Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add masking strategies to message transforms #2284

Open
wants to merge 4 commits into
base: main
Choose a base branch
from
Open

Add masking strategies to message transforms #2284

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ac6272f
Add masking strategies to message transforms
supreethmanyam Jan 20, 2025
4e97a57
Add tests for masking strategies in message transforms
supreethmanyam Jan 21, 2025
0381a13
Merge branch 'pytorch:main' into add-chat-loss-masking-strategies
supreethmanyam Jan 28, 2025
8ef107a
Convert MaskingStrategy to Enum and fix default strategy in transforms
supreethmanyam Jan 28, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions tests/test_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -55,14 +55,17 @@
Message(
role="system",
content=CHAT_SAMPLE["system"],
masked=True,
),
Message(
role="user",
content=CHAT_SAMPLE["user"],
masked=False,
),
Message(
role="assistant",
content=CHAT_SAMPLE["assistant"],
masked=False,
),
]

Expand All @@ -72,6 +75,7 @@
Message(
role="assistant",
content=CHAT_SAMPLE["assistant"],
masked=False,
),
]

Expand Down
157 changes: 157 additions & 0 deletions tests/torchtune/data/test_messages.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
from torchtune.data._messages import (
ChosenRejectedToMessages,
InputOutputToMessages,
mask_messages,
Message,
OpenAIToMessages,
ShareGPTToMessages,
Expand Down Expand Up @@ -186,6 +187,33 @@ def test_call_train_on_input(self, sample):
]
assert_dialogue_equal(actual["messages"], expected)

@pytest.mark.parametrize(
"masking_strategy, expected_masks",
[
("train_on_all", [False, False]),
("train_on_assistant", [True, False]),
("train_on_last", [True, False]),
],
)
def test_call_masking_strategy(self, sample, masking_strategy, expected_masks):
transform = InputOutputToMessages(
column_map={"input": "maybe_input", "output": "maybe_output"},
masking_strategy=masking_strategy,
)
actual = transform(sample)
expected = [
Message(
role="user", content="hello world", masked=expected_masks[0], eot=True
),
Message(
role="assistant",
content="hello world",
masked=expected_masks[1],
eot=True,
),
]
assert_dialogue_equal(actual["messages"], expected)

def test_system_prompt(self, sample):
transform = InputOutputToMessages(
column_map={"input": "maybe_input", "output": "maybe_output"},
Expand Down Expand Up @@ -267,6 +295,55 @@ def test_call_train_on_input(self, sample):
]
assert_dialogue_equal(actual["rejected"], expected_rejected)

@pytest.mark.parametrize(
"masking_strategy, expected_masks",
[
("train_on_all", [[False, False], [False, False]]),
("train_on_assistant", [[True, False], [True, False]]),
("train_on_last", [[True, False], [True, False]]),
],
)
def test_call_masking_strategy(self, sample, masking_strategy, expected_masks):
transform = ChosenRejectedToMessages(
column_map={
"chosen": "maybe_chosen",
"rejected": "maybe_rejected",
},
masking_strategy=masking_strategy,
)
actual = transform(sample)
expected_chosen = [
Message(
role="user",
content="hello world",
masked=expected_masks[0][0],
eot=True,
),
Message(
role="assistant",
content="hello world",
masked=expected_masks[0][1],
eot=True,
),
]
assert_dialogue_equal(actual["chosen"], expected_chosen)

expected_rejected = [
Message(
role="user",
content="hello world",
masked=expected_masks[1][0],
eot=True,
),
Message(
role="assistant",
content="bye world",
masked=expected_masks[1][1],
eot=True,
),
]
assert_dialogue_equal(actual["rejected"], expected_rejected)

def test_system_prompt(self, sample):
transform = ChosenRejectedToMessages(
column_map={
Expand Down Expand Up @@ -333,6 +410,19 @@ def test_call_train_on_input(self):
converted_messages["messages"], MESSAGE_SAMPLE_TRAIN_ON_INPUT
)

@pytest.mark.parametrize(
"masking_strategy, expected_message",
[
("train_on_all", MESSAGE_SAMPLE_TRAIN_ON_INPUT),
("train_on_assistant", MESSAGE_SAMPLE),
("train_on_last", MESSAGE_SAMPLE),
],
)
def test_call_masking_strategy(self, masking_strategy, expected_message):
transform = ShareGPTToMessages(masking_strategy=masking_strategy)
converted_messages = transform(self.samples)
assert_dialogue_equal(converted_messages["messages"], expected_message)

def test_system_prompt(self):
transform = ShareGPTToMessages(new_system_prompt="you are a robot")
converted_messages = transform(self.samples)
Expand Down Expand Up @@ -423,6 +513,45 @@ def test_call_train_on_input(self):
converted_messages["messages"], MESSAGE_SAMPLE_TRAIN_ON_INPUT
)

@mock.patch("torchtune.data._messages.load_image")
@pytest.mark.parametrize(
"masking_strategy, is_multimodal, expected_message",
[
("train_on_all", False, MESSAGE_SAMPLE_TRAIN_ON_INPUT),
("train_on_assistant", False, MESSAGE_SAMPLE),
("train_on_last", False, MESSAGE_SAMPLE),
("train_on_all", True, None),
("train_on_assistant", True, None),
("train_on_last", True, None),
],
)
def test_call_masking_strategy(
self, mock_load_image, masking_strategy, is_multimodal, expected_message
):
test_img = Image.new(mode="RGB", size=(4, 4))
mock_load_image.return_value = test_img
transform = OpenAIToMessages(masking_strategy=masking_strategy)
if is_multimodal:
converted_messages = transform(self.image_samples)
assert_dialogue_equal(
converted_messages["messages"],
[
MESSAGE_SAMPLE[0],
Message(
role="user",
content=[
{"type": "text", "content": CHAT_SAMPLE["user"]},
{"type": "image", "content": test_img},
],
masked=True,
),
MESSAGE_SAMPLE[2],
],
)
else:
converted_messages = transform(self.samples)
assert_dialogue_equal(converted_messages["messages"], expected_message)

def test_system_prompt(self):
transform = OpenAIToMessages(new_system_prompt="you are a robot")
converted_messages = transform(self.samples)
Expand Down Expand Up @@ -542,3 +671,31 @@ def test_validate_messages():
match="Assistant message before expected user message at index 0 in messages",
):
validate_messages(messages)


@pytest.mark.parametrize(
"masking_strategy, messages_count, expected_masks",
[
("train_on_all", 3, [True, False, False]),
("train_on_assistant", 3, [True, True, False]),
("train_on_last", 3, [True, True, False]),
("train_on_all", 5, [True, False, False, False, False]),
("train_on_assistant", 5, [True, True, False, True, False]),
("train_on_last", 5, [True, True, True, True, False]),
],
)
def test_mask_messages(masking_strategy, messages_count, expected_masks):
messages = [
Message(role="system", content="hello"),
Message(role="user", content="hello"),
Message(role="assistant", content="world"),
Message(role="user", content="hello"),
Message(role="assistant", content="world"),
]

input_messages = messages[:messages_count]
mask_messages(input_messages, masking_strategy=masking_strategy)
expected_messages = messages[:messages_count]
for index in range(messages_count):
expected_messages[index].masked = expected_masks[index]
assert_dialogue_equal(input_messages, expected_messages)
Loading
Loading