<ADD> optimize code

app.py

@@ -22,7 +22,12 @@ DISPLAY_MAPPING = {
 def do_infer_probs(model, exemplar_attn_kv, exemplar_attn_mask, batched_choices_input):
     batched_choices_logprobs = []
     for batched_one_choice_input in batched_choices_input:
-        batch_input_ids, batch_attention_mask, batch_choice_start, batch_choice_end = batched_one_choice_input
+        (
+            batch_input_ids,
+            batch_attention_mask,
+            batch_choice_start,
+            batch_choice_end,
+        ) = batched_one_choice_input
         bs = len(batch_input_ids)
 
         merged_attn_mask = torch.cat((exemplar_attn_mask.expand(bs, -1), batch_attention_mask), dim=1)
@@ -178,7 +183,7 @@ Directly employing ICL results in lower prediction accuracy. However, in our pro
             """
 <h2 style='text-align: center; margin-bottom: 1rem'>
 <a href='https://arxiv.org/abs/2305.13016' target="_blank" style='text-decoration: none'>[Paper]</a>
-<a href='https://arxiv.org/abs/2305.13016' target="_blank" style='text-decoration: none'>[Code]</a>
+<a href='https://arxiv.org/abs/2305.13016' target="_blank" style='text-decoration: none'>[Code]</a> 
 </h2>"""
         )
 
@@ -205,7 +210,7 @@ Directly employing ICL results in lower prediction accuracy. However, in our pro
 
             prompt = gr.Textbox(label="Demonstrations (Prompt template formatted)", value=demos)
             gr.Markdown("<h2 style='text-align: center; margin-bottom: 1rem'>👇 Run forward tuning once !</h2>")
-            step_button = gr.Button("Click here to train LLM ! Now Step: 1")
+            step_button = gr.Button("Click here to train LLM ! Now Step: 1", variant="primary")
             big_table = gr.DataFrame(
                 value=init_table_result,
                 elem_id="the-table",