Modelfile Syntax: The Language of Creation

To create a custom model, you need to understand the four primary "verbs" of a Modelfile. Each one controls a different part of the AI's behavior.

1. FROM (The Foundation)

Every Modelfile must start with a FROM instruction. This tells Ollama which base model to "inherit" from.

Examples:

• FROM llama3: Uses the standard 8B Llama 3 model.
• FROM ./my-model.gguf: Uses a raw file you downloaded from Hugging Face.

---

2. SYSTEM (The Persona)

The SYSTEM instruction sets the persistent instructions that the model will follow. This is where you define the model's role, tone, and knowledge boundaries.

Example:

SYSTEM """
You are a Cloud Architect. 
Always provide answers using AWS-specific services. 
Be concise and avoid conversational filler like 'Sure!' or 'Of course!'
"""

Note: Use triple quotes (""") for multi-line instructions.

---

3. PARAMETER (The Engine Settings)

Parameters allow you to fine-tune the technical performance of the model.

Common Parameters:

• PARAMETER temperature 0.5: Controls creativity (0 is factual, 1 is creative).
• PARAMETER num_ctx 8192: Sets the context window (memory size).
• PARAMETER stop "[DONE]": Tells the model to stop generating text when it sees this specific string.
• PARAMETER repeat_penalty 1.1: Prevents the model from repeating the same sentence over and over.

---

4. TEMPLATE (The Advanced Formatting)

Every model (Llama, Mistral, Google) expects prompts to be formatted in a specific way. For example:

• Llama 3: <|begin_of_text|><|start_header_id|>user<|end_header_id|>
• Mistral: [INST] ... [/INST]

Unless you are an expert, you usually don't need to change this. Ollama automatically applies the correct template based on the FROM line.

---

5. MESSAGE (The Pre-fill)

You can include "Few-shot" examples directly in the Modelfile so the model always sees them before your first question.

Example:

MESSAGE user "What is the capital of France?"
MESSAGE assistant "The capital of France is Paris. [Butler Mode Active]"

---

The Complete Picture

Here is a professional-grade Modelfile for a "Code Reviewer" bot:

FROM llama3
PARAMETER temperature 0.1
PARAMETER num_ctx 16384
SYSTEM "You are an expert Senior Security Engineer. Review the user's code for potential SQL injection and XSS vulnerabilities."

How to Apply It

Save that text as a file named SecurityBot. Then run: ollama create sec-bot -f SecurityBot

---

Key Takeaways

• FROM is the only mandatory instruction.
• SYSTEM defines the "Soul" of the model.
• PARAMETER defines the "Spec" of the engine.
• Use Triple Quotes for long instructions.
• The ollama create command brings the script to life.