update

app.py

@@ -146,29 +146,16 @@ def calculate_order_by_first_substring(selected_models):
     query_ids_df = first_columns[first_columns["Model Type"] == "Text Only"]
     query_ids_df = query_ids_df[query_ids_df["Model Name"].isin(selected_models)]
 
-    print(len(query_ids_df))
-
     query_ids_df = query_ids_df.groupby("query_id").filter(
         lambda x: x["parsed_judge_response"].eq(1).all()
     )
 
-    print(len(query_ids_df))
-
     query_ids = query_ids_df.query_id.unique()
     fsm_ids = query_ids_df.fsm_id.unique()
-    print(
-        "fsm_ids",
-        len(fsm_ids),
-        "Total of 25 FSM is solvable by everything on the first substring",
-    )
 
     text_only = all_data[all_data["Model Type"] == "Text Only"]
     text_only_filtered = text_only[text_only["fsm_id"].isin(fsm_ids)]
 
-    print(
-        f"Number of query_ids from text_only_filtered: {len(text_only_filtered.query_id.unique())}"
-    )
-
     text_only_filtered = (
         text_only_filtered.groupby(["Model Name"])["parsed_judge_response"]
         .mean()
@@ -182,7 +169,10 @@ def calculate_order_by_first_substring(selected_models):
     )
     text_only_filtered.sort_values("Accuracy", ascending=False, inplace=True)
 
-    return text_only_filtered
+    number_of_queries = len(query_ids)
+    number_of_fsms = len(fsm_ids)
+
+    return text_only_filtered, number_of_queries, number_of_fsms
 
 
 with gr.Blocks() as demo:
@@ -233,13 +223,20 @@ with gr.Blocks() as demo:
             label="Models to include",
             choices=all_text_only_model_names,
             value=all_text_only_model_names,
+            interactive=True,
         )
+        with gr.Row():
+            number_of_queries = gr.Textbox(label="Number of queries to include")
+
+            number_of_fsms = gr.Textbox(label="Number of FSMs to include")
+
         constrained_leader_board_text = gr.Dataframe()
 
-    included_models.input(
+    included_models.select(
         fn=calculate_order_by_first_substring,
         inputs=[included_models],
-        outputs=[constrained_leader_board_text],
+        outputs=[constrained_leader_board_text, number_of_queries, number_of_fsms],
+        queue=True,
     )
 
     demo.launch()