Update app.py

app.py

@@ -16,8 +16,9 @@ except Exception as e:
 
 # Load the model and tokenizer with fallback to FLAN-T5-Small
 model_name = "Qwen/Qwen2-0.5B-Instruct"
-fallback_model = "google/flan-t5-small"  # Smaller model (~250 MB quantized)
+fallback_model = "google/flan-t5-small"
 tokenizer = AutoTokenizer.from_pretrained(model_name)
+loaded_model_name = model_name  # Track loaded model
 
 try:
     model = AutoModelForCausalLM.from_pretrained(
@@ -29,12 +30,13 @@ try:
     print(f"Successfully loaded {model_name}.")
 except Exception as e:
     print(f"Failed to load {model_name} with quantization: {e}. Falling back to {fallback_model}.")
+    loaded_model_name = fallback_model
     model = AutoModelForCausalLM.from_pretrained(
         fallback_model,
         device_map="cpu",
         low_cpu_mem_usage=True,
     )
-    tokenizer = AutoTokenizer.from_pretrained(fallback_model)  # Update tokenizer for fallback
+    tokenizer = AutoTokenizer.from_pretrained(fallback_model)
 
 def generate_llm_response(message):
     """Generate response using the loaded model with multilingual prompting"""
@@ -44,16 +46,16 @@ def generate_llm_response(message):
     # Detect if the input is in Nepali
     is_nepali = bool(re.search(r'[\u0900-\u097F]', message))
     
-    # Craft a prompt based on language detection
+    # Craft a detailed prompt based on language detection
     if is_nepali:
-        prompt = f"तपाईं एक नेपाली च्याटबोट हुनुहुन्छ। प्रयोगकर्ताले भनेको कुराको जवाफ नेपालीमा दिनुहोस्: {message}"
+        prompt = f"तपाईं एक नेपाली च्याटबोट हुनुहुन्छ जसले नेपालीमा प्राकृतिक र पूर्ण जवाफ दिन्छ। प्रयोगकर्ताले भन्नुभयो: '{message}'। जवाफ दिनुहोस्:"
     else:
-        prompt = f"You are a friendly chatbot that can respond in English or Nepali. Respond to the user's message: {message}"
+        prompt = f"You are a friendly chatbot that responds naturally in English or Nepali. The user said: '{message}'. Please respond:"
     
     try:
         # Tokenize and generate response
         inputs = tokenizer(prompt, return_tensors="pt").to("cpu")
-        outputs = model.generate(**inputs, max_new_tokens=64, temperature=0.7, do_sample=True)
+        outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.7, do_sample=True)  # Increased tokens for completeness
         response = tokenizer.decode(outputs[0], skip_special_tokens=True)
         
         # Post-process to ensure a complete sentence
@@ -61,6 +63,10 @@ def generate_llm_response(message):
         if response and not response.endswith(('.', '!', '?')):
             response += "।" if is_nepali else "."
         
+        # Fallback for inadequate Nepali response
+        if is_nepali and len(response.split()) < 3:  # If response is too short
+            return "माफ गर्नुहोस्, मलाई थप जानकारी चाहिए। तपाईंले के बारेमा कुरा गर्न चाहनुहुन्छ?"
+        
         return response if response else "Sorry, I couldn't generate a response."
     
     except Exception as e:
@@ -107,7 +113,7 @@ css = """
     border-radius: 25px !important;
 }
 .input-container input {
-    color: #ffffff !important;
+    color: #1e1e1e !important;
     background: transparent !important;
 }
 .gradio-chatbot {