Decision Matrix: Which Type to Use and Why?

We have explored the four primary levers of model adaptation:

Supervised Fine-Tuning (SFT): Teaching commands.
Few-Shot Learning: Real-time steering.
Transfer Learning: Reusing task-based logic.
Domain-Specific Tuning (CPT): Learning new vocabulary.

But in a real engineering sprint, how do you choose? If you pick the wrong one, you waste weeks of work. If you pick the right one, you can outperform teams 10x your size. In this final lesson of Module 3, we provide the Architect’s Decision Matrix.

The Strategic Decision Matrix

To select your approach, evaluate your project across three dimensions: Data Availability, Target Precision, and Operational Budget.

1. Data Availability (The "What do I have?" Filter)

< 20 Examples: USE FEW-SHOT. You don't have enough data to move weights.
100 - 1,000 "Golden" Examples: USE SFT. This is the sweet spot for style and formatting.
> 100,000 Documents (Raw): USE CONTINUAL PRE-TRAINING (CPT). This is only for deep domain mastery.

2. Target Precision (The "How good does it need to be?" Filter)

Creativity & Flexiblity: USE FEW-SHOT. Let the model's base intelligence shine.
Rigid Schemas (JSON/XML): USE SFT. Prompts will eventually fail; weights won't.
Professional Persona: USE SFT. You want the "Vibe" to be instinctive.

3. Operational Budget (The "Can I afford it?" Filter)

Low Budget (< $100): USE FEW-SHOT. Zero upfront cost.
Medium Budget ($100 - $5,000): USE SFT (via LoRA). High ROI.
Enterprise Budget (> $50k): USE CPT -> SFT Pipeline. Only for building industry-leading proprietary models.

Visualizing the Technical Path

graph TD
    A["Project Start"] --> B["Do you have Labeled Examples?"]
    B -- No --> C["Do you have Raw Text?"]
    C -- Yes --> D["Continual Pre-training (CPT)"]
    C -- No --> E["Few-Shot / Prompting"]
    
    B -- Yes --> F["How many examples?"]
    F -- "< 50" --> E
    F -- "100 - 10,000" --> G["Supervised Fine-Tuning (SFT)"]
    F -- "> 1M" --> H["Full Fine-Tuning or RLHF"]
    
    D --> I["Refine with SFT"]
    G & E --> J["Production Endpoint"]
    I --> J

Combining Strategies: The Most Common Production Paths

In professional AI companies, you rarely use "Pure" versions. You use these Battle-Tested Combinations:

Path A: The "Brand Builder" (SFT + Few-Shot)

Goal: A bot that talks exactly like your brand but knows it's talking to this specific user.
Action: Fine-tune (SFT) the model on your brand voice. Then, use Few-Shot to put the user's specific history in the prompt.

Path B: The "Domain Expert" (CPT + SFT)

Goal: A bot that understands complex legal contracts and can summarize them for non-lawyers.
Action: Use CPT for 1,000,000 legal tokens. Then use SFT for 500 "Summary Commands."

Path C: The "Agentic Specialist" (SFT + RAG)

Goal: A coding assistant that knows your proprietary API and can generate valid requests.
Action: Fine-tune (SFT) on the API's syntax and style. Use RAG to fetch the current documentation for the specific endpoint.

The Decision Matrix Table

Requirement	Few-Shot	SFT	Transfer Learning	CPT
Stylistic Consistency	Poor	Excellent	Good	Poor
Real-time Data	Excellent	Poor	Poor	Poor
Strict JSON Schema	Fair	Excellent	Good	Poor
New Domain Vocab	Poor	Fair	Fair	Excellent
Latency Optimization	Poor	Excellent	Excellent	Good

Summary and Key Takeaways

Few-Shot is for the R&D and rapid prototyping phases.
SFT is the workhorse for production reliability and operational efficiency.
CPT is reserved for deep domain immersion (Bio, Law, Finance).
Decision Hack: Most projects that think they need CPT actually just need better SFT and a solid RAG pipeline.

You have now completed Module 3! You know exactly what path to take. In Module 4, we will explore the Use Cases That Justify Fine-Tuning, putting these strategies into real-world business contexts.

Final Module Reflection

Look at your Capstone Project goal (from Module 1). Which of these four techniques is your "Primary" path?
If you had to pick a "Secondary" path to augment it, which would it be?

SEO Metadata & Keywords

Focus Keywords: Fine-Tuning Decision Matrix, Choosing SFT vs CPT, When to use Few-Shot Learning, AI Model Strategy for Businesses, LLM Tuning Architecture. Meta Description: Master the architect's decision matrix for AI fine-tuning. Learn how to choose between SFT, CPT, and few-shot learning based on data, budget, and task goals.