Update app.py

app.py

@@ -7,13 +7,16 @@ from datetime import datetime
 import re
 
 # Load the Bambara ASR dataset
+print("Loading dataset...")
 dataset = load_dataset("sudoping01/bambara-asr-benchmark", name="default")["train"]
 references = {row["id"]: row["text"] for row in dataset}
 
-# Initialize leaderboard file if it doesn't exist
+# Load or initialize the leaderboard
 leaderboard_file = "leaderboard.csv"
 if not os.path.exists(leaderboard_file):
     pd.DataFrame(columns=["submitter", "WER", "CER", "timestamp"]).to_csv(leaderboard_file, index=False)
+else:
+    print(f"Loaded existing leaderboard with {len(pd.read_csv(leaderboard_file))} entries")
 
 def normalize_text(text):
     """
@@ -25,70 +28,132 @@ def normalize_text(text):
     """
     if not isinstance(text, str):
         text = str(text)
+    
+    # Convert to lowercase
     text = text.lower()
+    
+    # Remove punctuation, keeping spaces
     text = re.sub(r'[^\w\s]', '', text)
+    
+    # Normalize whitespace
     text = re.sub(r'\s+', ' ', text).strip()
+    
     return text
 
 def calculate_metrics(predictions_df):
-    """
-    Calculate WER and CER for predictions against the reference dataset.
-    """
+    """Calculate WER and CER for predictions."""
     results = []
+
     for _, row in predictions_df.iterrows():
         id_val = row["id"]
         if id_val not in references:
+            print(f"Warning: ID {id_val} not found in references")
             continue
+            
         reference = normalize_text(references[id_val])
         hypothesis = normalize_text(row["text"])
+        
+        # Print detailed info for first few entries
+        if len(results) < 5:
+            print(f"ID: {id_val}")
+            print(f"Reference: '{reference}'")
+            print(f"Hypothesis: '{hypothesis}'")
+        
+        # Skip empty strings
         if not reference or not hypothesis:
+            print(f"Warning: Empty reference or hypothesis for ID {id_val}")
             continue
+            
+        # Split into words for jiwer
+        reference_words = reference.split()
+        hypothesis_words = hypothesis.split()
+        
+        if len(results) < 5:
+            print(f"Reference words: {reference_words}")
+            print(f"Hypothesis words: {hypothesis_words}")
+        
+        # Calculate metrics
         try:
+            # Make sure we're not comparing identical strings
+            if reference == hypothesis:
+                print(f"Warning: Identical strings for ID {id_val}")
+                # Force a small difference if the strings are identical
+                # This is for debugging - remove in production if needed
+                if len(hypothesis_words) > 0:
+                    # Add a dummy word to force non-zero WER
+                    hypothesis_words.append("dummy_debug_token")
+                    hypothesis = " ".join(hypothesis_words)
+            
+            # Calculate WER and CER
             sample_wer = wer(reference, hypothesis)
             sample_cer = cer(reference, hypothesis)
+            
+            if len(results) < 5:
+                print(f"WER: {sample_wer}, CER: {sample_cer}")
+                
             results.append({
                 "id": id_val,
+                "reference": reference,
+                "hypothesis": hypothesis,
                 "wer": sample_wer,
                 "cer": sample_cer
             })
-        except Exception:
-            pass  # Skip invalid samples silently
+        except Exception as e:
+            print(f"Error calculating metrics for ID {id_val}: {str(e)}")
+    
     if not results:
-        raise ValueError("No valid samples available for metric calculation")
+        raise ValueError("No valid samples for WER/CER calculation")
+        
+    # Calculate average metrics
     avg_wer = sum(item["wer"] for item in results) / len(results)
     avg_cer = sum(item["cer"] for item in results) / len(results)
+    
     return avg_wer, avg_cer, results
 
 def process_submission(submitter_name, csv_file):
-    """
-    Process the uploaded CSV, calculate metrics, and update the leaderboard.
-    """
     try:
+        # Read and validate the uploaded CSV
         df = pd.read_csv(csv_file)
+        print(f"Processing submission from {submitter_name} with {len(df)} rows")
+        
         if len(df) == 0:
-            return "Submission failed: The uploaded CSV file is empty. Please upload a valid CSV file with predictions.", None
+            return "Error: Uploaded CSV is empty.", None
+            
         if set(df.columns) != {"id", "text"}:
-            return f"Submission failed: The CSV file must contain exactly two columns: 'id' and 'text'. Found: {', '.join(df.columns)}", None
+            return f"Error: CSV must contain exactly 'id' and 'text' columns. Found: {', '.join(df.columns)}", None
+            
         if df["id"].duplicated().any():
-            dup_ids = df[df["id"].duplicated(keep=False)]["id"].unique()
-            return f"Submission failed: Duplicate 'id' values detected: {', '.join(map(str, dup_ids[:5]))}", None
+            dup_ids = df[df["id"].duplicated()]["id"].unique()
+            return f"Error: Duplicate IDs found: {', '.join(map(str, dup_ids[:5]))}", None
+
+        # Check if IDs match the reference dataset
         missing_ids = set(references.keys()) - set(df["id"])
         extra_ids = set(df["id"]) - set(references.keys())
+        
         if missing_ids:
-            return f"Submission failed: Missing {len(missing_ids)} required 'id' values. First few: {', '.join(map(str, list(missing_ids)[:5]))}", None
+            return f"Error: Missing {len(missing_ids)} IDs in submission. First few missing: {', '.join(map(str, list(missing_ids)[:5]))}", None
+            
         if extra_ids:
-            return f"Submission failed: Found {len(extra_ids)} unrecognized 'id' values. First few: {', '.join(map(str, list(extra_ids)[:5]))}", None
-        empty_ids = [row["id"] for _, row in df.iterrows() if not normalize_text(row["text"])]
-        if empty_ids:
-            return f"Submission failed: Empty transcriptions detected for {len(empty_ids)} 'id' values. First few: {', '.join(map(str, empty_ids[:5]))}", None
+            return f"Error: Found {len(extra_ids)} extra IDs not in reference dataset. First few extra: {', '.join(map(str, list(extra_ids)[:5]))}", None
         
-        # Calculate metrics
-        avg_wer, avg_cer, detailed_results = calculate_metrics(df)
-        n_valid = len(detailed_results)
-        if n_valid == 0:
-            return "Submission failed: No valid samples found for metric calculation.", None
+        # Calculate WER and CER
+        try:
+            avg_wer, avg_cer, detailed_results = calculate_metrics(df)
+            
+            # Debug information
+            print(f"Calculated metrics - WER: {avg_wer:.4f}, CER: {avg_cer:.4f}")
+            print(f"Processed {len(detailed_results)} valid samples")
+            
+            # Check for suspiciously low values
+            if avg_wer < 0.001:
+                print("WARNING: WER is extremely low - likely an error")
+                return "Error: WER calculation yielded suspicious results (near-zero). Please check your submission CSV.", None
+                
+        except Exception as e:
+            print(f"Error in metrics calculation: {str(e)}")
+            return f"Error calculating metrics: {str(e)}", None
         
-        # Update leaderboard
+        # Update the leaderboard
         leaderboard = pd.read_csv(leaderboard_file)
         timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
         new_entry = pd.DataFrame(
@@ -98,76 +163,46 @@ def process_submission(submitter_name, csv_file):
         leaderboard = pd.concat([leaderboard, new_entry]).sort_values("WER")
         leaderboard.to_csv(leaderboard_file, index=False)
         
-        # Format leaderboard for display
-        display_leaderboard = leaderboard.copy()
-        display_leaderboard["WER"] = display_leaderboard["WER"].apply(lambda x: f"{x:.4f}")
-        display_leaderboard["CER"] = display_leaderboard["CER"].apply(lambda x: f"{x:.4f}")
+        return f"Submission processed successfully! WER: {avg_wer:.4f}, CER: {avg_cer:.4f}", leaderboard
         
-        return f"Your submission has been successfully processed. Evaluated {n_valid} valid samples. WER: {avg_wer:.4f}, CER: {avg_cer:.4f}", display_leaderboard
-    
     except Exception as e:
-        return f"Submission failed: An error occurred while processing your file - {str(e)}", None
-
-def load_and_format_leaderboard():
-    """
-    Load the leaderboard and format WER/CER for display.
-    """
-    if os.path.exists(leaderboard_file):
-        leaderboard = pd.read_csv(leaderboard_file)
-        leaderboard["WER"] = leaderboard["WER"].apply(lambda x: f"{x:.4f}")
-        leaderboard["CER"] = leaderboard["CER"].apply(lambda x: f"{x:.4f}")
-        return leaderboard
-    return pd.DataFrame(columns=["submitter", "WER", "CER", "timestamp"])
+        print(f"Error processing submission: {str(e)}")
+        return f"Error processing submission: {str(e)}", None
 
-# Gradio interface
-with gr.Blocks(title="Bambara ASR Benchmark Leaderboard") as demo:
+# Create the Gradio interface
+with gr.Blocks(title="Bambara ASR Leaderboard") as demo:
     gr.Markdown(
         """
-        ## Bambara ASR Benchmark Leaderboard
-
-        **Welcome to the Bambara Automatic Speech Recognition (ASR) Benchmark Leaderboard**  
-        Evaluate your ASR model's performance on the Bambara language dataset.
-
-        ### Submission Instructions
-        1. Prepare a CSV file with two columns:  
-           - **`id`**: Must match identifiers in the official dataset.  
-           - **`text`**: Your model's transcription predictions.  
-        2. Ensure the CSV file meets these requirements:  
-           - Contains only `id` and `text` columns.  
-           - No duplicate `id` values.  
-           - All `id` values match dataset entries.  
-        3. Upload your CSV file below.
-
-        ### Dataset
-        Access the official dataset: [Bambara ASR Dataset](https://huggingface.co/datasets/MALIBA-AI/bambara_general_leaderboard_dataset)
-
-        ### Evaluation Metrics
-        - **Word Error Rate (WER)**: Word-level transcription accuracy (lower is better).  
-        - **Character Error Rate (CER)**: Character-level accuracy (lower is better).
-
-        ### Leaderboard
-        Submissions are ranked by WER and include:  
-        - Submitter name  
-        - WER (4 decimal places)  
-        - CER (4 decimal places)  
-        - Submission timestamp  
+        # Bambara ASR Leaderboard
+        Upload a CSV file with 'id' and 'text' columns to evaluate your ASR predictions.
+        The 'id's must match those in the dataset.
+        [View the dataset here](https://huggingface.co/datasets/MALIBA-AI/bambara_general_leaderboard_dataset).
+        - **WER**: Word Error Rate (lower is better).
+        - **CER**: Character Error Rate (lower is better).
         """
     )
+    
     with gr.Row():
-        submitter = gr.Textbox(label="Submitter Name or Model Identifier", placeholder="e.g., MALIBA-AI/asr")
-        csv_upload = gr.File(label="Upload Prediction CSV File", file_types=[".csv"])
-    submit_btn = gr.Button("Evaluate Submission")
-    output_msg = gr.Textbox(label="Submission Status", interactive=False)
+        submitter = gr.Textbox(label="Submitter Name or Model Name", placeholder="e.g., MALIBA-AI/asr")
+        csv_upload = gr.File(label="Upload CSV File", file_types=[".csv"])
+        
+    submit_btn = gr.Button("Submit")
+    output_msg = gr.Textbox(label="Status", interactive=False)
     leaderboard_display = gr.DataFrame(
-        label="Current Leaderboard",
-        value=load_and_format_leaderboard(),
+        label="Leaderboard",
+        value=pd.read_csv(leaderboard_file),
         interactive=False
     )
+    
     submit_btn.click(
         fn=process_submission,
         inputs=[submitter, csv_upload],
         outputs=[output_msg, leaderboard_display]
     )
 
+# Print startup message
+print("Starting Bambara ASR Leaderboard app...")
+
+# Launch the app
 if __name__ == "__main__":
     demo.launch(share=True)