Update paraphraser.py

paraphraser.py

@@ -18,25 +18,26 @@ def paraphrase_comment(comment):
             "You are a content moderator tasked with rewriting toxic comments into neutral and constructive ones while maintaining the original meaning. "
             "Follow these guidelines:\n"
             "- Remove explicit hate speech, personal attacks, or offensive language.\n"
-            "- Keep the response neutral and professional.\n"
-            "- Ensure the rewritten comment retains the original intent but in a constructive tone.\n\n"
+            "- Keep the response neutral and conversational, suitable for a casual online platform.\n"
+            "- Ensure the rewritten comment retains the original intent but in a constructive tone, addressing the specific context of the comment (e.g., disagreement, frustration).\n\n"
             "Examples:\n"
             "Toxic: \"You're so dumb! You never understand anything!\"\n"
-            "Neutral: \"I think there's some misunderstanding. Let's clarify things.\"\n"
+            "Neutral: \"I think there might be a misunderstanding here. Can we go over this again to clear things up?\"\n"
             "Toxic: \"This is the worst idea ever. Only an idiot would suggest this.\"\n"
-            "Neutral: \"I don't think this idea works well. Maybe we can explore other options.\"\n\n"
+            "Neutral: \"I’m not sure this idea works for me. Could we look at some other options instead?\"\n"
+            "Toxic: \"You are an idiot and should leave this platform.\"\n"
+            "Neutral: \"It seems like you might not be enjoying this platform. Maybe we can talk about what’s not working for you?\"\n\n"
             f"Now, rewrite this comment: \"{comment}\""
         )
         inputs = tokenizer(prompt, return_tensors="pt", truncation=True, padding=True, max_length=512)
 
-        # Generate the paraphrased comment
+        # Generate the paraphrased comment with optimized parameters
         outputs = model.generate(
             **inputs,
-            max_length=512,
-            num_return_sequences=1,
-            temperature=0.7,
-            top_p=0.9,
-            do_sample=True
+            max_length=50,  # Reduced max_length for short comments
+            num_beams=4,  # Use beam search for faster and more consistent generation
+            early_stopping=True,  # Stop generation once a good sequence is found
+            do_sample=False  # Disable sampling to use beam search
         )
 
         paraphrased_comment = tokenizer.decode(outputs[0], skip_special_tokens=True)