Evaluating on My Own Datasets

[lyy1994]

Hi, I am trying to evaluate GLM-130B in our own datasets. I follow this guide to convert the CHID in FewCLUE to see whether my implementation is correct. However, I found that removing the tokenized inputs and choices in the released dataset does not lead to the same result reported in the paper. I think the inputs_pretokenized and choices_pretokenized are not properly tokenized and want to know how to ensure correct tokenization.