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 every turn when training a chat dataset with Alpaca format (#11053)

Nilaydc0a7dbd9a6e
diff --git a/studio/backend/tests/test_chatml_to_alpaca.py b/studio/backend/tests/test_chatml_to_alpaca.py
new file mode 100644
index 00000000000..67a05388a53
--- /dev/null
+++ b/studio/backend/tests/test_chatml_to_alpaca.py
@@ -0,0 +1,190 @@
+# 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, IterableDataset
+
+from utils.datasets import convert_chatml_to_alpaca, format_and_template_dataset
+
+SHAREGPT_ROLES = {"system": "system", "user": "human", "assistant": "gpt"}
+
+
+class _Tokenizer:
+ chat_template = "{{ messages }}"
+ eos_token = "</s>"
+
+
+def _messages(row):
+ return [
+ {"role": "system", "content": f"SYS-{row}"},
+ {"role": "user", "content": f"Q1-{row}"},
+ {"role": "assistant", "content": f"A1-{row}"},
+ {"role": "user", "content": f"Q2-{row}"},
+ {"role": "assistant", "content": f"A2-{row}"},
+ {"role": "user", "content": f"Q3-{row}"},
+ {"role": "assistant", "content": f"A3-{row}"},
+ ]
+
+
+def _sharegpt(row):
+ return [
+ {"from": SHAREGPT_ROLES[message["role"]], "value": message["content"]}
+ for message in _messages(row)
+ ]
+
+
+def _expected(row):
+ first = f"User: Q1-{row}\nAssistant: A1-{row}"
+ second = f"User: Q2-{row}\nAssistant: A2-{row}"
+ return [
+ {"instruction": f"Q1-{row}", "input": f"SYS-{row}", "output": f"A1-{row}"},
+ {"instruction": f"Q2-{row}", "input": f"SYS-{row}\n\n{first}", "output": f"A2-{row}"},
+ {
+ "instruction": f"Q3-{row}",
+ "input": f"SYS-{row}\n\n{first}\n{second}",
+ "output": f"A3-{row}",
+ },
+ ]
+
+
+ ("column", "conversation"),
+ [("messages", _messages), ("conversations", _sharegpt)],
+)
+def test_every_reply_becomes_a_row_with_the_system_prompt_and_earlier_exchanges(
+ column, conversation
+):
+ dataset = Dataset.from_dict({column: [conversation(1), conversation(2)]})
+
+ converted = convert_chatml_to_alpaca(dataset, batch_size = 1, num_proc = 1, chat_column = column)
+
+ assert converted.column_names == ["instruction", "input", "output"]
+ assert converted.to_list() == _expected(1) + _expected(2)
+
+
+def test_a_single_exchange_converts_as_before():
+ dataset = Dataset.from_dict(
+ {
+ "messages": [
+ [
+ {"role": "user", "content": "Hello"},
+ {"role": "assistant", "content": "Hi there"},
+ ]
+ ]
+ }
+ )
+
+ converted = convert_chatml_to_alpaca(dataset, num_proc = 1)
+
+ assert converted.select_columns(["instruction", "input", "output"]).to_list() == [
+ {"instruction": "Hello", "input": "", "output": "Hi there"}
+ ]
+
+
+def test_unpaired_empty_and_unknown_turns_are_skipped():
+ dataset = Dataset.from_dict(
+ {
+ "messages": [
+ None,
+ [],
+ [
+ {"role": "assistant", "content": "orphan"},
+ {"role": "user", "content": "Q"},
+ {"role": "tool", "content": "tool output"},
+ {"role": "user", "content": "more"},
+ {"role": "assistant", "content": ""},
+ {"role": "assistant", "content": "A"},
+ {"role": "assistant", "content": "A again"},
+ {"role": "user", "content": "trailing"},
+ ],
+ ]
+ }
+ )
+
+ converted = convert_chatml_to_alpaca(dataset, num_proc = 1)
+
+ assert converted.to_list() == [
+ {"instruction": "Q\n\nmore", "input": "", "output": "A\n\nA again"}
+ ]
+
+
+def test_parallel_conversion_keeps_every_reply_in_order(monkeypatch):
+ monkeypatch.setattr("utils.hardware.dataset_map_num_proc", lambda num_proc = None: num_proc)
+ dataset = Dataset.from_dict({"messages": [_messages(row) for row in range(1, 9)]})
+
+ converted = convert_chatml_to_alpaca(dataset, batch_size = 1, num_proc = 2)
+
+ assert converted.to_list() == [line for row in range(1, 9) for line in _expected(row)]
+
+
+ "stream",
+ [
+ lambda rows: IterableDataset.from_generator(lambda: iter(rows)),
+ lambda rows: Dataset.from_list(rows).to_iterable_dataset(),
+ ],
+ ids = ["generator", "typed"],
+)
+def test_streaming_conversion_splits_replies_the_same_way(stream):
+ rows = [{"id": row, "messages": _messages(row)} for row in (1, 2)]
+
+ converted = convert_chatml_to_alpaca(stream(rows), batch_size = 1)
+
+ assert list(converted) == _expected(1) + _expected(2)
+
+
+def test_structured_content_converts_to_text():
+ structured = [
+ {"role": message["role"], "content": [{"type": "text", "text": message["content"]}]}
+ for message in _messages(1)
+ ]
+ dataset = Dataset.from_dict({"messages": [structured]})
+
+ converted = convert_chatml_to_alpaca(dataset, num_proc = 1)
+
+ assert converted.to_list() == _expected(1)
+
+
+def test_alpaca_format_with_a_processor_keeps_the_system_prompt():
+ class _Processor(_Tokenizer):
+ image_processor = object()
+
+ dataset = Dataset.from_dict({"conversations": [_sharegpt(1)]})
+
+ result = format_and_template_dataset(
+ dataset,
+ model_name = "Gemma3ForConditionalGeneration",
+ tokenizer = _Processor(),
+ format_type = "alpaca",
+ batch_size = 1,
+ num_proc = 1,
+ )
+
+ assert result["success"] is True
+ assert result["final_format"] == "alpaca"
+ texts = list(result["dataset"]["text"])
+ assert len(texts) == 3
+ assert "### Input:\nSYS-1\n\nUser: Q1-1\nAssistant: A1-1\n" in texts[2]
+
+
+def test_alpaca_format_trains_on_every_sharegpt_exchange():
+ dataset = Dataset.from_dict({"conversations": [_sharegpt(1), _sharegpt(2)]})
+
+ result = format_and_template_dataset(
+ dataset,
+ model_name = "Qwen2ForCausalLM",
+ tokenizer = _Tokenizer(),
+ format_type = "alpaca",
+ batch_size = 1,
+ num_proc = 1,
+ )
+
+ assert result["success"] is True
+ assert result["detected_format"] == "sharegpt"
+ assert result["final_format"] == "alpaca"
+ texts = list(result["dataset"]["text"])
+ assert len(texts) == 6
+ assert texts[5].endswith(
+ "### Instruction:\nQ3-2\n\n### Input:\nSYS-2\n\nUser: Q1-2\nAssistant: A1-2\n"
+ "User: Q2-2\nAssistant: A2-2\n\n### Response:\nA3-2"
+ )
diff --git a/studio/backend/utils/datasets/format_conversion.py b/studio/backend/utils/datasets/format_conversion.py
index 4dfbedc619a..44cadc3b504 100644
--- a/studio/backend/utils/datasets/format_conversion.py
+++ b/studio/backend/utils/datasets/format_conversion.py
@@ -178,6 +178,16 @@ def standardize_chat_format(
return result
+def _content_text(content):
+ if isinstance(content, list):
+ return "\n".join(
+ part["text"]
+ for part in content
+ if isinstance(part, dict) and part.get("type") == "text" and part.get("text")
+ )
+ return cell_text(content)
+
+
def convert_chatml_to_alpaca(
dataset,
batch_size = 1000,
@@ -186,6 +196,15 @@ def convert_chatml_to_alpaca(
):
"""Convert ChatML to Alpaca format. Accepts a "messages" or "conversations" column with either standard "role"/"content" or ShareGPT "from"/"value" keys."""
is_iterable = is_streaming_dataset(dataset)
+ roles = {
+ "system": "system",
+ "user": "user",
+ "human": "user",
+ "input": "user",
+ "assistant": "assistant",
+ "gpt": "assistant",
+ "output": "assistant",
+ }
def _convert(examples):
chatml_data = examples.get(chat_column) if chat_column else None
@@ -202,28 +221,40 @@ def convert_chatml_to_alpaca(
inputs = []
for convo in chatml_data:
- instruction = ""
- output = ""
-
- for msg in convo:
- role = msg.get("role") or msg.get("from")
- content = msg.get("content") or msg.get("value")
-
- if role in ["user", "human", "input"] and not instruction:
+ turns = []
+ for msg in convo or []:
+ role = roles.get(msg.get("role") or msg.get("from"))
+ content = _content_text(msg.get("content") or msg.get("value"))
+ if role is None or not content:
+ continue
+ if turns and turns[-1][0] == role:
+ turns[-1][1] = f"{turns[-1][1]}\n\n{content}"
+ else:
+ turns.append([role, content])
+
+ system = ""
+ context = []
+ instruction = None
+ for role, content in turns:
+ if role == "system":
+ system = f"{system}\n\n{content}" if system else content
+ elif role == "user":
instruction = content
- elif role in ["assistant", "gpt", "output"] and not output:
- output = content
- break
-
- instructions.append(instruction)
- inputs.append("")
- outputs.append(output)
+ elif instruction is not None:
+ instructions.append(instruction)
+ inputs.append(
+ "\n\n".join(part for part in (system, "\n".join(context)) if part)
+ )
+ outputs.append(content)
+ context += [f"User: {instruction}", f"Assistant: {content}"]
+ instruction = None
return {"instruction": instructions, "input": inputs, "output": outputs}
dataset_map_kwargs = {
"batched": True,
"batch_size": batch_size,
+ "remove_columns": dataset.column_names or list(next(iter(dataset), {})),
}
if not is_iterable:
diff --git a/tests/python/test_studio_import_no_torch.py b/tests/python/test_studio_import_no_torch.py
index 503980b938e..3e37b6c7a5e 100644
--- a/tests/python/test_studio_import_no_torch.py
+++ b/tests/python/test_studio_import_no_torch.py
@@ -404,6 +404,8 @@ class TestFormatConversionNoTorchVenv:
# Test convert_chatml_to_alpaca with a simple dataset
class FakeDataset:
+ column_names = ['messages']
+
def map(self, fn, **kw):
result = fn({{
'messages': [[