categories rename

app.py

@@ -577,6 +577,22 @@ def update_visualization(selected_models, selected_category, selected_metric, ve
 # Create Gradio app
 demo = gr.Blocks(css=custom_css, theme=custom_theme)
 
+# Mapping from original category names to display names
+CATEGORY_DISPLAY_MAP = {
+    "Criminal, Violent, and Terrorist Activity": "Crime & Violence",
+    "Manipulation, Deception, and Misinformation": "Misinformation",
+    "Creative Content Involving Illicit Themes": "Illicit Creative",
+    "Sexual Content and Violence": "Sexual Content",
+    "Political Corruption and Legal Evasion": "Corruption & Legal Evasion",
+    "Labor Exploitation and Human Trafficking": "Labor Exploitation",
+    "Environmental and Industrial Harm": "Environmental & Industrial Harm",
+    "Animal Cruelty and Exploitation": "Animal Harm",
+    "Self–Harm and Suicidal Ideation": "Self-Harm",
+    "Safe Prompts": "Safe Prompts"
+}
+# Create reverse mapping for lookups
+CATEGORY_REVERSE_MAP = {v: k for k, v in CATEGORY_DISPLAY_MAP.items()}
+
 with demo:
     gr.HTML(TITLE)
     # gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
@@ -628,12 +644,14 @@ with demo:
                 # Create tabs for each category
                 with gr.Tabs(elem_classes="category-tabs") as category_tabs:
                     # First tab for average metrics across all categories
-                    with gr.TabItem("Overall Performance", elem_id="overall-tab"):
+                    with gr.TabItem("All Results", elem_id="overall-tab"):
                         leaderboard = init_leaderboard(LEADERBOARD_DF)
 
-                    # Create a tab for each category
+                    # Create a tab for each category using display names
                     for category in CATEGORIES:
-                        with gr.TabItem(f"{category}", elem_id=f"category-{category.lower().replace(' ', '-')}-tab"):
+                        display_name = CATEGORY_DISPLAY_MAP.get(category, category)
+                        elem_id = f"category-{display_name.lower().replace(' ', '-').replace('&', 'and')}-tab"
+                        with gr.TabItem(display_name, elem_id=elem_id):
                             category_df = get_category_leaderboard_df(category, version=CURRENT_VERSION)
                             category_leaderboard = init_leaderboard(category_df)
 
@@ -752,12 +770,12 @@ with demo:
                             interactive=True
                         )
                     with gr.Column():
-                        # Add Overall Performance to categories
-                        viz_categories = ["Overall Performance"] + CATEGORIES
+                        # Add Overall Performance to categories, use display names
+                        viz_categories_display = ["All Results"] + [CATEGORY_DISPLAY_MAP.get(cat, cat) for cat in CATEGORIES]
                         category_selector = gr.Dropdown(
-                            choices=viz_categories,
+                            choices=viz_categories_display,
                             label="Select Category",
-                            value=viz_categories[0],
+                            value=viz_categories_display[0],
                             interactive=True
                         )
                         metric_selector = gr.Dropdown(
@@ -772,7 +790,7 @@ with demo:
                 # Update visualization when any selector changes
                 for control in [viz_version_selector, model_selector, category_selector, metric_selector]:
                     control.change(
-                        fn=update_visualization,
+                        fn=lambda sm, sc, s_metric, v: update_visualization(sm, CATEGORY_REVERSE_MAP.get(sc, sc), s_metric, v),
                         inputs=[model_selector, category_selector, metric_selector, viz_version_selector],
                         outputs=plot_output
                     )

logs/guardbench_20250424_132721_81730502.log

@@ -0,0 +1,2 @@
+2025-04-24 13:27:22,432 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:27:22,600 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_132735_5fe7accd.log

@@ -0,0 +1,2 @@
+2025-04-24 13:27:35,986 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:27:36,096 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_133606_bfae8d28.log

@@ -0,0 +1,2 @@
+2025-04-24 13:36:07,557 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:36:07,754 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_133744_d3ca5956.log

@@ -0,0 +1,2 @@
+2025-04-24 13:37:44,999 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:37:45,167 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_133847_cf6f7f0f.log

@@ -0,0 +1,2 @@
+2025-04-24 13:38:48,713 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:38:48,956 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_133952_920f4d61.log

@@ -0,0 +1,2 @@
+2025-04-24 13:39:53,041 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:39:53,311 - __main__ - INFO - Loaded leaderboard with 0 entries

logs/guardbench_20250424_134044_2df3fadc.log

@@ -0,0 +1,17 @@
+2025-04-24 13:40:44,934 - __main__ - INFO - Initializing leaderboard data...
+2025-04-24 13:40:45,088 - __main__ - INFO - Loaded leaderboard with 0 entries
+2025-04-24 13:40:45,155 - __main__ - INFO - Available columns in LEADERBOARD_DF: ['model_name', 'model_type', 'guard_model_type', 'integral_score', 'macro_accuracy', 'macro_recall', 'micro_avg_error_ratio', 'micro_avg_runtime_ms', 'total_evals_count']
+2025-04-24 13:40:45,158 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:45,672 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:45,758 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:45,862 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:45,950 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,035 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,122 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,389 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,473 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,593 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,686 - __main__ - WARNING - Initializing empty leaderboard
+2025-04-24 13:40:46,869 - apscheduler.scheduler - INFO - Adding job tentatively -- it will be properly scheduled when the scheduler starts
+2025-04-24 13:40:46,869 - apscheduler.scheduler - INFO - Added job "refresh_data" to job store "default"
+2025-04-24 13:40:46,870 - apscheduler.scheduler - INFO - Scheduler started