Implement inheritance for each risk type

Signed-off-by: Martín Santillán Cooper <[email protected]>

examples/evaluate_granite_guardian_assistant_message_risks.py

@@ -3,7 +3,7 @@
 from unitxt import evaluate
 from unitxt.api import create_dataset
 from unitxt.blocks import Task
-from unitxt.metrics import GraniteGuardianWMLMetric, RiskType
+from unitxt.metrics import GraniteGuardianAssistantRisk, RiskType
 from unitxt.templates import NullTemplate
 
 print("Assistant response risks")
@@ -40,9 +40,7 @@
 
 risks = [
     f"metrics.granite_guardian.assistant_risk.{assistant_risk}"
-    for assistant_risk in GraniteGuardianWMLMetric.available_risks[
-        RiskType.ASSISTANT_MESSAGE
-    ]
+    for assistant_risk in GraniteGuardianAssistantRisk.get_available_risk_names()
 ]
 
 print(

examples/evaluate_granite_guardian_custom_risks.py

@@ -1,7 +1,7 @@
 from unitxt import evaluate
 from unitxt.api import create_dataset
 from unitxt.blocks import Task
-from unitxt.metrics import GraniteGuardianWMLMetric, RiskType
+from unitxt.metrics import GraniteGuardianCustomRisk, RiskType
 from unitxt.templates import NullTemplate
 
 print("Bring your own risk")
@@ -20,10 +20,9 @@
         default_template=NullTemplate(),
         # metrics=["metrics.granite_guardian.agentic.function_call[user_message_field=user_prompt, risk_definition=User message contains personal information or sensitive personal information that is included as a part of a prompt.]"],
         metrics=[
-            GraniteGuardianWMLMetric(
+            GraniteGuardianCustomRisk(
                 risk_name="personal_information",
                 risk_definition="User message contains personal information or sensitive personal information that is included as a part of a prompt.",
-                risk_type=RiskType.RAG.name,
                 user_message_field="user_prompt",
             )
         ],

examples/evaluate_granite_guardian_user_message_risks.py

@@ -3,7 +3,7 @@
 from unitxt import evaluate
 from unitxt.api import create_dataset
 from unitxt.blocks import Task
-from unitxt.metrics import GraniteGuardianMetric, RiskType
+from unitxt.metrics import GraniteGuardianUserRisk, RiskType
 from unitxt.templates import NullTemplate
 
 print("User prompt risks")
@@ -23,7 +23,7 @@
 
 risks = [
     f"metrics.granite_guardian.user_risk.{user_risk}"
-    for user_risk in GraniteGuardianMetric.available_risks[RiskType.USER_MESSAGE]
+    for user_risk in GraniteGuardianUserRisk.get_available_risk_names()
 ]
 
 print(

prepare/metrics/granite_guardian.py

@@ -1,8 +1,8 @@
 from unitxt import add_to_catalog
-from unitxt.metrics import GraniteGuardianMetric
+from unitxt.metrics import GraniteGuardianBase, RISK_TYPE_TO_CLASS
 
-for risk_type, risk_names in GraniteGuardianMetric.available_risks.items():
+for risk_type, risk_names in GraniteGuardianBase.available_risks.items():
     for risk_name in risk_names:
         metric_name = f"""granite_guardian.{risk_type.value}.{risk_name}"""
-        metric = GraniteGuardianMetric(risk_name=risk_name, risk_type=risk_type.name)
+        metric = RISK_TYPE_TO_CLASS[risk_type](risk_name=risk_name)
         add_to_catalog(metric, name=f"metrics.{metric_name}", overwrite=True)

prepare/metrics/rag_granite_guardian.py

@@ -1,5 +1,5 @@
 from unitxt import add_to_catalog
-from unitxt.metrics import GraniteGuardianMetric, MetricPipeline
+from unitxt.metrics import GraniteGuardianRagRisk, MetricPipeline
 from unitxt.operators import Copy, Set
 from unitxt.string_operators import Join
 
@@ -22,7 +22,7 @@
 
 for granite_risk_name, pred_field in risk_names_to_pred_field.items():
     metric_name = f"""granite_guardian_{granite_risk_name}"""
-    metric = GraniteGuardianMetric(
+    metric = GraniteGuardianRagRisk(
         main_score=metric_name,
         risk_name=granite_risk_name,
         user_message_field="question",

src/unitxt/catalog/metrics/granite_guardian/agentic_risk/function_call.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "function_call",
-    "risk_type": "AGENTIC"
+    "__type__": "granite_guardian_agentic_risk",
+    "risk_name": "function_call"
 }

src/unitxt/catalog/metrics/granite_guardian/assistant_risk/harm.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "harm",
-    "risk_type": "ASSISTANT_MESSAGE"
+    "__type__": "granite_guardian_assistant_risk",
+    "risk_name": "harm"
 }

src/unitxt/catalog/metrics/granite_guardian/assistant_risk/profanity.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "profanity",
-    "risk_type": "ASSISTANT_MESSAGE"
+    "__type__": "granite_guardian_assistant_risk",
+    "risk_name": "profanity"
 }

src/unitxt/catalog/metrics/granite_guardian/assistant_risk/social_bias.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "social_bias",
-    "risk_type": "ASSISTANT_MESSAGE"
+    "__type__": "granite_guardian_assistant_risk",
+    "risk_name": "social_bias"
 }

src/unitxt/catalog/metrics/granite_guardian/assistant_risk/unethical_behavior.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "unethical_behavior",
-    "risk_type": "ASSISTANT_MESSAGE"
+    "__type__": "granite_guardian_assistant_risk",
+    "risk_name": "unethical_behavior"
 }

src/unitxt/catalog/metrics/granite_guardian/assistant_risk/violence.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "violence",
-    "risk_type": "ASSISTANT_MESSAGE"
+    "__type__": "granite_guardian_assistant_risk",
+    "risk_name": "violence"
 }

src/unitxt/catalog/metrics/granite_guardian/rag_risk/answer_relevance.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "answer_relevance",
-    "risk_type": "RAG"
+    "__type__": "granite_guardian_rag_risk",
+    "risk_name": "answer_relevance"
 }

src/unitxt/catalog/metrics/granite_guardian/rag_risk/context_relevance.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "context_relevance",
-    "risk_type": "RAG"
+    "__type__": "granite_guardian_rag_risk",
+    "risk_name": "context_relevance"
 }

src/unitxt/catalog/metrics/granite_guardian/rag_risk/groundedness.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "groundedness",
-    "risk_type": "RAG"
+    "__type__": "granite_guardian_rag_risk",
+    "risk_name": "groundedness"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/harm.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "harm",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "harm"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/jailbreak.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "jailbreak",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "jailbreak"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/profanity.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "profanity",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "profanity"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/social_bias.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "social_bias",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "social_bias"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/unethical_behavior.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "unethical_behavior",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "unethical_behavior"
 }

src/unitxt/catalog/metrics/granite_guardian/user_risk/violence.json

@@ -1,5 +1,4 @@
 {
-    "__type__": "granite_guardian_metric",
-    "risk_name": "violence",
-    "risk_type": "USER_MESSAGE"
+    "__type__": "granite_guardian_user_risk",
+    "risk_name": "violence"
 }

src/unitxt/catalog/metrics/rag/granite_guardian_answer_relevance.json

@@ -2,7 +2,7 @@
     "__type__": "metric_pipeline",
     "main_score": "granite_guardian_answer_relevance",
     "metric": {
-        "__type__": "granite_guardian_metric",
+        "__type__": "granite_guardian_rag_risk",
         "main_score": "granite_guardian_answer_relevance",
         "risk_name": "answer_relevance",
         "user_message_field": "question",

src/unitxt/catalog/metrics/rag/granite_guardian_context_relevance.json

@@ -2,7 +2,7 @@
     "__type__": "metric_pipeline",
     "main_score": "granite_guardian_context_relevance",
     "metric": {
-        "__type__": "granite_guardian_metric",
+        "__type__": "granite_guardian_rag_risk",
         "main_score": "granite_guardian_context_relevance",
         "risk_name": "context_relevance",
         "user_message_field": "question",

src/unitxt/catalog/metrics/rag/granite_guardian_groundedness.json

@@ -2,7 +2,7 @@
     "__type__": "metric_pipeline",
     "main_score": "granite_guardian_groundedness",
     "metric": {
-        "__type__": "granite_guardian_metric",
+        "__type__": "granite_guardian_rag_risk",
         "main_score": "granite_guardian_groundedness",
         "risk_name": "groundedness",
         "user_message_field": "question",