more fixes

app.py

@@ -28,8 +28,8 @@ def load_models():
         )
     
     try:
-        # Use a smaller text generation model
-        generator_model = "distilgpt2"  # Smaller than opt-350m
+        # Use GPT-2 instead of DistilGPT-2
+        generator_model = "gpt2"
         generator_tokenizer = AutoTokenizer.from_pretrained(generator_model)
         generator = AutoModelForCausalLM.from_pretrained(generator_model)
         
@@ -71,37 +71,28 @@ def generate_content(
 ):
     char_limit = 280 if platform == "Twitter" else 500
     
-    # Create a more structured prompt that works for any product type
-    prompt = f"""Task: Write a {tone} {platform} post that promotes a product.
+    # Simpler, more direct prompt
+    prompt = f"""Marketing post for {platform}:
 
-Guidelines:
-- Highlight key benefits
-- Include a clear call-to-action
-- Be concise and engaging
-- Match the {tone} tone
-- Maximum {char_limit} characters
-
-Product Details:
-Name: {product_name}
-Target Audience: {target_audience}
-Key Benefits: {unique_benefits}
-Main Features: {key_features}
-
-Write the {platform} post here (do not include labels or prefixes):"""
+{product_name} helps {target_audience} by {unique_benefits}. {product_description}. Features include {key_features}.
 
+Marketing post in {tone} tone:
+"""
+    
     try:
-        # Generate multiple variations
-        inputs = generator_tokenizer(prompt, return_tensors="pt", max_length=256, truncation=True)
+        # Generate content with stricter parameters
+        inputs = generator_tokenizer(prompt, return_tensors="pt", max_length=128, truncation=True)
         outputs = generator.generate(
             inputs["input_ids"],
-            max_length=char_limit + 100,  # Extra space for cleaning
-            num_return_sequences=3,
-            temperature=0.7,
-            top_p=0.9,
+            max_length=char_limit // 2,  # Keep it concise
+            num_return_sequences=2,
+            temperature=0.9,
+            top_p=0.85,
             do_sample=True,
             pad_token_id=generator_tokenizer.eos_token_id,
             no_repeat_ngram_size=2,
-            min_length=50,
+            min_length=30,
+            repetition_penalty=1.2
         )
         
         generated_texts = [generator_tokenizer.decode(output, skip_special_tokens=True) for output in outputs]
@@ -109,66 +100,50 @@ Write the {platform} post here (do not include labels or prefixes):"""
         # Process and clean the generated content
         filtered_content = []
         for text in generated_texts:
-            # Extract only the post content (remove prompt)
-            post = text.split("Write the")[-1].split("post here")[-1].strip()
-            post = post.strip(':" ').strip()
+            # Extract only the part after "Marketing post in {tone} tone:"
+            try:
+                post = text.split("Marketing post in")[-1].split("tone:")[-1].strip()
+            except:
+                post = text
             
-            # Clean up common issues
-            post = post.replace("Product Details:", "")
-            post = post.replace("Guidelines:", "")
-            post = post.replace("[Post]:", "")
-            post = post.replace("Post:", "")
+            # Basic cleaning
+            post = ' '.join(post.split())  # Remove extra whitespace
             
-            # Skip if too short or contains prompt artifacts
-            if (len(post) < 30 or 
-                "Task:" in post or 
-                "Name:" in post or
-                "Target Audience:" in post):
+            # Skip if the post is too short or contains prompt artifacts
+            if len(post) < 20 or "Marketing post" in post or "tone:" in post:
                 continue
+                
+            # Ensure it starts with product name if it's not already included
+            if product_name not in post:
+                post = f"{product_name}: {post}"
             
-            # Truncate to character limit
-            post = post[:char_limit]
-            
-            # Add platform-specific formatting
-            if platform == "Instagram":
-                words = product_name.lower().split()
-                relevant_tags = " ".join([f"#{word}" for word in words if len(word) > 2])
-                if len(post) + len(relevant_tags) + 2 <= char_limit:
-                    post += f"\n{relevant_tags}"
+            # Ensure there's a call to action
+            if "learn more" not in post.lower() and len(post) + 15 <= char_limit:
+                post += " Learn more today!"
             
             # Check sentiment and safety
             try:
                 sentiment = sentiment_analyzer(post)[0]
                 safety_check = content_checker(post)[0]
                 
-                # Only add if content seems appropriate
-                if sentiment['label'] != 'negative' and float(safety_check.get('score', 0)) > 0.7:
-                    filtered_content.append({
-                        'text': post,
-                        'sentiment': sentiment['label'],
-                        'safety_score': f"{float(safety_check.get('score', 0)):.2f}"
-                    })
+                filtered_content.append({
+                    'text': post[:char_limit],
+                    'sentiment': sentiment['label'],
+                    'safety_score': f"{float(safety_check.get('score', 0)):.2f}"
+                })
             except Exception as e:
                 print(f"Error in content analysis: {str(e)}")
                 continue
         
-        # If no valid content was generated, create a generic but customized post
+        # If no valid content was generated, use a structured fallback
         if not filtered_content:
-            # Create a generic but customized format that works for any product
-            features_list = key_features.split(',')[0].strip()
-            benefits_list = unique_benefits.split(',')[0].strip()
+            benefit = unique_benefits.split(',')[0].strip()
+            feature = key_features.split(',')[0].strip()
             
-            generic_post = (
-                f"Discover {product_name}! "
-                f"{product_description} "
-                f"Featuring {features_list}. "
-                f"{benefits_list}. "
-                f"Perfect for {target_audience}. "
-                f"Learn more today!"
-            )[:char_limit]
+            fallback_post = f"{product_name}: {product_description} {benefit}. Featuring {feature}. Perfect for {target_audience}. Learn more today!"
             
             filtered_content.append({
-                'text': generic_post,
+                'text': fallback_post[:char_limit],
                 'sentiment': 'positive',
                 'safety_score': '1.00'
             })
@@ -178,7 +153,7 @@ Write the {platform} post here (do not include labels or prefixes):"""
     except Exception as e:
         print(f"Error in content generation: {str(e)}")
         return [{
-            'text': f"Discover {product_name}! {product_description[:100]}... Learn more today!",
+            'text': f"{product_name}: {product_description[:100]}... Learn more today!",
             'sentiment': 'neutral',
             'safety_score': '1.00'
         }]