Model Customization Guide - Which Method Should I Use?¶

Confused about how to customize your model? This guide helps you choose the right approach.

🎯 Quick Decision Tree¶

TL;DR: Change behavior → System Prompt (5 min). Have < 50 example Q&As → Few-Shot Learning (30 min). Have 100+ examples and need deep specialization → Fine-Tuning (hours/days, GPU).

For the full step-by-step decision tree, see Essential Concepts - Decision Tree.

📊 Comparison Table¶

Method	Time	Difficulty	When to Use	Cost
System Prompt	5 min	⭐ Easy	Change behavior, add rules	Free
Few-Shot	30 min	⭐⭐ Medium	Add 5-50 examples	Free
Fine-Tuning	Hours+	⭐⭐⭐⭐ Hard	100+ examples, deep specialization	GPU time

90% of users only need Methods 1 or 2!

🔧 Method 1: System Prompt (Easiest)¶

When to Use¶

Change model personality (formal, casual, technical, etc.)
Add specific rules or constraints
Define role (customer support, tutor, assistant)
Set output format (JSON, markdown, etc.)

What You Need¶

No training data required
Just clear instructions

Time Required¶

5-15 minutes

Example Use Cases¶

✅ Make a chatbot more professional ✅ Create a code assistant that explains step-by-step ✅ Build a translator with specific style ✅ Design a creative writer that uses certain themes

How to Do It¶

Step 1: Create a Modelfile

mkdir -p ./models/custom/my-assistant
nano ./models/custom/my-assistant/Modelfile

Step 2: Define behavior with SYSTEM prompt

FROM llama3.2:1b

PARAMETER temperature 0.7
PARAMETER num_ctx 4096
PARAMETER top_p 0.9

SYSTEM """
You are a helpful programming tutor.

Your role:
- Explain concepts in simple terms
- Provide code examples
- Ask clarifying questions if the request is unclear
- Break down complex topics into steps

Style:
- Use clear, beginner-friendly language
- Include emojis for readability
- Always test code before suggesting it
- Admit when you're unsure
"""

Step 3: Create the model

bash scripts/create-custom-model.sh my-assistant ./models/custom/my-assistant/Modelfile

Step 4: Test it

make chat
# Select "my-assistant" and start chatting

Pros¶

✅ Fastest method ✅ No training data needed ✅ Easy to iterate and improve ✅ Works immediately

Cons¶

❌ Limited knowledge beyond base model ❌ Can't add complex specialized knowledge ❌ May not follow instructions perfectly every time

📚 Method 2: Few-Shot Learning (Recommended)¶

When to Use¶

You have 5-50 example interactions
Want model to respond in a specific way
Need consistent tone/format
Teaching specific domain knowledge

What You Need¶

Question/answer pairs
Examples of ideal responses
JSONL file (optional, can use converter)

Time Required¶

30 minutes to 2 hours

Example Use Cases¶

✅ Customer support with FAQs ✅ Product documentation Q&A ✅ Style-specific content generation ✅ Domain-specific terminology

How to Do It¶

Step 1: Prepare your examples

Option A: Manual (for small datasets)

FROM llama3.2:1b

PARAMETER temperature 0.3
PARAMETER num_ctx 4096

SYSTEM """
You are TechCorp's customer support assistant.
Be helpful, professional, and concise.
"""

MESSAGE user "What are your business hours?"
MESSAGE assistant "We're open Monday-Friday, 9am-6pm EST. Weekend support is available via email."

MESSAGE user "How do I reset my password?"
MESSAGE assistant "To reset your password: 1) Click 'Forgot Password' on the login page. 2) Enter your email. 3) Check your inbox for a reset link. 4) Create a new password. If you don't receive the email within 5 minutes, check your spam folder."

MESSAGE user "What's your return policy?"
MESSAGE assistant "We offer a 30-day money-back guarantee. Items must be unused and in original packaging. Simply contact support to initiate a return and we'll email you a prepaid shipping label."

Option B: Using a JSONL file

# Use the Web UI converter
# 1. Go to http://localhost:8080/converter
# 2. Upload your Excel/CSV with questions and answers
# 3. Download the JSONL file
# 4. Save to ./data/training/my-data.jsonl

Then reference in Modelfile:

FROM llama3.2:1b

PARAMETER temperature 0.3
PARAMETER num_ctx 4096

SYSTEM """
You are a customer support assistant.
"""

# Load examples from JSONL (if using external data)
# Note: Currently requires manual MESSAGE blocks
# Future: We may support direct JSONL import

Step 2: Create the model

bash scripts/create-custom-model.sh support-bot ./models/custom/support-bot/Modelfile

Step 3: Test with real questions

docker compose exec ollama ollama run support-bot "How can I track my order?"

Pros¶

✅ More consistent than System Prompt alone ✅ Teaches specific knowledge ✅ Still fast and easy ✅ No GPU required

Cons¶

❌ Limited to ~50 examples (Modelfile size) ❌ Doesn't deeply "learn" patterns ❌ May not generalize well beyond examples

Tips¶

Start with 5-10 examples, test, then add more
Include edge cases and common variations
Make examples diverse (different question styles)
Test after adding each batch

🚀 Method 3: Fine-Tuning (Advanced)¶

When to Use¶

You have 100+ training examples
Need deep domain specialization (medical, legal, technical)
Want model to learn complex patterns
Require consistent structured outputs

What You Need¶

Large dataset (100-10,000+ examples)
GPU access (local or cloud)
Technical skills (Python, ML frameworks)
Time for training and evaluation

Time Required¶

Setup: 1-2 hours
Training: 30 minutes to 8+ hours (depends on size)
Evaluation: 1-2 hours

Example Use Cases¶

✅ Medical diagnosis assistant ✅ Legal document analyzer ✅ Code generation for specific framework ✅ Technical troubleshooting expert

High-Level Process¶

Ollama itself doesn't fine-tune models — you train with external tools and import the result:

Prepare a JSONL dataset (100+ examples)
Train externally with Unsloth or Hugging Face Transformers (locally or on Google Colab's free GPU), typically with LoRA/QLoRA
Export to GGUF format with llama.cpp
Import into Ollama: bash scripts/import-model.sh my-finetuned ./data/gguf/my-model.gguf

📖 Complete walkthrough with training scripts: Fine-Tuning Guide (canonical deep-dive). For a concrete worked example, see the Dataset Training Example.

Pros¶

✅ Deep learning of patterns ✅ Can handle complex domain knowledge ✅ Better generalization ✅ Consistent structured outputs

Cons¶

❌ Time-consuming ❌ Requires GPU ❌ Technical complexity ❌ Risk of overfitting ❌ Expensive (GPU time/cloud costs)

Important Notes¶

⚠️ Try Method 2 first! Few-shot learning often works surprisingly well. ⚠️ More data ≠ better - Quality matters more than quantity. ⚠️ Start small - Fine-tune on 100 examples, evaluate, then scale up.

🎓 Progressive Learning Path¶

Week 1: Start Simple¶

Use existing models from Ollama library
Test with Web UI
Understand basic prompting

Week 2: System Prompts¶

Create first custom Modelfile
Experiment with SYSTEM prompts
Adjust parameters (temperature, etc.)

Week 3: Few-Shot Learning¶

Collect 10-20 example Q&As
Add MESSAGE blocks to Modelfile
Test and refine

Week 4+: Advanced (Optional)¶

Consider fine-tuning for specialized needs
Explore LoRA adapters
Deploy to production

📖 Real-World Examples¶

Example 1: Personal Assistant (System Prompt)¶

Goal: Friendly helper that keeps responses concise

Method: System Prompt Time: 5 minutes

FROM llama3.2:1b
PARAMETER temperature 0.7
PARAMETER num_ctx 4096

SYSTEM """
You are a helpful personal assistant.
Keep responses under 3 sentences unless asked for details.
Be friendly and use a conversational tone.
"""

Example 2: FAQ Bot (Few-Shot)¶

Goal: Answer common questions about a product

Method: Few-Shot Learning Time: 30 minutes Data: 20 Q&A pairs from support tickets

FROM llama3.2:1b
PARAMETER temperature 0.3
PARAMETER num_ctx 4096

SYSTEM """
You are a support bot for WidgetPro 3000.
Answer questions based on the manual.
"""

MESSAGE user "How do I charge the battery?"
MESSAGE assistant "Connect the USB-C cable to the charging port on the bottom. A red light indicates charging; green means fully charged (takes ~2 hours)."

# ... 19 more examples ...

Example 3: Medical Coding Assistant (Fine-Tuning)¶

Goal: Convert medical notes to billing codes

Method: Fine-Tuning Time: 8 hours Data: 5,000 labeled examples

Reason for fine-tuning: - Requires learning complex ICD-10 code patterns - Needs high accuracy (billing consequences) - Large labeled dataset available - Structured output format

🤔 Still Not Sure?¶

Ask Yourself:¶

Q: Do I just want to change how it talks? → Use System Prompt

Q: Do I have specific examples of ideal responses? → Use Few-Shot Learning (if < 50 examples) → Use Fine-Tuning (if 100+ examples)

Q: Is it mission-critical with regulatory requirements? → Consider Fine-Tuning (after thorough testing)

Q: Am I on a tight deadline? → Use System Prompt or Few-Shot Learning

Q: Do I have a GPU and ML expertise? → Consider Fine-Tuning (but try simpler methods first!)

🚦 Traffic Light Guide¶

🟢 Start Here	🟡 Next Step	🔴 Advanced
System Prompt	Few-Shot Learning	Fine-Tuning
5 minutes	30 minutes	Hours/Days
No data needed	5-50 examples	100+ examples
Easy iteration	Medium complexity	High complexity
Use for 80% of cases	Use for 15% of cases	Use for 5% of cases

📚 Additional Resources¶

Essential Concepts - Learn the basics
Parameter Guide - Optimize your settings
Example Modelfiles - Ready-to-use templates
Fine-Tuning Guide - Complete fine-tuning deep-dive
Dataset Training Example - Worked dataset example

Remember: Start simple, test, iterate! Most users achieve their goals with just a well-crafted system prompt. Don't overcomplicate! 🎯