leaderboard: add description

leaderboard.py

@@ -23,6 +23,7 @@ def create_leaderboard_tab():
             row_count=0,
             col_count=0,
         )
+        legend = gr.Markdown(value="")
     def update_datasets(system_type: str):
         # When system type changes, clear dataset and table
         dataset_choices = get_datasets(system_type) if system_type else []
@@ -34,14 +35,17 @@ def create_leaderboard_tab():
     def update_table(system_type: str, dataset: str):
         # Only populate table when both selections are made
         if not system_type or not dataset:
-            return gr.DataFrame(value=None, row_count=0)
+            return [gr.DataFrame(value=None, row_count=0), gr.Markdown("")]
             
-        df, datatypes = get_leaderboard_data(system_type, dataset)
-        return gr.DataFrame(
-            value=df,
-            row_count=(len(df), "fixed"),
-            datatype=datatypes
-        )
+        df, datatypes, description = get_leaderboard_data(system_type, dataset)
+        return [
+            gr.DataFrame(
+                value=df,
+                row_count=(len(df), "fixed"),
+                datatype=datatypes
+            ),
+            gr.Markdown(description)
+        ]
     
     system_type.change(
         fn=update_datasets,
@@ -52,5 +56,5 @@ def create_leaderboard_tab():
     dataset.change(
         fn=update_table,
         inputs=[system_type, dataset],
-        outputs=[table]
+        outputs=[table, legend]
     )

leaderboard_data.py

@@ -17,6 +17,36 @@ from huggingface_hub.hf_api import RepoFolder
 from api import api
 
 
+def make_pretty(df: pd.DataFrame) -> pd.DataFrame:
+    """
+    Rename some columns
+    """
+    df["whisperv3_cer"] = df["whisperv3_cer"] * 100
+    rename_map = {
+        "whisperv3_cer": "WhisperV3 CER, %↓",
+        "utmos_mos": "UTMOS MOS, ↑", 
+        "aesthetics_enjoyment": "Enjoyment, ↑",
+        "aesthetics_usefullness": "Usefulness, ↑",
+        "aesthetics_complexity": "Complexity, ↑",
+        "aesthetics_quality": "Quality, ↑",
+        "ecapa_secs": "ECAPA SECS, ↑",
+    }
+    description_map = {
+        "whisperv3_cer": "Character Error Rate in %, measured with WhisperV3",
+        "utmos_mos": "Mean Opinion Score, showing how natural the speech is, measured with UMTOS.",
+        "ecapa_secs": "Speaker Embedding Cosine Similarity between reference audio and generated speech, measured with ECAPA.",
+    }
+    # compose a description for columns
+    description = ""
+    for k in df.columns:
+        if k in description_map:
+            description += f"* {description_map[k]}\n"
+    # Only rename columns that exist in the dataframe
+    existing_columns = {k: v for k, v in rename_map.items() if k in df.columns}
+    df = df.rename(columns=existing_columns)
+    return df, description
+
+
 def get_leaderboard_data(system_type: str, dataset: str) -> pd.DataFrame:
     """
     Fetches metrics.yaml for all systems of given type if they have evaluation for the given dataset.
@@ -64,8 +94,8 @@ def get_leaderboard_data(system_type: str, dataset: str) -> pd.DataFrame:
             continue
     df = pd.DataFrame(metrics_data)
     # Remove 'aesthetics_' prefix from column names where applicable
-    df.columns = [col.replace('aesthetics_', '') if col.startswith('aesthetics_') else col for col in df.columns]
+    df, description = make_pretty(df)
     # compose datatypes for the table: markdown for model name, and number for all other columns
     datatypes = ["markdown"] + ["number"] * (len(df.columns) - 1)
-    return df, datatypes
+    return df, datatypes, description