Skip to content

Fix Qwen rendering and parsing for tool calls #58

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

Merged
merged 3 commits into from
Oct 31, 2025
Merged

Fix Qwen rendering and parsing for tool calls #58

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
5695666
Fix Qwen rendering
eddie-scio Oct 30, 2025
ef123a5
precommit
eddie-scio Oct 30, 2025
a229698
bad merge
eddie-scio Oct 30, 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
20 changes: 13 additions & 7 deletions tinker_cookbook/renderers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -336,13 +336,20 @@ def _render_message(self, idx: int, message: Message) -> tuple[list[int], list[i
# <think> in the assistant messages, we so don't need to re-add it in those cases.
ob_str += "<think>\n"
# Observation (prompt) part
ac_str = f"{ac_content}<|im_end|>"
if "tool_calls" in message:
ac_content += "\n".join(
[
f"<tool_call>\n{json.dumps(tool_call)}\n</tool_call>"
for tool_call in message["tool_calls"]
]
)
ac_content += "<|im_end|>"
# Action part
ac_tail_str = "" # No action tail needed for Qwen format
# Action part that's only included in the last message in SFT
return (
self.tokenizer.encode(ob_str, add_special_tokens=False),
self.tokenizer.encode(ac_str, add_special_tokens=False),
self.tokenizer.encode(ac_content, add_special_tokens=False),
self.tokenizer.encode(ac_tail_str, add_special_tokens=False),
)

Expand Down Expand Up @@ -405,11 +412,10 @@ def parse_response(self, response: list[int]) -> tuple[Message, bool]:
if not parse_success:
return assistant_message, False

# NOTE:
# we use the <function_call>...</function_call> tag to wrap the tool call.
match = re.search(
r"<function_call>(.*?)</function_call>", assistant_message["content"], re.DOTALL
)
# Follow Qwen docs and Qwen-Agent's tool calling prompt to use <tool_call>...</tool_call> tags to wrap the tool call.
# - https://qwen.readthedocs.io/en/latest/getting_started/concepts.html#tool-calling
# - https://github.com/QwenLM/Qwen-Agent/blob/main/qwen_agent/llm/fncall_prompts/nous_fncall_prompt.py#L279-L282
match = re.search(r"<tool_call>(.*?)</tool_call>", assistant_message["content"], re.DOTALL)
if match:
tool_calls = self._parse_tool_call(match.group(1))
if tool_calls is None:
Expand Down