Update app.py

app.py

@@ -3,12 +3,11 @@ from fastapi.middleware.cors import CORSMiddleware
 from pydantic import BaseModel
 from llama_cpp import Llama
 import os
-import uvicorn # Required for running the app correctly on HF Spaces
+import uvicorn
 
 # -----------------------------
 # MODEL CONFIG
 # -----------------------------
-# Ensure these files are available in your HF Space repository
 MODEL_REPO = "bartowski/Phi-3.1-mini-4k-instruct-GGUF"
 MODEL_FILE = "Phi-3.1-mini-4k-instruct-IQ2_M.gguf"
 
@@ -17,16 +16,15 @@ try:
     llm = Llama.from_pretrained(
         repo_id=MODEL_REPO,
         filename=MODEL_FILE,
-        n_threads=4,      # match your CPU cores
-        n_ctx=1024,       # Context length for speed/memory
-        n_batch=512,
+        n_threads=4,      
+        n_ctx=1024,       
+        n_batch=512,      
         verbose=False,
     )
     print("Model loaded successfully.")
 except Exception as e:
-    # Crucial for debugging HF Spaces issues
     print(f"FATAL ERROR: Could not load model: {e}")
-    llm = None # Set to None so the server can start, but API calls will fail
+    llm = None 
 
 # -----------------------------
 # FASTAPI APP
@@ -43,7 +41,7 @@ app.add_middleware(
 class RequestBody(BaseModel):
     text: str
 
-# ⭐ ENHANCED SYSTEM PROMPT - Emphasizing natural, human-like output
+# Enhanced System Prompt (Same as before, it is strong enough)
 SYSTEM_PROMPT = """
 You are a highly skilled professional editor specializing in making academic or technical text sound natural, engaging, and indistinguishable from human writing.
 Your task is to rewrite the provided text following these strict rules:
@@ -84,18 +82,18 @@ async def humanize(body: RequestBody):
         output = llm(
             prompt,
             max_tokens=512,
-            temperature=0.45,           # Increased for more human-like variation
-            top_p=0.9,
+            # ⭐ CRITICAL TWEAK: Increase Temperature for randomness/creativity
+            temperature=0.8,         # AGGRESSIVELY increased from 0.45 to 0.8
+            # ⭐ CRITICAL TWEAK: Decrease top_p to focus on a smaller set of high-probability tokens
+            top_p=0.8,               # Decreased from 0.9 to 0.8
             repeat_penalty=1.05,
             stop=[END_TAG, "\nOriginal text:"], 
         )
         
+        # ... (rest of the robust post-processing remains the same)
         raw_result = output["choices"][0]["text"].strip()
-        
-        # Robust extraction: look for the closing tag and clean up
         result = raw_result.split(END_TAG)[0].strip()
         
-        # Remove the starting tag if the model included it in the output
         if result.startswith(START_TAG):
             result = result[len(START_TAG):].strip()
 
@@ -106,23 +104,18 @@ async def humanize(body: RequestBody):
 
     except Exception as e:
         print(f"Model execution error: {e}")
-        # Return a 500 status code, which matches your frontend's catch block logic
         raise HTTPException(status_code=500, detail="Model execution error")
 
 
 @app.get("/")
 def health():
-    # Simple health check, CRITICAL for HF Spaces to know the app is running
     return {"status": "ok"}
 
 # -----------------------------
 # DEPLOYMENT CONFIG FOR HUGGING FACE SPACES
 # -----------------------------
 if __name__ == "__main__":
-    # 🚨 CRITICAL: Use 0.0.0.0 host and 7860 port for HF Spaces deployments
     HOST = os.environ.get("HOST", "0.0.0.0")
     PORT = int(os.environ.get("PORT", 7860)) 
     print(f"Starting server on {HOST}:{PORT}")
-    
-    # You will need to install uvicorn: pip install uvicorn
     uvicorn.run(app, host=HOST, port=PORT)