Upload folder using huggingface_hub

hate_speech_demo.py

@@ -441,18 +441,108 @@ def get_openai_moderation(openai_client, user_input):
         return formatted_result, safety_level
     except Exception as e:
         return f"Safety Status: Error\nError: {str(e)}", "unsafe"
+    
+# Perspective AI
+def get_perspective_rating(user_input):
+    """
+    Get content rating from Google's Perspective API.
+    Returns formatted results and safety level.
+    """
+    import json
+    import requests
+    
+    # Use the API key provided
+    API_KEY = "AIzaSyDK8R5gn-Gm42dSyexhThbnBuzcglD6rgY"
+    
+    url = 'https://commentanalyzer.googleapis.com/v1alpha1/comments:analyze'
+    params = {
+        'key': API_KEY
+    }
+    data = {
+        'comment': {'text': user_input},
+        'requestedAttributes': {
+            'TOXICITY': {},
+            'SEVERE_TOXICITY': {},
+            'IDENTITY_ATTACK': {},
+            'INSULT': {},
+            'PROFANITY': {},
+            'THREAT': {},
+            'SEXUALLY_EXPLICIT': {}
+        }
+    }
+    
+    try:
+        start_time = time.time()
+        response = requests.post(url, params=params, data=json.dumps(data))
+        end_time = time.time()
+        response_time = round(end_time - start_time, 4)
+        
+        # Check for errors
+        if response.status_code != 200:
+            return f"Error: API returned status code {response.status_code}", "unsafe"
+            
+        response_json = response.json()
+        
+        # Check if we got the expected response structure
+        if 'attributeScores' not in response_json:
+            return f"Error in API response: {response_json}", "unsafe"
+        
+        # Extract all scores
+        scores = {}
+        flagged_attributes = []
+        
+        for attr, data in response_json['attributeScores'].items():
+            score = data['summaryScore']['value']
+            scores[attr] = score
+            
+            # Check if score is above threshold
+            if score > 0.5:
+                flagged_attributes.append(attr)
+        
+        # Determine safety level
+        if flagged_attributes:
+            safety_level = "unsafe"
+        elif any(score > 0.3 for score in scores.values()):
+            safety_level = "warning"
+        else:
+            safety_level = "safe"
+        
+        # Format the results
+        if flagged_attributes:
+            flagged_str = ", ".join(flagged_attributes)
+            formatted_result = f"Safety Status: Unsafe\nFlagged Attributes: {flagged_str}\n\n"
+        else:
+            formatted_result = "Safety Status: Safe\n\n"
+        
+        # Add all scores
+        formatted_result += "Attribute Scores:\n"
+        for attr, score in scores.items():
+            formatted_result += f"• {attr}: {score:.2f}\n"
+        
+        return formatted_result, safety_level
+    
+    except Exception as e:
+        return f"Error: {str(e)}", "unsafe"
 
 # Updated to only require one input
 def rate_user_input(user_input):
+    """
+    Function to rate a single user input using all four rating services:
+    - Contextual Safety Oracle
+    - LlamaGuard
+    - OpenAI Moderation
+    - Google Perspective API
+    """
     # Initialize APIs with hardcoded keys
     contextual_api = ContextualAPIUtils(api_key=ORACLE_API_KEY)
     together_client = Together(api_key=TOGETHER_API_KEY)
     openai_client = openai.OpenAI(api_key=OPENAI_API_KEY)
     
-    # Get ratings (now only using user_input)
-    llama_rating, llama_safety = get_llama_guard_rating(together_client, user_input)
-    contextual_rating, contextual_retrieval, contextual_safety = get_contextual_rating(contextual_api, user_input)
-    openai_rating, openai_safety = get_openai_moderation(openai_client, user_input)
+    # Get ratings from all four services
+    llama_rating, llama_safety = get_llama_guard_rating(together_client, user_input, user_input)
+    contextual_rating, contextual_retrieval, contextual_safety = get_contextual_rating(contextual_api, user_input, user_input)
+    openai_rating, openai_safety = get_openai_moderation(openai_client, user_input, user_input)
+    perspective_rating, perspective_safety = get_perspective_rating(user_input)
     
     # Format responses carefully to avoid random line breaks
     llama_rating = re.sub(r'\.(?=\s+[A-Z])', '.\n', llama_rating)
@@ -464,6 +554,7 @@ def rate_user_input(user_input):
     # Format results with HTML styling
     llama_html = f"""<div class="rating-box secondary-box {llama_safety}-rating">{llama_rating}</div>"""
     openai_html = f"""<div class="rating-box secondary-box {openai_safety}-rating">{openai_rating}</div>"""
+    perspective_html = f"""<div class="rating-box secondary-box {perspective_safety}-rating">{perspective_rating}</div>"""
     
     # Create the knowledge section (initially hidden) and button
     knowledge_html = ""
@@ -495,12 +586,12 @@ def rate_user_input(user_input):
         <div style="margin-top: 10px; margin-bottom: 5px;">
             <a href="#" id="btn-{popup_id}" class="knowledge-button" 
                onclick="document.getElementById('{popup_id}').style.display='block'; this.style.display='none'; return false;">
-               Show supporting evidence
+               Show Retrieved Knowledge
             </a>
         </div>
         """
     
-    # Format contextual results with HTML styling - button comes AFTER the results
+    # Format contextual results with HTML styling
     contextual_html = f"""
     <div class="rating-box contextual-box {contextual_safety}-rating">
         <button class="copy-button" onclick="navigator.clipboard.writeText(this.parentElement.innerText.replace('Copy', ''))">Copy</button>
@@ -510,7 +601,7 @@ def rate_user_input(user_input):
     {knowledge_html}
     """
     
-    return contextual_html, llama_html, openai_html, ""
+    return contextual_html, llama_html, openai_html, perspective_html, ""
 
 def random_test_case():
     try:
@@ -626,9 +717,15 @@ def create_gradio_app():
         background-color: #c4c4c3 !important;
         color: #000000 !important;
     }
+    
+    /* Perspective API styling */
+    .perspective-icon {
+        vertical-align: middle;
+        margin-right: 5px;
+    }
     """
     
-    with gr.Blocks(title="Hate Speech Rating Oracle", theme=theme, css=custom_css) as app:
+    with gr.Blocks(title="Hate Speech Policy Rating Oracle", theme=theme, css=custom_css) as app:
         # Add loading spinner
         loading_spinner = gr.HTML('<div id="loading-spinner"></div>')
         
@@ -713,23 +810,21 @@ def create_gradio_app():
         
         gr.HTML(policy_popup_html)
         
-        gr.Markdown("# Hate Speech Rating Oracle [BETA]")
+        gr.Markdown("# Hate Speech Policy Rating Oracle")
         gr.Markdown(
-    "Assess whether user-generated social content contains hate speech using Contextual AI's State-of-the-Art Agentic RAG system. Classifications are steerable and explainable as they are based on a policy document rather than parametric knowledge! This app also returns ratings from LlamaGuard 3.0 and the OpenAI Moderation API for you to compare. This is a demo from Contextual AI researchers. Feedback is welcome as we work with design partners to bring this to production.  \n"
-    "## Instructions \n" 
-    "Enter user-generated content to receive an assessment from all three models. Or use our random test case generator to have it pre-filled.  \n"
-    "## How it works\n" 
-    "* **Document-Grounded Evaluations**: Every rating is directly tied to our <a href='#' onclick='openPolicyPopup(); return false;'>hate speech policy document</a>, which makes our system far superior to other solutions that lack transparent decision criteria.\n"
-    "* **Adaptable Policies**: The policy document serves as a starting point and can be easily adjusted to meet your specific requirements. As policies evolve, the system immediately adapts without requiring retraining.\n"
-    "* **Clear Rationales**: Each evaluation includes a detailed explanation referencing specific policy sections, allowing users to understand exactly why content was flagged or approved.\n"
-    "* **Continuous Improvement**: The system learns from feedback, addressing any misclassifications by improving retrieval accuracy over time.\n\n"
-    "Our approach combines Contextual's state-of-the-art <a href='https://contextual.ai/blog/introducing-instruction-following-reranker/' target='_blank'>steerable reranker</a>, <a href='https://contextual.ai/blog/introducing-grounded-language-model/' target='_blank'>world's most grounded language model</a>, and <a href='https://contextual.ai/blog/combining-rag-and-specialization/' target='_blank'>tuning for agent specialization</a> to achieve superhuman performance in content evaluation tasks. This technology enables consistent, fine-grained assessments across any content type and format.\n\n"
-    
-    "## Contact info \n"
-    "Reach out to Bertie Vidgen, Head of Human Data at Contextual AI, to find out more or sign up as a design partner at bertie@contextual.ai \n"
-    "## SAFETY WARNING \n" 
-    "Some of the randomly generated test cases contain hateful language that you might find offensive or upsetting."
-)
+            "Assess whether user-generated social content contains hate speech using Contextual AI's State-of-the-Art Agentic RAG system. Classifications are steerable and explainable as they are based on a policy document rather than parametric knowledge! This app also returns ratings from LlamaGuard 3.0, the OpenAI Moderation API, and Google's Perspective API for you to compare. This is a demo from Contextual AI researchers. Feedback is welcome as we work with design partners to bring this to production.  \n"
+            "## Instructions \n" 
+            "Enter user-generated content to receive an assessment from all four models. Or use our random test case generator to have it pre-filled.  \n"
+            "## How it works\n" 
+            "* **Document-Grounded Evaluations**: Every rating is directly tied to our <a href='#' onclick='openPolicyPopup(); return false;'>hate speech policy document</a>, which makes our system far superior to other solutions that lack transparent decision criteria.\n"
+            "* **Adaptable Policies**: The policy document serves as a starting point and can be easily adjusted to meet your specific requirements. As policies evolve, the system immediately adapts without requiring retraining.\n"
+            "* **Clear Rationales**: Each evaluation includes a detailed explanation referencing specific policy sections, allowing users to understand exactly why content was flagged or approved.\n"
+            "* **Continuous Improvement**: The system learns from feedback, addressing any misclassifications by improving retrieval accuracy over time.\n\n"
+            "Our approach combines Contextual's state-of-the-art <a href='https://contextual.ai/blog/introducing-instruction-following-reranker/' target='_blank'>steerable reranker</a>, <a href='https://contextual.ai/blog/introducing-grounded-language-model/' target='_blank'>world's most grounded language model</a>, and <a href='https://contextual.ai/blog/combining-rag-and-specialization/' target='_blank'>tuning for agent specialization</a> to achieve superhuman performance in content evaluation tasks. This technology enables consistent, fine-grained assessments across any content type and format.\n\n"
+            
+            "<h3 style='color:red; font-weight:bold;'>SAFETY WARNING</h3>" 
+            "Some of the randomly generated test cases contain hateful language that you might find offensive or upsetting."
+        )
         
         with gr.Row():
             with gr.Column(scale=1):
@@ -762,7 +857,7 @@ def create_gradio_app():
                         # LlamaGuard section with permanent model card link
                         gr.HTML("""
                         <div>
-                            <h3 class="result-header">🦙 LlamaGuard 3.0</h3>
+                            <h3 class="result-header">🦙 LlamaGuard Rating</h3>
                             <div style="margin-top: -10px; margin-bottom: 10px;">
                                 <a href="https://github.com/meta-llama/PurpleLlama/blob/main/Llama-Guard3/8B/MODEL_CARD.md" 
                                    target="_blank" class="knowledge-button">View model card</a>
@@ -782,6 +877,21 @@ def create_gradio_app():
                         </div>
                         """)
                         openai_results = gr.HTML('<div class="rating-box secondary-box empty-rating">Rating will appear here</div>')
+                
+                # Add Perspective API section
+                with gr.Row():
+                    with gr.Column():
+                        # Perspective API section with model card link
+                        gr.HTML("""
+                        <div>
+                            <h3 class="result-header">🔍 Google Perspective API</h3>
+                            <div style="margin-top: -10px; margin-bottom: 10px;">
+                                <a href="https://developers.perspectiveapi.com/s/about-the-api" 
+                                   target="_blank" class="knowledge-button">View API info</a>
+                            </div>
+                        </div>
+                        """)
+                        perspective_results = gr.HTML('<div class="rating-box secondary-box empty-rating">Rating will appear here</div>')
         
         # Define show/hide loading indicator functions
         def show_loading():
@@ -819,7 +929,7 @@ def create_gradio_app():
         ).then(
             rate_user_input,
             inputs=[user_input],
-            outputs=[contextual_results, llama_results, openai_results, retrieved_knowledge]
+            outputs=[contextual_results, llama_results, openai_results, perspective_results, retrieved_knowledge]
         ).then(
             hide_loading,
             inputs=None,