Update app.py

app.py

@@ -87,75 +87,39 @@ class HumanizeRequest(BaseModel):
 # ==================== STYLE PROMPTS ====================
 STYLE_PROMPTS = {
     "professional": """<|system|>
-You are a professional writing assistant. Rewrite the text below in formal, corporate business language.
-
-IMPORTANT RULES:
-1. Output ONLY the rewritten text
-2. No explanations, no notes
-3. Keep the same meaning
-4. Use formal vocabulary
-5. Proper grammar and structure
-<|end|>
+You are a professional writing assistant. Rewrite the following text in formal, corporate business language.
+IMPORTANT: Output ONLY the rewritten text, nothing else.<|end|>
 <|user|>
-Text to rewrite: {text}
-
-Rewritten (professional):<|end|>
+Text to rewrite: {text}<|end|>
 <|assistant|>""",
 
     "casual": """<|system|>
-You are a casual writing assistant. Rewrite the text below in friendly, natural, conversational English.
-
-IMPORTANT RULES:
-1. Output ONLY the rewritten text
-2. No explanations, no notes
-3. Keep the same meaning
-4. Use contractions (I'm, don't, etc.)
-5. Sound like a real person speaking
-<|end|>
+You are a casual writing assistant. Rewrite the following text in friendly, natural, conversational English.
+IMPORTANT: Output ONLY the rewritten text, nothing else.<|end|>
 <|user|>
-Text to rewrite: {text}
-
-Rewritten (casual):<|end|>
+Text to rewrite: {text}<|end|>
 <|assistant|>""",
 
     "academic": """<|system|>
-You are an academic writing assistant. Rewrite the text below in formal scholarly language.
-
-IMPORTANT RULES:
-1. Output ONLY the rewritten text
-2. No explanations, no notes
-3. Keep the same meaning
-4. Use precise academic vocabulary
-5. Maintain formal structure
-<|end|>
+You are an academic writing assistant. Rewrite the following text in formal scholarly language.
+IMPORTANT: Output ONLY the rewritten text, nothing else.<|end|>
 <|user|>
-Text to rewrite: {text}
-
-Rewritten (academic):<|end|>
+Text to rewrite: {text}<|end|>
 <|assistant|>""",
 
     "marketing": """<|system|>
-You are a marketing copywriter. Rewrite the text below into persuasive marketing language.
-
-IMPORTANT RULES:
-1. Output ONLY the rewritten text
-2. No explanations, no notes
-3. Keep the same meaning
-4. Use emotional hooks and benefits
-5. Make it engaging and compelling
-<|end|>
+You are a marketing copywriter. Rewrite the following text into persuasive marketing language.
+IMPORTANT: Output ONLY the rewritten text, nothing else.<|end|>
 <|user|>
-Text to rewrite: {text}
-
-Rewritten (marketing):<|end|>
+Text to rewrite: {text}<|end|>
 <|assistant|>"""
 }
 
 STYLE_TEMPERATURES = {
     "professional": 0.3,
-    "casual": 0.6,
+    "casual": 0.5,
     "academic": 0.4,
-    "marketing": 0.7
+    "marketing": 0.6
 }
 
 # ==================== HELPER FUNCTIONS ====================
@@ -164,10 +128,31 @@ def clean_output(text: str) -> str:
     if not text:
         return ""
     
-    # Remove Phi-3.1 special tokens
-    clean = re.sub(r'<\|[^>]+\|>', '', text)
-    clean = re.sub(r'Rewritten\s*\([^)]+\):', '', clean, flags=re.IGNORECASE)
-    clean = re.sub(r'Text to rewrite:.*', '', clean, flags=re.DOTALL)
+    # Remove emojis and special characters
+    emoji_pattern = re.compile("["
+        u"\U0001F600-\U0001F64F"  # emoticons
+        u"\U0001F300-\U0001F5FF"  # symbols & pictographs
+        u"\U0001F680-\U0001F6FF"  # transport & map symbols
+        u"\U0001F1E0-\U0001F1FF"  # flags
+        "]+", flags=re.UNICODE)
+    
+    clean = emoji_pattern.sub(r'', text)
+    
+    # Remove Phi-3.1 special tokens if any remain
+    clean = re.sub(r'<\|[^>]+\|>', '', clean)
+    
+    # Remove "Text to rewrite:" and similar instruction text
+    clean = re.sub(r'(Text to rewrite|Rewritten|Humanized|Assistant:|Assistant response:).*?:', '', clean, flags=re.IGNORECASE)
+    
+    # Remove any remaining instruction words
+    instruction_words = [
+        "here is the", "here's the", "here is", "here's",
+        "the rewritten text is:", "the rewritten text:",
+        "rewritten text:", "humanized text:", "output:"
+    ]
+    
+    for word in instruction_words:
+        clean = re.sub(fr'\b{re.escape(word)}\b', '', clean, flags=re.IGNORECASE)
     
     # Clean whitespace
     clean = re.sub(r'\s+', ' ', clean)
@@ -197,16 +182,19 @@ async def transform_with_model(text: str, style: str) -> str:
         prompt = format_prompt(text, style)
         temperature = STYLE_TEMPERATURES[style]
         
+        # Debug: Print prompt (remove in production)
+        print(f"DEBUG - Prompt for {style}: {prompt[:100]}...")
+        
         # Run inference
         start_time = time.time()
         
         output = llm(
             prompt,
-            max_tokens=min(400, len(text) + 100),
+            max_tokens=min(200, len(text) + 50),
             temperature=temperature,
-            top_p=0.9,
+            top_p=0.95,
             repeat_penalty=1.1,
-            stop=[END_TOKEN, "<|end|>", "\n\n"],
+            stop=["<|end|>", "\n"],
             echo=False,
         )
         
@@ -218,9 +206,15 @@ async def transform_with_model(text: str, style: str) -> str:
         else:
             result = str(output)
         
+        # Debug: Print raw output
+        print(f"DEBUG - Raw output: {result[:100]}...")
+        
         # Clean result
         cleaned = clean_output(result)
         
+        # Debug: Print cleaned output
+        print(f"DEBUG - Cleaned output: {cleaned}")
+        
         # Fallback if output is empty
         if not cleaned or cleaned.isspace():
             cleaned = f"[{style.capitalize()} Version]: {text}"
@@ -332,7 +326,7 @@ async def get_styles():
         styles_info[style] = {
             "description": first_line,
             "temperature": STYLE_TEMPERATURES[style],
-            "max_tokens": 400
+            "max_tokens": 200
         }
     
     return {