2.2.8

Question	Answer & Explanation
### Question 1 - Choosing Customization Techniques for Output Formatting A retail company is using an Amazon Bedrock-hosted foundation model to generate product descriptions. The marketing team requires the model to consistently output responses in a specific JSON schema and adopt a playful, brand-specific tone. The ML team wants to implement this with the least operational overhead and avoid provisioning any training infrastructure. Which customization technique should the team apply to meet these requirements? - A) Provision an Amazon SageMaker AI training instance and perform supervised fine-tuning using a dataset of 10,000 JSON-formatted product descriptions. - B) Use task-specific prompt engineering by defining explicit instructions, desired output format, and a few worked examples within the prompt. - C) Implement Reinforcement Fine-Tuning (RFFT) in Amazon Bedrock using a reward model that scores the JSON validity of the outputs. - D) Implement knowledge distillation by having a larger teacher model generate product descriptions to train a smaller student model.	Correct Answer: B Correct Answer Analysis: - B is Correct: Task-specific prompt engineering (providing explicit instructions, specifying the JSON format, and offering few-shot examples) is the lowest-overhead method to guide a foundation model's output style and structure. It requires no training jobs, compute provisioning, or custom data preparation pipelines. Distractor Analysis: - A is Incorrect: Supervised fine-tuning introduces significant operational overhead (data preparation, managing training jobs) and is generally overkill for formatting and tone adjustments when prompt engineering can achieve the same result natively. - C is Incorrect: Reinforcement fine-tuning is used to improve reasoning and complex task execution based on a reward function, not for basic structural formatting like JSON schemas. It also requires running training jobs, violating the overhead constraint. - D is Incorrect: Distillation is used to compress a large model into a smaller, faster one. It does not inherently solve JSON formatting constraints and requires heavy operational setup. > Exam Tip / AWS Best Practice: Always exhaust prompt engineering (instructions, few-shot examples, output formatting) for style and structural changes before considering compute-heavy fine-tuning.
### Question 2 - Preparing Data for Supervised Fine-Tuning in Bedrock A healthcare startup wants to customize an Amazon Bedrock foundation model to answer patient queries based on internal, proprietary medical guidelines. The team has compiled 5,000 verified expert-reviewed question-and-answer pairs. They need the model to learn this specific domain knowledge and phrasing style, improving accuracy over what prompt engineering alone can achieve. Which data artifact is specifically required to perform this customization technique in Amazon Bedrock? - A) A set of prompts to be used by a teacher model to generate synthetic training data. - B) A reward function that assigns a numerical score to the model's generated responses. - C) Labeled input and output pairs (e.g., User Prompt and Assistant Response). - D) A large corpus of unstructured text documents stored in Amazon S3 for continued pre-training.	Correct Answer: C Correct Answer Analysis: - C is Correct: Supervised fine-tuning requires a dataset of labeled input-output pairs. In the context of an LLM, this typically means pairs of user prompts/instructions and the desired expert assistant responses. This teaches the model to map specific domain inputs to the preferred output style and knowledge. Distractor Analysis: - A is Incorrect: A set of prompts for a teacher model to generate synthetic data is the artifact required for distillation, not supervised fine-tuning. - B is Incorrect: A reward function that scores responses is the artifact required for Reinforcement Fine-Tuning (RFFT), where the model learns by trial and error to maximize a reward. - D is Incorrect: Unstructured text documents are used for continued pre-training (domain-adaptive pre-training) or for RAG ingestion via Bedrock Knowledge Bases, not for supervised fine-tuning. > Exam Tip / AWS Best Practice: Match the data artifact to the customization technique: Input/Output pairs = Supervised Fine-Tuning; Reward Functions = Reinforcement Fine-Tuning; Prompt Sets = Distillation.
### Question 3 - Addressing Catastrophic Forgetting During Fine-Tuning An AI engineering team is fine-tuning a foundation model on a highly specialized, narrow dataset of legal contracts. After the customization job completes, they notice the model performs exceptionally well on legal tasks but has lost its general conversational capabilities and struggles with basic reasoning tasks it previously handled easily. Which technique should the team apply to prevent this issue while still customizing the model for legal contracts? - A) Increase the learning rate during the fine-tuning job to converge faster. - B) Implement early stopping based on the validation loss of the legal dataset. - C) Blend broader, general-purpose data into the specialized training mixture. - D) Switch to a smaller student model using knowledge distillation.	Correct Answer: C Correct Answer Analysis: - C is Correct: Catastrophic forgetting occurs when a model is fine-tuned on a narrow domain and overwrites its previously learned general capabilities. Blending broader, general-purpose data into the training mixture helps the model retain its general reasoning and conversational abilities while learning the new domain-specific features. Distractor Analysis: - A is Incorrect: Increasing the learning rate often exacerbates catastrophic forgetting by making larger, more destructive updates to the model's weights, overwriting general knowledge more aggressively. - B is Incorrect: Early stopping prevents overfitting on the training data, but it does not explicitly ensure the retention of general capabilities outside the domain of the training/validation set. - D is Incorrect: Distillation is a model compression technique, not a mitigation strategy for catastrophic forgetting. > Exam Tip / AWS Best Practice: To prevent catastrophic forgetting during fine-tuning, mix a percentage of general instruction-following data into your specialized training dataset.
### Question 4 - Implementing Knowledge Distillation A gaming company is using a large, highly capable foundation model via Amazon Bedrock to generate dynamic NPC dialogue. While the quality is excellent, the latency and inference costs are too high for real-time game interactions. The team has a large dataset of historical game logs containing thousands of unique player prompts. Which customization technique should the team use to create a smaller, faster model that mimics the large model's quality with lower latency? - A) Supervised fine-tuning on the player prompts using a small model. - B) Knowledge distillation using the large model as a teacher to label the prompt set for a smaller student model. - C) Reinforcement fine-tuning using a reward function based on player engagement metrics. - D) Implementing an Amazon Bedrock Knowledge Base to cache the large model's responses.	Correct Answer: B Correct Answer Analysis: - B is Correct: Knowledge distillation is the process of transferring knowledge from a large, complex "teacher" model to a smaller, faster "student" model. The managed path uses a prompt set; the teacher model generates the responses, and this input/output pair dataset is then used to fine-tune the smaller student model, achieving lower latency and cost. Distractor Analysis: - A is Incorrect: Supervised fine-tuning on just the player prompts without high-quality target responses from the large model would not transfer the large model's specific capabilities to the smaller model. - C is Incorrect: Reinforcement fine-tuning optimizes a model against a reward function but does not inherently compress the model into a smaller, faster architecture. - D is Incorrect: Caching responses via a Knowledge Base (RAG) might reduce latency for repeated queries, but it does not create a customized, smaller model capable of generating novel responses faster. > Exam Tip / AWS Best Practice: Distillation is the go-to technique when you need to balance the quality of a massive LLM with the latency and cost constraints of a smaller model.
### Question 5 - Task-Specific Prompt Engineering and Version Management A data science team is iterating on prompts for a customer service chatbot using Amazon Bedrock. The team needs to test different phrasing, inject dynamic variables (like customer names and order numbers), and compare performance across different model versions. They want to manage this process without modifying the underlying application code every time a prompt is updated. Which AWS feature provides versioning, variable injection, and A/B testing for task-specific prompt engineering with the least operational overhead? - A) Storing prompts as JSON files in an Amazon S3 bucket and fetching them via an AWS Lambda function. - B) Using Amazon Bedrock Prompt Management to create, version, and test dynamic prompts. - C) Fine-tuning a custom model in SageMaker AI where the prompt text is hardcoded into the training data. - D) Deploying an Amazon API Gateway endpoint that routes requests to different hardcoded prompts based on headers.	Correct Answer: B Correct Answer Analysis: - B is Correct: Amazon Bedrock Prompt Management allows teams to create, save, and version prompts. It supports dynamic variables (injection) and enables A/B testing of prompts against different foundation models directly within the console or API, completely decoupling prompt iteration from application code. Distractor Analysis: - A is Incorrect: While storing prompts in S3 decouples them from code, it requires building custom logic (e.g., Lambda) to handle variable injection, versioning, and A/B testing, resulting in higher operational overhead. - C is Incorrect: Hardcoding prompts into fine-tuning training data is a customization technique, not a prompt engineering technique. It is inflexible and introduces massive overhead for simple prompt tweaks. - D is Incorrect: API Gateway routing is a network-level solution and does not natively provide prompt versioning or dynamic variable injection for LLMs. > Exam Tip / AWS Best Practice: Treat prompts as code. Use Amazon Bedrock Prompt Management for versioning, variable injection, and testing to avoid hardcoding strings in your application logic.
### Question 6 - Identifying Retrieval vs. Generation Failures in RAG An enterprise RAG application uses Amazon Bedrock Knowledge Bases to query internal HR policies. Users complain that the chatbot frequently provides incorrect or hallucinated answers. The ML team inspects the application logs and observes that the vector search retriever is returning documents that are only loosely related to the user's question, completely missing the correct HR policy page. Where should the team focus their customization efforts to fix this issue most effectively? - A) Fine-tune the generating LLM on the HR policy documents to improve its reasoning. - B) Apply task-specific prompt engineering to instruct the LLM to "only use the provided context." - C) Optimize the retrieval components by adjusting the chunking strategy and evaluating the embedding model. - D) Increase the max_tokens generation parameter so the model has more room to explain its answer.	Correct Answer: C Correct Answer Analysis: - C is Correct: If the retrieved passages do not contain the answer, no amount of prompt engineering or model customization will fix the hallucination. The failure is in the retrieval step. Adjusting the chunking strategy (e.g., size, overlap) or swapping to a domain-specific embedding model directly improves vector search relevance. Distractor Analysis: - A is Incorrect: Fine-tuning the generator model does not fix the vector search returning irrelevant documents. The model cannot generate the correct answer if it isn't provided the correct context in the prompt. - B is Incorrect: Instructing the model to "only use the provided context" will likely cause it to output "I don't know" because the provided context (loosely related docs) doesn't contain the answer. The root cause is retrieval, not generation. - D is Incorrect: Increasing max_tokens allows for longer responses but has no effect on the accuracy of the retrieved context or the model's reasoning ability. > Exam Tip / AWS Best Practice: In RAG systems, if the retriever fails to fetch the right passages, fix the embeddings and chunking. If the passages are right but the answer is wrong, fix the prompt or the model.
### Question 7 - Reinforcement Fine-Tuning (RFFT) Use Cases A software company wants to customize a foundation model to generate Python code that solves complex algorithmic challenges. They do not have a large dataset of pre-written, expert-level solutions, but they do have an automated testing environment that can execute Python code and determine if the unit tests pass or fail. Which customization technique in Amazon Bedrock is best suited for this scenario? - A) Supervised fine-tuning using the problem descriptions as input and placeholder code as output. - B) Task-specific prompt engineering with few-shot examples of pseudocode. - C) Reinforcement fine-tuning using a reward function that scores responses based on unit test execution results. - D) Knowledge distillation from a smaller model to a larger model to improve code generation.	Correct Answer: C Correct Answer Analysis: - C is Correct: Reinforcement Fine-Tuning (RFFT) is ideal when you have a task that can be scored objectively but lack a large dataset of labeled examples. A reward function (in this case, executing the code and checking unit tests) provides the signal for the model to learn from, optimizing the model's behavior to generate code that passes the tests. Distractor Analysis: - A is Incorrect: Supervised fine-tuning requires high-quality input-output pairs. Placeholder code would teach the model to generate bad code, not solve the algorithmic challenges. - B is Incorrect: Prompt engineering might help, but it relies on the model's zero-shot or few-shot capabilities. It does not update the model weights to fundamentally improve its algorithmic coding skills based on feedback. - D is Incorrect: Distillation transfers knowledge from a teacher to a student. It is not the mechanism for using a reward function to improve task-specific performance. > Exam Tip / AWS Best Practice: Reinforcement Fine-Tuning (RFFT) uses a reward function to score model outputs, making it perfect for tasks with verifiable outcomes (like math, coding, or logic) where ground truth labels are hard to source.
### Question 8 - Deciding Between Fine-Tuning and RAG for Proprietary Knowledge A financial institution has a completely proprietary, highly complex risk-assessment API that was developed in-house. The syntax, parameters, and error codes are entirely unique and not documented anywhere on the public internet. The ML team needs a foundation model to write accurate API calls for this internal system. The API schema is too large to fit in the context window of the model. Which customization technique is the most effective way to teach the model this proprietary syntax? - A) Using Amazon Bedrock Knowledge Bases to chunk the API schema and retrieve relevant sections dynamically. - B) Applying task-specific prompt engineering with a summary of the API rules. - C) Performing supervised fine-tuning on a dataset of thousands of correct API call examples. - D) Increasing the temperature parameter to encourage more creative API calls.	Correct Answer: C Correct Answer Analysis: - C is Correct: When teaching a model a completely new, proprietary syntax or format that is absent from its pre-training data, and the schema is too large for the context window, supervised fine-tuning is the best approach. It embeds the specific syntax rules and patterns directly into the model's weights, allowing it to generate correct API calls without needing the full schema in the prompt. Distractor Analysis: - A is Incorrect: While RAG can retrieve chunks of the schema, complex code generation often requires a holistic understanding of multiple parameters and interactions that can be lost when chunked. Furthermore, RAG adds latency and retrieval overhead. Fine-tuning internalizes the knowledge. - B is Incorrect: Prompt engineering cannot teach a model a new syntax if it has never seen it before and cannot fit the full schema in its context window. - D is Incorrect: Increasing temperature increases randomness, which is the exact opposite of what is needed for strict, syntactically correct API code generation. > Exam Tip / AWS Best Practice: If a task requires the model to learn new syntax, formats, or styles not present in its pre-training data, and the context limit is a bottleneck, fine-tuning is preferred over RAG or prompt engineering.
### Question 9 - Optimizing the Generator in RAG An enterprise search application uses Amazon Bedrock Knowledge Bases to retrieve internal wikis. The ML team verifies that the vector store is consistently retrieving the exact paragraphs containing the answers to user queries. However, the foundation model often ignores the retrieved context and generates a generic, unhelpful response. What is the most direct and least costly customization technique to fix this issue? - A) Fine-tuning the embedding model used in the vector store. - B) Re-chunking the source documents into smaller paragraphs. - C) Adjusting the prompt template in the Bedrock Knowledge Base to explicitly instruct the model to "Answer the question using only the provided context." - D) Running a supervised fine-tuning job on the foundation model using the internal wikis as training data.	Correct Answer: C Correct Answer Analysis: - C is Correct: The retrieval is working perfectly (correct paragraphs are fetched), but the generation is failing (the model ignores the context). This is a generator issue. Modifying the prompt template in Bedrock Knowledge Bases to explicitly instruct the model to rely solely on the retrieved context is the most direct, least costly, and operationally efficient fix. Distractor Analysis: - A is Incorrect: The embedding model is already retrieving the correct paragraphs; adjusting it will not fix the generator's failure to use the provided context. - B is Incorrect: Re-chunking is a retrieval optimization. Since retrieval is already accurate, changing chunk sizes is unnecessary and could actually degrade current retrieval performance. - D is Incorrect: Supervised fine-tuning is costly, time-consuming, and operationally heavy. A simple instruction prompt adjustment is the standard first step before resorting to model weight updates. > Exam Tip / AWS Best Practice: Diagnose RAG failures precisely: If retrieval is accurate but the answer is wrong, fix the prompt template (generator). If retrieval is inaccurate, fix embeddings/chunking (retriever).
### Question 10 - Combining Models for Performance and Cost Optimization A large e-commerce platform uses an ensemble of three XGBoost models to predict customer churn. The models are currently hosted on three separate Amazon SageMaker AI endpoints. The engineering team wants to reduce inference costs and improve endpoint utilization without significantly compromising prediction accuracy. The models have similar latency profiles and memory footprints. Which architectural change will best meet these requirements? - A) Use Amazon SageMaker AI Multi-Model Endpoints (MME) to host all three models behind a single endpoint and serving container. - B) Distill the three XGBoost models into a single large neural network using Amazon Bedrock. - C) Fine-tune one of the XGBoost models on the combined datasets and delete the other two. - D) Convert the models to a serverless architecture using AWS Lambda and store them in Amazon ECR.	Correct Answer: A Correct Answer Analysis: - A is Correct: Amazon SageMaker AI Multi-Model Endpoints (MME) allow you to deploy multiple models on the same endpoint and serving container. This improves endpoint utilization and reduces costs compared to maintaining separate endpoints for each model, while keeping the ensemble intact for accuracy. Distractor Analysis: - B is Incorrect: Knowledge distillation in Bedrock is a technique for foundation models (LLMs), not an optimization technique for an ensemble of XGBoost models. Converting XGBoost to a neural network would likely increase costs and complexity. - C is Incorrect: Keeping only one model destroys the ensemble, which could lead to a significant compromise in prediction accuracy, violating the constraint. - D is Incorrect: Using AWS Lambda for machine learning inference requires building custom serving logic, handling model loading manually, and dealing with Lambda's payload limits. It is a higher-overhead pattern compared to SageMaker MME. > Exam Tip / AWS Best Practice: To reduce costs and increase utilization for multiple models with similar footprints, consolidate them using SageMaker AI Multi-Model Endpoints (MME) or Multi-Container Endpoints.