[Finetune] Fix evaluation

[minmingzhu]

After trying several solutions, I finally used the tokenize_func from neural_chat:

1. Using the transformer tokenizer with padding and attention_mask, and copying input_ids to labels.
2. Using the transformer tokenizer with padding and attention_mask, and masking the input in the prompt.
3. Using the transformer tokenizer with padding and attention_mask, and masking the response in the prompt.
4. Using neural_chat's tokenize_func for padding, filling attention_mask and labels which masking the input in the prompt.

<style> </style>

Case	Average	ARC	HellaSwag	MMLU	TruthfulQA	Winogrande
Case 1	62.116	55.97	78.39	60.75	45.86	69.61
Case 2	58.99	54.18	72.85	60.23	42.18	65.51
Case 3	59.064	53.24	73.21	60.85	41.56	66.46
Case 4	63.03	53.67	81.47	60.27	43.02	76.72