feat: Update win rate calculation in GenAI Evaluation for rubric based evaluation

PiperOrigin-RevId: 747474971

Author: vertex-sdk-bot

vertexai/preview/evaluation/_evaluation.py

@@ -317,11 +317,31 @@ def _aggregate_summary_metrics(
                     f"{metric.metric_name}/{constants.MetricResult.PAIRWISE_CHOICE_KEY}"
                 )
                 if pairwise_choice_col_name in metrics_table:
+                    candidate_model_win_rate_choices = [
+                        "CANDIDATE",
+                        utils.RATING_TO_VERDICT["B>>A"],
+                        utils.RATING_TO_VERDICT["A<<B"],
+                        utils.RATING_TO_VERDICT["B>A"],
+                        utils.RATING_TO_VERDICT["A<B"],
+                    ]
+                    baseline_model_win_rate_choices = [
+                        "BASELINE",
+                        utils.RATING_TO_VERDICT["A>>B"],
+                        utils.RATING_TO_VERDICT["B<<A"],
+                        utils.RATING_TO_VERDICT["B<A"],
+                        utils.RATING_TO_VERDICT["A>B"],
+                    ]
                     summary_metrics[
                         f"{metric.metric_name}/candidate_model_win_rate"
-                    ] = (metrics_table[pairwise_choice_col_name] == "CANDIDATE").mean()
+                    ] = (
+                        metrics_table[pairwise_choice_col_name].isin(
+                            candidate_model_win_rate_choices
+                        )
+                    ).mean()
                     summary_metrics[f"{metric.metric_name}/baseline_model_win_rate"] = (
-                        metrics_table[pairwise_choice_col_name] == "BASELINE"
+                        metrics_table[pairwise_choice_col_name].isin(
+                            baseline_model_win_rate_choices
+                        )
                     ).mean()
             else:
                 score_col_name = f"{str(metric)}/{constants.MetricResult.SCORE_KEY}"

vertexai/preview/evaluation/utils.py

@@ -73,7 +73,7 @@
 )
 _SXS_RATING_REGEX = re.compile(r"\[\[(SxSRating:[AB<>=]+)\]\]", re.DOTALL)
 
-_RATING_TO_VERDICT = {
+RATING_TO_VERDICT = {
     "B>>A": "Candidate response is better than the baseline response.",
     "A<<B": "Candiate response is better than the baseline response.",
     "B>A": "Candidate response is slightly better than the baseline response.",
@@ -554,8 +554,8 @@ def parse_pairwise_rubric_result(
 
     return {
         "pairwise_choice": (
-            _RATING_TO_VERDICT[rating_str]
-            if rating_str in _RATING_TO_VERDICT
+            RATING_TO_VERDICT[rating_str]
+            if rating_str in RATING_TO_VERDICT
             else rating_str
         ),
         "score": (