openhands commited on
Commit
d17eff0
·
1 Parent(s): 49f9739

Only show 'Detailed Benchmark Results' when more than one benchmark exists

Browse files

- Modified create_benchmark_details_display to check benchmark count
- If category has 1 or fewer benchmarks, skip showing the detailed section
- Updated docstring to reflect this behavior

Files changed (1) hide show
  1. ui_components.py +3 -2
ui_components.py CHANGED
@@ -1074,6 +1074,7 @@ def create_benchmark_details_display(
1074
  """
1075
  Generates a detailed breakdown for each benchmark within a given category.
1076
  For each benchmark, it creates a title, a filtered table, and a scatter plot.
 
1077
  Args:
1078
  full_df (pd.DataFrame): The complete, "pretty" dataframe for the entire split.
1079
  tag_map (dict): The "pretty" tag map to find the list of benchmarks.
@@ -1082,8 +1083,8 @@ def create_benchmark_details_display(
1082
  # 1. Get the list of benchmarks for the selected category
1083
  benchmark_names = tag_map.get(category_name, [])
1084
 
1085
- if not benchmark_names:
1086
- gr.Markdown(f"No detailed benchmarks found for the category: {category_name}")
1087
  return
1088
 
1089
  gr.HTML(f'<h2 class="benchmark-main-subtitle">{category_name} Detailed Benchmark Results</h2>')
 
1074
  """
1075
  Generates a detailed breakdown for each benchmark within a given category.
1076
  For each benchmark, it creates a title, a filtered table, and a scatter plot.
1077
+ Only shows the detailed results if there is more than one benchmark in the category.
1078
  Args:
1079
  full_df (pd.DataFrame): The complete, "pretty" dataframe for the entire split.
1080
  tag_map (dict): The "pretty" tag map to find the list of benchmarks.
 
1083
  # 1. Get the list of benchmarks for the selected category
1084
  benchmark_names = tag_map.get(category_name, [])
1085
 
1086
+ # Only show detailed results if there is more than one benchmark
1087
+ if len(benchmark_names) <= 1:
1088
  return
1089
 
1090
  gr.HTML(f'<h2 class="benchmark-main-subtitle">{category_name} Detailed Benchmark Results</h2>')