Bedrock and GenAI Flashcards
What is Bedrock used for?
To create GenAI applications on AWS
Is Bedrock fully managed?
Yes
What is Amazon’s foundation model?
Amazon Titan
How many foundation models does Bedrock support?
At least seven including their own.
When you choose a foundation model in Bedrock, is it shared?
No, it makes a copy of the foundation model which is only available to you which you can further fine tune with your own data.
When using a foundation model in Bedrock, is your data ever sent back to train the model?
No. Your data is never used to train the model?
What is the playground in Bedrock?
A place for you to test a specific model based on your input. It helps you select your
Can you access Titan through an API?
Yes
What types of inputs does Amazon Titan support?
Image, Text, and multmodal
Can Amazon Titan be customized with your own data?
Yes
What are the pros and cons of smaller foundation models?
They are cheaper, but know less.
What is the maximum amount of tokens for Amazon Titan?
8K
What is a continued pre-training job in Bedrock?
It continuously trains your model on the newest data.
What is a fine-tuning job in Bedrock?
It is a one-time training job on your custom data.
Are all foundation models able to be fine-tuned or trained on custom data?
No
Where does custom data for fine-tuning jobs for Bedrock need to be stored?
In S3
What are hyperparameters?
They are configuration parameters
Where is validation data for a fine-tuning job stored?
In S3
Does fine-tuning require provisioned throughput?
Yes. This is expensive.
What does model fin-tuning do under the hood?
It changes the weight of the base foundation model.
What is instruction-based fine-tuning?
It improves the performance of a pre-trained model on domain specific tasks. It uses labeled examples that are prompt-response pairs.
What is continued pre-training?
This is also called domain-adaption fine-tuning. It makes a model an expert in a specific domain. E.g., ingesting all AWS documentation.
What is single-turn messaging?
Part of instruction-based fine-tuning.
What is multi-turn messaging?
Part of instruction-based fine-tuning. Good for larger context.
What is Transfer Learning?
Use a pre-trained model to adapt it to a new related task. Wiidely used for image classification and NLP.
What is Automatic Evaluation in Amazon Bedrock?
It allows you to evaluate a model for quality control. You can use your own questions or AWS’
What are benchmark datasets?
They are curated collections of data designed to evaluate the performance of a language model.
What do benchmark datasets help measure?
accuracy, speed, efficiency, and scalability.
Can a benchmark dataset detect bias?
Yes, some can.
What is the metric ROUGE used for?
Evaluating automatic summarization and machine translation systems in a foundation model.
What is the metric BLEU for?
Evaluates the quality of generated text, especially for translations in a foundation model.
What is the metric BERTscore for ?
Semantic similarity between generated text in a foundation model.
Where can you perform model evaluation?
Within Bedrock
What is RAG?
Retrieval-Augmented Generation.. This allows a foundation model to reference a datasource outside of its training data.
Where does Bedrock store RAG data that was loaded in S3?
Into a Knowledge Base, which is a vector database for retrieval.
Does RAG require foundation model fine-tuning?
No. It is not required.
What is a good use case for RAG?
When the model needs access to real-time data.
What AWS services can be used for the vector database required for RAG?
OpenSearch and DocumentDB, Aurora, and RDS for PGSQL
What is an embedding model in the context of RAG?
It converts the data into vectors for the vector database.
What are some data sources for RAG?
S3, Confluence, Sharepoint, Salesforce, Web Pages
What is a good use case for RAG?
A customer service chatbot that will have access to your products, features, and specifications.
Can you create a knowledge base with a root user?
No. You must use an IAM user
What is tokenization?
It turns raw text into a sequence of tokens.
What is word-based tokenization?
Text is split into individual words?
What is subword tokenization?
Some words can be split. This is useful for long words.
What is a context window?
It is the number of tokens an LLM can consider when generating text.
Do large context windows cause high memory and processing spikes?
Yes
What is embedding?
It creates vectors out of text, images, or audio using embedding models to insert into a vector database.
What are embedding models good for?
They can power search applications.
What are Bedrock Guardrails?
They control the interactions between users and foundation models.
Can Bedrock guardrails filter undesirable and harmful content?
Yes.
Can Bedrock Guardrails remove PII?
Yes
Can you create multiple Bedrock Guardrails?
Yes
Can you view all the user inputs that can violate Bedrock Guardrails?
Yes
Can Bedrock Guardrails mask emails?
Yes
What are Bedrock Agents?
They manage and carry out various multi-step tasks related to infrastructure provisioning, application deployment, and operational activities.
What are Agent Action Groups?
They are pre-defined action groups for agents to perform. For example, one action group can query a database for a contact name and another can be an API call to get recent orders.
What is model invocation logging?
It sends logs of all invocations to the CloudWatch logs. It can include text, images, and embeddings.
Are Bedrock metrics sent to CloudWatch?
Yes
What is Bedrock Studio?
It gives access to your team so they can easily create AI powered applications.
What is Batch Mode in Bedrock?
Allows you to make multiple predictions at a time and can provide discounts up to 50%. Takes longer for a response.
What is the main driver of costs in Bedrock?
Number of input and output tokens
