtual assistants
  2. Journalism and research
  3. Content marketing
134
Q

What is Amazon Bedrock knowledge base?

A

Provide you the capability of amassing data sources into a repository of information

135
Q

What model optmization technique used by Amazon Bedrock knowledge bases?

A

RAG without customization

136
Q

Pricing of Amazon Bedrock

A

Pay as you go
Based on no of tokens in input and response for text based models

Pay as you go
Based on no of images in input and response for image based models

Batch
Multiple predictions at a time and sent as 1 file to S3
50% discount

Provisioned Throughput
Based on no of input and response tokens processed each minute for text based models and is called Provisioned Throughput

137
Q

What is Fine-tuning?

A

Taking a pre-trained language model and further training it on a specific task or domain-specific dataset

138
Q

What is another name of instruction based fine tuning?

A

Prompt Tuning

139
Q

Instruction based find tuning features

A

Labeled examples
Prompt-response pairs

140
Q

Continued Pre-training features

A

Particular field or area of knowledge
Unlabeled data
Domain specific training

141
Q

What is another name of human feedback data used in fine tuning?

A

Reinforcement learning from human feedback (RLHF)

142
Q

Which type of learning creates a reward model?

143
Q

Steps in creating a model from scratch?

A
  1. Selecting the appropriate neural network architecture, layers, and hyperparameters
  2. A large and diverse dataset must be curated, cleaned, and preprocessed
  3. Model is initialized with random weights and trained using various optimization algorithms
144
Q

What are agents used during performance improvement of a model?

A

Carry out various multi-step tasks related to infrastructure provisioning, application deployment, and operational activities.

Task coordination
Reporting and logging
Scalability and concurrency
Integration and communication

145
Q

Model improvement techniques and pricing impact

A

Cheapest to Costly

Prompt Engineering - No model training needed

RAG - Use external knowledge but no FM changes. Cost for using vector dbs

Instruction based fine tuning - FM is fine tuned with instructions and change the tone of the model.

Domain Adaption fine tuning - Domain specific model training

146
Q

Does changing temperature, Top K, Top P affects model pricing?

147
Q

Does temperature, Top K, Top P affects model latency?

148
Q

Types of benchmark datasets for generativeAI evaluation?

A

GLUE - text classification, question answering, and natural language inference
SuperGLUE - compositional language understanding
SQuAD - question-answering capabilities
WMT - machine translation systems

149
Q

What are three types of automated metrics?

A

Perplexity (a measure of how well the model predicts the next token)
BLEU score (for evaluating machine translation)
F1 score (for evaluating classification or entity recognition tasks)

150
Q

Formula for F1?

A

(2xPrecisionxRecall) / (Precision+Recall)

151
Q

What is the advantage of automated metrics?

A

Automated metrics can be useful for rapid iterations and fine-tuning during model development

152
Q

What are the disadvatage of automated metrics?

A

They fail to capture the nuances and complexities of human language and might not align perfectly with human judgments

153
Q

ROUGE automated metrics definition?

A

Automatic summarization and machine translation systems
The main idea behind ROUGE is to count the number of overlapping units

154
Q

BLEU automated metrics definition?

A

Similarity between a generated text and one or more reference translations

Used to evaluate the quality of text that has been machine-translated from one natural language to another

155
Q

BERTScore automated metrics definition?

A

Compute contextualized embeddings for the input texts, and then calculates the cosine similarity between them

It relies on semantic similarity rather than relying on exact lexical matches

156
Q

ROUGE Vs BLEU Vs BERTScore Vs Perplexity

A

Compare N-gram matches
Vs
Evaluate Quality(Prcesion and penalizes)
Vs
Semantic similarity(Compare embeddings)
Vs
How confident the model to predict next token(lower is better)

157
Q

What is automated metrics used for?

A

Assess the performace of a FM in text summarization, machine translation, and open-ended text generation

158
Q

What is negative prompting?

A

Negative prompting is used to guide the model away from producing certain types of content or exhibiting specific behaviors

159
Q

What are the key elements in a prompt?

A

Intructions, Context, Input data and desired output

160
Q

What is used to influence the model response?

A

Inference parameters

161
Q

Randomness and diversity parameters?

A

Temperature - A higher temperature makes the output more diverse and unpredictable, and a lower temperature makes it more focused and predictable

Top P - With a low top p setting, like 0.250, the model will only consider words that make up the top 25 percent of the total probability distribution. Higher P means more diverse

Top K - Set to 50, the model will only consider the 50 most likely words for the next word in the sequence

162
Q

Length and stop sequence in inference parameter?

A

Maximum length - Used in text summarization and translation
Stop Sequence - When the model encounters a stop sequence during the inference process, it will terminate the generation regardless of the maximum length setting

163
Q

Inferencing at the edge?

A

Running Small Language Models on an edge device

164
Q

Zero-shot, few shots and CoT prompting?

A

Zero-shot - Present task to generative model w/o and example
Few-shot - Present task to generative model with some examples
Chain of Thought - Divides intricate reasoning tasks into smaller, intermediary steps

165
Q

Prompt misuse - Poisoning, hijacking, and prompt injection

A

Poisoning - intentional introduction of malicious or biased data

Hijacking, and prompt injection - influencing the outputs of generative models by embedding specific instructions

166
Q

Prompt misuse - Exposure

A

Risk of exposing sensitive or confidential information from its training corpus

167
Q

Prompt misuse - Prompt leaking

A

Exposing the prompt or inputs used within the model or data used by the model

168
Q

Prompt misuse - Jailbreaking

A

Modifying or circumventing the constraints and safety measures implemented in a generative model or AI assistant to gain unauthorized access or functionality

169
Q

The most common algorithms used to perform the similarity search are?

A

k-NN or cosine similarity

170
Q

Name a few vector databases offered by Amazon?

A

Amazon Opesearch
pgvector extension in RDS
Amazon Kendra

171
Q

What is measured by benchmark databases in model evaluation?

A

Accuracy
Speed and efficiency
Scalability

172
Q

Is it true - Creating a benchmark dataset is a manual process?

A

True. It is a set of questions and answers provided by the SME. Model’s response to the same questions is compared with benchmark datasets answers and model performance is scored

173
Q

What type of fine tuning is this - retraining the model on a new dataset that consists of prompts followed by the desired outputs?

A

Instruction tuning

174
Q

What type of fine tuning is this - model is refined through a reinforcement learning process, where a reward model built from human feedback guides the model?

A

Reinforcement learning from human feedback (RLHF)

175
Q

What type of fine tuning is this - the approach involves extending the training phase of a pre-trained model by continuously feeding it new and emerging data?

A

Continuous pretraining

176
Q

What are the steps in fine tuning?

A

Data curation
Labeling
Governance and compliance
Representativeness and bias checking
Feedback integration

177
Q

What are the two types of ROUGE metrics

A

ROUGE-N - This metric primarily assesses the fluency of the text and the extent to which it includes key ideas from the reference. Compare N-gram matches between required vs actual output

ROUGE-L - It is good at evaluating the coherence and order of the narrative in the outputs. Compare the longest sequence of words matche between required vs actual output

178
Q

What does BLEU do?

A

Measures the precision of N-grams in the machine-generated text that appears in the reference texts and applies a penalty for overly short translations (brevity penalty)

179
Q

Services that provide infrastructure protection in AWS?

A

iAM and NACLs

180
Q

Advantages of AI governance

A

Managing, optimizing, and scaling the organizational AI initiative
Maintaining responsible and trustworthy AI practices
Establish clear policies, guidelines, and oversight mechanisms

181
Q

Proximity of data to the compute resources used for training and inference, which AI data governance data management concepts?

A

Data residency

182
Q

Model performance metrics and System events, which AI data governance data management concepts?

A

Data logging

183
Q

Data visualization and Exploratory data analysis (EDA), which AI data governance data management concepts?

A

Data analysis

184
Q

Performance metrics used for monitoring under AI governance

A

Accuracy
Precision
Recall
F1-score
Latency

185
Q

Fine tuned models vs Self trained models

A

Fine tuning a model using your data vs training a model from scratch using your data

186
Q

What is source citation in GenerativeAI?

A

It refers to the act of properly attributing and acknowledging the sources of the data used to train the model.

Datasets
Databases
Other sources

187
Q

What is Documenting data origins in GenerativeAI?

A

It provides detailed information about the provenance, or the place of origin of the data used to train the model.

Details about the data collection process
The methods used to curate and clean the data
Any preprocessing or transformations applied to the data

188
Q

What is the technique used to track the history of data, including its origin, transformation, and movement through different systems?

A

Data lineage

189
Q

What is the systematic organization and documentation of the datasets, models, and other resources used in the development of a generative AI system?

A

Cataloging

190
Q

What is called the standardized format for documenting the key details about an ML model, including its intended use, performance characteristics, and potential limitations?

A

Model cards

191
Q

GenerativeAI scoping matrix(From low to high ownership)

A

Scope 1 : Consumer App (ChatGpt)
Scope 2 : Enterprise App (SaaS like Amazon Q developer)
Scope 3: Pre-trained models (Amazon Bedrock)
Scope 4: Fine-tuned models (Amazon Bedrock customized or SageMaker Jumpstart)
Scope 5: Self trained models (SageMaker )

192
Q

What is data engineering life cycle

A

Automation and access control - AWS Glue
Data collection - Kinesis, DMS, Glue
Data prep and cleaning - EMR or Glue
Data quality check - Glue data brew or Glue data quality check
Data visualization and analysis - Quicksight or Neptune
IaC deployment - CloudFormation
Monitoring and Debugging - CloudWatch

193
Q

4 ways to build your AI solutions by chosing FM?

A

Reuse
Adapt
Customize
Start from scratch

194
Q

Reuse and Adapt only in which AWS GenAI tool?

195
Q

Amazon Q pricing

A

Subscription based - Lite and Pro + Data storage for client documents

196
Q

Amazon Q Admin controls?

A

Same as Guardrails

197
Q

Reuse, Adapt and Customize only in which AWS AL/ML service?

A

Amazon Bedrock

198
Q

Reuse, Adapt, Customize and Start from scratch

A

Amazon SageMaker

199
Q

Foundation Models comparison

A

Amazon Titan Text Express, LLAMA 2, Claude, stability.ai

Claude can take maximum tokens - 200K
stability.ai is for image

Content creation is by Titan
Text generation and customer service by LLAMA 2
Analysis and Forecasting by Claude
Image creation by stability.ai

200
Q

What is PartyRock

A

Its a playground on Amazon Bedrock to build GenAI apps
You can access without having AWS account

201
Q

Which model used in classification?

202
Q

Custom classfication and custom entity recognition in which AWS AI managed service?

A

Amazon Comprehend

203
Q

Redaction is used in which AWS AI managed service?

A

Amazon Transcribe to remove PII information

204
Q

Custome vocabularies and custome language models are used in which AWS AI managed service?

A

Amazon Transcribe to transcribe technical terms and jargons and context

205
Q

Lexicons, SSML are used in which AWS AI managed service?

A

Amazon Polly to read specific type of text and add break, whisper etc

206
Q

What are SLOTs use in which AWS AI managed service?

A

Amazon Lex to provide input parameters

207
Q

Amazon Lex can integrate with what AWS services?

A

Comprehend, Connect, Lambda function & Kendra