ToolLocal LLMs

Ollama Modelfile builder

Fill in the base model, a system prompt and the parameters you actually want to change. You get a Modelfile you can save, and the two commands that turn it into a model you can run by name. Nothing leaves your browser.

Model

Name for ollama create

Lowercase letters, digits, - _ . — an optional :tag after it. This is what you type after ollama run.

FROM — base model

A name from ollama.com/library (add :tag for a specific size — the model's own page lists them), a Safetensors model directory, or a path to a GGUF file such as ./model.gguf.

SYSTEM — system prompt

The instruction baked into every chat. More than one line is wrapped in the documented """…""" form.

Parameters

Leave a box empty to leave that parameter alone — the model then uses the default shown under it.

temperature

float · default 0.8 — higher answers more creatively

num_ctx

int · default 2048 — context window used to generate the next token

top_p

float · default 0.9 — higher is more diverse, lower more focused

top_k

int · default 40 — 100 is more diverse, 10 more conservative

repeat_penalty

float · default 1.0 — 1.5 penalizes repetition harder, 0.9 is lenient

seed

int · default 0 — a fixed seed repeats the same text for the same prompt

num_predict

int · default -1 (infinite) — max tokens to predict

stop — one sequence per line

Each line becomes its own PARAMETER stop line. Generation stops when one is produced.

Advanced — template, adapter, license, message history

TEMPLATE

Leave this empty unless you know the base model's chat format. The template is the exact string sent to the model; the syntax is model specific, so a wrong one breaks chat formatting — the model rambles, never stops, or ignores your system prompt. Empty means your model keeps the base model's template. Copy the real one first with ollama show --modelfile <model>. Whitespace is passed through exactly as you type it, trailing newline included.

ADAPTER — path to a (Q)LoRA adapter

Absolute, or relative to the Modelfile. It must match the base model in FROM, or the results are erratic.

LICENSE

The legal license the model is shared under. Only worth filling in if you plan to publish it.

MESSAGE — example history

A short conversation that shows the model how you want it to answer. Rows with no text are skipped.

Modelfile

JavaScript is off, so this is a worked example rather than your Modelfile. Edit the lines by hand — every instruction is explained in the table below.

FROM llama3.2

SYSTEM """You are a blunt code reviewer. Point out bugs and risks first.
Be terse. Use bullet points. No praise."""

PARAMETER temperature 0.4
PARAMETER num_ctx 4096

Save it as a file called Modelfile (no extension), then, in that folder:

ollama create reviewer -f Modelfile
ollama run reviewer

The builder runs entirely in your browser. The link in your address bar carries the whole form, so you can bookmark a Modelfile or send it to someone.

What each line does

Straight from the Ollama Modelfile reference — only the instructions and parameters this builder can write.

docs.ollama.com/modelfile →

Line	Type · default	What it does
`FROM`	instruction	Defines the base model to use. Required. A library name, a Safetensors model directory, or a path to a .gguf file (absolute, or relative to the Modelfile).
`SYSTEM`	instruction	Specifies the system message that will be set in the template — the instruction baked into every chat.
`PARAMETER temperature`	float · 0.8	The temperature of the model. Increasing the temperature will make the model answer more creatively.
`PARAMETER num_ctx`	int · 2048	Sets the size of the context window used to generate the next token.
`PARAMETER top_p`	float · 0.9	Works together with top-k. A higher value (e.g. 0.95) leads to more diverse text; a lower value (e.g. 0.5) is more focused and conservative.
`PARAMETER top_k`	int · 40	Reduces the probability of generating nonsense. A higher value (e.g. 100) gives more diverse answers, a lower value (e.g. 10) is more conservative.
`PARAMETER repeat_penalty`	float · 1.0	Sets how strongly to penalize repetitions. A higher value (e.g. 1.5) penalizes repetitions more strongly, a lower value (e.g. 0.9) is more lenient.
`PARAMETER seed`	int · 0	Sets the random number seed to use for generation. Setting this to a specific number will make the model generate the same text for the same prompt.
`PARAMETER num_predict`	int · -1	Maximum number of tokens to predict when generating text. -1 means infinite generation.
`PARAMETER stop`	string · repeats	Sets the stop sequences to use. When this pattern is encountered the LLM will stop generating text and return. Multiple stop patterns are set as multiple separate stop parameters.
`TEMPLATE`	instruction	The full prompt template to be sent to the model, in Go template syntax. Variables: .System, .Prompt, .Response. Syntax may be model specific.
`ADAPTER`	instruction	Defines the (Q)LoRA adapters to apply to the base model. Absolute path, or relative to the Modelfile. If the base model is not the one the adapter was tuned from, behaviour will be erratic.
`LICENSE`	instruction	Specifies the legal license under which the model is shared or distributed.
`MESSAGE`	instruction	Specify message history. Repeat it to build a short conversation that guides the model to answer in a similar way. Valid roles: system, user, assistant.

The Modelfile is not case sensitive and instructions can be in any order — uppercase and FROM first are conventions that keep it readable.

Guides that go with it

The plates that cover installing Ollama, calling it from code, and the Modelfile itself.

All Local LLM plates →