fix: Fetch new results properly

app.py

@@ -670,6 +670,9 @@ def fetch_results() -> dict[Language, pd.DataFrame]:
         data_dict = defaultdict(dict)
         for record in records:
             model_name = record["model"]
+            raw_results = record["results"]["raw"]
+            if isinstance(raw_results, dict) and "test" in raw_results:
+                raw_results = raw_results.get("test", raw_results)
 
             dataset_name = record["dataset"]
             if dataset_name in possible_dataset_names:
@@ -681,7 +684,7 @@ def fetch_results() -> dict[Language, pd.DataFrame]:
                         f"test_{dataset.task.metric}",
                         test_score_dict.get(dataset.task.metric)
                     )
-                    for test_score_dict in record["results"]["raw"]
+                    for test_score_dict in raw_results
                 ]
                 if dataset.task in data_dict[model_name]:
                     data_dict[model_name][dataset.task].append(scores)