[data] fix: parse nested JSON literals in LFM2 tools (#10940)

Co-authored-by: LimbC-C <329404569+LimbC-C@users.noreply.github.com>

src/llamafactory/data/tool_utils.py

@@ -934,9 +934,7 @@ class Xing4ToolUtils(ToolUtils):
             if params_content[position:].strip():
                 return content
 
-            results.append(
-                FunctionCall(name_match.group(1), json.dumps(arguments, ensure_ascii=False))
-            )
+            results.append(FunctionCall(name_match.group(1), json.dumps(arguments, ensure_ascii=False)))
 
         return results
 
@@ -1038,19 +1036,16 @@ class LFM2ToolUtils(ToolUtils):
     @staticmethod
     def _ast_to_value(node: ast.AST) -> Any:
         """Convert an AST node to a Python value, handling JSON-style booleans/null."""
-        # Handle JSON-style true/false/null as Name nodes
-        if isinstance(node, ast.Name):
-            if node.id == "true":
-                return True
-            elif node.id == "false":
-                return False
-            elif node.id == "null":
-                return None
-            else:
-                raise ValueError(f"Unknown identifier: {node.id}")
 
-        # Use literal_eval for other cases (strings, numbers, lists, dicts)
-        return ast.literal_eval(node)
+        class JsonLiteralTransformer(ast.NodeTransformer):
+            def visit_Name(self, name: ast.Name) -> ast.AST:
+                literals = {"true": True, "false": False, "null": None}
+                if name.id in literals:
+                    return ast.copy_location(ast.Constant(value=literals[name.id]), name)
+                return name
+
+        # Normalize literals at every depth without evaluating names or expressions.
+        return ast.literal_eval(JsonLiteralTransformer().visit(node))
 
     @override
     @staticmethod

tests/data/test_formatter.py

@@ -449,6 +449,30 @@ def test_lfm2_tool_round_trip():
     assert json.loads(extracted[0][1]) == original["arguments"]
 
 
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
+@pytest.mark.parametrize(
+    "arguments",
+    [
+        {"options": {"enabled": True, "fallback": None}},
+        {"items": [True, False, None]},
+        {"options": [{"active": [True, False, None], "label": "true false null"}]},
+        {"enabled": True, "fallback": None},
+    ],
+)
+def test_lfm2_tool_round_trip_nested_json_literals(arguments):
+    formatter = FunctionFormatter(slots=["{{content}}"], tool_format="lfm2")
+    extractor = ToolFormatter(tool_format="lfm2")
+    rendered = formatter.apply(content=json.dumps({"name": "test_tool", "arguments": arguments}))[0]
+    assert extractor.extract(rendered) == [("test_tool", json.dumps(arguments, ensure_ascii=False))]
+
+
+@pytest.mark.runs_on(["cpu", "mps", "xpu"])
+@pytest.mark.parametrize("value", ['{"enabled": unknown}', "[1 + 2]"])
+def test_lfm2_tool_extractor_rejects_nonliteral_values(value):
+    content = f"<|tool_call_start|>[test_tool(options={value})]<|tool_call_end|>"
+    assert ToolFormatter(tool_format="lfm2").extract(content) == content
+
+
 @pytest.mark.runs_on(["cpu", "mps", "xpu"])
 def test_minicpm5_function_formatter():
     formatter = FunctionFormatter(slots=["{{content}}<|im_end|>\n"], tool_format="minicpm5")