tidy up

app.py

@@ -12,11 +12,11 @@ column_names = {
     "MODEL": "Model",
     "Avg. WER": "Average WER ⬇️",    
     "RTF": "RTF (1e-3) ⬇️",
-    "Common Voice WER": "Common Voice",
+    "Common Voice WER": "Common Voice WER ⬇️",
     "D_AVG_CV_WER": "Delta AVG-CV WER",
 }
 
-# Skipping testings just uing the numbers computed in the original space.
+# Skipping testings just uing the numbers computed in the original space for my sanity sake
 # eval_queue_repo, requested_models, csv_results = load_all_info_from_dataset_hub()
 
 # if not csv_results.exists():
@@ -111,12 +111,10 @@ data = [
      1.2, 26.41, 41.75]
 ]
 
-# Noms de colonnes mis à jour
 columns = [
-    "model", "RTF", "Avrg. WER", "Common Voice"
+    "Model", "RTF (1e-3) ⬇️", "Average WER ⬇️", "Common Voice WER ⬇️"
 ]
 
-# Création du DataFrame avec les noms de colonnes mis à jour
 original_df = pd.DataFrame(data, columns=columns)
 
 # Formats the columns
@@ -134,7 +132,8 @@ original_df.rename(columns=column_names, inplace=True)
 original_df.sort_values(by='Common Voice', inplace=True)
 
 # Compute delta between average WER and CV WER
-original_df['Detla'] = original_df['Avrg. WER'] - original_df['Common Voice']
+original_df['Detla Avg. C.V. WER'] = original_df['Average WER ⬇️'] - original_df['Common Voice WER ⬇️']
+original_df['Detla Avg. C.V. WER'] = original_df['Detla Avg. C.V. WER'].apply(formatter)
 
 COLS = [c.name for c in fields(AutoEvalColumn)]
 TYPES = [c.type for c in fields(AutoEvalColumn)]
@@ -194,6 +193,21 @@ def request_model(model_text, chbcoco2017):
 with gr.Blocks() as demo:
     gr.HTML(BANNER, elem_id="banner")
     gr.Markdown(INTRODUCTION_TEXT, elem_classes="markdown-text")
+    CUSTOM_MESSAGE="""Legend:
+This space is a fork of the original [hf-audio/open_asr_leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard).
+It aims to show how CommonVoice Test Set is large enough for most languages to give a relativly good approximation of the average WER/CER but at a much lower computational cost.
+#### Why is this useful?
+Because, it gives us a standardized test set for most languages allowing us to programatically choose a relatively good model for any CV supported languages.
+
+`Model`, `RTF (1e-3) ⬇️` and`Average WER ⬇️` were reported from [hf-audio/open_asr_leaderboard](https://huggingface.co/spaces/hf-audio/open_asr_leaderboard) the 9 using version from the 7 Sept. 2023.
+
+### Results
+CommonVoice Test test give a word error rate (WER) within less than 20 points of the average WER.
+
+It's not good. Don't use only CommonVoice to choose the most adequate architecture.
+But to quickly find a suitable ASR model for a large panel of lanugages in a programatic fashion, it's close enough."""
+    gr.Markdown(CUSTOM_MESSAGE, elem_classes="markdown-text")
+    
 
     with gr.Tabs(elem_classes="tab-buttons") as tabs:
         with gr.TabItem("🏅 Leaderboard", elem_id="od-benchmark-tab-table", id=0):