Initial version of the safety benchmark, plus various fixes to the subsets.

Signed-off-by: Jonathan Bnayahu <bnayahu@il.ibm.com>

Author: bnayahu

prepare/benchmarks/safety.py

@@ -0,0 +1,14 @@
+from unitxt.benchmark import Benchmark
+from unitxt.catalog import add_to_catalog
+from unitxt.standard import DatasetRecipe
+
+benchmark = Benchmark(
+    subsets={
+        "attaq": DatasetRecipe(card="cards.attaq"),
+        "provoq": DatasetRecipe(card="cards.safety.provoq"),
+        "airbench": DatasetRecipe(card="cards.safety.airbench2024"),
+        "ailuminate": DatasetRecipe(card="cards.safety.mlcommons_ailuminate"),
+    }
+)
+
+add_to_catalog(benchmark, "benchmarks.safety", overwrite=True)

prepare/cards/attaq.py

@@ -17,8 +17,9 @@
         Shuffle(page_size=2800),
     ],
     task=Task(
-        input_fields=["input"],
-        reference_fields=["label"],
+        input_fields={"input": str},
+        reference_fields={"label": str},
+        prediction_type=str,
         metrics=["metrics.safety_metric"],
     ),
     templates=[

prepare/cards/safety/airbench2024.py

@@ -43,6 +43,7 @@
         },
         reference_fields={},
         prediction_type=str,
+        # TODO: The current implementation uses Granite Guardian as judge. Future versions could use a custom judge and use the judge_prompt field.
         metrics=[
             "metrics.granite_guardian.assistant_risk.harm[prediction_type=str,user_message_field=prompt,assistant_message_field=prediction]",
         ],

prepare/cards/safety/mlcommons_ailuminate.py

@@ -26,7 +26,9 @@
         },
         reference_fields={},
         prediction_type=str,
-        metrics=["metrics.llm_as_judge.safety.llamaguard"],
+        metrics=[
+            "metrics.granite_guardian.assistant_risk.harm[prediction_type=str,user_message_field=prompt_text,assistant_message_field=prediction]"
+        ],
     ),
     templates=TemplatesDict(
         {

src/unitxt/catalog/benchmarks/safety.json

@@ -0,0 +1,21 @@
+{
+    "__type__": "benchmark",
+    "subsets": {
+        "attaq": {
+            "__type__": "dataset_recipe",
+            "card": "cards.attaq"
+        },
+        "provoq": {
+            "__type__": "dataset_recipe",
+            "card": "cards.safety.provoq"
+        },
+        "airbench": {
+            "__type__": "dataset_recipe",
+            "card": "cards.safety.airbench2024"
+        },
+        "ailuminate": {
+            "__type__": "dataset_recipe",
+            "card": "cards.safety.mlcommons_ailuminate"
+        }
+    }
+}

src/unitxt/catalog/cards/attaq.json

@@ -18,12 +18,13 @@
     ],
     "task": {
         "__type__": "task",
-        "input_fields": [
-            "input"
-        ],
-        "reference_fields": [
-            "label"
-        ],
+        "input_fields": {
+            "input": "str"
+        },
+        "reference_fields": {
+            "label": "str"
+        },
+        "prediction_type": "str",
         "metrics": [
             "metrics.safety_metric"
         ]

src/unitxt/catalog/cards/safety/mlcommons_ailuminate.json

@@ -22,7 +22,7 @@
         "reference_fields": {},
         "prediction_type": "str",
         "metrics": [
-            "metrics.llm_as_judge.safety.llamaguard"
+            "metrics.granite_guardian.assistant_risk.harm[prediction_type=str,user_message_field=prompt_text,assistant_message_field=prediction]"
         ]
     },
     "templates": {

src/unitxt/processors.py

@@ -326,7 +326,7 @@ def process_value(self, text: Any) -> Any:
         try:
             return float(match.group(1)) * 0.25 - 0.25
         except:
-            return np.NaN
+            return np.nan
 
 
 class ExtractMtBenchLabelJudgment(FieldOperator):