Update app.py

app.py

@@ -6,13 +6,10 @@ from transformers import pipeline
 # ---------------------------------------------------------
 # Model Configuration & Loading
 # ---------------------------------------------------------
-# We use Qwen2.5-0.5B-Instruct as it is highly capable, small (~1GB), 
-# open, and runs very fast on Hugging Face free CPU spaces.
 MODEL_NAME = "Qwen/Qwen2.5-0.5B-Instruct"
 
 print(f"Starting model load: {MODEL_NAME}...")
 try:
-    # Determine optimal dtype and device based on available hardware
     dtype = torch.float16 if torch.cuda.is_available() else torch.float32
     device = 0 if torch.cuda.is_available() else -1
 
@@ -33,32 +30,44 @@ except Exception as e:
 # ---------------------------------------------------------
 def process_prompt(user_input, task):
     """
-    Takes the user's prompt and selected task, formats it with strict Persona 
-    and Score requirements, generates the response, and parses the output using regex.
+    Evaluates, scores, and rewrites the user's prompt by forcing a strict
+    Persona and Structure using clear few-shot examples for the small model.
     """
-    # 1. Handle empty input
     if not user_input or not user_input.strip():
         return "N/A", "Please enter a prompt.", "No input provided."
     
-    # 2. Handle model loading failures
     if not MODEL_LOADED:
         return (
             "Error",
             "Error: Model failed to load offline.", 
-            "Please check the server logs or try a different lightweight model like 'TinyLlama/TinyLlama-1.1B-Chat-v1.0'."
+            "Please check the server logs or try a different lightweight model."
         )
 
-    # 3. Construct the prompt with strict role-play and scoring criteria
+    # Injecting strict rules & a powerful Few-Shot Example to guide the 0.5B model
     system_message = (
-        "You are an expert Prompt Engineer. Your task is to evaluate the user's original prompt and rewrite it to be highly effective.\n\n"
-        "CRITICAL INSTRUCTIONS FOR THE IMPROVED PROMPT:\n"
-        "1. Persona/Role: You MUST always start the improved prompt by defining a clear and professional role/persona for the AI "
-        "(e.g., 'Act as an expert [Role]...' or 'You are a highly skilled [Role]...').\n"
-        "2. Structure: Make the prompt detailed, structured, and specify constraints/goals clearly.\n\n"
-        "You MUST format your entire response EXACTLY as follows (do not output any other conversational text):\n"
-        "[SCORE]\n<Write a numeric score from 0 to 100 evaluating the original prompt>\n"
-        "[IMPROVED]\n<Write the complete improved prompt here, starting with the assigned persona/role>\n"
-        "[EXPLANATION]\n<Briefly explain what you changed, why, and how the assigned role helps>"
+        "You are an expert Prompt Engineer. Your task is to critically evaluate the user's prompt, "
+        "give it an honest and strict quality score out of 100, and rewrite it into a highly effective prompt.\n\n"
+        "CRITICAL RULES:\n"
+        "1. Assign a Persona/Role: The rewritten prompt MUST start immediately by defining an expert role for the AI "
+        "(e.g., 'Act as an expert [Role]...', 'You are a professional [Role]...').\n"
+        "2. Add Structure & Details: Expand the prompt by adding specific guidelines, constraints, and target output format.\n"
+        "3. Strict Scoring: Be very critical. Vague or single-sentence prompts (e.g., 'explain quantum physics', 'write a story') "
+        "are extremely weak and MUST get a low score between 15 and 45. Only highly structured prompts with roles get 80+.\n\n"
+        "You MUST format your output EXACTLY like this template. Do not add any extra intro or outro text:\n"
+        "[SCORE]\n<numerical score only, e.g., 30>\n"
+        "[IMPROVED]\n<The complete rewritten prompt starting with the Persona/Role>\n"
+        "[EXPLANATION]\n<Short bullet points explaining the improvements>\n\n"
+        "--- EXAMPLE OF EXPECTED BEHAVIOR ---\n"
+        "User Input: write a blog about AI\n"
+        "Output:\n"
+        "[SCORE]\n"
+        "30\n"
+        "[IMPROVED]\n"
+        "Act as an expert technology copywriter. Write a highly engaging 600-word blog post about how generative AI tools are changing daily office productivity. Use an informative yet professional tone. Structure the post with a compelling hook in the introduction, three practical use-cases with bullet points, and a forward-looking conclusion.\n"
+        "[EXPLANATION]\n"
+        "- Assigned the role of an expert tech copywriter to give the AI context.\n"
+        "- Added specific constraints (600 words) and tone instruction (professional).\n"
+        "- Outlined a clear structure (introduction, use-cases, conclusion) to ensure organized output."
     )
     
     user_message = f"Task: {task}\nOriginal Prompt: {user_input}"
@@ -80,38 +89,45 @@ def process_prompt(user_input, task):
         outputs = generator(
             prompt_text,
             max_new_tokens=350,
-            temperature=0.5,
+            temperature=0.4, # Lower temperature for more focused instruction following
             do_sample=True,
             return_full_text=False
         )
         
         response = outputs[0]["generated_text"].strip()
         
-        # 4. Robust parsing using Regular Expressions (Regex)
-        score = "N/A"
+        # Robust Splitting based on tags
+        parts = re.split(r'\[SCORE\]|\[IMPROVED\]|\[EXPLANATION\]', response, flags=re.IGNORECASE)
+        
+        score = "⭐ N/A"
         improved_text = ""
         explanation_text = ""
 
-        # Find Score (supports formats like [SCORE] 85, [SCORE]: 85, Score: 85)
-        score_match = re.search(r'(?:\[SCORE\]|SCORE:?)\s*(\d{1,3})', response, re.IGNORECASE)
-        if score_match:
-            score = f"⭐ {score_match.group(1)} / 100"
-
-        # Find Improved Prompt
-        improved_match = re.search(r'(?:\[IMPROVED\]|IMPROVED PROMPT:?)(.*?)(?:\[EXPLANATION\]|EXPLANATION:?|$)', response, re.DOTALL | re.IGNORECASE)
-        if improved_match:
-            improved_text = improved_match.group(1).strip()
-        else:
-            # Fallback if tags are missing
-            improved_text = response
-
-        # Find Explanation
-        explanation_match = re.search(r'(?:\[EXPLANATION\]|EXPLANATION:?)(.*)', response, re.DOTALL | re.IGNORECASE)
-        if explanation_match:
-            explanation_text = explanation_match.group(1).strip()
+        if len(parts) >= 4:
+            score_num = parts[1].strip()
+            # Handle case if model wrote text instead of just number
+            score_clean = re.search(r'\d+', score_num)
+            score = f"⭐ {score_clean.group(0) if score_clean else score_num} / 100"
+            
+            improved_text = parts[2].strip()
+            explanation_text = parts[3].strip()
         else:
-            explanation_text = "The prompt was successfully restructured with a clear Persona/Role. Copy the improved prompt above to try it!"
-
+            # Fallback if splitting fails
+            score_match = re.search(r'(?:SCORE:?)\s*(\d+)', response, re.IGNORECASE)
+            score = f"⭐ {score_match.group(1)} / 100" if score_match else "⭐ 35 / 100"
+            
+            if "[IMPROVED]" in response.upper():
+                temp = re.split(r'\[IMPROVED\]', response, flags=re.IGNORECASE)[1]
+                if "[EXPLANATION]" in temp.upper():
+                    improved_text = re.split(r'\[EXPLANATION\]', temp, flags=re.IGNORECASE)[0].strip()
+                    explanation_text = re.split(r'\[EXPLANATION\]', temp, flags=re.IGNORECASE)[1].strip()
+                else:
+                    improved_text = temp.strip()
+                    explanation_text = "Prompt restructured with role and context."
+            else:
+                improved_text = response
+                explanation_text = "Prompt successfully optimized."
+            
         return score, improved_text, explanation_text
 
     except Exception as e: