Handeled the issue of prompt and instructions covering the space of actual response

app.py

@@ -72,11 +72,18 @@ def generate_text(prompt, tone, max_length, temperature=0.7, top_p=0.9, repetiti
     input_text = tone_prompts.get(tone, prompt)
     # This picks the right instruction from the dictionary based on the tone.
     inputs = tokenizer(input_text, return_tensors="pt")
+    input_ids = inputs["input_ids"]
     # This turns our input text (with the tone instruction) into a format (tensors) that the model can process using the tokenizer.
+    input_token_length = input_ids.shape[1]  # Get the number of tokens in the input
+    # Store the length of the input
+    
     outputs = model.generate(
         inputs["input_ids"],
-        max_length=max_length + len(input_text.split()),
+        # max_length=max_length + len(input_text.split()),
         # This sets how long the generated text can be. We add the number of words in our input text (len(input_text.split())) to the max_length the user picked, so the model knows how many total words to create.
+        # CHANGE: Use max_new_tokens for clarity instead of calculating total length
+        max_new_tokens=max_length
+        # Generate THIS many NEW tokens
         temperature=temperature,
         # This controls how creative the model gets. A lower temperature (e.g., 0.7) keeps things more predictable, while a higher one makes it wilder and more random—think of it like adjusting the spice level!
         top_p=top_p,
@@ -86,9 +93,14 @@ def generate_text(prompt, tone, max_length, temperature=0.7, top_p=0.9, repetiti
         num_return_sequences=1,
         # This tells the model to give us just one version of the text. If we wanted more options, we could change
         do_sample=True
+        pad_token_id=tokenizer.eos_token_id # Good practice for generation
     )
+    # --- Decode ONLY the generated part ---
+    # Slice the output tensor to get only the tokens AFTER the input tokens
     # This tells the model to generate text: it uses the input IDs, sets a max length, and adjusts creativity with temperature, top_p, and repetition_penalty.
-    return tokenizer.decode(outputs[0], skip_special_tokens=True)
+    generated_token_ids = outputs[0, input_token_length:] 
+    generated_text = tokenizer.decode(generated_token_ids, skip_special_tokens=True)
+    return generated_text # Return only the newly generated text
     # This turns again the model's output back into readable form, skipping any extra tokens we don’t need.
     
 # Clean and Solid UI for our Project, keeping the blue theme of gemini.