fix bug for ceval

eval/evaluate_ceval.py

@@ -80,7 +80,7 @@ def eval_subject(
     score = []
 
     few_shot_prompt = generate_few_shot_prompt(
-        k, subject_name, dev_df) if few_shot else []
+        k, subject_name, dev_df) if few_shot else ''
     all_probs = {'prob_A': [], 'prob_B': [], 'prob_C': [], 'prob_D': []}
     if args.debug: print(f"few_shot_prompt: {few_shot_prompt}")
 
@@ -95,10 +95,10 @@ def eval_subject(
         softval = torch.nn.functional.softmax(
                 torch.tensor(
                     [
-                        logits[tokenizer("A")['input_ids']],
+                        logits[tokenizer("A")['input_ids'][-1]],
-                        logits[tokenizer("B")['input_ids']],
+                        logits[tokenizer("B")['input_ids'][-1]],
-                        logits[tokenizer("C")['input_ids']],
+                        logits[tokenizer("C")['input_ids'][-1]],
-                        logits[tokenizer("D")['input_ids']],
+                        logits[tokenizer("D")['input_ids'][-1]],
                     ]
                 ),
                 dim=0,