[MODIFY] Med-Safety: Average -> Harmfulness Score

src/display/utils.py

@@ -38,8 +38,9 @@ auto_eval_column_dict = []
 auto_eval_column_dict.append(["model_type_symbol", ColumnContent, ColumnContent("T", "str", True, never_hidden=True)])
 auto_eval_column_dict.append(["model", ColumnContent, ColumnContent("Model", "markdown", True, never_hidden=True)])
 auto_eval_column_dict.append(["revision", ColumnContent, ColumnContent("Model sha", "str", False, True)])
-auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average", "number", True, False, dataset_task_col=True, med_safety_col=True, invariant=False)])
+auto_eval_column_dict.append(["average", ColumnContent, ColumnContent("Average", "number", True, False, dataset_task_col=True, invariant=False)])
 auto_eval_column_dict.append(["overall", ColumnContent, ColumnContent("Overall Score", "number", True, False, medical_summarization_col=True, aci_col=True, soap_col=True, invariant=False)])
+auto_eval_column_dict.append(["harmfulness", ColumnContent, ColumnContent("Harmfulness Score", "number", True, False, med_safety_col=True, invariant=False)])
 for task in HarnessTasks:
     auto_eval_column_dict.append([task.name, ColumnContent, ColumnContent(task.value.col_name, "number", True, False, dataset_task_col=True, invariant=False)])
 for column in OpenEndedColumns:

src/leaderboard/read_evals.py

@@ -265,7 +265,7 @@ class EvalResult:
         # changes to be made here
         if subset == "med_safety":
             average = sum([v for v in self.med_safety_results.values() if v is not None]) / len(MedSafetyColumns)
-            data_dict[AutoEvalColumn.average.name] = average
+            data_dict[AutoEvalColumn.harmfulness.name] = average
             if len(self.med_safety_results) > 0:
                 for task in MedSafetyColumns:
                     data_dict[task.value.col_name] = self.med_safety_results[task.value.benchmark]

src/populate.py

@@ -21,15 +21,15 @@ def get_leaderboard_df(results_path: str, requests_path: str, cols: list, benchm
     if subset == "datasets":
         df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=False)
     elif subset == "med_safety":
-        df = df.sort_values(by=[AutoEvalColumn.average.name], ascending=True)
+        df = df.sort_values(by=[AutoEvalColumn.harmfulness.name], ascending=True)
     elif subset == "open_ended":
         df = df.sort_values(by=["ELO"], ascending=False)
     elif subset == "medical_summarization":
-        df = df.sort_values(by=["Overall Score"], ascending=False)
+        df = df.sort_values(by=[AutoEvalColumn.overall.name], ascending=False)
     elif subset == "aci":
-        df = df.sort_values(by=["Overall Score"], ascending=False)
+        df = df.sort_values(by=[AutoEvalColumn.overall.name], ascending=False)
     elif subset == "soap":
-        df = df.sort_values(by=["Overall Score"], ascending=False)
+        df = df.sort_values(by=[AutoEvalColumn.overall.name], ascending=False)
     cols = list(set(df.columns).intersection(set(cols)))
     df = df[cols].round(decimals=2)
     # filter out if any of the benchmarks have not been produced