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 single-string-column datasets as raw text (#15899)

Yakshithffd1a184c683
diff --git a/studio/backend/hub/utils/dataset_format.py b/studio/backend/hub/utils/dataset_format.py
index fa558e06c61..0a8507b8cf9 100644
--- a/studio/backend/hub/utils/dataset_format.py
+++ b/studio/backend/hub/utils/dataset_format.py
@@ -7,7 +7,11 @@ import re
from typing import Any, Optional
from utils.datasets.cells import text_cell_check, unwrap_message_column
-from utils.datasets.format_detection import detect_dataset_format, has_message_prompt_completion
+from utils.datasets.format_detection import (
+ detect_dataset_format,
+ has_message_prompt_completion,
+ is_plain_text_dataset,
+)
from utils.datasets.pretokenized import PRETOKENIZED_FORMAT, is_pretokenized_dataset
@@ -734,6 +738,23 @@ def check_dataset_format(dataset, is_vlm: bool = False) -> dict:
"detected_speaker_column": multimodal_info.get("detected_speaker_column"),
}
+ if is_plain_text_dataset(dataset):
+ # Before the VLM branch: text-only data trains as raw text on a vision model too.
+ # Media flags off: a name match (a "speech" column) would route an audio model to its audio path.
+ return {
+ "requires_manual_mapping": False,
+ "detected_format": "raw",
+ "columns": columns,
+ "suggested_mapping": None,
+ "detected_image_column": None,
+ "detected_text_column": None,
+ "is_image": False,
+ "multimodal_columns": None,
+ "is_audio": False,
+ "detected_audio_column": None,
+ "detected_speaker_column": None,
+ }
+
if is_vlm:
vlm_structure = detect_vlm_dataset_structure(dataset)
requires_mapping = vlm_structure["format"] == "unknown"
diff --git a/studio/backend/tests/test_plain_text_dataset.py b/studio/backend/tests/test_plain_text_dataset.py
new file mode 100644
index 00000000000..8d52d8728bb
--- /dev/null
+++ b/studio/backend/tests/test_plain_text_dataset.py
@@ -0,0 +1,117 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""One string column is raw training text: the preview and the formatter must agree on it."""
+
+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
+from utils.datasets.dataset_utils import format_dataset # noqa: E402
+
+
+def _text_dataset(column: str = "text"):
+ return Dataset.from_list(
+ [
+ {column: "## What happened\nWrote the fix."},
+ {column: "## What happened\nReviewed the diff."},
+ ]
+ )
+
+
[email protected]("column", ["text", "content", "body"], ids = ["text", "content", "body"])
[email protected]("is_vlm", [False, True], ids = ["text_model", "vision_model"])
+def test_single_string_column_needs_no_mapping(column, is_vlm):
+ result = dataset_format.check_dataset_format(_text_dataset(column), is_vlm = is_vlm)
+
+ assert result["requires_manual_mapping"] is False
+ assert result["detected_format"] == "raw"
+ assert result["suggested_mapping"] is None
+ assert result["columns"] == [column]
+
+
+def test_text_column_keeps_detection_flags_for_other_branches():
+ result = dataset_format.check_dataset_format(_text_dataset())
+
+ assert result["is_image"] is False
+ assert result["is_audio"] is False
+ assert result["detected_image_column"] is None
+
+
+def test_media_named_text_column_is_not_routed_as_media():
+ result = dataset_format.check_dataset_format(_text_dataset("speech"), is_vlm = True)
+
+ assert result["detected_format"] == "raw"
+ assert result["is_audio"] is False
+ assert result["detected_audio_column"] is None
+ assert result["is_image"] is False
+
+
+ "rows",
+ [
+ pytest.param([{"text": "a", "id": 1}], id = "text_plus_metadata"),
+ pytest.param([{"text": " "}], id = "blank_text"),
+ pytest.param([{"text": 5}], id = "non_string_text"),
+ ],
+)
+def test_other_shapes_still_require_mapping(rows):
+ result = dataset_format.check_dataset_format(Dataset.from_list(rows))
+
+ assert result["detected_format"] == "unknown"
+ assert result["requires_manual_mapping"] is True
+
+
+ "row", [{"image": "photos/sloth.jpg"}, {"audio": "clips/sloth.wav"}], ids = ["image", "audio"]
+)
+def test_media_path_column_is_not_raw_text(row):
+ result = dataset_format.check_dataset_format(Dataset.from_list([row]))
+
+ assert result["detected_format"] != "raw"
+ assert result["requires_manual_mapping"] is True
+
+
+def test_plain_text_dataset_helper_rejects_other_shapes():
+ assert format_detection.is_plain_text_dataset(_text_dataset()) is True
+ assert format_detection.is_plain_text_dataset(_text_dataset("content")) is True
+ assert (
+ format_detection.is_plain_text_dataset(Dataset.from_list([{"text": "a", "id": 1}])) is False
+ )
+
+
[email protected]("column", ["text", "content"], ids = ["text", "content"])
+def test_auto_resolves_a_single_string_column_to_raw_text(column):
+ info = format_dataset(_text_dataset(column), format_type = "auto")
+
+ assert info["final_format"] == "raw_text"
+ assert info["requires_manual_mapping"] is False
+ assert info["dataset"].column_names == ["text"]
+ assert info["dataset"]["text"] == [
+ "## What happened\nWrote the fix.",
+ "## What happened\nReviewed the diff.",
+ ]
+
+
+def test_renamed_column_is_reported():
+ info = format_dataset(_text_dataset("content"), format_type = "auto")
+
+ assert any("content" in warning for warning in info["warnings"])
+
+
+def test_auto_leaves_a_mapped_dataset_alone():
+ info = format_dataset(
+ Dataset.from_list([{"text": "a", "id": 1}]),
+ format_type = "auto",
+ custom_format_mapping = {"text": "user"},
+ )
+
+ assert info["final_format"] != "raw_text"
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index 038f9875e87..eef0f6e24b0 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -10,6 +10,7 @@ from .format_detection import (
detect_multimodal_dataset,
detect_vlm_dataset_structure,
detect_custom_format_heuristic,
+ is_plain_text_dataset,
)
from .format_conversion import (
standardize_chat_format,
@@ -435,6 +436,10 @@ def _format_dataset(
multimodal_info = detect_multimodal_dataset(dataset)
+ if format_type == "auto" and not custom_format_mapping and is_plain_text_dataset(dataset):
+ # Matches the preview's "raw" verdict; auto would otherwise fail as "unknown".
+ format_type = "raw"
+
if format_type == "raw":
raw_result = prepare_raw_text_dataset(
dataset, split_name = split_name, text_column = raw_text_column
diff --git a/studio/backend/utils/datasets/format_detection.py b/studio/backend/utils/datasets/format_detection.py
index 2ff229bd5d4..30fd5ebd345 100644
--- a/studio/backend/utils/datasets/format_detection.py
+++ b/studio/backend/utils/datasets/format_detection.py
@@ -17,6 +17,7 @@ CONVERSATION_COLUMNS = ("messages", "conversations", "texts")
_CHATML_KEYS = frozenset({"role", "content"})
_SHAREGPT_KEYS = frozenset({"from", "value"})
_TRACE_SUFFIXES = ("__trace", "_trace")
+_PLAIN_TEXT_SAMPLE_ROWS = 10
def _sample_dataset_rows(dataset, limit: int = 100) -> list[dict]:
@@ -159,6 +160,27 @@ def _detect_conversation_column(
return None
+def is_plain_text_dataset(dataset) -> bool:
+ """One column of non-blank strings: raw text, whatever the column is called (prepare_raw_text_dataset renames it)."""
+ column_names = getattr(dataset, "column_names", None)
+ if column_names is not None and len(list(column_names)) != 1:
+ return False
+ rows = _sample_dataset_rows(dataset, limit = _PLAIN_TEXT_SAMPLE_ROWS)
+ if not rows:
+ return False
+ return all(_is_single_string_row(row) for row in rows)
+
+
+def _is_single_string_row(row) -> bool:
+ if not isinstance(row, dict) or len(row) != 1:
+ return False
+ value = next(iter(row.values()))
+ if not isinstance(value, str) or not value.strip() or _is_image_value(value):
+ return False
+ # A column of media paths still needs its other roles mapped.
+ return not value.strip().lower().endswith(_AUDIO_EXTENSIONS)
+
+
def detect_dataset_format(dataset):
"""Detect dataset format by inspecting structure: {"format": alpaca/sharegpt/chatml/unknown, "chat_column": str or None, "needs_standardization": bool, "sample_keys": keys found in messages}."""
sample_rows = _sample_dataset_rows(dataset)