Understanding Modelfiles in Ollama

Introduction 

As we already know that we download, run and configure the models in ollama.  But a main question arises ‘can we customize the behavior of the AI model?’.  For example making any model to perform only specific task like coding assistant, education, company-specific, etc.  So, to manage these things ollama provides Modelfiles.

What is Modelfiles?

Modelfile is a configuration file which defines tha behavior and setting of any AI model.  In simple word Modelfile work as instruction or a blueprint for any AI model.  It also defines which base model should be used by the model and how they will generate the response for any users prompt. 

A Modelfile allows users to customize the behavior of an AI model without retraining it.  By using Modelfiles we can define personality, response style, system instruction and other parameter of model that control response generation.  Buy the help of Modelfiles we can create a task specific AI models.

Instead of modifying the actual knowledge of the model Modelfiles provides additional instruction which helps to generate the response accordingly.  Modelfiles provide flexibility, consistency and better control over AI model's behavior. 

Why do we need modelfiles?

Normally when we run any model in ollama then AI model generates the response according to there behavior.

Example:

user → llama 3 → response 

here AI model woks as a general-purpose AI which means it try to answers every types of question asked to the AI model but in real world we don't need a general purpose AI every time.  Suppose a company wants to create a customer support assistant AI model or any developer wants to create a coding assistant model, in these case we have to control the behavior of the AI model and this is done my Modelfiles.  We can give instructions to the model to generate specific field response using Modelfiles.  By using Modelfiles we can give instructions to the AI model like:

• Behave like a coding assistant 
• Generate detailed explanations, etc. 

How Modelfiles work?

Modelfiles do not train the model and it also do not change the knowledge of any AI model.  It only provides additional instructions to the model that hoe the response should be generated by the AI model.

workflow:

Base model → modelfile Instructions → customized AI model → response 

example:

suppose we are using llama 3 model now how the normal model response: 

User: What is Python?

Llama 3:
Python is a programming language.

Now if we modify the Modelfile like:

You are an expert Python teacher.
Explain every concept in detail.

Then the response will be like :

Python is a high-level, interpreted, general-purpose programming language created by Guido van Rossum and first released in 1991.

In simple terms:
Python is a programming language that allows humans to write instructions that a computer can understand and execute.

Python is designed to be easy to read, easy to learn, and highly versatile, which is why it is one of the most popular programming languages in the world.

Here, the model and knowledge of the model is same but only behavior is changed.  That's why Modelfiles provides lightweight customization.

Components of a Modelfile

A Modelfile is generally made up of some important components which are:

1. FROM

FROM instruction defines that the which base model is used to create the custom model.

Example:

FROM llama3

here we define to ollama that we are using llama3 model.  It is a mandatory component of Modelfiles.  without FROM instruction ollama will not be able to understand that which model is being customized.

2. SYSTEM

System Instruction Defines the personality and behavior of the model.

Example: 

SYSTEM you are an AI teacher.

Now the model will try to answer the users prompt like an AI teacher.  SYSTEM prompt is one of the powerful component of Modelfiles.

3. PARAMETER

PARAMETER controls the response generation setting of any AI model.

Example:

PARAMETER temperature 0.7

PARAMETER decides the creativity, Randomness, Response length, Prediction behavior of any model.  PARAMETER is used for advanced customization of any model.

Benefits of Modelfiles

there are various advantages of using Modelfiles like:

1. Easy Customization
2. Reusability
3. Consistency 
4. Faster Development
5. No additional training cost

Read more about ollama:

previous topic: Understanding Quantization and ollama model storage
next topic: Creating Custom AI Models with Modelfiles