📚 Amazon Bedrock Fine-Tuning & Model Selection

1. Different Providers & Model Capabilities

  • Providers: Anthropic, Amazon, DeepSeek, Stability AI, etc.
  • Models vary in strengths:
    • Claude 3.5 Haiku → Best for text tasks.
    • Amazon Nova Reel → Text-to-video / Image-to-video.
  • Exam Tip: You will not be tested on which is best, only on what each can or cannot do.

2. Comparing Models

  • Compare Mode: Test models side-by-side in Bedrock playground.
  • Compare by:
    • ✅ Capabilities (text, image, video)
    • ✅ Output style/format
    • ✅ Speed (latency)
    • ✅ Cost (token usage)
  • Example:
    • Nova Micro: ❌ No image upload, faster, shorter responses.
    • Claude 3.5 Sonnet: ✅ Image support, longer/more detailed answers.

3. Fine-Tuning Methods – Comparison Table

Feature Instruction-Based Fine-Tuning Continued Pre-Training Transfer Learning
Data Type Labeled (prompt–response pairs) Unlabeled (raw text) Labeled or Unlabeled
Goal Improve performance on domain-specific tasks Make model expert in a specific domain Adapt a pre-trained model to a new but related task
Example Train chatbot to respond in a specific tone Feed all AWS docs to become AWS expert Adapt GPT for medical text classification
Changes Model Weights? ✅ Yes ✅ Yes ✅ Yes
Complexity Medium High Varies
Cost Lower (less data) Higher (more data) Varies
Exam Keyword “Labeled data”, “prompt-response” “Unlabeled data”, “domain adaptation” “Adapt model to new task”
Bedrock Support Supported on some models Supported on some models General ML concept (not Bedrock-specific)

Instruction-based Fine Tuning

Continued Pre-training

---

4. Messaging Fine-Tuning

  • Single-Turn Messaging:
    • One Q → One A
    • Optional system context

  • Multi-Turn Messaging:
    • Full conversations, alternating user and assistant
    • Used for chatbot training in multi-step dialog

5. Transfer Learning

  • Definition: Using a pre-trained model for a new but related task.
  • Common in:
    • Image classification
    • NLP (BERT, GPT)
  • Exam Tip: If question is general ML → choose Transfer Learning as the answer.
    If it’s about Bedrock & domain-specific → choose Fine-Tuning.

6. Fine-Tuning Requirements in Amazon Bedrock

  • Training data must:
    • Be in Amazon S3
    • Follow specific formatting

  • Provisioned Throughput is required for:
    • Creating the custom model
    • Using the custom model

  • Not all models can be fine-tuned (usually open-source models are supported).

7. Common Use Cases

  • Chatbot with specific persona, tone, or target audience
  • Adding up-to-date knowledge
  • Integrating exclusive private data (customer logs, internal documents)
  • Improving categorization, accuracy, or response style

8. Exam Tips

  • Keyword Mapping:
    • “Labeled data” → Instruction-based Fine-Tuning
    • “Unlabeled data” / “Domain adaptation” → Continued Pre-Training
    • “Adapt to a new related task” → Transfer Learning
  • Provisioned Throughput is a must for custom models in Bedrock.
  • Fine-tuning changes weights of the base model → creates a private version.
  • Compare models not just on quality, but also latency and token cost.

9. Good to know

  • Re-training an FM requires a higher budget
  • Instruction-based fine-tuning is usually cheaper as computations are less intense and the amount of data required usually less
  • It also requires experienced ML engineers to perform the task
  • You must prepare the data, do the fine-tuning, evaluate the model
  • Running a fine-tuned model is also more expensive (provisioned throughput)

10. Provisioned Throughput (AWS Bedrock)

Definition:
Reserving a fixed amount of processing capacity for your custom (fine-tuned) model.

Why Needed

  • Fine-tuned models run on dedicated resources.
  • Ensures consistent speed, low latency, and predictable costs.
  • Prevents slowdowns during high demand.

With vs Without

  • Without: Performance can drop when traffic spikes.
  • With: Guaranteed performance for the reserved capacity.

Exam Tip: For custom models in Bedrock, provisioned throughput is required.