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: keep AI Assist label names and system prompt when the dataset format is Alpaca (#11597)

Nilay43e0921ab937
diff --git a/studio/backend/tests/test_alpaca_ai_assist_mapping.py b/studio/backend/tests/test_alpaca_ai_assist_mapping.py
new file mode 100644
index 00000000000..dcab1c7e10d
--- /dev/null
+++ b/studio/backend/tests/test_alpaca_ai_assist_mapping.py
@@ -0,0 +1,142 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+from datasets import Dataset
+
+from utils.datasets import format_dataset
+
+PROMPT = "Classify the sentiment."
+LABEL_MAPPING = {"label": {"0": "neg", "1": "pos"}}
+
+
+def _dataset():
+ return Dataset.from_dict({"text": ["Loved it", "Hated it"], "label": [1, 0]})
+
+
+def _format(format_type, mapping):
+ return format_dataset(
+ _dataset(),
+ format_type = format_type,
+ batch_size = 2,
+ custom_format_mapping = {"text": "instruction", "label": "output", **mapping},
+ )
+
+
+def test_alpaca_mapping_keeps_label_names_and_system_prompt():
+ result = _format("alpaca", {"__label_mapping": LABEL_MAPPING, "__system_prompt": PROMPT})
+
+ assert result["final_format"] == "alpaca", result["warnings"]
+ rows = result["dataset"]
+ assert list(rows["output"]) == ["pos", "neg"]
+ assert list(rows["instruction"]) == [f"{PROMPT}\n\nLoved it", f"{PROMPT}\n\nHated it"]
+ assert list(rows["input"]) == ["", ""]
+
+
+def test_alpaca_mapping_uses_system_prompt_alone_for_a_blank_instruction():
+ result = format_dataset(
+ Dataset.from_dict({"text": ["Loved it", ""], "label": [1, 0]}),
+ format_type = "alpaca",
+ batch_size = 2,
+ custom_format_mapping = {"text": "instruction", "label": "output", "__system_prompt": PROMPT},
+ )
+
+ assert result["final_format"] == "alpaca", result["warnings"]
+ assert list(result["dataset"]["instruction"]) == [f"{PROMPT}\n\nLoved it", PROMPT]
+
+
+def test_alpaca_mapping_keeps_a_zero_label_and_blanks_a_missing_one():
+ result = format_dataset(
+ Dataset.from_dict({"text": ["a", "b", "c"], "label": [1, 0, None]}),
+ format_type = "alpaca",
+ batch_size = 3,
+ custom_format_mapping = {"text": "instruction", "label": "output"},
+ )
+
+ assert list(result["dataset"]["output"]) == ["1", "0", ""]
+
+
+def test_alpaca_mapping_keeps_every_column_the_advisor_assigns_to_a_role():
+ result = format_dataset(
+ Dataset.from_dict({"premise": ["P1", "P2"], "hypothesis": ["H1", "H2"], "label": [0, 1]}),
+ format_type = "alpaca",
+ batch_size = 2,
+ custom_format_mapping = {
+ "premise": "instruction",
+ "hypothesis": "instruction",
+ "label": "output",
+ "__label_mapping": LABEL_MAPPING,
+ "__system_prompt": PROMPT,
+ },
+ )
+
+ assert result["final_format"] == "alpaca", result["warnings"]
+ assert list(result["dataset"]["instruction"]) == [
+ f"{PROMPT}\n\nP1\nH1",
+ f"{PROMPT}\n\nP2\nH2",
+ ]
+ assert list(result["dataset"]["output"]) == ["neg", "pos"]
+
+
+def test_alpaca_mapping_renders_list_and_text_struct_cells_as_text():
+ result = format_dataset(
+ Dataset.from_dict(
+ {
+ "text": ["a", "b"],
+ "labels": [[0], [1, 2]],
+ "answers": [{"text": ["Paris", "paris"]}, {"text": ["Rome"]}],
+ }
+ ),
+ format_type = "alpaca",
+ batch_size = 2,
+ custom_format_mapping = {"text": "instruction", "labels": "input", "answers": "output"},
+ )
+
+ assert result["final_format"] == "alpaca", result["warnings"]
+ assert list(result["dataset"]["input"]) == ["0", "1, 2"]
+ assert list(result["dataset"]["output"]) == ["Paris", "Rome"]
+
+
+def test_alpaca_mapping_names_each_label_in_a_list_cell():
+ result = format_dataset(
+ Dataset.from_dict({"text": ["a", "b"], "labels": [[0], [1, 2]]}),
+ format_type = "alpaca",
+ batch_size = 2,
+ custom_format_mapping = {
+ "text": "instruction",
+ "labels": "output",
+ "__label_mapping": {"labels": {"0": "joy", "1": "anger", "2": "fear"}},
+ },
+ )
+
+ assert result["final_format"] == "alpaca", result["warnings"]
+ assert list(result["dataset"]["output"]) == ["joy", "anger, fear"]
+
+
+def test_chatml_advisor_mapping_adds_system_turn_and_label_names():
+ result = _format("chatml", {"__label_mapping": LABEL_MAPPING, "__system_prompt": PROMPT})
+
+ assert result["final_format"] == "chatml_conversations", result["warnings"]
+ assert list(result["dataset"]["conversations"]) == [
+ [
+ {"role": "system", "content": PROMPT},
+ {"role": "user", "content": "Loved it"},
+ {"role": "assistant", "content": "pos"},
+ ],
+ [
+ {"role": "system", "content": PROMPT},
+ {"role": "user", "content": "Hated it"},
+ {"role": "assistant", "content": "neg"},
+ ],
+ ]
+
+
+def test_chatml_advisor_mapping_renders_list_cells_as_text():
+ result = format_dataset(
+ Dataset.from_dict({"text": ["a", "b"], "labels": [[0], [1, 2]]}),
+ format_type = "chatml",
+ batch_size = 2,
+ custom_format_mapping = {"text": "user", "labels": "assistant", "__system_prompt": PROMPT},
+ )
+
+ assert result["final_format"] == "chatml_conversations", result["warnings"]
+ assert [convo[-1]["content"] for convo in result["dataset"]["conversations"]] == ["0", "1, 2"]
diff --git a/studio/backend/utils/datasets/dataset_utils.py b/studio/backend/utils/datasets/dataset_utils.py
index 8f0c3b79e78..f90f79a98dc 100644
--- a/studio/backend/utils/datasets/dataset_utils.py
+++ b/studio/backend/utils/datasets/dataset_utils.py
@@ -26,6 +26,7 @@ from .chat_templates import (
get_tokenizer_chat_template,
DEFAULT_ALPACA_TEMPLATE,
)
+from .cells import cell_text
from .raw_text import prepare_raw_text_dataset
from .vlm_processing import generate_smart_vlm_instruction
from .data_collators import DeepSeekOCRDataCollator, VLMDataCollator
@@ -222,20 +223,16 @@ def _apply_user_mapping(
def _extract_column_value(val, col: str, label_mapping: dict) -> str:
"""Extract a string value from a column, handling complex types and label mapping."""
if isinstance(val, dict):
- if "text" in val:
- inner = val["text"]
- str_val = inner[0] if isinstance(inner, list) and inner else str(inner)
- else:
- str_val = json.dumps(val, ensure_ascii = False)
- elif isinstance(val, list):
- str_val = val[0] if len(val) == 1 else ", ".join(str(v) for v in val)
- else:
- str_val = str(val) if val is not None else ""
-
- if col in label_mapping and isinstance(label_mapping[col], dict):
- str_val = label_mapping[col].get(str_val, str_val)
+ if "text" not in val:
+ return json.dumps(val, ensure_ascii = False)
+ inner = val["text"]
+ val = inner[0] if isinstance(inner, list) and inner else inner
- return str_val
+ names = label_mapping.get(col)
+ if not isinstance(names, dict):
+ names = {}
+ values = val if isinstance(val, list) else [val]
+ return ", ".join(names.get(text, text) for text in map(cell_text, values))
def _apply_template_mapping(
@@ -301,17 +298,22 @@ def _apply_user_mapping_alpaca(
mapping: dict,
batch_size: int = 1000,
):
- """Apply user-provided column mapping to convert dataset to Alpaca format. Accepts any format's role names, normalises via _TO_CHATML, then maps user -> instruction, system -> input, assistant -> output. Returns a dataset with instruction/input/output columns."""
- col_for: dict[str, str | None] = {
- "instruction": None,
- "input": None,
- "output": None,
+ """Apply user-provided column mapping to convert dataset to Alpaca format. Accepts any format's role names, normalises via _TO_CHATML, then maps user -> instruction, system -> input, assistant -> output. Advisor ``__label_mapping`` names label values; ``__system_prompt`` is prepended to instruction, since system-role columns already fill input. Returns a dataset with instruction/input/output columns."""
+ meta = {k: v for k, v in mapping.items() if k.startswith("__")}
+ column_roles = {k: v for k, v in mapping.items() if not k.startswith("__")}
+ system_prompt = meta.get("__system_prompt", "")
+ label_mapping = meta.get("__label_mapping", {})
+
+ cols_for: dict[str, list[str]] = {
+ "instruction": [],
+ "input": [],
+ "output": [],
}
- for col_name, role in mapping.items():
+ for col_name, role in column_roles.items():
canonical = _TO_CHATML.get(role)
alpaca_field = _CHATML_TO_ALPACA.get(canonical) if canonical else None
if alpaca_field:
- col_for[alpaca_field] = col_name
+ cols_for[alpaca_field].append(col_name)
def _convert(examples):
num = len(next(iter(examples.values())))
@@ -322,8 +324,13 @@ def _apply_user_mapping_alpaca(
("input", inputs),
("output", outputs),
):
- col = col_for[field]
- val = str(examples[col][i]) if col and col in examples and examples[col][i] else ""
+ val = "\n".join(
+ _extract_column_value(examples[col][i], col, label_mapping)
+ for col in cols_for[field]
+ if col in examples
+ )
+ if field == "instruction" and system_prompt:
+ val = "\n\n".join(part for part in (system_prompt, val) if part)
dest.append(val)
return {"instruction": instructions, "input": inputs, "output": outputs}