FAQ

Language Bias

As we mentioned in our paper, the language output of the model will to some extent affect our evaluation results. For the longformer model, its performance in Chinese is relatively poor. Therefore, when calculating RtA, we consider responses with a Chinese character ratio greater than \(\alpha\) as invalid sample (the default setting of \(\alpha\) is 0.3)