Skip to content

Skill 2.1.2: Identify fine-tuning strategies for pre-trained FMs to meet business needs

Question 1 - Prompt Engineering vs. Fine-tuning Decision

A healthcare technology company is building an application that summarizes clinical notes for physicians. The team selects Amazon Bedrock with the Claude 3 model. During initial evaluation, they notice the model produces coherent summaries but fails to consistently use the company's standardized medical abbreviations (e.g., "HTN" for hypertension, "CBC" for complete blood count). Engineers have added a system prompt instructing the model to use standard abbreviations, but evaluations show the model still inconsistently expands certain abbreviations.

The ML team needs to resolve this terminology inconsistency while minimizing training costs and operational overhead. The solution must also ensure model behavior stays consistently on-domain without degrading general language capabilities.

Which approach should the team implement?

A) Enable continuous pre-training on the company's entire electronic health record (EHR) corpus B) Perform supervised fine-tuning on a dataset of query-response pairs demonstrating correct abbreviation usage C) Switch to a larger foundation model with more parameters D) Add more few-shot examples to the prompt with explicit abbreviation mappings

Click to reveal Answer & Explanation

Correct Answer: B

Correct Answer Analysis

  • B is Correct: Supervised fine-tuning (SFT) is the appropriate next step when prompt engineering (instructions + few-shot examples) fails to consistently enforce domain-specific terminology and formatting requirements. By training on labeled JSONL pairs showing inputs alongside correctly abbreviated outputs, the model's weights are updated to consistently adopt the target clinical abbreviations. This directly addresses the persistent gap.

Distractor Analysis

  • A is Incorrect: Continuous pre-training on raw EHR data would update the model's general knowledge but is not designed to enforce specific output formatting rules like abbreviation usage. It also requires massive compute resources and risks introducing uncurated clinical content.
  • C is Incorrect: Increasing model size may improve general reasoning but does not guarantee adherence to organization-specific terminology conventions. A larger model with the same prompt would likely exhibit the same abbreviation inconsistency while significantly increasing inference cost and latency.
  • D is Incorrect: The scenario explicitly states that prompt instructions and few-shot examples have already been tried and evaluations still show inconsistent behavior. Adding more examples without changing weights is unlikely to close the gap and may hit context window limits.

Key Takeaway: Follow the progression: Prompt Engineering → Evaluation → Fine-tuning. Fine-tuning is warranted only when evaluations demonstrate a consistent gap that prompting cannot close.

Question 2 - Fine-tuning Strategy Selection: SFT vs. RLHF

A financial services firm wants to deploy a customer service chatbot using Amazon Bedrock. The bot must:

  1. Decline requests outside the financial domain politely
  2. Follow the company's compliance-approved response templates verbatim for sensitive topics
  3. Maintain consistent tone across all interactions

The team has prepared 500 labeled examples of "good" and "bad" responses. They also have a scoring rubric that evaluates responses on helpfulness, safety, and compliance adherence.

Which fine-tuning strategy should the ML team select to BEST meet these requirements?

A) Reinforcement fine-tuning (RLHF) using the scoring rubric as a reward function B) Supervised fine-tuning using only the 500 labeled "good" responses as response targets C) Distillation from a larger teacher model that already follows compliance rules D) No fine-tuning; add the rubric as a system prompt with the 500 examples appended

Click to reveal Answer & Explanation

Correct Answer: B

Correct Answer Analysis

  • B is Correct: Supervised fine-tuning with labeled input-output pairs is the correct approach when you have examples of expected responses. The 500 labeled "good" responses serve as training targets, teaching the model to produce compliant, on-tone answers for financial queries. This is the standard SFT path.

Distractor Analysis

  • A is Incorrect: RLHF requires a reward model or human preference ranking (chosen vs. rejected responses), not just a rubric. Additionally, RLHF is typically more complex and compute-intensive than SFT, and should be reserved for cases where defining explicit correct responses is difficult. Here, the team already has labeled "good" responses.
  • C is Incorrect: Distillation requires a teacher model that already exhibits the desired behavior. The scenario does not mention an existing compliant teacher model, and distillation would add unnecessary complexity when labeled data already exists for SFT.
  • D is Incorrect: The scenario implies that the rubric and examples are evaluation materials, not prompt content. Appending 500 examples to every prompt would be extremely expensive (high token usage, high latency) and may exceed context window limits.

Key Takeaway: SFT is the default fine-tuning choice when you have labeled examples of expected behavior. RLHF and distillation are specialized paths for when explicit desired outputs are hard to define or when a compliant teacher model already exists.

Question 3 - Distillation for Latency-Critical Deployment

A real-time fraud detection system uses a large language model through Amazon Bedrock to analyze transaction descriptions and flag suspicious patterns. The current model (Claude 3 Opus) achieves high accuracy but exceeds the 200ms response-time SLA for high-volume transaction screening.

The ML team wants to maintain accuracy while reducing inference latency and cost. They have collected 100,000 transaction descriptions paired with the current model's high-quality classifications.

Which approach achieves the business requirements with the LEAST additional training complexity?

A) Enable Model Distillation in Amazon Bedrock, using the Claude 3 Opus as the teacher model and a smaller model (e.g., Claude 3 Haiku) as the student B) Continue using Claude 3 Opus but enable provisioned throughput with higher commitment C) Perform supervised fine-tuning on Claude 3 Sonnet using the same 100,000 labeled examples D) Switch to on-demand throughput and batch similar transactions before sending them to the model

Click to reveal Answer & Explanation

Correct Answer: A

Correct Answer Analysis

  • A is Correct: Amazon Bedrock Model Distillation is purpose-built for this scenario: transferring knowledge from a larger, high-performing teacher model to a smaller, faster student model. The 100,000 transaction-prompt pairs can be used to generate teacher responses, and the student model (Haiku) learns to mimic the teacher's outputs, achieving comparable accuracy with significantly lower latency and cost—critical for the sub-200ms SLA.

Distractor Analysis

  • B is Incorrect: Provisioned throughput guarantees capacity but does not reduce the inherent inference latency of a large model like Opus. It also increases cost, which does nothing to solve the latency violation.
  • C is Incorrect: Supervised fine-tuning on Claude 3 Sonnet still uses a relatively large model with higher latency than a lightweight model like Haiku. It may not meet the strict latency SLA, and SFT is less purpose-built for transferring existing model behavior than distillation.
  • D is Incorrect: Batching transactions introduces artificial delays because individual transactions must wait for enough volume to accumulate before processing, directly conflicting with real-time SLA requirements.

Key Takeaway: Model distillation transfers knowledge from large models (teacher) to smaller, faster models (student) and is ideal for latency-critical or cost-sensitive deployments.

Question 4 - Fine-tuning Data Preparation

A media streaming company wants to fine-tune a Claude 3 Sonnet model for generating movie recommendation explanations. Their data scientists have collected:

  • 10,000 user reviews from their platform
  • 5,000 conversations where users asked for recommendations
  • Movie metadata from an internal database
  • 1,000 curated pairs of "user query → ideal recommendation response" written by content editors

What is the correct data preparation step for supervised fine-tuning in Amazon Bedrock?

A) Upload the raw user reviews and movie metadata as the training dataset in CSV format B) Combine all four data sources into a single corpus and enable continuous pre-training C) Format only the 1,000 curated pairs as JSONL files with prompt and completion (or messages) fields for the fine-tuning job D) Use the 5,000 user conversations directly as the fine-tuning dataset without labeling

Click to reveal Answer & Explanation

Correct Answer: C

Correct Answer Analysis

  • C is Correct: Supervised fine-tuning in Amazon Bedrock requires labeled input-output pairs in JSONL format, typically structured with prompt/completion or a conversational messages array. The 1,000 curated pairs written by content editors represent clean, high-quality labeled examples of the desired behavior, making them the ideal SFT dataset.

Distractor Analysis

  • A is Incorrect: Raw user reviews and metadata are unstructured or tabular data, not labeled prompt-response pairs. SFT requires examples showing what the ideal output should look like for a given input, not raw content dumps.
  • B is Incorrect: Continuous pre-training feeds a large corpus of raw text to broaden a model's general knowledge, but it does not teach the model how to follow instruction-response patterns for specific tasks like generating recommendation explanations.
  • D is Incorrect: User conversations may contain inconsistent, off-topic, or low-quality responses. Using them without filtering or labeling would teach the model noisy behavior. Curated, editor-approved examples are essential for reliable SFT.

Key Takeaway: Supervised fine-tuning requires a clean, labeled dataset of high-quality input-output pairs in JSONL format. Raw conversational logs or metadata are not a substitute for curated training data.

Question 5 - Catastrophic Forgetting Prevention

A legal tech startup has fine-tuned a foundation model on 50,000 court opinions to generate legal document summaries. After deployment, users report that while the model handles legal summarization well, it has become confused when asked to handle general conversations or perform basic tasks like drafting polite emails.

The ML team must restore general capability while preserving the legal summarization performance.

Which mitigation strategy should they implement during their next fine-tuning iteration?

A) Increase the fine-tuning learning rate and run additional epochs on the legal dataset to strengthen domain performance B) Mix general-domain examples (from a diverse corpus) with the legal summarization examples in the fine-tuning dataset C) Remove the fine-tuned model and revert to the base foundation model, using only prompt engineering for legal tasks D) Fine-tune exclusively on a larger set of court opinions to further specialize the model

Click to reveal Answer & Explanation

Correct Answer: B

Correct Answer Analysis

  • B is Correct: Catastrophic forgetting occurs when fine-tuning is performed exclusively on a narrow domain, causing the model to lose previously acquired general capabilities. The standard mitigation is to blend a subset of general-domain examples with the specialized task data during SFT. This mixed dataset allows the model to learn the legal summarization task while maintaining general language understanding.

Distractor Analysis

  • A is Incorrect: Increasing the learning rate and epochs would intensify specialization, making catastrophic forgetting worse by overfitting the model to legal data.
  • C is Incorrect: Reverting loses the valuable fine-tuning investment and forces reliance on prompting alone, which may be insufficient for the specialized legal summarization quality the team achieved.
  • D is Incorrect: Training on more legal data without mixing in general examples will further degrade general language ability, worsening the problem.

Key Takeaway: Prevent catastrophic forgetting by mixing domain-specific and general-domain examples during fine-tuning. A narrow training corpus overfits the model to that domain and degrades general capability.

Question 6 - Fine-tuning vs. RAG Selection

An e-commerce platform wants to build a customer support assistant that answers questions about return policies, shipping timelines, and product availability. The policies change frequently (approximately weekly), and the assistant must cite the specific policy document for every answer.

The development team is considering fine-tuning a foundation model on their policy documents versus implementing a RAG approach using Amazon Bedrock Knowledge Bases.

Which approach BEST satisfies the requirements with minimal operational overhead?

A) Fine-tune a foundation model on all current policy documents, scheduling weekly fine-tuning jobs to refresh the model B) Fine-tune on historical policies and use RAG only for novel queries that the fine-tuned model cannot answer C) Use RAG with Amazon Bedrock Knowledge Bases, syncing the knowledge base whenever policies are updated D) Perform continuous pre-training on policy document changes to keep the model parameters current

Click to reveal Answer & Explanation

Correct Answer: C

Correct Answer Analysis

  • C is Correct: RAG is the ideal architecture for scenarios with frequently changing content and citation requirements. Amazon Bedrock Knowledge Bases automatically handles document ingestion, chunking, embedding, and vector storage. When policies change, the team only needs to trigger a sync to update the vector index—no training jobs required. Each retrieved chunk carries source metadata, enabling the model to cite the specific policy document behind each answer.

Distractor Analysis

  • A is Incorrect: Weekly fine-tuning jobs require significant manual effort, GPU compute, and model version management. This creates unnecessary operational overhead compared to a simple knowledge base sync. Additionally, fine-tuned models cannot easily cite sources.
  • B is Incorrect: This hybrid adds unnecessary complexity. RAG alone handles all the requirements without the overhead of maintaining a fine-tuned model. Combining both doubles infrastructure complexity without clear benefit.
  • D is Incorrect: Continuous pre-training is designed for expanding a model's general knowledge breadth, not for transactional content updates like policy changes. It is compute-intensive and does not provide source attribution.

Key Takeaway: Choose RAG for frequently changing factual content and citation requirements. Choose fine-tuning for teaching consistent style, format, or reasoning associated with relatively stable domain patterns.

Question 7 - Fine-tuning Hyperparameter Selection

A machine learning team is fine-tuning a foundation model on a specialized medical transcription dataset containing 10,000 high-quality examples. They want to ensure the model learns the domain-specific formatting and terminology without overfitting.

They plan to run supervised fine-tuning in Amazon Bedrock.

Which hyperparameter configuration should they select for the training job?

A) Set a very high learning rate (e.g., 0.1) over a short number of epochs to force rapid convergence on domain terms B) Disable fine-tuning validation entirely and use the full dataset for training to maximize training signal C) Use a moderate learning rate (default or slightly lower) with a limited number of epochs and enable validation to monitor for divergence from the base model D) Set the epoch count as high as possible to ensure the model thoroughly memorizes all 10,000 examples

Click to reveal Answer & Explanation

Correct Answer: C

Correct Answer Analysis

  • C is Correct: Fine-tuning best practice is to use a moderate learning rate that is lower than the original pre-training rate to avoid disrupting the model's pre-trained representations. A limited number of epochs (typically 2-5) prevents overfitting, and enabling validation helps monitor whether the model is learning the domain task without degrading general capacity.

Distractor Analysis

  • A is Incorrect: A very high learning rate can cause catastrophic forgetting by aggressively overriding pre-trained weights. It may produce unstable training and poor generalization.
  • B is Incorrect: Validation data is critical for monitoring training health and detecting overfitting. Using all data for training removes the ability to evaluate whether the model is learning patterns versus memorizing examples. Amazon Bedrock automatically holds out a validation split if configured.
  • D is Incorrect: Running many epochs on a small dataset causes overfitting. The model memorizes the training examples rather than learning generalizable rules, resulting in poor performance on unseen queries.

Key Takeaway: Fine-tune with a lower-than-pretraining learning rate and limit epochs to prevent overfitting and catastrophic forgetting. Validation monitoring is essential for fine-tuning success.

Question 8 - Evaluating Fine-tuned Model Baselines

An ML team has completed a supervised fine-tuning job on a foundation model to improve instruction-following for a custom code-generation task. Before deploying to production, they need to demonstrate that fine-tuning actually improved performance.

They plan to use SageMaker Clarify to evaluate the fine-tuned model. However, the evaluation results require a comparison baseline.

Which two evaluation baselines should the team establish? (Select TWO)

A) A baseline using the base (non-fine-tuned) foundation model on the same evaluation dataset B) A baseline using a completely different ML algorithm (e.g., XGBoost) applied to the same task C) A human-evaluated baseline from expert annotators scored on the same task D) A baseline using the fine-tuned model itself on a truncated evaluation dataset to establish internal consistency E) A baseline using the fine-tuned model with a randomly shuffled version of the evaluation dataset

Click to reveal Answer & Explanation

Correct Answer: A, C

Correct Answer Analysis

  • A is Correct: Comparing the fine-tuned model against the base foundation model is the standard evaluation methodology. It demonstrates that fine-tuning contributed measurable improvement beyond what the original model could already achieve.
  • C is Correct: Human-annotated ground truth or expert evaluation provides the gold standard baseline. For tasks like code generation, expert reviews can establish what the expected quality should be. Comparing model output to human benchmarks shows how close the model is to professional-level performance.

Distractor Analysis

  • B is Incorrect: XGBoost is a traditional ML algorithm, not a generative language model. It cannot perform code generation and is an inappropriate baseline for evaluating a fine-tuned LLM.
  • D is Incorrect: Using the fine-tuned model to evaluate itself on a truncated dataset does not establish whether fine-tuning improved anything. It only tests internal consistency and provides no comparative signal.
  • E is Incorrect: Shuffled evaluation data as a baseline tests whether the model relies on patterns within the data structure, but it is not a standard or meaningful comparison for demonstrating fine-tuning gains.

Key Takeaway: Always compare a fine-tuned model against the base foundation model and human benchmarks when available. This isolates the value added by fine-tuning and anchors it to real expected quality.

Question 9 - Provisioned Throughput for Fine-tuned Deployment

After completing fine-tuning on a Claude 3 Haiku model using Amazon Bedrock, a logistics company needs to deploy the model to handle real-time shipment tracking queries. Their API receives approximately 500 concurrent requests during peak hours, and each request must be processed within 500ms.

The team must ensure consistent, low-latency inference for the fine-tuned model.

What should they do?

A) Use Amazon Bedrock on-demand throughput with default quotas and rely on API retries to handle throttling B) Purchase provisioned throughput for the fine-tuned model with sufficient model units to cover the peak concurrent request volume C) Deploy the model to Amazon SageMaker Endpoint using a multi-model endpoint with autoscaling D) Use AWS Lambda functions to queue requests and batch them before sending to Bedrock

Click to reveal Answer & Explanation

Correct Answer: B

Correct Answer Analysis

  • B is Correct: For fine-tuned Amazon Bedrock models, provisioned throughput is the correct deployment option for production workloads requiring consistent performance. It provides guaranteed capacity and latency, ideal for handling 500 concurrent requests with a 500ms SLA. The team can estimate required model units based on concurrent token generation needs.

Distractor Analysis

  • A is Incorrect: On-demand throughput is subject to throttling under high concurrency and provides no latency SLAs. Relying on retries would likely violate the 500ms response time requirement under peak load.
  • C is Incorrect: Fine-tuned Bedrock models are served through Amazon Bedrock provisioned throughput, not SageMaker endpoints. Moving the custom model to SageMaker requires arbitrary model deployment infrastructure and loses Bedrock-specific optimizations.
  • D is Incorrect: Batching queued requests in Lambda functions adds latency because requests wait for batch accumulation, conflicting with real-time requirements. Lambda's 15-minute max execution time and cold starts also introduce unpredictable latency.

Key Takeaway: Use Bedrock provisioned throughput for production fine-tuned model deployments requiring guaranteed concurrency and latency. On-demand throughput is suitable for development or low-volume testing only.

Question 10 - Fine-tuning Mode Selection (Full vs. Parameter-Efficient)

A startup with limited GPU budget wants to fine-tune a large language model to generate product descriptions in their brand's voice. They do not need to modify the model's general language understanding—only its stylistic output patterns. The model has billions of parameters.

They are concerned about training costs and want to minimize the compute resources required while achieving acceptable domain adaptation.

Which fine-tuning approach should they choose?

A) Full fine-tuning using SageMaker distributed training to modify all model parameters B) Parameter-efficient fine-tuning (PEFT) using Amazon Bedrock's supported fine-tuning modes (e.g., LoRA-based approaches) C) Fine-tune only the model's embedding layer while keeping transformer blocks frozen D) Use in-context learning with an expanded few-shot example set instead of fine-tuning

Click to reveal Answer & Explanation

Correct Answer: B

Correct Answer Analysis

  • B is Correct: Parameter-efficient fine-tuning (PEFT) modifies only a small subset of model parameters while keeping the base model frozen. This dramatically reduces training compute, memory, and cost while still achieving domain adaptation for tasks like adjusting output style. Amazon Bedrock supports fine-tuning modes optimized for this use case.

Distractor Analysis

  • A is Incorrect: Full fine-tuning modifies all model parameters and requires significant GPU resources—typically out of budget for startups with limited compute. It is unnecessary when only surface-level stylistic changes are needed.
  • C is Incorrect: Fine-tuning only the embedding layer is too limited because embedding layers have relatively few parameters and would not effectively capture instruction-following or stylistic patterns. This option is not a supported or effective fine-tuning method in Amazon Bedrock.
  • D is Incorrect: While in-context learning is cheap and fast, it has limitations on how many examples fit in the context window. The scenario indicates fine-tuning is desired because the team wants persistent domain adaptation; adding more examples is unlikely to consistently achieve the stylistic consistency needed.

Key Takeaway: Choose parameter-efficient fine-tuning when the adaptation is surface-level (style, format, tone) or when computational resources are constrained. Full fine-tuning is rarely the default choice for modern LLMs with billions of parameters.

End of Question Set