Open luoweb opened 1 year ago
Regarding the pre-training process of StarCoder and the SFT process of our WizardCoder, it's important to note that a majority of the training data used is in English. Consequently, WizardCoder's proficiency in handling non-English inputs is relatively limited. To enhance its capability in dealing with non-English languages, one potential approach is to incorporate additional training with non-English data.
model wizardcode Chinese output response repeat frequently , how to configure the parameter avoid the problem?