Skip to content

fix: merge ShareGPT tool observations #10578

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

Open
he-yufeng wants to merge 1 commit into
base: main
Choose a base branch
from
+65 −9
Open

fix: merge ShareGPT tool observations #10578

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
42 changes: 33 additions & 9 deletions src/llamafactory/data/converter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -141,9 +141,6 @@ def __call__(self, example: dict[str, Any]) -> dict[str, Any]:
self.dataset_attr.function_tag: Role.FUNCTION.value,
self.dataset_attr.system_tag: Role.SYSTEM.value,
}
odd_tags = (self.dataset_attr.user_tag, self.dataset_attr.observation_tag)
even_tags = (self.dataset_attr.assistant_tag, self.dataset_attr.function_tag)
accept_tags = (odd_tags, even_tags)
messages = example[self.dataset_attr.messages]
if (
self.dataset_attr.system_tag
Expand All @@ -156,20 +153,46 @@ def __call__(self, example: dict[str, Any]) -> dict[str, Any]:
system = example[self.dataset_attr.system] if self.dataset_attr.system else ""

aligned_messages = []
tool_responses = []
broken_data = False
for turn_idx, message in enumerate(messages):
if message[self.dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
for message in messages:
role = message[self.dataset_attr.role_tag]
content = message[self.dataset_attr.content_tag]

if role not in tag_mapping:
logger.warning_rank0(f"Invalid role tag in {messages}.")
broken_data = True
break

if role == self.dataset_attr.observation_tag:
tool_responses.append(content)
continue
elif len(tool_responses) > 0:
observation_content = "\n</tool_response>\n<tool_response>\n".join(tool_responses)
aligned_messages.append(
{
"role": Role.OBSERVATION.value,
"content": observation_content,
}
)
tool_responses = []

aligned_messages.append(
{
"role": tag_mapping[message[self.dataset_attr.role_tag]],
"content": message[self.dataset_attr.content_tag],
"role": tag_mapping[role],
"content": content,
}
)

@gemini-code-assist gemini-code-assist Bot Jun 14, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

high

If the conversation history ends with an observation message (which is common in pairwise/ranking datasets where the final assistant response is stored in chosen/rejected instead of the conversations list), any pending tool_responses will not be flushed because the loop terminates before encountering a non-observation message. We should flush any remaining tool_responses after the loop finishes.

        if len(tool_responses) > 0:
            observation_content = "\n</tool_response>\n<tool_response>\n".join(tool_responses)
            aligned_messages.append(
                {
                    "role": Role.OBSERVATION.value,
                    "content": observation_content,
                }
            )

odd_tags = (Role.USER.value, Role.OBSERVATION.value)
even_tags = (Role.ASSISTANT.value, Role.FUNCTION.value)
accept_tags = (odd_tags, even_tags)
for turn_idx, message in enumerate(aligned_messages):
if message["role"] not in accept_tags[turn_idx % 2]:
logger.warning_rank0(f"Invalid role tag in {messages}.")
broken_data = True
break

if (not self.dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
self.dataset_attr.ranking and len(aligned_messages) % 2 == 0
):
Expand All @@ -193,9 +216,10 @@ def __call__(self, example: dict[str, Any]) -> dict[str, Any]:
): # pairwise example
chosen = example[self.dataset_attr.chosen]
rejected = example[self.dataset_attr.rejected]
response_tags = (self.dataset_attr.assistant_tag, self.dataset_attr.function_tag)
if (
chosen[self.dataset_attr.role_tag] not in accept_tags[-1]
or rejected[self.dataset_attr.role_tag] not in accept_tags[-1]
chosen[self.dataset_attr.role_tag] not in response_tags
or rejected[self.dataset_attr.role_tag] not in response_tags
):
logger.warning_rank0(f"Invalid role tag in {[chosen, rejected]}.")
broken_data = True
Expand Down
32 changes: 32 additions & 0 deletions tests/data/test_converter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,3 +62,35 @@ def test_sharegpt_converter():
"_videos": None,
"_audios": None,
}


@pytest.mark.runs_on(["cpu", "mps"])
def test_sharegpt_converter_merges_consecutive_observations():
dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
data_args = DataArguments()
example = {
"conversations": [
{"from": "human", "value": "Check both symbols."},
{"from": "function_call", "value": "<tool_call>lookup_a</tool_call>\n<tool_call>lookup_b</tool_call>"},
{"from": "observation", "value": '{"symbol": "A", "price": 10}'},
{"from": "observation", "value": '{"symbol": "B", "price": 20}'},
{"from": "gpt", "value": "A is 10 and B is 20."},
]
}
dataset_converter = get_dataset_converter("sharegpt", dataset_attr, data_args)
assert dataset_converter(example) == {
"_prompt": [
{"role": Role.USER.value, "content": "Check both symbols."},
{"role": Role.FUNCTION.value, "content": "<tool_call>lookup_a</tool_call>\n<tool_call>lookup_b</tool_call>"},
{
"role": Role.OBSERVATION.value,
"content": '{"symbol": "A", "price": 10}\n</tool_response>\n<tool_response>\n{"symbol": "B", "price": 20}',
},
],
"_response": [{"role": Role.ASSISTANT.value, "content": "A is 10 and B is 20."}],
"_system": "",
"_tools": "",
"_images": None,
"_videos": None,
"_audios": None,
}