1.2.9

Question	Answer & Explanation
### Question 1 - Continuous Pre-training Data Preparation A financial institution wants to adapt a foundation model to understand complex, proprietary financial jargon and writing styles. They have a massive corpus of unlabeled financial research reports, historical memos, and proprietary notes. The ML team wants to update the model's internal weights to absorb this domain-specific vocabulary. Which data preparation approach matches this goal with the least operational overhead? (Select ONE) - A) Create JSONL files of prompt-completion pairs. - B) Assemble the unlabeled domain text in plain text or JSONL format. - C) Extract embeddings for the text and store in OpenSearch. - D) Build a prompt set for a teacher model to generate responses.	Correct Answer: B Correct Option Analysis: - B is Correct: Continuous pre-training (also known as domain adaptation) uses self-supervised learning on large volumes of unlabeled text. Assembling the raw, unlabeled text without creating prompt-response pairs is the correct and lowest-overhead preparation method for this customization path. Distractor Analysis: - A is Incorrect: Prompt-completion pairs are required for supervised fine-tuning (SFT), not continuous pre-training. Pre-training ignores the pairing and just uses the text distribution. - C is Incorrect: Embedding text and storing it in OpenSearch is the preparation step for a Retrieval-Augmented Generation (RAG) pipeline, which does not update the model's weights. - D is Incorrect: Building a prompt set for a teacher model is the data preparation step for model distillation, not continuous pre-training. > Exam Tip: Continuous pre-training absorbs vocabulary and style from unlabeled text; fine-tuning learns tasks from labeled pairs.
### Question 2 - Model Distillation Data Artifacts An enterprise is building a customer service chatbot and wants to use a smaller, cost-effective foundation model. They have a larger, highly capable model that performs well but is too expensive to run in production. The team decides to use Amazon Bedrock model distillation. The team must prepare the training artifacts for the smaller model with minimal manual labeling effort. What data preparation step is required for this distillation process? (Select ONE) - A) Prepare a large dataset of unlabeled domain text. - B) Manually label thousands of prompt-response pairs for the smaller model. - C) Assemble a diverse set of prompts covering the target use cases. - D) Convert the customer service documents into vector embeddings.	Correct Answer: C Correct Option Analysis: - C is Correct: For model distillation, the primary data artifact prepared by the user is a set of prompts covering the target use cases. The managed distillation path uses the larger "teacher" model to generate responses to these prompts, which are then used to fine-tune the smaller "student" model, avoiding manual response labeling. Distractor Analysis: - A is Incorrect: Unlabeled domain text is used for continuous pre-training, not distillation. - B is Incorrect: The key advantage of distillation is avoiding manual labeling of responses; the teacher model generates them. - D is Incorrect: Vector embeddings are for RAG pipelines, not for distilling model weights. > Exam Tip: Distillation requires a prompt set; the teacher model generates the responses to train the student model.
### Question 3 - Fine-Tuning Data Formatting A healthcare startup is preparing data to fine-tune an Amazon Bedrock foundation model for a specific task: extracting patient symptoms from clinical notes and outputting them in a structured JSON format. The startup has 10,000 examples of clinical notes and the corresponding correct JSON outputs. The data must be formatted according to Bedrock fine-tuning requirements. How should the data be prepared? (Select ONE) - A) Save the dataset as CSV files with separate columns for prompts and completions. - B) Store the unlabeled clinical notes in an S3 bucket for continued pre-training. - C) Format the data as JSONL files with prompt-completion pairs. - D) Chunk the clinical notes and generate embeddings using Amazon Titan.	Correct Answer: C Correct Option Analysis: - C is Correct: Amazon Bedrock supervised fine-tuning requires labeled examples supplied as JSONL (JSON Lines) files in a specific schema, typically containing prompt and completion pairs. Distractor Analysis: - A is Incorrect: While CSV can hold tabular data, Amazon Bedrock fine-tuning natively expects JSONL for prompt-completion pairs, not CSV. - B is Incorrect: Storing unlabeled notes is for continuous pre-training; this scenario requires supervised fine-tuning since they have the corresponding correct JSON outputs (labels). - D is Incorrect: Chunking and generating embeddings is for RAG, not for fine-tuning a model to output structured JSON. > Exam Tip: Amazon Bedrock supervised fine-tuning natively consumes JSONL files containing prompt and completion pairs.
### Question 4 - Redacting PII for Pre-training A legal firm is preparing data for continuous pre-training of a foundation model on proprietary legal case files. The data contains highly sensitive Personally Identifiable Information (PII) of clients. Compliance policies mandate that PII must be completely removed before the data is used for any model training, ensuring it cannot surface in model outputs. Which approach satisfies this zero data retention compliance requirement? (Select ONE) - A) Mask the PII using AWS Glue DataBrew, replacing names with "[REDACTED]". - B) Use Amazon Comprehend to detect PII entities, then drop the records containing PII or remove the entities completely before continuous pre-training. - C) Fine-tune the model instead of continuous pre-training, as fine-tuning inherently prevents PII leakage. - D) Use AWS Lambda to replace PII with hashed values, as hashing guarantees PII cannot be reverse-engineered by the model.	Correct Answer: B Correct Option Analysis: - B is Correct: Redaction removes the PII entirely or drops the record. Amazon Comprehend can detect PII entities in text, allowing the engineering team to strip the sensitive data out before the text is sent for continuous pre-training. Distractor Analysis: - A is Incorrect: Masking hides a value but keeps its shape (e.g., replacing a name with [REDACTED]). The prompt mentions the need to completely remove PII (redaction), and continuous pre-training might still learn patterns from the masked tokens or surrounding context. - C is Incorrect: Fine-tuning does not inherently prevent PII leakage; if PII is in the fine-tuning data, it can be memorized and regurgitated. - D is Incorrect: Hashing replaces PII with deterministic strings, but the model might learn to associate the hashes with specific contexts, and hashing does not remove the underlying identity pattern. True redaction is safer. > Exam Tip: Redaction removes sensitive data entirely, while masking only obscures it. Always run PII detection and redaction before any FM customization.
### Question 5 - Behavior Tuning vs Knowledge Retrieval An ML team is preparing data for Amazon Bedrock fine-tuning. They have a dataset of 5,000 customer support interactions. The team wants the model to learn a specific behavior: apologizing empathetically before providing a solution. The team must ensure the model learns this behavior accurately and consistently updates its internal weights. Which data preparation strategy is the MOST effective? (Select ONE) - A) Provide unlabeled text of all customer support interactions for continuous pre-training. - B) Assemble prompt-response pairs where the prompt is the customer query and the completion is the empathetic apology followed by the solution. - C) Create a prompt set without responses and let the model generate its own empathetic responses during distillation. - D) Store the interactions in a vector store and use RAG to inject empathetic examples at inference.	Correct Answer: B Correct Option Analysis: - B is Correct: Supervised fine-tuning (SFT) teaches a model a specific behavior or task using labeled prompt-response pairs. Providing the exact desired output (apology + solution) trains the model to adopt this behavior in its weights. Distractor Analysis: - A is Incorrect: Unlabeled text for continuous pre-training absorbs vocabulary and style, but it does not explicitly teach the behavior of pairing an apology with a solution. - C is Incorrect: Distillation uses a teacher model to generate responses. If the teacher doesn't already exhibit the empathetic behavior, the student won't learn it. - D is Incorrect: RAG injects context at inference time but does not update the model's internal weights to inherently adopt the tone. > Exam Tip: To teach an FM a specific behavior (like tone or formatting), prepare JSONL prompt-response pairs for supervised fine-tuning.
### Question 6 - Distillation Training Data Generation An e-commerce company wants to compress their large foundation model into a smaller one for edge deployment using Amazon Bedrock model distillation. They have collected a wide variety of user queries that the model needs to handle. The team must prepare the training data for the smaller model with the least operational overhead. What is the correct data preparation step for this distillation task? (Select ONE) - A) Use the prompt set to generate responses from the larger teacher model, which will then be used to fine-tune the smaller student model. - B) Manually write the ideal responses for each prompt in the set to train the student model. - C) Train a separate classification model to filter the prompts before passing them to the student model. - D) Embed the prompts using Amazon Titan Embeddings and store them in a vector store for the student model to retrieve.	Correct Answer: A Correct Option Analysis: - A is Correct: In Bedrock model distillation, the user provides the prompt set. The managed service runs the larger teacher model against these prompts to generate high-quality responses, and these synthetic prompt-response pairs are then used to fine-tune the smaller student model. Distractor Analysis: - B is Incorrect: Manually writing responses defeats the purpose of distillation, which is to transfer knowledge from the teacher model automatically, minimizing manual labeling overhead. - C is Incorrect: Training a separate classification model is unnecessary and introduces high operational overhead. - D is Incorrect: Vector stores are used for RAG architectures, not for training a distilled model's weights. > Exam Tip: Distillation automates response generation via the teacher model, drastically reducing the manual labeling overhead for fine-tuning the student.
### Question 7 - Continuous Pre-training Data Structure A manufacturing company wants to adapt a foundation model to understand technical manuals, equipment schematics, and proprietary maintenance logs. They have millions of pages of unstructured text. They decide to use continuous pre-training. The data must be prepared efficiently for continuous pre-training. How should the ML engineer structure this data? (Select ONE) - A) Format the data as JSONL with prompt-completion pairs. - B) Convert the text into dense vector embeddings using an embedding model. - C) Assemble the raw, unlabeled text documents into a single corpus without pairing prompts and responses. - D) Use AWS Glue DataBrew to one-hot encode the text categories.	Correct Answer: C Correct Option Analysis: - C is Correct: Continuous pre-training requires a large body of unlabeled domain text. The model learns the patterns, vocabulary, and structure of the text through self-supervised learning without any prompt-completion pairing. Distractor Analysis: - A is Incorrect: JSONL with prompt-completion pairs is for supervised fine-tuning. - B is Incorrect: Generating embeddings is the first step in building a RAG pipeline, not for continuous pre-training. - D is Incorrect: One-hot encoding is a tabular machine learning technique for categorical features, not for unstructured text foundation models. > Exam Tip: Continuous pre-training consumes massive amounts of raw, unlabeled text to teach an FM domain-specific jargon.
### Question 8 - Fine-Tuning Code Generation Data Steps A tech company is preparing data for fine-tuning a model to perform code generation in a proprietary language. They have a repository of code snippets and their corresponding natural language descriptions. The data must be split into training and validation sets, and formatted for Amazon Bedrock. Which steps should the engineer take to prepare this data? (Select TWO) - A) Convert the snippets and descriptions into JSONL prompt-completion pairs. - B) Split the JSONL dataset into training and validation files. - C) Embed the code snippets into a vector database for retrieval. - D) Provide the code snippets as unlabeled text for continued pre-training. - E) Distill a smaller model using the code snippets as the prompt set.	Correct Answers: A and B Correct Option Analysis: - A is Correct: Amazon Bedrock fine-tuning requires the data to be structured as JSONL files containing prompt-completion pairs (in this case, description -> code snippet). - B is Correct: To evaluate the fine-tuning job's performance and prevent overfitting, the dataset must be split into separate training and validation JSONL files. Distractor Analysis: - C is Incorrect: Embedding code snippets is for semantic search/RAG, not for supervised fine-tuning. - D is Incorrect: Using the data as unlabeled text for continued pre-training would not teach the model the specific mapping of description to code snippet. - E is Incorrect: While distillation is a valid technique, the question specifically asks for the steps to prepare data for fine-tuning the model. > Exam Tip: Always prepare JSONL files split into training and validation sets for Amazon Bedrock supervised fine-tuning.
### Question 9 - Detecting Sensitive Data in Text An ML engineer is preparing a dataset for fine-tuning a foundation model. The dataset contains conversational chat logs. The engineer notices that some chat logs contain highly sensitive user credentials and API keys. The data must be redacted before fine-tuning to prevent the model from memorizing and regurgitating secrets. Which AWS service can automatically detect and assist in redacting these sensitive data types from the unstructured text? (Select ONE) - A) Amazon Macie - B) Amazon Comprehend - C) Amazon SageMaker Feature Store - D) AWS Glue DataBrew	Correct Answer: B Correct Option Analysis: - B is Correct: Amazon Comprehend provides PII detection capabilities specifically for unstructured text. It can identify entities like credentials, API keys, names, and addresses, allowing engineers to programmatically redact or mask them before model training. Distractor Analysis: - A is Incorrect: Amazon Macie is a data security service that discovers and protects sensitive data in Amazon S3, but it does not parse and redact PII within unstructured text documents for ML pipelines the way Comprehend does. - C is Incorrect: SageMaker Feature Store is for storing engineered tabular features, not for text PII detection. - D is Incorrect: AWS Glue DataBrew is for visual data preparation of tabular data; it does not have native NLP PII detection for unstructured conversational logs. > Exam Tip: Use Amazon Comprehend to detect and redact PII in unstructured text before preparing it for FM customization.
### Question 10 - Choosing the Right Customization Artifact A retail company wants to customize a foundation model to adopt their specific brand voice and tone when responding to customer reviews. They have 1,000 examples of customer reviews paired with the ideal brand-approved responses. The data is static and unlikely to change frequently. The solution must update the model's weights to ensure the tone is consistently applied without relying on prompt engineering. Which data preparation approach fulfills this requirement? (Select ONE) - A) Prepare the 1,000 review-response pairs as JSONL files for supervised fine-tuning. - B) Assemble a large corpus of unlabeled brand content for continuous pre-training. - C) Chunk the brand responses and add them to a vector store for RAG. - D) Generate a prompt set for distillation using the larger model.	Correct Answer: A Correct Option Analysis: - A is Correct: Supervised fine-tuning with prompt-response pairs (JSONL) is the correct way to update model weights to learn a specific behavior or tone, especially when the dataset is static and prompt engineering alone is deemed insufficient. Distractor Analysis: - B is Incorrect: Continuous pre-training with unlabeled text teaches vocabulary and domain knowledge, but it is inefficient for teaching specific prompt-response behaviors like tone matching. - C is Incorrect: RAG stores knowledge in a vector store for retrieval but does not update the model's internal weights to adopt a specific tone. - D is Incorrect: Distillation is for transferring capabilities from a teacher to a student model to reduce size, not specifically for learning brand tone from labeled examples. > Exam Tip: When prompt engineering isn't enough to teach a static behavior or tone, prepare JSONL prompt-response pairs for supervised fine-tuning.