Current Implementation

Overview

The Modelfile parser is implemented in src/core/modelfile.rs using the nom parser combinator library. It parses a text-based configuration format for defining model configurations, parameters, templates, and system prompts.

Grammar

The Modelfile grammar follows this structure:


command -> Instruction arguments*

Instruction -> "FROM" | "PARAMETER" | "TEMPLATE" | "SYSTEM" | "ADAPTER" | "LICENSE" | "MESSAGE" | "#"

arguments -> WORD | quoted_string | multiline_string

quoted_string -> "<str>"

multiline_string -> """<str>"""

Supported Instructions

FROM (Required)

The FROM instruction specifies the base model to use. This is the only required instruction in a Modelfile.

Syntax:


FROM <model_name>

Examples:


FROM llama3.2

FROM llama3.2:latest

FROM /path/to/model.gguf

FROM mlx-community/dolphin3.0-llama3.2-1B-4Bit

Validation:

Only one FROM instruction is allowed per Modelfile
A Modelfile without a FROM instruction will fail validation

PARAMETER

The PARAMETER instruction sets model inference parameters. Multiple parameters can be specified.

Syntax:


PARAMETER <parameter_name> <value>

Supported Parameters:

Integer Parameters:

num_ctx - Context window size
repeat_last_n - Number of tokens to consider for repeat penalty
seed - Random seed for generation
num_predict - Maximum number of tokens to predict
top_k - Top-k sampling parameter

Float Parameters:

temperature - Sampling temperature (0.0 to 1.0+)
repeat_penalty - Penalty for repeating tokens
top_p - Nucleus sampling parameter
min_p - Minimum probability threshold

String Parameters:

stop - Stop sequence (can be specified multiple times)

Examples:


PARAMETER num_ctx 4096

PARAMETER temperature 0.7

PARAMETER stop "<|eot_id|>"

PARAMETER stop "</s>"

PARAMETER top_p 0.9

PARAMETER seed 42

Validation:

Parameter names are case-insensitive
Integer parameters must parse as valid i32 values
Float parameters must parse as valid f32 values
Invalid parameter types will result in an error

TEMPLATE

The TEMPLATE instruction defines the prompt template for the model. Only one template is allowed.

Syntax:


TEMPLATE <template_string>

Examples:


TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|>

{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>

{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>

{{ .Response }}<|eot_id|>"""

Validation:

Only one TEMPLATE instruction is allowed per Modelfile
Template can be specified as a single-line string, quoted string, or multiline string

SYSTEM

The SYSTEM instruction sets the system prompt for the model. Only one system prompt is allowed.

Syntax:


SYSTEM <system_prompt>

Examples:


SYSTEM "You are a helpful assistant."

SYSTEM """

You are a bot

You also not a bot

"""

Validation:

Only one SYSTEM instruction is allowed per Modelfile
System prompt can be specified as a single-line string, quoted string, or multiline string

ADAPTER

The ADAPTER instruction specifies a path to a LoRA adapter file. Only one adapter is allowed.

Syntax:


ADAPTER <adapter_path>

Examples:


ADAPTER /path/to/adapter.bin

ADAPTER ./adapters/my_adapter.safetensors

Validation:

Only one ADAPTER instruction is allowed per Modelfile

LICENSE

The LICENSE instruction specifies the license text for the model. Only one license is allowed.

Syntax:


LICENSE <license_text>

Examples:


LICENSE "MIT"

LICENSE """

Apache License

Version 2.0, January 2004

...

"""

Validation:

Only one LICENSE instruction is allowed per Modelfile
License text can be specified as a single-line string, quoted string, or multiline string

MESSAGE

The MESSAGE instruction defines example messages for few-shot learning or conversation examples. Multiple messages can be specified.

Syntax:


MESSAGE <role> <message_content>

Supported Roles:

system - System message
user - User message
assistant - Assistant message

Examples:


MESSAGE user """Is Toronto in Canada?

or Is cologne in france"""

MESSAGE assistant yes

MESSAGE user Is Sacramento in Canada?

MESSAGE assistant no

Validation:

Role must be one of: system, user, or assistant (case-insensitive)
Invalid roles will result in an error

Comments

Comments are supported using the # character. Comments are preserved in the parsed Modelfile.

Syntax:


# <comment_text>

Examples:


# Modelfile generated by "ollama show"

# To build a new Modelfile based on this one, replace the FROM line with:

FROM llama3.2

# This is a comment

String Formats

The parser supports three string formats for instruction arguments:

Single-line strings: Plain text until end of line
```
FROM llama3.2
```
Quoted strings: Text enclosed in double quotes "
```
SYSTEM "You are a helpful assistant."
```

Multiline strings: Text enclosed in triple quotes """


SYSTEM """

You are a bot

You also not a bot

"""