Update app.py

redefine based on open asr leaderboard package

app.py

@@ -8,46 +8,59 @@ import re
 
 from huggingface_hub import login
 
+# Authentication setup
 token = os.environ.get("HG_TOKEN")
-
+print(f"Token exists: {token is not None}")
+if token:
+    print(f"Token length: {len(token)}")
+    print(f"Token first few chars: {token[:4]}...")
 login(token)
 
 print("Loading dataset...")
-dataset = load_dataset("sudoping01/bambara-speech-recognition-benchmark", name="default")["eval"]
-references = {row["id"]: row["text"] for row in dataset}
+try:
+    dataset = load_dataset("sudoping01/bambara-speech-recognition-benchmark", name="default", use_auth_token=token)["eval"]
+    print(f"Successfully loaded dataset with {len(dataset)} samples")
+    references = {row["id"]: row["text"] for row in dataset}
+except Exception as e:
+    print(f"Error loading dataset: {str(e)}")
+    # Fallback in case dataset can't be loaded
+    references = {}
+    print("WARNING: Using empty references dictionary due to dataset loading error")
 
+# Initialize leaderboard file
 leaderboard_file = "leaderboard.csv"
 if not os.path.exists(leaderboard_file):
-    pd.DataFrame(columns=["submitter", "WER", "CER", "timestamp"]).to_csv(leaderboard_file, index=False)
+    pd.DataFrame(columns=["submitter", "WER", "CER", "weighted_WER", "weighted_CER", "samples_evaluated", "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):
     """
-    Normalize text for WER/CER calculation:
-    - Convert to lowercase
-    - Remove punctuation
-    - Replace multiple spaces with single space
-    - Strip leading/trailing spaces
+    Normalize text by converting to lowercase, removing punctuation, and normalizing whitespace.
     """
     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)
+    # Remove punctuation, keeping spaces
+    text = re.sub(r'[^\w\s]', '', text)
     
-    # # Normalize whitespace
-    # text = re.sub(r'\s+', ' ', text).strip()
+    # Normalize whitespace
+    text = re.sub(r'\s+', ' ', text).strip()
     
     return text
 
 def calculate_metrics(predictions_df):
-    """Calculate WER and CER for predictions."""
-    results = []
+    """
+    Calculate WER and CER for each sample and return averages and per-sample results.
+    Uses both standard average and length-weighted average.
+    """
+    per_sample_metrics = []
+    total_ref_words = 0
+    total_ref_chars = 0
 
+    # Process each sample
     for _, row in predictions_df.iterrows():
         id_val = row["id"]
         if id_val not in references:
@@ -57,146 +70,262 @@ def calculate_metrics(predictions_df):
         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()
+        reference_chars = list(reference)
+        
+        # Skip very short references for more stable metrics
+        if len(reference_words) < 2:
+            print(f"Warning: Reference too short for ID {id_val}, skipping")
+            continue
         
-        if len(results) < 5:
+        # Store sample info for debugging (first few samples)
+        if len(per_sample_metrics) < 5:
+            print(f"ID: {id_val}")
+            print(f"Reference: '{reference}'")
+            print(f"Hypothesis: '{hypothesis}'")
             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:
+            # Cap metrics at sensible values to prevent outliers
+            sample_wer = min(sample_wer, 2.0)  # Cap at 200% WER
+            sample_cer = min(sample_cer, 2.0)  # Cap at 200% CER
+            
+            # For weighted calculations
+            total_ref_words += len(reference_words)
+            total_ref_chars += len(reference_chars)
+            
+            if len(per_sample_metrics) < 5:
                 print(f"WER: {sample_wer}, CER: {sample_cer}")
-                
-            results.append({
+            
+            per_sample_metrics.append({
                 "id": id_val,
                 "reference": reference,
                 "hypothesis": hypothesis,
+                "ref_word_count": len(reference_words),
+                "ref_char_count": len(reference_chars),
                 "wer": sample_wer,
                 "cer": sample_cer
             })
         except Exception as e:
             print(f"Error calculating metrics for ID {id_val}: {str(e)}")
     
-    if not results:
+    if not per_sample_metrics:
         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
+    # Calculate standard average metrics
+    avg_wer = sum(item["wer"] for item in per_sample_metrics) / len(per_sample_metrics)
+    avg_cer = sum(item["cer"] for item in per_sample_metrics) / len(per_sample_metrics)
+    
+    # Calculate weighted average metrics based on reference length
+    weighted_wer = sum(item["wer"] * item["ref_word_count"] for item in per_sample_metrics) / total_ref_words
+    weighted_cer = sum(item["cer"] * item["ref_char_count"] for item in per_sample_metrics) / total_ref_chars
+    
+    print(f"Simple average WER: {avg_wer:.4f}, CER: {avg_cer:.4f}")
+    print(f"Weighted average WER: {weighted_wer:.4f}, CER: {weighted_cer:.4f}")
+    print(f"Processed {len(per_sample_metrics)} valid samples")
+    
+    return avg_wer, avg_cer, weighted_wer, weighted_cer, per_sample_metrics
+
+def styled_error(message):
+    """Format error messages with red styling"""
+    return f"<div style='color: red; font-weight: bold; padding: 10px; border-radius: 5px; background-color: #ffe0e0;'>{message}</div>"
+
+def styled_success(message):
+    """Format success messages with green styling"""
+    return f"<div style='color: green; font-weight: bold; padding: 10px; border-radius: 5px; background-color: #e0ffe0;'>{message}</div>"
+
+def styled_info(message):
+    """Format informational messages with blue styling"""
+    return f"<div style='color: #004080; padding: 10px; border-radius: 5px; background-color: #e0f0ff;'>{message}</div>"
 
 def process_submission(submitter_name, csv_file):
+    """
+    Process a submission CSV, calculate metrics, and update the leaderboard.
+    Returns a status message and updated leaderboard.
+    """
     try:
+        # Validate submitter name
+        if not submitter_name or len(submitter_name.strip()) < 3:
+            return styled_error("Please provide a valid submitter name (at least 3 characters)"), None
+        
         # Read and validate the uploaded CSV
         df = pd.read_csv(csv_file)
         print(f"Processing submission from {submitter_name} with {len(df)} rows")
         
+        # Basic validation
         if len(df) == 0:
-            return "Error: Uploaded CSV is empty.", None
+            return styled_error("Error: Uploaded CSV is empty."), None
+        
+        if len(df) < 10:
+            return styled_error("Error: Submission contains too few samples (minimum 10 required)."), None
             
         if set(df.columns) != {"id", "text"}:
-            return f"Error: CSV must contain exactly 'id' and 'text' columns. Found: {', '.join(df.columns)}", None
+            return styled_error(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()]["id"].unique()
-            return f"Error: Duplicate IDs found: {', '.join(map(str, dup_ids[:5]))}", None
+            return styled_error(f"Error: Duplicate IDs found: {', '.join(map(str, dup_ids[:5]))}."), None
+
+        # Ensure text column contains strings
+        df["text"] = df["text"].astype(str)
+
+        # Check for valid references
+        if not references:
+            return styled_error("Error: Reference dataset could not be loaded. Please try again later."), 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"Error: Missing {len(missing_ids)} IDs in submission. First few missing: {', '.join(map(str, list(missing_ids)[:5]))}", None
+            return styled_error(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"Error: Found {len(extra_ids)} extra IDs not in reference dataset. First few extra: {', '.join(map(str, list(extra_ids)[:5]))}", None
+            return styled_error(f"Error: Found {len(extra_ids)} extra IDs not in reference dataset. First few extra: {', '.join(map(str, list(extra_ids)[:5]))}."), None
+        
+        # Check for suspicious submissions (high percentage of exact matches)
+        exact_matches = 0
+        for _, row in df.iterrows():
+            if normalize_text(row["text"]) == normalize_text(references[row["id"]]):
+                exact_matches += 1
         
-        # Calculate WER and CER
+        exact_match_ratio = exact_matches / len(df)
+        if exact_match_ratio > 0.95:  # If 95% exact matches, likely copying reference
+            return styled_error("Suspicious submission: Too many exact matches with reference texts."), None
+        
+        # Calculate metrics
         try:
-            avg_wer, avg_cer, detailed_results = calculate_metrics(df)
+            avg_wer, avg_cer, weighted_wer, weighted_cer, detailed_results = calculate_metrics(df)
             
             # Debug information
             print(f"Calculated metrics - WER: {avg_wer:.4f}, CER: {avg_cer:.4f}")
+            print(f"Weighted metrics - WER: {weighted_wer:.4f}, CER: {weighted_cer:.4f}")
             print(f"Processed {len(detailed_results)} valid samples")
             
             # Check for suspiciously low values
-            if avg_wer < 0.001:
+            if avg_wer < 0.001 or weighted_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
+                return styled_error("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
+            return styled_error(f"Error calculating metrics: {str(e)}"), None
         
         # Update the leaderboard
         leaderboard = pd.read_csv(leaderboard_file)
         timestamp = datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+        
         new_entry = pd.DataFrame(
-            [[submitter_name, avg_wer, avg_cer, timestamp]],
-            columns=["submitter", "WER", "CER", "timestamp"]
+            [[submitter_name, avg_wer, avg_cer, weighted_wer, weighted_cer, len(detailed_results), timestamp]],
+            columns=["submitter", "WER", "CER", "weighted_WER", "weighted_CER", "samples_evaluated", "timestamp"]
         )
-        leaderboard = pd.concat([leaderboard, new_entry]).sort_values("WER")
-        leaderboard.to_csv(leaderboard_file, index=False)
         
-        return f"Submission processed successfully! WER: {avg_wer:.4f}, CER: {avg_cer:.4f}", leaderboard
+        # Combine with existing leaderboard and keep only the best submission per submitter
+        combined = pd.concat([leaderboard, new_entry])
+        # Sort by WER (ascending) and get first entry for each submitter
+        best_entries = combined.sort_values("WER").groupby("submitter").first().reset_index()
+        # Sort the resulting dataframe by WER
+        updated_leaderboard = best_entries.sort_values("WER")
+        updated_leaderboard.to_csv(leaderboard_file, index=False)
+        
+        # Create detailed metrics summary
+        metrics_summary = f"""
+        <h3>Submission Results</h3>
+        <table>
+            <tr><td><b>Submitter:</b></td><td>{submitter_name}</td></tr>
+            <tr><td><b>Word Error Rate (WER):</b></td><td>{avg_wer:.4f}</td></tr>
+            <tr><td><b>Character Error Rate (CER):</b></td><td>{avg_cer:.4f}</td></tr>
+            <tr><td><b>Weighted WER:</b></td><td>{weighted_wer:.4f}</td></tr>
+            <tr><td><b>Weighted CER:</b></td><td>{weighted_cer:.4f}</td></tr>
+            <tr><td><b>Samples Evaluated:</b></td><td>{len(detailed_results)}</td></tr>
+            <tr><td><b>Submission Time:</b></td><td>{timestamp}</td></tr>
+        </table>
+        """
+        
+        return styled_success(f"Submission processed successfully!") + styled_info(metrics_summary), updated_leaderboard
         
     except Exception as e:
         print(f"Error processing submission: {str(e)}")
-        return f"Error processing submission: {str(e)}", None
+        return styled_error(f"Error processing submission: {str(e)}"), None
 
 # Create the Gradio interface
 with gr.Blocks(title="Bambara ASR Leaderboard") as demo:
     gr.Markdown(
         """
         # 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).
+        
+        ## Metrics
+        - **WER**: Word Error Rate (lower is better) - measures word-level accuracy
+        - **CER**: Character Error Rate (lower is better) - measures character-level accuracy
+        
+        We report both standard averages and length-weighted averages (where longer samples have more influence on the final score).
         """
     )
     
     with gr.Row():
-        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="Leaderboard",
-        value=pd.read_csv(leaderboard_file),
-        interactive=False
-    )
+        with gr.Column(scale=1):
+            submitter = gr.Textbox(
+                label="Submitter Name or Model Name", 
+                placeholder="e.g., MALIBA-AI/asr",
+                info="Name to appear on the leaderboard"
+            )
+            csv_upload = gr.File(
+                label="Upload CSV File", 
+                file_types=[".csv"],
+                info="CSV must have 'id' and 'text' columns"
+            )
+            submit_btn = gr.Button("Submit", variant="primary")
+            
+        with gr.Column(scale=2):
+            with gr.Accordion("Submission Format", open=False):
+                gr.Markdown(
+                    """
+                    ### CSV Format Requirements
+                    
+                    Your CSV file must:
+                    - Have exactly two columns: `id` and `text`
+                    - The `id` column must match the IDs in the reference dataset
+                    - The `text` column should contain your model's transcriptions
+                    
+                    Example:
+                    ```
+                    id,text
+                    audio_001,n ye foro ka taa
+                    audio_002,i ni ce
+                    ```
+                    
+                    ### Evaluation Process
+                    
+                    Your submissions are evaluated by:
+                    1. Normalizing both reference and predicted text (lowercase, punctuation removal)
+                    2. Calculating Word Error Rate (WER) and Character Error Rate (CER)
+                    3. Computing both simple average and length-weighted average
+                    4. Ranking on the leaderboard by WER (lower is better)
+                    
+                    Only your best submission is kept on the leaderboard.
+                    """
+                )
+    
+    output_msg = gr.HTML(label="Status")
+    
+    # Leaderboard display
+    with gr.Accordion("Leaderboard", open=True):
+        leaderboard_display = gr.DataFrame(
+            label="Current Standings",
+            value=pd.read_csv(leaderboard_file),
+            interactive=False
+        )
     
     submit_btn.click(
         fn=process_submission,