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: accept the chat columns the trainer accepts in the dataset check (#11849)

Nilay4c42e20ca4ad
diff --git a/studio/backend/hub/utils/dataset_format.py b/studio/backend/hub/utils/dataset_format.py
index af6d80cfb4e..9a48437c564 100644
--- a/studio/backend/hub/utils/dataset_format.py
+++ b/studio/backend/hub/utils/dataset_format.py
@@ -6,6 +6,8 @@ from __future__ import annotations
import re
from typing import Any, Optional
+from utils.datasets.format_detection import detect_dataset_format
+
def _first_row(dataset) -> Optional[dict]:
try:
@@ -26,66 +28,6 @@ def _keyword_in_column(keyword: str, col_name: str) -> bool:
return re.search(r"\b" + re.escape(keyword) + r"\b", col_name, re.IGNORECASE) is not None
-def _unknown_dataset_format(
- chat_column: Optional[str] = None, sample_keys: Optional[list[str]] = None
-) -> dict:
- return {
- "format": "unknown",
- "chat_column": chat_column,
- "needs_standardization": None,
- "sample_keys": sample_keys or [],
- }
-
-
-def detect_dataset_format(dataset) -> dict:
- sample = _first_row(dataset)
- if sample is None:
- return _unknown_dataset_format()
- column_names = set(sample.keys())
- if {"instruction", "output"}.issubset(column_names):
- return {
- "format": "alpaca",
- "chat_column": None,
- "needs_standardization": False,
- "sample_keys": [],
- }
-
- chat_column = None
- if "messages" in column_names:
- chat_column = "messages"
- elif "conversations" in column_names:
- chat_column = "conversations"
- elif "texts" in column_names:
- chat_column = "texts"
-
- if not chat_column:
- return _unknown_dataset_format()
-
- chat_data = sample.get(chat_column)
- if not isinstance(chat_data, (list, tuple)) or not chat_data:
- return _unknown_dataset_format(chat_column)
- first_msg = chat_data[0]
- if not isinstance(first_msg, dict):
- return _unknown_dataset_format(chat_column)
- msg_keys = set(first_msg.keys())
- sample_keys = [str(key) for key in msg_keys]
- if "from" in msg_keys or "value" in msg_keys:
- return {
- "format": "sharegpt",
- "chat_column": chat_column,
- "needs_standardization": True,
- "sample_keys": sample_keys,
- }
- if "role" in msg_keys and "content" in msg_keys:
- return {
- "format": "chatml",
- "chat_column": chat_column,
- "needs_standardization": False,
- "sample_keys": sample_keys,
- }
- return _unknown_dataset_format(chat_column, sample_keys)
-
-
def detect_custom_format_heuristic(dataset):
sample = _first_row(dataset)
if sample is None:
@@ -742,6 +684,7 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"suggested_mapping": None,
"detected_image_column": None,
"detected_text_column": None,
+ "chat_column": detected.get("chat_column"),
"is_image": multimodal_info["is_image"],
"multimodal_columns": multimodal_info.get("multimodal_columns"),
**audio_fields,
diff --git a/studio/backend/tests/test_dataset_check_chat_column.py b/studio/backend/tests/test_dataset_check_chat_column.py
new file mode 100644
index 00000000000..34df3489b72
--- /dev/null
+++ b/studio/backend/tests/test_dataset_check_chat_column.py
@@ -0,0 +1,76 @@
+# 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 sys
+from pathlib import Path
+
+import pytest
+from datasets import Dataset
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+ sys.path.insert(0, str(_BACKEND_ROOT))
+
+from hub.utils import dataset_format # noqa: E402
+from utils.datasets import format_detection # noqa: E402
+
+_CHATML = [
+ {"role": "user", "content": "a"},
+ {"role": "assistant", "content": "b"},
+]
+_SHAREGPT = [
+ {"from": "human", "value": "a"},
+ {"from": "gpt", "value": "b"},
+]
+
+_CASES = [
+ pytest.param([{"messages": _CHATML}], "chatml", "messages", id = "messages"),
+ pytest.param([{"conversations": _SHAREGPT}], "sharegpt", "conversations", id = "conversations"),
+ pytest.param([{"texts": _CHATML}], "chatml", "texts", id = "texts"),
+ pytest.param([{"dialog": _CHATML}], "chatml", "dialog", id = "dialog"),
+ pytest.param([{"conversation": _SHAREGPT}], "sharegpt", "conversation", id = "conversation"),
+ pytest.param([{"id": "1", "agent_trace": _CHATML}], "chatml", "agent_trace", id = "agent_trace"),
+ pytest.param(
+ [{"messages": []}, {"messages": _CHATML}], "chatml", "messages", id = "empty-first-row"
+ ),
+]
+
+
[email protected]("rows, expected_format, expected_column", _CASES)
+def test_check_accepts_the_chat_column_the_trainer_uses(rows, expected_format, expected_column):
+ dataset = Dataset.from_list(rows)
+ trainer = format_detection.detect_dataset_format(dataset)
+ result = dataset_format.check_dataset_format(dataset)
+
+ assert trainer["format"] == expected_format
+ assert trainer["chat_column"] == expected_column
+ assert result["requires_manual_mapping"] is False
+ assert result["detected_format"] == expected_format
+ assert result["chat_column"] == expected_column
+
+
+def test_alpaca_still_passes_without_a_chat_column():
+ dataset = Dataset.from_list([{"instruction": "a", "input": "", "output": "b"}])
+ result = dataset_format.check_dataset_format(dataset)
+
+ assert result["requires_manual_mapping"] is False
+ assert result["detected_format"] == "alpaca"
+ assert result["chat_column"] is None
+
+
+def test_plain_columns_still_fall_back_to_the_heuristic():
+ dataset = Dataset.from_list([{"question": "a", "answer": "b"}])
+ result = dataset_format.check_dataset_format(dataset)
+
+ assert result["detected_format"] == "custom_heuristic"
+ assert result["suggested_mapping"] == {"question": "user", "answer": "assistant"}
+
+
+def test_preview_standardizes_a_sharegpt_column_under_another_name():
+ dataset = Dataset.from_list([{"conversation": _SHAREGPT}])
+ preview = dataset_format.format_dataset_preview(dataset)
+
+ assert preview[0]["conversation"] == [
+ {"role": "user", "content": "a"},
+ {"role": "assistant", "content": "b"},
+ ]