What does the TEMPLATE directive do in a Modelfile?

TEMPLATE defines the full prompt string Ollama passes to the model. It uses Go template syntax to place the system prompt, the user message, and control tokens in the exact layout the model expects.

What variables can I use in an Ollama template?

The core variables are {{ .System }} for the system prompt, {{ .Prompt }} for the user message, and {{ .Response }} for where the reply begins. You wrap them in {{ if }} blocks to include parts only when present.

Why is my custom model looping or rambling?

The TEMPLATE usually has the wrong control tokens or a missing stop sequence. Export the model's original template, compare the tokens, and add a matching PARAMETER stop so generation ends cleanly.

How do I see a model's default template?

Run ollama show --modelfile to print the full Modelfile, including the TEMPLATE block. Copy that as your starting point instead of guessing the format.

Build / AI / Ollama

Modelfile Prompt Templates

Customize the Modelfile TEMPLATE directive and control tokens so a model's chat format works correctly.

TL;DR

TEMPLATE controls the exact prompt sent to the model.
Insert content with .System, .Prompt, and .Response variables.
Match control tokens like <|im_end|> to the model's format.

Jump to section

The TEMPLATE Directive
Control Tokens
A Full Template
View And Edit
Tips
Warnings
In Practice
FAQ

The TEMPLATE Directive

TEMPLATE

Define the full prompt string sent to the model.

TEMPLATE """{{ .Prompt }}"""

.System

Insert the system prompt into the template.

{{ if .System }}{{ .System }}{{ end }}

.Prompt

Insert the user's message into the template.

{{ .Prompt }}

.Response

Marks where the model's reply begins.

{{ .Response }}

Control Tokens

Special Tokens

Models use tokens to mark turn boundaries.

<|im_start|>user
{{ .Prompt }}<|im_end|>

Match Training

Use the exact tokens the model was trained on.

# Wrong tokens break the chat format

Stop Sequences

Set a stop token so generation ends cleanly.

PARAMETER stop "<|im_end|>"

A Full Template

System Block

Render the system prompt only when present.

{{ if .System }}<|system|>
{{ .System }}
{{ end }}

User Block

Wrap the user prompt in the model's tags.

<|user|>
{{ .Prompt }}

Assistant Tag

End with the assistant tag to cue the reply.

<|assistant|>

View And Edit

Show Template

Print the model's current Modelfile and template.

ollama show llama3.2 --modelfile

Edit And Rebuild

Change the template, then rebuild the model.

ollama create my-model -f Modelfile

Test It

Run the model and check the formatting.

ollama run my-model "Say hi"

Tips

Start from the model's official template with ollama show --modelfile, then edit that rather than writing a TEMPLATE from scratch.
Keep any special control tokens exactly as the model was trained, since even a small change can break the chat format.

Warnings

A broken TEMPLATE can make a model ramble, loop, or ignore the system prompt; verify the format before relying on the model.
Text placed after .Response in a TEMPLATE is dropped during generation, so keep the assistant tag before that point.

In Practice

Export a model's template, correct the control tokens, rebuild it, and confirm the chat format works.

Exporting the real Modelfile gives you a correct starting point.
Editing the tokens in place avoids typos that break the format.
ollama create bakes the corrected template into a new model.
A quick test run confirms the reply ends cleanly without looping.

# 1. Save the current Modelfile
ollama show llama3.2 --modelfile > Modelfile

# 2. Edit the TEMPLATE block to match tokens,
#    e.g. <|start|> ... <|end|> per the model

# 3. Rebuild with the corrected template
ollama create llama3-fixed -f Modelfile

# 4. Verify the reply is clean, no loops
ollama run llama3-fixed "Hello"