Custom Models With Modelfile

Baking in what you'd otherwise repeat every call

Every lesson so far that used a system prompt (Lessons 4, 9, 12) sent it fresh on every single call, as part of messages. That's fine for a program you control, but sometimes you want a model that's permanently a certain persona or configuration, so that any call to it, from any script, gets that behavior automatically, without every caller having to remember to include the right system message.

A Modelfile does exactly that: it defines a new, named model built on top of an existing one, with a system prompt and default parameters baked in. It's the same idea as a Docker image built FROM a base image with some setup already applied.

The Modelfile format, briefly

Written to a text file and built with ollama create <name> -f Modelfile on the CLI, a Modelfile looks like this:

FROM llama3.2
SYSTEM You are Pirate Pete, a pirate-themed coding assistant. Always answer in pirate speak, but keep technical accuracy intact.
PARAMETER temperature 0.7

FROM picks the base model, SYSTEM bakes in a fixed system prompt, PARAMETER bakes in default options (Lesson 6). This lesson never writes that file directly, ollama.create() takes the same three pieces as Python keyword arguments instead, useful when you want a program (not a human with a text editor) to define custom models.

The code, piece by piece

for _ in ollama.create(
    model=CUSTOM_MODEL_NAME,
    from_="llama3.2",
    system=SYSTEM_PROMPT,
    parameters={"temperature": 0.7},
):
    pass

ollama.create() streams progress updates as it builds the new model (note from_, with a trailing underscore, since from is a reserved Python keyword). The loop just drains that progress stream; a real program might print or log each update instead of discarding it.

response = ollama.chat(
    model=CUSTOM_MODEL_NAME,
    messages=[{"role": "user", "content": "What is a variable in programming?"}],
)

No system message in this call at all, and yet the reply comes back fully in character. That's the entire point: the persona now lives on the model itself, not in every caller's code.

info = ollama.show(CUSTOM_MODEL_NAME)
system_line = next(line for line in info.modelfile.splitlines() if line.startswith("SYSTEM"))

ollama.show(), from Lesson 2, returns the model's full generated Modelfile as text on .modelfile; the SYSTEM line inside it proves the prompt really is stored on the model, not something this script quietly re-sent.

ollama.delete(CUSTOM_MODEL_NAME)

Same cleanup pattern as Lesson 7: this model was created just for the lesson, so it's removed again at the end rather than left sitting on disk.