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.

Studio: train a dataset's system column as the system prompt (#12614)

Daniel Han0599726f6f4f
diff --git a/studio/backend/tests/test_dataset_system_column.py b/studio/backend/tests/test_dataset_system_column.py
new file mode 100644
index 00000000000..6c5b890c880
--- /dev/null
+++ b/studio/backend/tests/test_dataset_system_column.py
@@ -0,0 +1,110 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import pytest
+from datasets import Dataset
+
+from utils.datasets import format_and_template_dataset
+
+
+class _TurnTokenizer:
+ chat_template = "{{ messages }}"
+ eos_token = "</s>"
+
+ def apply_chat_template(self, conversation, **_kwargs):
+ return "".join(f"<{turn['role']}>{turn['content']}" for turn in conversation)
+
+
+def _sharegpt(turns):
+ return "conversations", [{"from": role, "value": content} for role, content in turns]
+
+
+def _chatml(turns):
+ return "messages", [{"role": role, "content": content} for role, content in turns]
+
+
+def _format(rows):
+ result = format_and_template_dataset(
+ Dataset.from_list(rows), model_name = "stub-model", tokenizer = _TurnTokenizer()
+ )
+ assert result["success"], result["errors"]
+ return list(result["dataset"]["text"])
+
+
[email protected]("shape", [_sharegpt, _chatml], ids = ["sharegpt", "chatml"])
+def test_system_column_is_trained_as_the_system_turn(shape):
+ column, convo = shape([("user", "q"), ("assistant", "a")])
+
+ texts = _format([{"system": "Here is a list of functions", column: convo}])
+
+ assert texts == ["<system>Here is a list of functions<user>q<assistant>a"]
+
+
[email protected]("shape", [_sharegpt, _chatml], ids = ["sharegpt", "chatml"])
+def test_system_column_does_not_override_a_leading_system_turn_or_add_a_blank_one(shape):
+ column, with_system = shape([("system", "inline"), ("user", "q"), ("assistant", "a")])
+ _, without_system = shape([("user", "q"), ("assistant", "a")])
+
+ texts = _format(
+ [
+ {"system": "column", column: with_system},
+ {"system": " ", column: without_system},
+ ]
+ )
+
+ assert texts == ["<system>inline<user>q<assistant>a", "<user>q<assistant>a"]
+
+
+class _NoSystemTokenizer(_TurnTokenizer):
+ def apply_chat_template(self, conversation, **kwargs):
+ if any(turn["role"] == "system" for turn in conversation):
+ raise ValueError("System role not supported")
+ return super().apply_chat_template(conversation, **kwargs)
+
+
[email protected]("shape", [_sharegpt, _chatml], ids = ["sharegpt", "chatml"])
+def test_system_column_falls_back_when_the_template_rejects_a_system_turn(shape):
+ column, convo = shape([("user", "q"), ("assistant", "a")])
+
+ result = format_and_template_dataset(
+ Dataset.from_list([{"system": "Here is a list of functions", column: convo}]),
+ model_name = "stub-model",
+ tokenizer = _NoSystemTokenizer(),
+ )
+
+ assert result["success"], result["errors"]
+ assert list(result["dataset"]["text"]) == ["<user>q<assistant>a"]
+
+
+def test_a_malformed_row_is_dropped_on_its_own_when_a_system_column_is_present():
+ _, convo = _chatml([("user", "q"), ("assistant", "a")])
+
+ texts = _format(
+ [
+ {"system": "column", "messages": convo},
+ {"system": "column", "messages": [None, *convo]},
+ ]
+ )
+
+ assert texts == ["<system>column<user>q<assistant>a"]
+
+
[email protected]("shape", [_sharegpt, _chatml], ids = ["sharegpt", "chatml"])
+def test_system_column_is_kept_when_converting_to_alpaca(shape):
+ column, convo = shape([("user", "q"), ("assistant", "a")])
+ _, with_system = shape([("system", "inline"), ("user", "q"), ("assistant", "a")])
+
+ result = format_and_template_dataset(
+ Dataset.from_list(
+ [
+ {"system": "Here is a list of functions", column: convo},
+ {"system": "column", column: with_system},
+ ]
+ ),
+ model_name = "stub-model",
+ tokenizer = _TurnTokenizer(),
+ format_type = "alpaca",
+ )
+
+ assert result["success"], result["errors"]
+ assert list(result["dataset"]["input"]) == ["Here is a list of functions", "inline"]
diff --git a/studio/backend/utils/datasets/chat_templates.py b/studio/backend/utils/datasets/chat_templates.py
index 741a0b7d47c..963a0b90f5a 100644
--- a/studio/backend/utils/datasets/chat_templates.py
+++ b/studio/backend/utils/datasets/chat_templates.py
@@ -297,6 +297,18 @@ def get_dataset_info_summary(dataset_info):
}
+def _with_system_turn(convo, system):
+ if (
+ isinstance(system, str)
+ and system.strip()
+ and convo
+ and isinstance(convo[0], dict)
+ and convo[0].get("role") != "system"
+ ):
+ return [{"role": "system", "content": system}, *convo]
+ return convo
+
+
def apply_chat_template_to_dataset(
dataset_info,
tokenizer,
@@ -516,12 +528,20 @@ def apply_chat_template_to_dataset(
def _format_chatml(examples):
convos = examples[chat_column]
+ systems = examples.get("system") or [None] * len(convos)
texts = []
row_errors = []
- for convo in convos:
+ for convo, system in zip(convos, systems):
try:
- text = _render_conversation(tokenizer, convo)
+ with_system = _with_system_turn(convo, system)
+ try:
+ text = _render_conversation(tokenizer, with_system)
+ except Exception:
+ # A template without a system role still trains the conversation.
+ if with_system is convo:
+ raise
+ text = _render_conversation(tokenizer, convo)
if remove_bos_prefix:
text = text.removeprefix('<bos>')
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index e5714187efe..bc0dd5868db 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -236,8 +236,9 @@ def convert_chatml_to_alpaca(
instructions = []
outputs = []
inputs = []
+ column_systems = examples.get("system") or [None] * len(chatml_data)
- for convo in chatml_data:
+ for convo, column_system in zip(chatml_data, column_systems):
turns = []
for msg in convo or []:
role = roles.get(msg.get("role") or msg.get("from"))
@@ -248,6 +249,13 @@ def convert_chatml_to_alpaca(
turns[-1][1] = f"{turns[-1][1]}\n\n{content}"
else:
turns.append([role, content])
+ # A separate system column fills in only when the conversation has no leading system turn.
+ if (
+ isinstance(column_system, str)
+ and column_system.strip()
+ and not (turns and turns[0][0] == "system")
+ ):
+ turns.insert(0, ["system", column_system])
system = ""
context = []