Add CodeVulnerabilityEvaluator and UngroundedAttributesEvaluator to _SafetyEvaluation (Azure#40635)

Author: slister1001

sdk/evaluation/azure-ai-evaluation/azure/ai/evaluation/_safety_evaluation/_safety_evaluation.py

@@ -21,6 +21,8 @@
     _fluency,
     _xpia,
     _coherence,
+    _code_vulnerability,
+    _ungrounded_attributes,
 )
 from azure.ai.evaluation._evaluators._eci._eci import ECIEvaluator
 from azure.ai.evaluation._evaluate import _evaluate
@@ -32,7 +34,7 @@
     AdversarialScenario,
     AdversarialScenarioJailbreak,
     IndirectAttackSimulator,
-    DirectAttackSimulator ,
+    DirectAttackSimulator,
 )
 from azure.ai.evaluation.simulator._adversarial_scenario import _UnstableAdversarialScenario
 from azure.ai.evaluation.simulator._utils import JsonLineList
@@ -72,6 +74,7 @@ class _SafetyEvaluator(Enum):
     """
 
     CONTENT_SAFETY = "content_safety"
+    CODE_VULNERABILITY = "code_vulnerability"
     GROUNDEDNESS = "groundedness"
     PROTECTED_MATERIAL = "protected_material"
     RELEVANCE = "relevance"
@@ -81,6 +84,7 @@ class _SafetyEvaluator(Enum):
     INDIRECT_ATTACK = "indirect_attack"
     DIRECT_ATTACK = "direct_attack"
     ECI = "eci"
+    UNGROUNDED_ATTRIBUTES = "ungrounded_attributes"
 
 
 @experimental
@@ -380,6 +384,10 @@ def _get_scenario(
                 )
             if evaluator == _SafetyEvaluator.ECI:
                 return _UnstableAdversarialScenario.ECI
+            if evaluator == _SafetyEvaluator.CODE_VULNERABILITY:
+                return AdversarialScenario.ADVERSARIAL_CODE_VULNERABILITY
+            if evaluator == _SafetyEvaluator.UNGROUNDED_ATTRIBUTES:
+                return AdversarialScenario.ADVERSARIAL_UNGROUNDED_ATTRIBUTES
             if evaluator in [
                 _SafetyEvaluator.GROUNDEDNESS,
                 _SafetyEvaluator.RELEVANCE,
@@ -461,6 +469,14 @@ def _get_evaluators(
                 evaluators_dict["eci"] = ECIEvaluator(
                     azure_ai_project=self.azure_ai_project, credential=self.credential
                 )
+            elif evaluator == _SafetyEvaluator.CODE_VULNERABILITY:
+                evaluators_dict["code_vulnerability"] = _code_vulnerability.CodeVulnerabilityEvaluator(
+                    azure_ai_project=self.azure_ai_project, credential=self.credential
+                )
+            elif evaluator == _SafetyEvaluator.UNGROUNDED_ATTRIBUTES:
+                evaluators_dict["ungrounded_attributes"] = _ungrounded_attributes.UngroundedAttributesEvaluator(
+                    azure_ai_project=self.azure_ai_project, credential=self.credential
+                )
             else:
                 msg = (
                     f"Invalid evaluator: {evaluator}. Supported evaluators are: {_SafetyEvaluator.__members__.values()}"
@@ -597,7 +613,28 @@ def _validate_inputs(
                 category=ErrorCategory.INVALID_VALUE,
                 blame=ErrorBlame.USER_ERROR,
             )
-    
+        
+        if _SafetyEvaluator.CODE_VULNERABILITY in evaluators and num_turns > 1:
+            self.logger.error("Code vulnerability evaluation only supports single-turn conversations.")
+            msg = "Code vulnerability evaluation only supports single-turn conversations."
+            raise EvaluationException(
+                message=msg,
+                internal_message=msg,
+                target=ErrorTarget.UNKNOWN,
+                category=ErrorCategory.INVALID_VALUE,
+                blame=ErrorBlame.USER_ERROR,
+            )
+        if _SafetyEvaluator.UNGROUNDED_ATTRIBUTES in evaluators and num_turns > 1:
+            self.logger.error("Ungrounded attributes evaluation only supports single-turn conversations.")
+            msg = "Ungrounded attributes evaluation only supports single-turn conversations."
+            raise EvaluationException(
+                message=msg,
+                internal_message=msg,
+                target=ErrorTarget.UNKNOWN,
+                category=ErrorCategory.INVALID_VALUE,
+                blame=ErrorBlame.USER_ERROR,
+            )
+
         if _SafetyEvaluator.CONTENT_SAFETY in evaluators and scenario and num_turns > 1 and scenario != AdversarialScenario.ADVERSARIAL_CONVERSATION:
             self.logger.error(f"Adversarial scenario {scenario} is not supported for content safety evaluation with more than 1 turn.")
             msg = f"Adversarial scenario {scenario} is not supported for content safety evaluation with more than 1 turn."

sdk/evaluation/azure-ai-evaluation/tests/unittests/test_safety_evaluation.py

@@ -298,3 +298,39 @@ async def test_call_with_async_target(self, mock_evaluate, mock_simulate, safety
         # Verify that _simulate was called with the async target
         mock_simulate.assert_called_once()
         assert mock_simulate.call_args[1]["target"] == mock_async_target
+
+    def test_get_scenario_code_vulnerability(self, safety_eval):
+        scenario = safety_eval._get_scenario([_SafetyEvaluator.CODE_VULNERABILITY])
+        assert scenario == AdversarialScenario.ADVERSARIAL_CODE_VULNERABILITY
+
+    def test_get_scenario_ungrounded_attributes(self, safety_eval):
+        scenario = safety_eval._get_scenario([_SafetyEvaluator.UNGROUNDED_ATTRIBUTES])
+        assert scenario == AdversarialScenario.ADVERSARIAL_UNGROUNDED_ATTRIBUTES
+
+    def test_get_evaluators_code_vulnerability(self, safety_eval):
+        evaluators = safety_eval._get_evaluators([_SafetyEvaluator.CODE_VULNERABILITY])
+        assert "code_vulnerability" in evaluators
+        assert evaluators["code_vulnerability"].__class__.__name__ == "CodeVulnerabilityEvaluator"
+
+    def test_get_evaluators_ungrounded_attributes(self, safety_eval):
+        evaluators = safety_eval._get_evaluators([_SafetyEvaluator.UNGROUNDED_ATTRIBUTES])
+        assert "ungrounded_attributes" in evaluators
+        assert evaluators["ungrounded_attributes"].__class__.__name__ == "UngroundedAttributesEvaluator"
+
+    def test_validate_inputs_code_vulnerability_multi_turn(self, safety_eval, mock_target):
+        with pytest.raises(EvaluationException) as exc_info:
+            safety_eval._validate_inputs(
+                target=mock_target,
+                evaluators=[_SafetyEvaluator.CODE_VULNERABILITY],
+                num_turns=3,
+            )
+        assert "Code vulnerability evaluation only supports single-turn conversations" in str(exc_info.value)
+
+    def test_validate_inputs_ungrounded_attributes_multi_turn(self, safety_eval, mock_target):
+        with pytest.raises(EvaluationException) as exc_info:
+            safety_eval._validate_inputs(
+                target=mock_target,
+                evaluators=[_SafetyEvaluator.UNGROUNDED_ATTRIBUTES],
+                num_turns=3,
+            )
+        assert "Ungrounded attributes evaluation only supports single-turn conversations" in str(exc_info.value)