fix(data): return content when llama3/mistral tool extractors find no calls

[he-yufeng]

What does this PR do?

tool_extractor returns a str (the original content) when the assistant message contains no function calls, and a list[FunctionCall] otherwise. Llama3ToolUtils.tool_extractor and MistralToolUtils.tool_extractor returned the list comprehension directly, so a payload that parses to an empty list (e.g. "[]") produced [] instead of the content string:

Llama3ToolUtils.tool_extractor("[]")   # -> []   (should be "[]")
MistralToolUtils.tool_extractor("[]")  # -> []   (should be "[]")

The Qwen3.5, Seed, and LFM2 extractors already guard this with return results if results else content (Seed via #10408). This PR applies the same guard to the llama3 and mistral extractors so they stay consistent with the documented str | list[FunctionCall] contract.

Tests

Added test_llama3_tool_extractor_empty_returns_content and test_mistral_tool_extractor_empty_returns_content in tests/data/test_formatter.py, asserting ToolFormatter(tool_format=...).extract("[]") == "[]". Both fail before the change (they return []) and pass after.

PYTHONPATH=src python -m pytest tests/data/test_formatter.py   # 36 passed
ruff check . && ruff format --check .                          # clean on the changed files

[gemini-code-assist]

Code Review

This pull request modifies the tool extractors in src/llamafactory/data/tool_utils.py to return the original content string instead of an empty list when no tool calls are extracted, and adds corresponding unit tests in tests/data/test_formatter.py. The review feedback correctly identifies potential TypeError vulnerabilities if the parsed tools are not dictionaries, as the current try-except block only catches KeyError. It is recommended to validate the structure of the tools before processing them.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

If tools contains elements that are not dictionaries (e.g., if the model outputs a list of integers or strings, or nested lists), accessing tool["name"] or tool["parameters"] will raise a TypeError (e.g., 'int' object is not subscriptable). Since the try-except block only catches KeyError, this will cause an unhandled exception and crash the extraction process.

To make this more robust and adhere to defensive programming practices, we should validate that each tool in tools is a dictionary and contains the required keys before attempting to construct the FunctionCall objects.

Suggested change

	results = [
	FunctionCall(tool["name"], json.dumps(tool["parameters"], ensure_ascii=False)) for tool in tools
	]
	if not all(isinstance(tool, dict) and "name" in tool and "parameters" in tool for tool in tools):
	return content
	results = [
	FunctionCall(tool["name"], json.dumps(tool["parameters"], ensure_ascii=False)) for tool in tools
	]

[gemini-code-assist]

Similarly, if tools contains elements that are not dictionaries, accessing tool["name"] or tool["arguments"] will raise a TypeError which is not caught by the except KeyError: block.

We should validate that each tool in tools is a dictionary and contains the required keys before attempting to construct the FunctionCall objects.

Suggested change

	results = [FunctionCall(tool["name"], json.dumps(tool["arguments"], ensure_ascii=False)) for tool in tools]
	if not all(isinstance(tool, dict) and "name" in tool and "arguments" in tool for tool in tools):
	return content
	results = [FunctionCall(tool["name"], json.dumps(tool["arguments"], ensure_ascii=False)) for tool in tools]