Beyond the Prompt

Prompt engineering can take you far, but sometimes you need a model that fundamentally understands a specific "Brand Voice," a complex internal JSON format, or a specialized technical vocabulary not found on the internet. This is where Fine-tuning comes in.

In the AWS Certified Generative AI Developer – Professional exam, you must demonstrate when and how to fine-tune a model, specifically focusing on Parameter-Efficient Fine-Tuning (PEFT).

---

1. Fine-tuning vs. RAG (The Professional Choice)

This is the most common decision-point in Domain 4.

Feature	RAG	Fine-Tuning
Why?	To give the model Facts and New Knowledge.	To give the model Style, Format, or Domain Dialect.
Cost	Low (Pay-per-token).	High (Training costs + Provisioned Throughput).
Latency	Medium (Searching the DB takes time).	Low (The model just 'knows' how to respond).
Complexity	Database management.	Dataset curation and training.

The Golden Rule: Start with RAG. Only fine-tune if RAG cannot achieve the required Formatting or Tone.

---

2. Parameter-Efficient Fine-Tuning (PEFT)

Traditional fine-tuning updates all the billions of parameters in a model. This is incredibly slow and expensive. PEFT techniques only update a small fraction of the parameters, making it possible to train models in hours instead of weeks.

LoRA (Low-Rank Adaptation)

The most popular PEFT technique.

• Instead of modifying the massive weight matrices of the model, you add small "Adapter" layers on top.
• During training, only the weights in these small adapters change.
• The Benefit: You keep the base model "frozen" and only swap the tiny adapters for different tasks.

---

3. Fine-tuning in Amazon Bedrock

AWS provides a serverless way to fine-tune models like Amazon Titan and Meta Llama.

The Workflow:

1. Prepare Data: Create a .jsonl file with "Prompt" and "Completion" pairs.
2. Upload to S3: The training service needs access to the data.
3. Start Job: Use the Bedrock console or API to start a Model Customization Job.
4. Provisioned Throughput: To use your fine-tuned model, you MUST purchase provisioned throughput (a dedicated model instance).

graph LR
    D[JSONL Dataset] --> S3[S3 Bucket]
    S3 --> B[Bedrock Customization Job]
    B --> M[Custom Model: 'My-Titan-v1']
    M --> P[Provisioned Throughput]
    P --> U[User Request]

---

4. Dataset Curation (The Hard Part)

A fine-tuned model is only as good as its training data.

• Quality over Quantity: 1,000 high-quality, human-verified examples are better than 10,000 messy ones.
• Diversity: If you only show the model "Summary" examples, it will lose its ability to "Chat."

Code Example: Training Data Format (JSONL)

{"prompt": "Summarize this ticket: [TICKET_1]", "completion": "Resolution: Hardware failure."}
{"prompt": "Summarize this ticket: [TICKET_2]", "completion": "Resolution: Software bug."}

---

5. Decision Factors for the Exam

You will be asked to choose between Bedrock and SageMaker for fine-tuning.

• Choose Bedrock for simplicity and supported models (Titan, Llama).
• Choose SageMaker if you need full control over the training script, hyper-parameters, or if you are using a model from Hugging Face that Bedrock doesn't support.

---

6. Pro-Tip: The "Task-Specific Adapter"

In a professional enterprise architecture, you can have one "Base" model and ten different LoRA Adapters:

• One for Customer Support.
• One for Sales Emails.
• One for Code Generation. This saves massive amounts of storage and compute compared to having ten different 70B parameter models.

---

Knowledge Check: Test Your Tuning Knowledge

---

Summary

Fine-tuning is the "Precision Surgery" of AI. It's expensive and technical, but it provides a level of control that prompting cannot match. In the next lesson, we will look at Continued Pre-training—the move from "style" to "deep domain knowledge."

---

Next Lesson: Scaling the Mountain: Continued Pre-training