Exams / AIP-C01 / Study guide / Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry)
Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry): AIP-C01 study notes and practice questions
What to know about Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry) for the AIP-C01 exam, from the Foundation Model Integration, Data Management, and Compliance domain (31% of the exam guide), then 1 free question to test it. Part of the AIP-C01 study guide.
These questions describe a team that wants a model to behave a certain way, or wants one tuned model shared across teams, and ask which customization and which lifecycle controls fit.
Fine-tuning or RAG
Use RAG when the model needs facts that change or are private, and answers must cite sources. Updating facts means re-syncing documents.
Use fine-tuning when the model must learn a behavior: a house style, a strict format, domain terminology or a task pattern that prompting does not hold reliably.
Combine them when needed: fine-tune for behavior, retrieve for facts.
Fine-tuning is not the answer to “add this quarter’s price list.”
Bedrock customization options
Fine-tuning trains on labeled prompt and response examples in S3, usually JSONL. Continued pre-training uses unlabeled domain text.
Distillation trains a smaller student model on the outputs of a larger teacher to cut cost and latency per request.
Custom Model Import brings an open-weight model fine-tuned elsewhere, for example on SageMaker AI, into Bedrock for managed serving. See custom model import.
LoRA adapters
LoRA trains small adapter weights on top of a frozen base model. Many domain variants can share one base, so training and storage stay cheap. On SageMaker AI, LMI containers with inference components can serve many adapters on one endpoint that shares the base model. To serve through Bedrock instead, merge the adapter into the base weights and use Custom Model Import.
Lifecycle controls
SageMaker Model Registry groups model versions in model package groups. Each version carries metadata and an approval status. Deployment pipelines should deploy only approved versions.
Automated retraining runs a pipeline that trains on new data, evaluates against a baseline, registers the version and deploys with deployment guardrails. Blue/green deployment with traffic shifting and automatic rollback on CloudWatch alarms limits the blast radius.
Base model lifecycle: a customized model depends on its base model. When the base model reaches Legacy, plan a re-customization on a supported base, then retire the old custom model after cutover.
How to choose
Scenario wording
Choose
“Answers must always follow our approved legal template”
Fine-tuning on approved examples
“Answers change weekly from policy documents”
RAG, not fine-tuning
“Fifty teams need small domain variants of one model”
LoRA adapters on a shared base
“Open-weight model tuned on SageMaker AI, served through Bedrock”
Custom Model Import
“Reduce the cost of a large model’s outputs”
Distillation to a smaller student model
“Promote only tested versions to production”
SageMaker Model Registry approval gate
Common traps
Fine-tuning to add knowledge, when retrieval is the right tool and keeps answers current.
Deploying the newest training run without an approval step in the registry.
Skipping data review: training data with personal data needs governance before use.
Ignoring base model deprecation, which strands custom models.
Expecting fine-tuning to stop hallucinations about facts that were never in the training data.
A furniture retailer writes about 40,000 product listings a year, and its copywriters have approved about 10,000 finished listings as examples of the brand voice and the required JSON layout.
The listings must sound like the brand and follow the same JSON structure every time, so the storefront can parse them. Each listing also contains facts such as dimensions, price and stock status, and these change daily.
The team wants a single approach that keeps the brand voice and layout consistent, keeps the facts current, and keeps each request under the model's context limit.
Which approach meets these requirements?
Choose one answer.
AA. Fine-tune a model on the approved listings to learn the voice and JSON layout, and pass each day's product facts in the request from the catalog API.
BB. Train a model every night on the full catalog, so that it learns both the brand voice and the day's facts, and skip any lookup step at request time.
CC. Use prompt engineering alone, placing all 10,000 approved listings in the system prompt of every request so that the model copies their voice and layout.
DD. Fine-tune a model on the product catalog's facts, so that it memorizes each item's price and dimensions, and repeat the fine-tuning every week to keep those facts up to date.
B. Wrong: A nightly job means facts for products added during the day are unknown until the next run, and the training cost repeats every night. It also mixes changing facts into the model, which is the wrong place for them.
C. Wrong: Ten thousand listings are far too many for one context window, and repeating them in every request would be slow and costly. The approach misses the cost and limit requirements.
D. Wrong: Fine-tuning does not reliably memorize individual facts, and prices and stock change daily, so weekly training would still leave stale answers. The approach misses the need for current facts.
Explanation
Style and output structure are patterns that fine-tuning learns well from examples, while changing facts belong in the request, pulled from the catalog. Training nightly or weekly keeps facts stale and costly, and putting all examples in every prompt exceeds the context limit.
Want more on Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry)? The 5 full-length practice sets have 5 more Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry) questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →
Study Fine-tuning, LoRA adapters and custom model lifecycle (Bedrock, SageMaker AI, Model Registry)