Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

get_chat_template: write the chosen system message into the Ollama Modelfile (#16863)

Mohammad Hijjawi5249097dc765
diff --git a/tests/saving/test_ollama_modelfile_trained_template.py b/tests/saving/test_ollama_modelfile_trained_template.py
index 1f50ccddef2..1bc97757ae6 100644
--- a/tests/saving/test_ollama_modelfile_trained_template.py
+++ b/tests/saving/test_ollama_modelfile_trained_template.py
@@ -1,11 +1,13 @@
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+import pytest
from datasets import Dataset
from tokenizers import Tokenizer, models
from transformers import CLIPImageProcessor, LlavaProcessor, PreTrainedTokenizerFast
from unsloth.chat_templates import apply_chat_template, get_chat_template
+from unsloth.ollama_template_mappers import OLLAMA_TEMPLATES
from unsloth.save import create_ollama_modelfile
ALPACA = """Below are some instructions that describe some tasks. Write responses that appropriately complete each request.
@@ -90,3 +92,149 @@ def test_processor_template_survives_the_gguf_unwrap():
assert modelfile is not None
assert "### Instruction:" in modelfile
+
+
+def test_system_message_passed_to_get_chat_template_reaches_the_modelfile():
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "alpaca", system_message = "You are a pirate."
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert modelfile.endswith('SYSTEM """You are a pirate."""')
+ assert modelfile.count("SYSTEM ") == 1
+ assert "Below are some instructions" not in modelfile
+
+
+def test_system_message_replaces_the_default_of_a_mapped_model():
+ name = "unsloth/Qwen2.5-0.5B-Instruct"
+ tokenizer = get_chat_template(
+ _tokenizer(name), chat_template = "qwen-2.5", system_message = "Be brief."
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, name, "x.gguf")
+
+ assert modelfile.endswith('SYSTEM """Be brief."""')
+ assert "You are Qwen" not in modelfile
+ assert "<|im_start|>system" in modelfile
+
+
+def test_qwen_tool_prompt_uses_the_same_system_message_as_the_modelfile():
+ name = "unsloth/Qwen2.5-0.5B-Instruct"
+ tools = [{"type": "function", "function": {"name": "f", "parameters": {}}}]
+ tokenizer = get_chat_template(
+ _tokenizer(name), chat_template = "qwen-2.5", system_message = "Be brief."
+ )
+ stock = get_chat_template(_tokenizer(name), chat_template = "qwen-2.5")
+
+ prompt = tokenizer.apply_chat_template(CONVERSATION, tools = tools, tokenize = False)
+
+ assert prompt.startswith("<|im_start|>system\nBe brief.\n\n# Tools")
+ assert create_ollama_modelfile(tokenizer, name, "x.gguf").endswith('SYSTEM """Be brief."""')
+ assert stock.apply_chat_template(CONVERSATION, tools = tools, tokenize = False).startswith(
+ "<|im_start|>system\nYou are Qwen, created by Alibaba Cloud. You are a helpful assistant.\n\n"
+ )
+
+
+def test_system_message_is_added_when_the_modelfile_has_none():
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "llama-3.1", system_message = "Be brief."
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert modelfile.endswith('PARAMETER min_p 0.1\nSYSTEM """Be brief."""')
+ assert modelfile.count("SYSTEM ") == 1
+
+
+def test_empty_system_message_drops_the_default_from_the_modelfile():
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "unsloth", system_message = ""
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert "SYSTEM " not in modelfile
+ assert ">>> Assistant: {{ .Response }}" in modelfile
+
+
+def test_default_system_message_leaves_the_stock_modelfile():
+ tokenizer = get_chat_template(_tokenizer("some/model"), chat_template = "alpaca")
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert tokenizer._ollama_modelfile == OLLAMA_TEMPLATES["alpaca"]
+ assert tokenizer._ollama_system_message is None
+ assert modelfile.endswith(
+ 'SYSTEM """Below are some instructions that describe some tasks. '
+ 'Write responses that appropriately complete each request."""'
+ )
+
+
+def test_system_message_without_a_template_default_leaves_the_modelfile():
+ # chatml has no default, so get_chat_template only warns and the Jinja template ignores it.
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "chatml", system_message = "Be brief."
+ )
+ stock = get_chat_template(_tokenizer("some/model"), chat_template = "chatml")
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert modelfile == create_ollama_modelfile(stock, "some/model", "x.gguf")
+
+
+def test_system_message_keeps_the_mapped_template_choice():
+ # qwen-2.5 and qwen-2.5-coder train the same; the mapper picks the coder Modelfile, which adds FIM.
+ name = "unsloth/Qwen2.5-Coder-7B-Instruct"
+ tokenizer = get_chat_template(
+ _tokenizer(name), chat_template = "qwen-2.5", system_message = "Be brief."
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, name, "x.gguf")
+
+ assert "<|fim_prefix|>" in modelfile
+ assert modelfile.endswith('SYSTEM """Be brief."""')
+ assert modelfile.count("SYSTEM ") == 1
+
+
+def test_system_message_reaches_the_modelfile_of_a_processor():
+ processor = LlavaProcessor(
+ image_processor = CLIPImageProcessor(), tokenizer = _tokenizer("some/vlm"), patch_size = 14
+ )
+ processor = get_chat_template(processor, chat_template = "alpaca", system_message = "Be brief.")
+
+ modelfile = create_ollama_modelfile(processor.tokenizer, "some/vlm", "x.gguf")
+
+ assert modelfile.endswith('SYSTEM """Be brief."""')
+
+
+def test_later_get_chat_template_without_system_message_restores_the_default():
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "alpaca", system_message = "Be brief."
+ )
+ tokenizer = get_chat_template(tokenizer, chat_template = "alpaca")
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert "Be brief." not in modelfile
+ assert 'SYSTEM """Below are some instructions' in modelfile
+
+
+def test_system_message_is_written_verbatim():
+ message = 'Say "hi" and use {braces}, {__EOS_TOKEN__}, \\1 and """ freely.'
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "alpaca", system_message = message
+ )
+
+ modelfile = create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
+
+ assert modelfile.endswith(f'SYSTEM """{message}"""')
+
+
+def test_system_message_that_ollama_would_cut_short_is_rejected():
+ tokenizer = get_chat_template(
+ _tokenizer("some/model"), chat_template = "alpaca", system_message = 'Use """ \nfor code.'
+ )
+
+ with pytest.raises(ValueError, match = "Ollama Modelfile"):
+ create_ollama_modelfile(tokenizer, "some/model", "x.gguf")
diff --git a/tests/test_get_chat_template_processor.py b/tests/test_get_chat_template_processor.py
index 5d8a9677b81..6fe2bcf7ee9 100644
--- a/tests/test_get_chat_template_processor.py
+++ b/tests/test_get_chat_template_processor.py
@@ -122,6 +122,7 @@ def _reattach(processor, tokenizer):
tokenizer = tokenizer,
chat_template = "TEMPLATE",
ollama_modelfile = "MODELFILE",
+ ollama_system_message = None,
)
diff --git a/unsloth/chat_templates.py b/unsloth/chat_templates.py
index 5de363f73c2..56b379d0f8e 100644
--- a/unsloth/chat_templates.py
+++ b/unsloth/chat_templates.py
@@ -2894,6 +2894,8 @@ def get_chat_template(
"{%- endif %}" + \
chat_template
+ # Only templates that bake the message in pass it on; applied after create_ollama_modelfile picks the template.
+ ollama_system_message = system_message if "{system_message}" in chat_template else None
chat_template, system_message = _change_system_message(chat_template, type_chat_template, system_message)
tokenizer.chat_template = chat_template
@@ -2917,6 +2919,7 @@ def get_chat_template(
if _processor is not None:
# GGUF export unwraps the processor before it builds the Ollama Modelfile.
tokenizer._ollama_modelfile = ollama_modelfile
+ tokenizer._ollama_system_message = ollama_system_message
_processor.tokenizer = tokenizer
_processor.chat_template = chat_template
for _token in ("bos_token", "eos_token", "pad_token",):
@@ -2930,6 +2933,7 @@ def get_chat_template(
tokenizer = patch_saving_functions(tokenizer, vision = _processor is not None)
tokenizer._ollama_modelfile = ollama_modelfile
+ tokenizer._ollama_system_message = ollama_system_message
tokenizer._system_message = system_message
tokenizer._unsloth_chat_template_name = (
(template_name, chat_template) if template_name is not None else None
@@ -3645,11 +3649,14 @@ extra_eos_tokens = None,
tokenizer.chat_template = jinja_template
tokenizer._ollama_modelfile = modelfile
+ # This Modelfile already holds default_system_message; drop one left by an earlier get_chat_template.
+ tokenizer._ollama_system_message = None
tokenizer._unsloth_input_part = input_part
tokenizer._unsloth_output_part = output_part
if hasattr(tokenizer, "tokenizer"):
tokenizer.tokenizer.chat_template = jinja_template
tokenizer.tokenizer._ollama_modelfile = modelfile
+ tokenizer.tokenizer._ollama_system_message = None
tokenizer.tokenizer._unsloth_input_part = input_part
tokenizer.tokenizer._unsloth_output_part = output_part
diff --git a/unsloth/save.py b/unsloth/save.py
index ea3addd93ec..3cb5c3a48f8 100644
--- a/unsloth/save.py
+++ b/unsloth/save.py
@@ -3608,6 +3608,23 @@ def fix_tokenizer_bos_token(tokenizer):
return fix_bos_token, chat_template
+_OLLAMA_SYSTEM_RE = re.compile(r'^SYSTEM """.*?"""[ \t]*\n?', flags = re.DOTALL | re.MULTILINE)
+# Ollama ends a """ value at the first line break preceded by """, so such a message would be cut short.
+_OLLAMA_SYSTEM_CUT_RE = re.compile(r'"""\s*\n')
+
+
+def _set_ollama_system_message(modelfile, system_message):
+ if _OLLAMA_SYSTEM_CUT_RE.search(system_message):
+ raise ValueError(
+ 'Unsloth: The system message has a line ending in """, which an Ollama Modelfile cannot hold.'
+ )
+ system_line = f'SYSTEM """{system_message}"""\n' if system_message != "" else ""
+ modelfile, n_found = _OLLAMA_SYSTEM_RE.subn(lambda _: system_line, modelfile, count = 1)
+ if n_found == 0 and system_line:
+ modelfile = modelfile.rstrip("\n") + "\n" + system_line
+ return modelfile
+
+
def create_ollama_modelfile(tokenizer, base_model_name, model_location):
"""
Creates an Ollama Modelfile.
@@ -3651,9 +3668,14 @@ def create_ollama_modelfile(tokenizer, base_model_name, model_location):
__FILE_LOCATION__ = model_location,
)
- modelfile = modelfile.replace("⚫@✅#🦥", "{").replace("⚡@🦥#⛵", "}").rstrip()
+ modelfile = modelfile.replace("⚫@✅#🦥", "{").replace("⚡@🦥#⛵", "}")
- return modelfile
+ # Applied after formatting, so the message is written verbatim.
+ system_message = getattr(tokenizer, "_ollama_system_message", None)
+ if system_message is not None:
+ modelfile = _set_ollama_system_message(modelfile, system_message)
+
+ return modelfile.rstrip()
def create_ollama_model(username: str, model_name: str, tag: str, modelfile_path: str):