Update app.py

app.py

@@ -30,7 +30,10 @@ async def predict(request: Request):
             inputs.input_ids,
             max_length=40,
             do_sample=True,
-            num_return_sequences=1
+            num_return_sequences=1,
+            temperature=0.7,
+            top_p=0.9,
+            repetition_penalty=1.2,
         )
         for token_id in outputs[0]:
             token = tokenizer.decode(token_id, skip_special_tokens=True)