Skip to content

[data] fix KeyError on invalid role tag in ranking dataset converters #10601

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 2 commits into
base: main
Choose a base branch
from
+69 −26
Open

[data] fix KeyError on invalid role tag in ranking dataset converters #10601

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
92e42c4
[data] fix KeyError on invalid role tag in ranking dataset converters
he-yufeng Jun 23, 2026
e15ce17
[data] also guard OpenAI converter main loop against unknown role tags
he-yufeng Jun 26, 2026
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
52 changes: 26 additions & 26 deletions src/llamafactory/data/converter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -198,19 +198,19 @@ def __call__(self, example: dict[str, Any]) -> dict[str, Any]:
or rejected[self.dataset_attr.role_tag] not in accept_tags[-1]
):
logger.warning_rank0(f"Invalid role tag in {[chosen, rejected]}.")
broken_data = True

prompt = aligned_messages
response = [
{
"role": tag_mapping[chosen[self.dataset_attr.role_tag]],
"content": chosen[self.dataset_attr.content_tag],
},
{
"role": tag_mapping[rejected[self.dataset_attr.role_tag]],
"content": rejected[self.dataset_attr.content_tag],
},
]
prompt, response = [], []
else:
prompt = aligned_messages
response = [
{
"role": tag_mapping[chosen[self.dataset_attr.role_tag]],
"content": chosen[self.dataset_attr.content_tag],
},
{
"role": tag_mapping[rejected[self.dataset_attr.role_tag]],
"content": rejected[self.dataset_attr.content_tag],
},
]
else: # normal example
prompt = aligned_messages[:-1]
response = aligned_messages[-1:]
Expand Down Expand Up @@ -319,19 +319,19 @@ def __call__(self, example: dict[str, Any]) -> dict[str, Any]:
or rejected[self.dataset_attr.role_tag] not in accept_tags[-1]

@gemini-code-assist gemini-code-assist Bot Jun 23, 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

In OpenAIDatasetConverter, accept_tags contains standardized roles (Role.ASSISTANT.value, Role.FUNCTION.value), whereas chosen and rejected contain raw role tags from the dataset (e.g., "gpt"). Comparing the raw role tags directly against accept_tags[-1] will always fail if custom role tags are used, causing valid examples to be incorrectly skipped as having an invalid role tag.

To fix this, we should look up the raw role tags in tag_mapping using .get() before comparing them to accept_tags[-1].

            if (
                tag_mapping.get(chosen[self.dataset_attr.role_tag]) not in accept_tags[-1]
                or tag_mapping.get(rejected[self.dataset_attr.role_tag]) not in accept_tags[-1]
            ):

):
logger.warning_rank0(f"Invalid role tag in {[chosen, rejected]}.")
broken_data = True

prompt = aligned_messages
response = [
{
"role": tag_mapping[chosen[self.dataset_attr.role_tag]],
"content": chosen[self.dataset_attr.content_tag],
},
{
"role": tag_mapping[rejected[self.dataset_attr.role_tag]],
"content": rejected[self.dataset_attr.content_tag],
},
]
prompt, response = [], []
else:
prompt = aligned_messages
response = [
{
"role": tag_mapping[chosen[self.dataset_attr.role_tag]],
"content": chosen[self.dataset_attr.content_tag],
},
{
"role": tag_mapping[rejected[self.dataset_attr.role_tag]],
"content": rejected[self.dataset_attr.content_tag],
},
]
else: # normal example
prompt = aligned_messages[:-1]
response = aligned_messages[-1:]
Expand Down
20 changes: 20 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 +62,23 @@ def test_sharegpt_converter():
"_videos": None,
"_audios": None,
}


@pytest.mark.runs_on(["cpu", "mps"])
def test_sharegpt_converter_ranking_skips_invalid_role():
# A pairwise (ranking) example whose chosen/rejected carries a role tag not
# in the accepted set must be skipped, not crash with KeyError on tag_mapping.
dataset_attr = DatasetAttr("hf_hub", "llamafactory/tiny-supervised-dataset")
dataset_attr.ranking = True
dataset_attr.chosen = "chosen"
dataset_attr.rejected = "rejected"
data_args = DataArguments()
example = {
"conversations": [{"from": "human", "value": "Solve the math problem.\n3 + 4"}],
"chosen": {"from": "not_a_role", "value": "The answer is 7."},
"rejected": {"from": "gpt", "value": "The answer is 8."},
}
dataset_converter = get_dataset_converter("sharegpt", dataset_attr, data_args)
result = dataset_converter(example)
assert result["_prompt"] == []
assert result["_response"] == []