Update app.py

app.py

@@ -32,6 +32,8 @@ def inference_func(prompt, max_length=64, sampling_temperature=0.7):
         max_length=max_length,
         sampling_topk=20,
         sampling_temperature=sampling_temperature,
+        repetition_penalty=1.1,
+        end_token=[26168, 27, 208, 14719, 9078, 18482, 27, 208],
         include_prompt_in_result=False,
     )
     output = tokenizer.decode(results[0].sequences_ids[0])