Update app.py

app.py

@@ -1,56 +1,14 @@
 import gradio as gr
 import requests
-from typing import Dict, Tuple
+from typing import Dict, Tuple, List
 
-# Custom CSS for Open Sans font and color theme
-css = """
-@import url('https://fonts.googleapis.com/css2?family=Open+Sans:wght@300;400;600;700&display=swap');
-
-body { 
-    font-family: 'Open Sans', sans-serif !important;
-}
-
-.primary-btn { 
-    background-color: #3452db !important;
-}
-
-.primary-btn:hover { 
-    background-color: #2a41af !important;
-}
-"""
-
-# Create custom theme with specific color
-theme = gr.themes.Soft(
-    primary_hue=gr.themes.colors.Color(
-        name="blue",
-        c50="#eef1ff",
-        c100="#e0e5ff",
-        c200="#c3cbff",
-        c300="#a5b2ff",
-        c400="#8798ff",
-        c500="#6a7eff",
-        c600="#3452db",  # Our main color
-        c700="#2a41af",
-        c800="#1f3183",
-        c900="#152156",
-        c950="#0a102b",
-    )
-)
-
-BRAND_EXAMPLES = [
-    "Nike - Just Do It. The power of determination.",
-    "Apple - Think Different. Innovation redefined.",
-    "McDonald's - I'm Lovin' It. Creating joy.",
-    "BMW - The Ultimate Driving Machine.",
-    "L'Oréal - Because You're Worth It."
+SEMANTIC_CATEGORIES = [
+    "references to the term 'word' and its variations in different contexts",
+    "the start of a document",
+    "references to global events and organizations",
+    "elements related to search engine optimization (SEO)"
 ]
 
-def get_feature_url(feature_id: int, dashboard: bool = False) -> str:
-    base_url = f"https://www.neuronpedia.org/gemma-2-2b/0-gemmascope-mlp-16k/{feature_id}"
-    if dashboard:
-        return f"{base_url}?embed=true&embedexplanation=true&embedplots=true&embedtest=true&height=300"
-    return base_url
-
 def get_top_features(text: str, k: int = 5) -> Dict:
     url = "https://www.neuronpedia.org/api/search-with-topk"
     payload = {
@@ -70,53 +28,44 @@ def get_top_features(text: str, k: int = 5) -> Dict:
     )
     return response.json() if response.status_code == 200 else None
 
+def categorize_features(features: List[Dict]) -> List[str]:
+    # This is a placeholder - in practice, you'd want to analyze the features 
+    # and map them to semantic categories based on their activations
+    categories = []
+    for cat in SEMANTIC_CATEGORIES:
+        # Here you would check if the features match this category
+        if len(categories) < 5:  # Limit to top 5 categories
+            categories.append(cat)
+    return categories
+
 def format_output(data: Dict) -> Tuple[str, str, str]:
     if not data:
         return "Error analyzing text", "", ""
-        
-    output = (
-        "# Neural Feature Analysis\n\n"
-        "*Analyzing neural patterns in the text using Gemma's interpretable features*\n\n"
-    )
     
-    # Format token-feature analysis
+    output = "# Semantic Analysis\n\n"
+    output += "*Analyzing semantic patterns in the text using Gemma's neural features*\n\n"
+    
+    all_features = []
     for result in data['results']:
-        token = result['token']
-        if token == '<bos>':  # Skip BOS token
+        if result['token'] == '<bos>':
             continue
-            
-        features = result['top_features']
-        if features:
-            output += f"\n## 🔍 Token: '{token}'\n"
-            for feat in features:
-                feat_index = feat['feature_index']
-                activation = feat['activation_value']
-                feature_url = get_feature_url(feat_index)
-                
-                # Try to get feature info/label
-                feature_info = ""
-                if 'feature' in feat and isinstance(feat['feature'], dict):
-                    if 'name' in feat['feature']:
-                        feature_info = f" - {feat['feature']['name']}"
-                    elif 'description' in feat['feature']:
-                        feature_info = f" - {feat['feature']['description']}"
-                
-                output += f"### Feature [{feat_index}]({feature_url}){feature_info}\n"
-                output += f"- **Activation Score:** {activation:.2f}\n"
-                output += f"- [📊 View Analysis Dashboard]({feature_url})\n"
+        all_features.extend(result['top_features'])
     
-    # Get highest activation feature for dashboard
-    max_activation = 0
-    max_feature = None
+    # Sort all features by activation value
+    top_features = sorted(all_features, key=lambda x: x['activation_value'], reverse=True)[:5]
     
-    for result in data['results']:
-        for feature in result['top_features']:
-            if feature['activation_value'] > max_activation:
-                max_activation = feature['activation_value']
-                max_feature = feature['feature_index']
+    # Get semantic categories
+    categories = categorize_features(top_features)
     
-    if max_feature:
-        dashboard_url = get_feature_url(max_feature, dashboard=True)
+    # Format output with semantic categories
+    output += "## 🔍 Key Semantic Patterns\n\n"
+    for cat in categories:
+        output += f"• {cat}\n"
+    
+    # Add feature dashboard for highest activation feature
+    if top_features:
+        max_feature = top_features[0]
+        dashboard_url = f"https://www.neuronpedia.org/gemma-2-2b/0-gemmascope-mlp-16k/{max_feature['feature_index']}?embed=true&embedexplanation=true&embedplots=true&embedtest=true&height=300"
         iframe = f'''
             <div style="border:1px solid #eee;border-radius:8px;padding:1px;background:#fff;">
                 <iframe 
@@ -128,51 +77,36 @@ def format_output(data: Dict) -> Tuple[str, str, str]:
                 ></iframe>
             </div>
         '''
-        feature_label = f"📊 Feature {max_feature} Analysis (Peak Activation: {max_activation:.2f})"
+        feature_label = f"📊 Top Neural Pattern Analysis"
     else:
         iframe = ""
-        feature_label = "No significant features found"
+        feature_label = "No significant patterns found"
     
     return output, iframe, feature_label
 
 def create_interface():
-    with gr.Blocks(theme=theme, css=css) as interface:
-        gr.Markdown(
-            "# 🧠 Neural Feature Analyzer",
-            elem_classes="text-center"
-        )
-        gr.Markdown(
-            "*Analyze text using Gemma's interpretable neural features*\n\n"
-            "Shows top 5 most activated features for each token with density < 1%",
-            elem_classes="text-center"
-        )
+    with gr.Blocks() as interface:
+        gr.Markdown("# 🧠 Semantic Pattern Analyzer")
+        gr.Markdown("*Analyze semantic patterns in text using neural features*")
         
         with gr.Row():
             with gr.Column():
                 input_text = gr.Textbox(
                     lines=5,
-                    placeholder="Enter text to analyze...",
-                    label="Input Text",
-                    elem_classes="input-text"
-                )
-                analyze_btn = gr.Button(
-                    "🔍 Analyze Neural Features", 
-                    variant="primary",
-                    elem_classes="primary-btn"
-                )
-                gr.Examples(
-                    BRAND_EXAMPLES, 
-                    inputs=input_text,
-                    elem_classes="examples"
+                    placeholder="Enter text to analyze semantic patterns...",
+                    label="Input Text"
                 )
+                analyze_btn = gr.Button("🔍 Analyze Patterns", variant="primary")
+                gr.Examples([
+                    "WordLift is an AI-powered SEO tool that optimizes content",
+                    "This document outlines the global health initiatives",
+                    "Using key words and terms to improve search rankings"
+                ], inputs=input_text)
             
             with gr.Column():
-                output_text = gr.Markdown(elem_classes="output-text")
-                feature_label = gr.Text(
-                    show_label=False,
-                    elem_classes="feature-label"
-                )
-                dashboard = gr.HTML(elem_classes="dashboard")
+                output_text = gr.Markdown()
+                feature_label = gr.Text(show_label=False)
+                dashboard = gr.HTML()
         
         analyze_btn.click(
             fn=lambda text: format_output(get_top_features(text)),