Domain 3 of 5 · Chapter 3 of 4

Training & Fine-Tuning

Unlock the complete study guide + 1,040 practice questions across 16 full exams.

Bundled into the existing AWS Certified AI Practitioner AIF-C01 premium course — no separate purchase.

Included in this chapter:

  • The training spectrum: from scratch to surgical
  • Data prep and the AWS mechanics of customization
  • Decide: fine-tune vs RAG vs prompt engineering

FM customization approaches compared

AspectPre-trainingFine-tuningContinued pre-trainingRAG
Changes model weights?Yes, creates the model from scratchYes, adjusts existing weightsYes, adjusts existing weightsNo, adds context at inference time
Data neededVast, broad, mostly unlabeled corpusSmaller labeled task-specific examplesLarge unlabeled domain-specific textExternal knowledge source / documents to retrieve
CostExtremely high (rarely done by customers)Moderate; per-token training + model storageHigher than fine-tuning; per-token training + storageLow one-time setup; pay per retrieval and tokens
Best forBuilding a new foundation modelSpecialized tone, format, or task behaviorDeepening knowledge of an entire domainKeeping answers current with changing or proprietary facts

Decision tree

How to customize a foundation model? Must you change the model's weights? Need fresh or private facts at query time? No — inference only Have labeled prompt- completion examples? Yes — bake it in Prompt engineering cheapest, try first; no data, inference-time No RAG inject current/proprietary facts; weights unchanged Yes Align to human preferences? Yes — labeled Continued pre-training unlabeled corpus No Fine-tuning (instruction tuning) labeled task behavior No RLHF human preference rankings align model Yes
Cheat sheet

Cheat sheet

  • Training stages are told apart by their data, not their algorithm
  • Pre-training builds a foundation model from scratch on unlabeled data
  • Fine-tuning adjusts an existing model's weights using labeled task data
  • Continued pre-training deepens domain knowledge from large unlabeled text
  • Distillation transfers a teacher's capability into a smaller, cheaper student
  • Every fine-tuning method changes the model's parameters
  • Instruction tuning teaches a model to follow directions, tone, and format
  • Domain adaptation specializes a general model for one field
  • Transfer learning reuses pre-trained knowledge instead of starting over
  • Parameter-efficient fine-tuning updates only a small slice of parameters
  • RLHF aligns a model to human preferences via a reward model
  • RLHF requires a dataset of human preference judgments
  • A fine-tuned model is only as good as its training data
  • Bad training data makes a fine-tuned model worse, not better
  • Training data must be governed and representative
  • Prompt first, RAG for facts, fine-tune for durable behavior
  • Don't fine-tune to add changing or factual knowledge
  • Fine-tuning carries recurring training, storage, and serving costs

Unlock with Premium — includes all practice exams and the complete study guide.

References

  1. What are Foundation Models?
  2. Customize models in Amazon Bedrock with your own data using fine-tuning and continued pre-training Blog
  3. Custom models, Amazon Bedrock
  4. Model distillation, Amazon Bedrock
  5. What is Transfer Learning?
  6. Customize your model to improve its performance, Amazon Bedrock
  7. Guidelines for model customization, Amazon Bedrock
  8. What is RLHF (Reinforcement Learning from Human Feedback)?
  9. Prepare the datasets, Amazon Bedrock
  10. Train a model with Amazon SageMaker AI
  11. Train models, Amazon SageMaker AI
  12. Amazon Bedrock pricing
  13. Knowledge Bases, Amazon Bedrock